diff options
| author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-15 20:36:56 +0000 |
|---|---|---|
| committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-15 20:36:56 +0000 |
| commit | 51de1d8436100f725f3576aefa24a2bd2057bc28 (patch) | |
| tree | c6d1d5264b6d40a8d7ca34129f36b7d61e188af3 /video/out/gpu | |
| parent | Initial commit. (diff) | |
| download | mpv-51de1d8436100f725f3576aefa24a2bd2057bc28.tar.xz mpv-51de1d8436100f725f3576aefa24a2bd2057bc28.zip | |
Adding upstream version 0.37.0.upstream/0.37.0
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'video/out/gpu')
29 files changed, 12648 insertions, 0 deletions
diff --git a/video/out/gpu/context.c b/video/out/gpu/context.c new file mode 100644 index 0000000..5ce18af --- /dev/null +++ b/video/out/gpu/context.c @@ -0,0 +1,277 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <stddef.h> +#include <stdlib.h> +#include <stdio.h> +#include <string.h> +#include <stdbool.h> +#include <math.h> +#include <assert.h> + +#include "config.h" +#include "common/common.h" +#include "common/msg.h" +#include "options/options.h" +#include "options/m_option.h" +#include "video/out/vo.h" + +#include "context.h" +#include "spirv.h" + +/* OpenGL */ +extern const struct ra_ctx_fns ra_ctx_glx; +extern const struct ra_ctx_fns ra_ctx_x11_egl; +extern const struct ra_ctx_fns ra_ctx_drm_egl; +extern const struct ra_ctx_fns ra_ctx_wayland_egl; +extern const struct ra_ctx_fns ra_ctx_wgl; +extern const struct ra_ctx_fns ra_ctx_angle; +extern const struct ra_ctx_fns ra_ctx_dxgl; +extern const struct ra_ctx_fns ra_ctx_rpi; +extern const struct ra_ctx_fns ra_ctx_android; + +/* Vulkan */ +extern const struct ra_ctx_fns ra_ctx_vulkan_wayland; +extern const struct ra_ctx_fns ra_ctx_vulkan_win; +extern const struct ra_ctx_fns ra_ctx_vulkan_xlib; +extern const struct ra_ctx_fns ra_ctx_vulkan_android; +extern const struct ra_ctx_fns ra_ctx_vulkan_display; +extern const struct ra_ctx_fns ra_ctx_vulkan_mac; + +/* Direct3D 11 */ +extern const struct ra_ctx_fns ra_ctx_d3d11; + +/* No API */ +extern const struct ra_ctx_fns ra_ctx_wldmabuf; + +static const struct ra_ctx_fns *contexts[] = { +#if HAVE_D3D11 + &ra_ctx_d3d11, +#endif + +// OpenGL contexts: +#if HAVE_EGL_ANDROID + &ra_ctx_android, +#endif +#if HAVE_RPI + &ra_ctx_rpi, +#endif +#if HAVE_EGL_ANGLE_WIN32 + &ra_ctx_angle, +#endif +#if HAVE_GL_WIN32 + &ra_ctx_wgl, +#endif +#if HAVE_GL_DXINTEROP + &ra_ctx_dxgl, +#endif +#if HAVE_EGL_WAYLAND + &ra_ctx_wayland_egl, +#endif +#if HAVE_EGL_X11 + &ra_ctx_x11_egl, +#endif +#if HAVE_GL_X11 + &ra_ctx_glx, +#endif +#if HAVE_EGL_DRM + &ra_ctx_drm_egl, +#endif + +// Vulkan contexts: +#if HAVE_VULKAN + +#if HAVE_ANDROID + &ra_ctx_vulkan_android, +#endif +#if HAVE_WIN32_DESKTOP + &ra_ctx_vulkan_win, +#endif +#if HAVE_WAYLAND + &ra_ctx_vulkan_wayland, +#endif +#if HAVE_X11 + &ra_ctx_vulkan_xlib, +#endif +#if HAVE_VK_KHR_DISPLAY + &ra_ctx_vulkan_display, +#endif +#if HAVE_COCOA && HAVE_SWIFT + &ra_ctx_vulkan_mac, +#endif +#endif + +/* No API contexts: */ +#if HAVE_DMABUF_WAYLAND + &ra_ctx_wldmabuf, +#endif +}; + +static int ra_ctx_api_help(struct mp_log *log, const struct m_option *opt, + struct bstr name) +{ + mp_info(log, "GPU APIs (contexts):\n"); + mp_info(log, " auto (autodetect)\n"); + for (int n = 0; n < MP_ARRAY_SIZE(contexts); n++) { + if (!contexts[n]->hidden) + mp_info(log, " %s (%s)\n", contexts[n]->type, contexts[n]->name); + } + return M_OPT_EXIT; +} + +static int ra_ctx_validate_api(struct mp_log *log, const struct m_option *opt, + struct bstr name, const char **value) +{ + struct bstr param = bstr0(*value); + if (bstr_equals0(param, "auto")) + return 1; + for (int i = 0; i < MP_ARRAY_SIZE(contexts); i++) { + if (bstr_equals0(param, contexts[i]->type) && !contexts[i]->hidden) + return 1; + } + return M_OPT_INVALID; +} + +static int ra_ctx_context_help(struct mp_log *log, const struct m_option *opt, + struct bstr name) +{ + mp_info(log, "GPU contexts (APIs):\n"); + mp_info(log, " auto (autodetect)\n"); + for (int n = 0; n < MP_ARRAY_SIZE(contexts); n++) { + if (!contexts[n]->hidden) + mp_info(log, " %s (%s)\n", contexts[n]->name, contexts[n]->type); + } + return M_OPT_EXIT; +} + +static int ra_ctx_validate_context(struct mp_log *log, const struct m_option *opt, + struct bstr name, const char **value) +{ + struct bstr param = bstr0(*value); + if (bstr_equals0(param, "auto")) + return 1; + for (int i = 0; i < MP_ARRAY_SIZE(contexts); i++) { + if (bstr_equals0(param, contexts[i]->name) && !contexts[i]->hidden) + return 1; + } + return M_OPT_INVALID; +} + +// Create a VO window and create a RA context on it. +// vo_flags: passed to the backend's create window function +struct ra_ctx *ra_ctx_create(struct vo *vo, struct ra_ctx_opts opts) +{ + bool api_auto = !opts.context_type || strcmp(opts.context_type, "auto") == 0; + bool ctx_auto = !opts.context_name || strcmp(opts.context_name, "auto") == 0; + + if (ctx_auto) { + MP_VERBOSE(vo, "Probing for best GPU context.\n"); + opts.probing = true; + } + + // Hack to silence backend (X11/Wayland/etc.) errors. Kill it once backends + // are separate from `struct vo` + bool old_probing = vo->probing; + vo->probing = opts.probing; + + for (int i = 0; i < MP_ARRAY_SIZE(contexts); i++) { + if (contexts[i]->hidden) + continue; + if (!opts.probing && strcmp(contexts[i]->name, opts.context_name) != 0) + continue; + if (!api_auto && strcmp(contexts[i]->type, opts.context_type) != 0) + continue; + + struct ra_ctx *ctx = talloc_ptrtype(NULL, ctx); + *ctx = (struct ra_ctx) { + .vo = vo, + .global = vo->global, + .log = mp_log_new(ctx, vo->log, contexts[i]->type), + .opts = opts, + .fns = contexts[i], + }; + + MP_VERBOSE(ctx, "Initializing GPU context '%s'\n", ctx->fns->name); + if (contexts[i]->init(ctx)) { + vo->probing = old_probing; + return ctx; + } + + talloc_free(ctx); + } + + vo->probing = old_probing; + + // If we've reached this point, then none of the contexts matched the name + // requested, or the backend creation failed for all of them. + if (!vo->probing) + MP_ERR(vo, "Failed initializing any suitable GPU context!\n"); + return NULL; +} + +struct ra_ctx *ra_ctx_create_by_name(struct vo *vo, const char *name) +{ + for (int i = 0; i < MP_ARRAY_SIZE(contexts); i++) { + if (strcmp(name, contexts[i]->name) != 0) + continue; + + struct ra_ctx *ctx = talloc_ptrtype(NULL, ctx); + *ctx = (struct ra_ctx) { + .vo = vo, + .global = vo->global, + .log = mp_log_new(ctx, vo->log, contexts[i]->type), + .fns = contexts[i], + }; + + MP_VERBOSE(ctx, "Initializing GPU context '%s'\n", ctx->fns->name); + if (contexts[i]->init(ctx)) + return ctx; + talloc_free(ctx); + } + return NULL; +} + +void ra_ctx_destroy(struct ra_ctx **ctx_ptr) +{ + struct ra_ctx *ctx = *ctx_ptr; + if (!ctx) + return; + + if (ctx->spirv && ctx->spirv->fns->uninit) + ctx->spirv->fns->uninit(ctx); + + ctx->fns->uninit(ctx); + talloc_free(ctx); + + *ctx_ptr = NULL; +} + +#define OPT_BASE_STRUCT struct ra_ctx_opts +const struct m_sub_options ra_ctx_conf = { + .opts = (const m_option_t[]) { + {"gpu-context", + OPT_STRING_VALIDATE(context_name, ra_ctx_validate_context), + .help = ra_ctx_context_help}, + {"gpu-api", + OPT_STRING_VALIDATE(context_type, ra_ctx_validate_api), + .help = ra_ctx_api_help}, + {"gpu-debug", OPT_BOOL(debug)}, + {"gpu-sw", OPT_BOOL(allow_sw)}, + {0} + }, + .size = sizeof(struct ra_ctx_opts), +}; diff --git a/video/out/gpu/context.h b/video/out/gpu/context.h new file mode 100644 index 0000000..6788e6f --- /dev/null +++ b/video/out/gpu/context.h @@ -0,0 +1,107 @@ +#pragma once + +#include "video/out/vo.h" +#include "video/csputils.h" + +#include "ra.h" + +struct ra_ctx_opts { + bool allow_sw; // allow software renderers + bool want_alpha; // create an alpha framebuffer if possible + bool debug; // enable debugging layers/callbacks etc. + bool probing; // the backend was auto-probed + char *context_name; // filter by `ra_ctx_fns.name` + char *context_type; // filter by `ra_ctx_fns.type` +}; + +extern const struct m_sub_options ra_ctx_conf; + +struct ra_ctx { + struct vo *vo; + struct ra *ra; + struct mpv_global *global; + struct mp_log *log; + + struct ra_ctx_opts opts; + const struct ra_ctx_fns *fns; + struct ra_swapchain *swapchain; + struct spirv_compiler *spirv; + + void *priv; +}; + +// The functions that make up a ra_ctx. +struct ra_ctx_fns { + const char *type; // API type (for --gpu-api) + const char *name; // name (for --gpu-context) + + bool hidden; // hide the ra_ctx from users + + // Resize the window, or create a new window if there isn't one yet. + // Currently, there is an unfortunate interaction with ctx->vo, and + // display size etc. are determined by it. + bool (*reconfig)(struct ra_ctx *ctx); + + // This behaves exactly like vo_driver.control(). + int (*control)(struct ra_ctx *ctx, int *events, int request, void *arg); + + // These behave exactly like vo_driver.wakeup/wait_events. They are + // optional. + void (*wakeup)(struct ra_ctx *ctx); + void (*wait_events)(struct ra_ctx *ctx, int64_t until_time_ns); + void (*update_render_opts)(struct ra_ctx *ctx); + + // Initialize/destroy the 'struct ra' and possibly the underlying VO backend. + // Not normally called by the user of the ra_ctx. + bool (*init)(struct ra_ctx *ctx); + void (*uninit)(struct ra_ctx *ctx); +}; + +// Extra struct for the swapchain-related functions so they can be easily +// inherited from helpers. +struct ra_swapchain { + struct ra_ctx *ctx; + struct priv *priv; + const struct ra_swapchain_fns *fns; +}; + +// Represents a framebuffer / render target +struct ra_fbo { + struct ra_tex *tex; + bool flip; // rendering needs to be inverted + + // Host system's colorspace that it will be interpreting + // the frame buffer as. + struct mp_colorspace color_space; +}; + +struct ra_swapchain_fns { + // Gets the current framebuffer depth in bits (0 if unknown). Optional. + int (*color_depth)(struct ra_swapchain *sw); + + // Called when rendering starts. Returns NULL on failure. This must be + // followed by submit_frame, to submit the rendered frame. This function + // can also fail sporadically, and such errors should be ignored unless + // they persist. + bool (*start_frame)(struct ra_swapchain *sw, struct ra_fbo *out_fbo); + + // Present the frame. Issued in lockstep with start_frame, with rendering + // commands in between. The `frame` is just there for timing data, for + // swapchains smart enough to do something with it. + bool (*submit_frame)(struct ra_swapchain *sw, const struct vo_frame *frame); + + // Performs a buffer swap. This blocks for as long as necessary to meet + // params.swapchain_depth, or until the next vblank (for vsynced contexts) + void (*swap_buffers)(struct ra_swapchain *sw); + + // See vo. Usually called after swap_buffers(). + void (*get_vsync)(struct ra_swapchain *sw, struct vo_vsync_info *info); +}; + +// Create and destroy a ra_ctx. This also takes care of creating and destroying +// the underlying `struct ra`, and perhaps the underlying VO backend. +struct ra_ctx *ra_ctx_create(struct vo *vo, struct ra_ctx_opts opts); +void ra_ctx_destroy(struct ra_ctx **ctx); + +// Special case of creating a ra_ctx while specifying a specific context by name. +struct ra_ctx *ra_ctx_create_by_name(struct vo *vo, const char *name); diff --git a/video/out/gpu/d3d11_helpers.c b/video/out/gpu/d3d11_helpers.c new file mode 100644 index 0000000..30d9eae --- /dev/null +++ b/video/out/gpu/d3d11_helpers.c @@ -0,0 +1,966 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <windows.h> +#include <d3d11.h> +#include <dxgi1_6.h> +#include <versionhelpers.h> + +#include "common/common.h" +#include "common/msg.h" +#include "misc/bstr.h" +#include "osdep/io.h" +#include "osdep/threads.h" +#include "osdep/windows_utils.h" + +#include "d3d11_helpers.h" + +// Windows 8 enum value, not present in mingw-w64 headers +#define DXGI_ADAPTER_FLAG_SOFTWARE (2) +typedef HRESULT(WINAPI *PFN_CREATE_DXGI_FACTORY)(REFIID riid, void **ppFactory); + +static mp_once d3d11_once = MP_STATIC_ONCE_INITIALIZER; +static PFN_D3D11_CREATE_DEVICE pD3D11CreateDevice = NULL; +static PFN_CREATE_DXGI_FACTORY pCreateDXGIFactory1 = NULL; +static void d3d11_load(void) +{ + HMODULE d3d11 = LoadLibraryW(L"d3d11.dll"); + HMODULE dxgilib = LoadLibraryW(L"dxgi.dll"); + if (!d3d11 || !dxgilib) + return; + + pD3D11CreateDevice = (PFN_D3D11_CREATE_DEVICE) + GetProcAddress(d3d11, "D3D11CreateDevice"); + pCreateDXGIFactory1 = (PFN_CREATE_DXGI_FACTORY) + GetProcAddress(dxgilib, "CreateDXGIFactory1"); +} + +static bool load_d3d11_functions(struct mp_log *log) +{ + mp_exec_once(&d3d11_once, d3d11_load); + if (!pD3D11CreateDevice || !pCreateDXGIFactory1) { + mp_fatal(log, "Failed to load base d3d11 functionality: " + "CreateDevice: %s, CreateDXGIFactory1: %s\n", + pD3D11CreateDevice ? "success" : "failure", + pCreateDXGIFactory1 ? "success": "failure"); + return false; + } + + return true; +} + +#define D3D11_DXGI_ENUM(prefix, define) { case prefix ## define: return #define; } + +static const char *d3d11_get_format_name(DXGI_FORMAT fmt) +{ + switch (fmt) { + D3D11_DXGI_ENUM(DXGI_FORMAT_, UNKNOWN); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32A32_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32A32_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32A32_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32A32_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32B32_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16B16A16_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16B16A16_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16B16A16_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16B16A16_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16B16A16_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16B16A16_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G32_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32G8X24_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, D32_FLOAT_S8X24_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32_FLOAT_X8X24_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, X32_TYPELESS_G8X24_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R10G10B10A2_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R10G10B10A2_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R10G10B10A2_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R11G11B10_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8B8A8_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8B8A8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8B8A8_UNORM_SRGB); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8B8A8_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8B8A8_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8B8A8_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16G16_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, D32_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R32_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R24G8_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, D24_UNORM_S8_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R24_UNORM_X8_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, X24_TYPELESS_G8_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16_FLOAT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, D16_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R16_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8_UINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8_SINT); + D3D11_DXGI_ENUM(DXGI_FORMAT_, A8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R1_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R9G9B9E5_SHAREDEXP); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R8G8_B8G8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, G8R8_G8B8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC1_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC1_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC1_UNORM_SRGB); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC2_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC2_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC2_UNORM_SRGB); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC3_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC3_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC3_UNORM_SRGB); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC4_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC4_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC4_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC5_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC5_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC5_SNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B5G6R5_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B5G5R5A1_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B8G8R8A8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B8G8R8X8_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, R10G10B10_XR_BIAS_A2_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B8G8R8A8_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B8G8R8A8_UNORM_SRGB); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B8G8R8X8_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B8G8R8X8_UNORM_SRGB); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC6H_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC6H_UF16); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC6H_SF16); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC7_TYPELESS); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC7_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, BC7_UNORM_SRGB); + D3D11_DXGI_ENUM(DXGI_FORMAT_, AYUV); + D3D11_DXGI_ENUM(DXGI_FORMAT_, Y410); + D3D11_DXGI_ENUM(DXGI_FORMAT_, Y416); + D3D11_DXGI_ENUM(DXGI_FORMAT_, NV12); + D3D11_DXGI_ENUM(DXGI_FORMAT_, P010); + D3D11_DXGI_ENUM(DXGI_FORMAT_, P016); + D3D11_DXGI_ENUM(DXGI_FORMAT_, 420_OPAQUE); + D3D11_DXGI_ENUM(DXGI_FORMAT_, YUY2); + D3D11_DXGI_ENUM(DXGI_FORMAT_, Y210); + D3D11_DXGI_ENUM(DXGI_FORMAT_, Y216); + D3D11_DXGI_ENUM(DXGI_FORMAT_, NV11); + D3D11_DXGI_ENUM(DXGI_FORMAT_, AI44); + D3D11_DXGI_ENUM(DXGI_FORMAT_, IA44); + D3D11_DXGI_ENUM(DXGI_FORMAT_, P8); + D3D11_DXGI_ENUM(DXGI_FORMAT_, A8P8); + D3D11_DXGI_ENUM(DXGI_FORMAT_, B4G4R4A4_UNORM); + D3D11_DXGI_ENUM(DXGI_FORMAT_, P208); + D3D11_DXGI_ENUM(DXGI_FORMAT_, V208); + D3D11_DXGI_ENUM(DXGI_FORMAT_, V408); + D3D11_DXGI_ENUM(DXGI_FORMAT_, FORCE_UINT); + default: + return "<Unknown>"; + } +} + +static const char *d3d11_get_csp_name(DXGI_COLOR_SPACE_TYPE csp) +{ + switch (csp) { + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_FULL_G22_NONE_P709); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_FULL_G10_NONE_P709); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_STUDIO_G22_NONE_P709); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_STUDIO_G22_NONE_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RESERVED); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_FULL_G22_NONE_P709_X601); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G22_LEFT_P601); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_FULL_G22_LEFT_P601); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G22_LEFT_P709); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_FULL_G22_LEFT_P709); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G22_LEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_FULL_G22_LEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_FULL_G2084_NONE_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G2084_LEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_STUDIO_G2084_NONE_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G22_TOPLEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G2084_TOPLEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_FULL_G22_NONE_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_GHLG_TOPLEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_FULL_GHLG_TOPLEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_STUDIO_G24_NONE_P709); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, RGB_STUDIO_G24_NONE_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G24_LEFT_P709); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G24_LEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, YCBCR_STUDIO_G24_TOPLEFT_P2020); + D3D11_DXGI_ENUM(DXGI_COLOR_SPACE_, CUSTOM); + default: + return "<Unknown>"; + } +} + +static bool d3d11_get_mp_csp(DXGI_COLOR_SPACE_TYPE csp, + struct mp_colorspace *mp_csp) +{ + if (!mp_csp) + return false; + + // Colorspaces utilizing gamma 2.2 (G22) are set to + // AUTO as that keeps the current default flow regarding + // SDR transfer function handling. + // (no adjustment is done unless the user has a CMS LUT). + // + // Additionally, only set primary information with colorspaces + // utilizing non-709 primaries to keep the current behavior + // regarding not doing conversion from BT.601 to BT.709. + switch (csp) { + case DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P709: + *mp_csp = (struct mp_colorspace){ + .gamma = MP_CSP_TRC_AUTO, + .primaries = MP_CSP_PRIM_AUTO, + }; + break; + case DXGI_COLOR_SPACE_RGB_FULL_G10_NONE_P709: + *mp_csp = (struct mp_colorspace) { + .gamma = MP_CSP_TRC_LINEAR, + .primaries = MP_CSP_PRIM_AUTO, + }; + break; + case DXGI_COLOR_SPACE_RGB_FULL_G2084_NONE_P2020: + *mp_csp = (struct mp_colorspace) { + .gamma = MP_CSP_TRC_PQ, + .primaries = MP_CSP_PRIM_BT_2020, + }; + break; + case DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P2020: + *mp_csp = (struct mp_colorspace) { + .gamma = MP_CSP_TRC_AUTO, + .primaries = MP_CSP_PRIM_BT_2020, + }; + break; + default: + return false; + } + + return true; +} + +static bool query_output_format_and_colorspace(struct mp_log *log, + IDXGISwapChain *swapchain, + DXGI_FORMAT *out_fmt, + DXGI_COLOR_SPACE_TYPE *out_cspace) +{ + IDXGIOutput *output = NULL; + IDXGIOutput6 *output6 = NULL; + DXGI_OUTPUT_DESC1 desc = { 0 }; + char *monitor_name = NULL; + bool success = false; + + if (!out_fmt || !out_cspace) + return false; + + HRESULT hr = IDXGISwapChain_GetContainingOutput(swapchain, &output); + if (FAILED(hr)) { + mp_err(log, "Failed to get swap chain's containing output: %s!\n", + mp_HRESULT_to_str(hr)); + goto done; + } + + hr = IDXGIOutput_QueryInterface(output, &IID_IDXGIOutput6, + (void**)&output6); + if (FAILED(hr)) { + // point where systems older than Windows 10 would fail, + // thus utilizing error log level only with windows 10+ + mp_msg(log, IsWindows10OrGreater() ? MSGL_ERR : MSGL_V, + "Failed to create a DXGI 1.6 output interface: %s\n", + mp_HRESULT_to_str(hr)); + goto done; + } + + hr = IDXGIOutput6_GetDesc1(output6, &desc); + if (FAILED(hr)) { + mp_err(log, "Failed to query swap chain's output information: %s\n", + mp_HRESULT_to_str(hr)); + goto done; + } + + monitor_name = mp_to_utf8(NULL, desc.DeviceName); + + mp_verbose(log, "Queried output: %s, %ldx%ld @ %d bits, colorspace: %s (%d)\n", + monitor_name, + desc.DesktopCoordinates.right - desc.DesktopCoordinates.left, + desc.DesktopCoordinates.bottom - desc.DesktopCoordinates.top, + desc.BitsPerColor, + d3d11_get_csp_name(desc.ColorSpace), + desc.ColorSpace); + + *out_cspace = desc.ColorSpace; + + // limit ourselves to the 8bit and 10bit formats for now. + // while the 16bit float format would be preferable as something + // to default to, it seems to be hard-coded to linear transfer + // in windowed mode, and follows configured colorspace in full screen. + *out_fmt = desc.BitsPerColor > 8 ? + DXGI_FORMAT_R10G10B10A2_UNORM : DXGI_FORMAT_R8G8B8A8_UNORM; + + success = true; + +done: + talloc_free(monitor_name); + SAFE_RELEASE(output6); + SAFE_RELEASE(output); + return success; +} + +// Get a const array of D3D_FEATURE_LEVELs from max_fl to min_fl (inclusive) +static int get_feature_levels(int max_fl, int min_fl, + const D3D_FEATURE_LEVEL **out) +{ + static const D3D_FEATURE_LEVEL levels[] = { + D3D_FEATURE_LEVEL_12_1, + D3D_FEATURE_LEVEL_12_0, + D3D_FEATURE_LEVEL_11_1, + D3D_FEATURE_LEVEL_11_0, + D3D_FEATURE_LEVEL_10_1, + D3D_FEATURE_LEVEL_10_0, + D3D_FEATURE_LEVEL_9_3, + D3D_FEATURE_LEVEL_9_2, + D3D_FEATURE_LEVEL_9_1, + }; + static const int levels_len = MP_ARRAY_SIZE(levels); + + int start = 0; + for (; start < levels_len; start++) { + if (levels[start] <= max_fl) + break; + } + int len = 0; + for (; start + len < levels_len; len++) { + if (levels[start + len] < min_fl) + break; + } + *out = &levels[start]; + return len; +} + +static IDXGIAdapter1 *get_d3d11_adapter(struct mp_log *log, + struct bstr requested_adapter_name, + struct bstr *listing) +{ + HRESULT hr = S_OK; + IDXGIFactory1 *factory; + IDXGIAdapter1 *picked_adapter = NULL; + + hr = pCreateDXGIFactory1(&IID_IDXGIFactory1, (void **)&factory); + if (FAILED(hr)) { + mp_fatal(log, "Failed to create a DXGI factory: %s\n", + mp_HRESULT_to_str(hr)); + return NULL; + } + + for (unsigned int adapter_num = 0; hr != DXGI_ERROR_NOT_FOUND; adapter_num++) + { + IDXGIAdapter1 *adapter = NULL; + DXGI_ADAPTER_DESC1 desc = { 0 }; + char *adapter_description = NULL; + + hr = IDXGIFactory1_EnumAdapters1(factory, adapter_num, &adapter); + if (FAILED(hr)) { + if (hr != DXGI_ERROR_NOT_FOUND) { + mp_fatal(log, "Failed to enumerate at adapter %u\n", + adapter_num); + } + continue; + } + + if (FAILED(IDXGIAdapter1_GetDesc1(adapter, &desc))) { + mp_fatal(log, "Failed to get adapter description when listing at adapter %u\n", + adapter_num); + continue; + } + + adapter_description = mp_to_utf8(NULL, desc.Description); + + if (listing) { + bstr_xappend_asprintf(NULL, listing, + "Adapter %u: vendor: %u, description: %s\n", + adapter_num, desc.VendorId, + adapter_description); + } + + if (requested_adapter_name.len && + bstr_case_startswith(bstr0(adapter_description), + requested_adapter_name)) + { + picked_adapter = adapter; + } + + talloc_free(adapter_description); + + if (picked_adapter) { + break; + } + + SAFE_RELEASE(adapter); + } + + SAFE_RELEASE(factory); + + return picked_adapter; +} + +static HRESULT create_device(struct mp_log *log, IDXGIAdapter1 *adapter, + bool warp, bool debug, int max_fl, int min_fl, + ID3D11Device **dev) +{ + const D3D_FEATURE_LEVEL *levels; + int levels_len = get_feature_levels(max_fl, min_fl, &levels); + if (!levels_len) { + mp_fatal(log, "No suitable Direct3D feature level found\n"); + return E_FAIL; + } + + D3D_DRIVER_TYPE type = warp ? D3D_DRIVER_TYPE_WARP + : D3D_DRIVER_TYPE_HARDWARE; + UINT flags = debug ? D3D11_CREATE_DEVICE_DEBUG : 0; + return pD3D11CreateDevice((IDXGIAdapter *)adapter, adapter ? D3D_DRIVER_TYPE_UNKNOWN : type, + NULL, flags, levels, levels_len, D3D11_SDK_VERSION, dev, NULL, NULL); +} + +bool mp_d3d11_list_or_verify_adapters(struct mp_log *log, + bstr adapter_name, + bstr *listing) +{ + IDXGIAdapter1 *picked_adapter = NULL; + + if (!load_d3d11_functions(log)) { + return false; + } + + if ((picked_adapter = get_d3d11_adapter(log, adapter_name, listing))) { + SAFE_RELEASE(picked_adapter); + return true; + } + + return false; +} + +// Create a Direct3D 11 device for rendering and presentation. This is meant to +// reduce boilerplate in backends that D3D11, while also making sure they share +// the same device creation logic and log the same information. +bool mp_d3d11_create_present_device(struct mp_log *log, + struct d3d11_device_opts *opts, + ID3D11Device **dev_out) +{ + bool debug = opts->debug; + bool warp = opts->force_warp; + int max_fl = opts->max_feature_level; + int min_fl = opts->min_feature_level; + // Normalize nullptr and an empty string to nullptr to simplify handling. + char *adapter_name = (opts->adapter_name && *(opts->adapter_name)) ? + opts->adapter_name : NULL; + ID3D11Device *dev = NULL; + IDXGIDevice1 *dxgi_dev = NULL; + IDXGIAdapter1 *adapter = NULL; + bool success = false; + HRESULT hr; + + if (!load_d3d11_functions(log)) { + goto done; + } + + adapter = get_d3d11_adapter(log, bstr0(adapter_name), NULL); + + if (adapter_name && !adapter) { + mp_warn(log, "Adapter matching '%s' was not found in the system! " + "Will fall back to the default adapter.\n", + adapter_name); + } + + // Return here to retry creating the device + do { + // Use these default feature levels if they are not set + max_fl = max_fl ? max_fl : D3D_FEATURE_LEVEL_11_0; + min_fl = min_fl ? min_fl : D3D_FEATURE_LEVEL_9_1; + + hr = create_device(log, adapter, warp, debug, max_fl, min_fl, &dev); + + // Retry without debug, if SDK is not available + if (debug && hr == DXGI_ERROR_SDK_COMPONENT_MISSING) { + mp_warn(log, "gpu-debug disabled due to error: %s\n", mp_HRESULT_to_str(hr)); + debug = false; + continue; + } + + if (SUCCEEDED(hr)) + break; + + // Trying to create a D3D_FEATURE_LEVEL_12_0 device on Windows 8.1 or + // below will not succeed. Try an 11_1 device. + if (max_fl >= D3D_FEATURE_LEVEL_12_0 && + min_fl <= D3D_FEATURE_LEVEL_11_1) + { + mp_dbg(log, "Failed to create 12_0+ device, trying 11_1\n"); + max_fl = D3D_FEATURE_LEVEL_11_1; + continue; + } + + // Trying to create a D3D_FEATURE_LEVEL_11_1 device on Windows 7 + // without the platform update will not succeed. Try an 11_0 device. + if (max_fl >= D3D_FEATURE_LEVEL_11_1 && + min_fl <= D3D_FEATURE_LEVEL_11_0) + { + mp_dbg(log, "Failed to create 11_1+ device, trying 11_0\n"); + max_fl = D3D_FEATURE_LEVEL_11_0; + continue; + } + + // Retry with WARP if allowed + if (!warp && opts->allow_warp) { + mp_dbg(log, "Failed to create hardware device, trying WARP\n"); + warp = true; + max_fl = opts->max_feature_level; + min_fl = opts->min_feature_level; + continue; + } + + mp_fatal(log, "Failed to create Direct3D 11 device: %s\n", + mp_HRESULT_to_str(hr)); + goto done; + } while (true); + + // if we picked an adapter, release it here - we're taking another + // from the device. + SAFE_RELEASE(adapter); + + hr = ID3D11Device_QueryInterface(dev, &IID_IDXGIDevice1, (void**)&dxgi_dev); + if (FAILED(hr)) { + mp_fatal(log, "Failed to get DXGI device\n"); + goto done; + } + hr = IDXGIDevice1_GetParent(dxgi_dev, &IID_IDXGIAdapter1, (void**)&adapter); + if (FAILED(hr)) { + mp_fatal(log, "Failed to get DXGI adapter\n"); + goto done; + } + + IDXGIDevice1_SetMaximumFrameLatency(dxgi_dev, opts->max_frame_latency); + + DXGI_ADAPTER_DESC1 desc; + hr = IDXGIAdapter1_GetDesc1(adapter, &desc); + if (FAILED(hr)) { + mp_fatal(log, "Failed to get adapter description\n"); + goto done; + } + + D3D_FEATURE_LEVEL selected_level = ID3D11Device_GetFeatureLevel(dev); + mp_verbose(log, "Using Direct3D 11 feature level %u_%u\n", + ((unsigned)selected_level) >> 12, + (((unsigned)selected_level) >> 8) & 0xf); + + char *dev_name = mp_to_utf8(NULL, desc.Description); + mp_verbose(log, "Device Name: %s\n" + "Device ID: %04x:%04x (rev %02x)\n" + "Subsystem ID: %04x:%04x\n" + "LUID: %08lx%08lx\n", + dev_name, + desc.VendorId, desc.DeviceId, desc.Revision, + LOWORD(desc.SubSysId), HIWORD(desc.SubSysId), + desc.AdapterLuid.HighPart, desc.AdapterLuid.LowPart); + talloc_free(dev_name); + + if (desc.Flags & DXGI_ADAPTER_FLAG_SOFTWARE) + warp = true; + // If the primary display adapter is a software adapter, the + // DXGI_ADAPTER_FLAG_SOFTWARE flag won't be set, but the device IDs should + // still match the Microsoft Basic Render Driver + if (desc.VendorId == 0x1414 && desc.DeviceId == 0x8c) + warp = true; + if (warp) { + mp_msg(log, opts->force_warp ? MSGL_V : MSGL_WARN, + "Using a software adapter\n"); + } + + *dev_out = dev; + dev = NULL; + success = true; + +done: + SAFE_RELEASE(adapter); + SAFE_RELEASE(dxgi_dev); + SAFE_RELEASE(dev); + return success; +} + +static HRESULT create_swapchain_1_2(ID3D11Device *dev, IDXGIFactory2 *factory, + struct mp_log *log, + struct d3d11_swapchain_opts *opts, + bool flip, DXGI_FORMAT format, + IDXGISwapChain **swapchain_out) +{ + IDXGISwapChain *swapchain = NULL; + IDXGISwapChain1 *swapchain1 = NULL; + HRESULT hr; + + DXGI_SWAP_CHAIN_DESC1 desc = { + .Width = opts->width ? opts->width : 1, + .Height = opts->height ? opts->height : 1, + .Format = format, + .SampleDesc = { .Count = 1 }, + .BufferUsage = opts->usage, + }; + + if (flip) { + // UNORDERED_ACCESS with FLIP_SEQUENTIAL seems to be buggy with + // Windows 7 drivers + if ((desc.BufferUsage & DXGI_USAGE_UNORDERED_ACCESS) && + !IsWindows8OrGreater()) + { + mp_verbose(log, "Disabling UNORDERED_ACCESS for flip-model " + "swapchain backbuffers in Windows 7\n"); + desc.BufferUsage &= ~DXGI_USAGE_UNORDERED_ACCESS; + } + + if (IsWindows10OrGreater()) { + desc.SwapEffect = DXGI_SWAP_EFFECT_FLIP_DISCARD; + } else { + desc.SwapEffect = DXGI_SWAP_EFFECT_FLIP_SEQUENTIAL; + } + desc.BufferCount = opts->length; + } else { + desc.SwapEffect = DXGI_SWAP_EFFECT_DISCARD; + desc.BufferCount = 1; + } + + hr = IDXGIFactory2_CreateSwapChainForHwnd(factory, (IUnknown*)dev, + opts->window, &desc, NULL, NULL, &swapchain1); + if (FAILED(hr)) + goto done; + hr = IDXGISwapChain1_QueryInterface(swapchain1, &IID_IDXGISwapChain, + (void**)&swapchain); + if (FAILED(hr)) + goto done; + + *swapchain_out = swapchain; + swapchain = NULL; + +done: + SAFE_RELEASE(swapchain1); + SAFE_RELEASE(swapchain); + return hr; +} + +static HRESULT create_swapchain_1_1(ID3D11Device *dev, IDXGIFactory1 *factory, + struct mp_log *log, + struct d3d11_swapchain_opts *opts, + DXGI_FORMAT format, + IDXGISwapChain **swapchain_out) +{ + DXGI_SWAP_CHAIN_DESC desc = { + .BufferDesc = { + .Width = opts->width ? opts->width : 1, + .Height = opts->height ? opts->height : 1, + .Format = format, + }, + .SampleDesc = { .Count = 1 }, + .BufferUsage = opts->usage, + .BufferCount = 1, + .OutputWindow = opts->window, + .Windowed = TRUE, + .SwapEffect = DXGI_SWAP_EFFECT_DISCARD, + }; + + return IDXGIFactory1_CreateSwapChain(factory, (IUnknown*)dev, &desc, + swapchain_out); +} + +static bool update_swapchain_format(struct mp_log *log, + IDXGISwapChain *swapchain, + DXGI_FORMAT format) +{ + DXGI_SWAP_CHAIN_DESC desc; + + HRESULT hr = IDXGISwapChain_GetDesc(swapchain, &desc); + if (FAILED(hr)) { + mp_fatal(log, "Failed to query swap chain's current state: %s\n", + mp_HRESULT_to_str(hr)); + return false; + } + + hr = IDXGISwapChain_ResizeBuffers(swapchain, 0, desc.BufferDesc.Width, + desc.BufferDesc.Height, + format, 0); + if (FAILED(hr)) { + mp_fatal(log, "Couldn't update swapchain format: %s\n", + mp_HRESULT_to_str(hr)); + return false; + } + + return true; +} + +static bool update_swapchain_color_space(struct mp_log *log, + IDXGISwapChain *swapchain, + DXGI_COLOR_SPACE_TYPE color_space) +{ + IDXGISwapChain4 *swapchain4 = NULL; + const char *csp_name = d3d11_get_csp_name(color_space); + bool success = false; + HRESULT hr = E_FAIL; + unsigned int csp_support_flags; + + hr = IDXGISwapChain_QueryInterface(swapchain, &IID_IDXGISwapChain4, + (void *)&(swapchain4)); + if (FAILED(hr)) { + mp_err(log, "Failed to create v4 swapchain for color space " + "configuration (%s)!\n", + mp_HRESULT_to_str(hr)); + goto done; + } + + hr = IDXGISwapChain4_CheckColorSpaceSupport(swapchain4, + color_space, + &csp_support_flags); + if (FAILED(hr)) { + mp_err(log, "Failed to check color space support for color space " + "%s (%d): %s!\n", + csp_name, color_space, mp_HRESULT_to_str(hr)); + goto done; + } + + mp_verbose(log, + "Swapchain capabilities for color space %s (%d): " + "normal: %s, overlay: %s\n", + csp_name, color_space, + (csp_support_flags & DXGI_SWAP_CHAIN_COLOR_SPACE_SUPPORT_FLAG_PRESENT) ? + "yes" : "no", + (csp_support_flags & DXGI_SWAP_CHAIN_COLOR_SPACE_SUPPORT_FLAG_OVERLAY_PRESENT) ? + "yes" : "no"); + + if (!(csp_support_flags & DXGI_SWAP_CHAIN_COLOR_SPACE_SUPPORT_FLAG_PRESENT)) { + mp_err(log, "Color space %s (%d) is not supported by this swapchain!\n", + csp_name, color_space); + goto done; + } + + hr = IDXGISwapChain4_SetColorSpace1(swapchain4, color_space); + if (FAILED(hr)) { + mp_err(log, "Failed to set color space %s (%d) for this swapchain " + "(%s)!\n", + csp_name, color_space, mp_HRESULT_to_str(hr)); + goto done; + } + + mp_verbose(log, "Swapchain successfully configured to color space %s (%d)!\n", + csp_name, color_space); + + success = true; + +done: + SAFE_RELEASE(swapchain4); + return success; +} + +static bool configure_created_swapchain(struct mp_log *log, + IDXGISwapChain *swapchain, + DXGI_FORMAT requested_format, + DXGI_COLOR_SPACE_TYPE requested_csp, + struct mp_colorspace *configured_csp) +{ + DXGI_FORMAT probed_format = DXGI_FORMAT_UNKNOWN; + DXGI_FORMAT selected_format = DXGI_FORMAT_UNKNOWN; + DXGI_COLOR_SPACE_TYPE probed_colorspace = DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P709; + DXGI_COLOR_SPACE_TYPE selected_colorspace; + const char *format_name = NULL; + const char *csp_name = NULL; + struct mp_colorspace mp_csp = { 0 }; + bool mp_csp_mapped = false; + + query_output_format_and_colorspace(log, swapchain, + &probed_format, + &probed_colorspace); + + + selected_format = requested_format != DXGI_FORMAT_UNKNOWN ? + requested_format : + (probed_format != DXGI_FORMAT_UNKNOWN ? + probed_format : DXGI_FORMAT_R8G8B8A8_UNORM); + selected_colorspace = requested_csp != -1 ? + requested_csp : probed_colorspace; + format_name = d3d11_get_format_name(selected_format); + csp_name = d3d11_get_csp_name(selected_colorspace); + mp_csp_mapped = d3d11_get_mp_csp(selected_colorspace, &mp_csp); + + mp_verbose(log, "Selected swapchain format %s (%d), attempting " + "to utilize it.\n", + format_name, selected_format); + + if (!update_swapchain_format(log, swapchain, selected_format)) { + return false; + } + + if (!IsWindows10OrGreater()) { + // On older than Windows 10, query_output_format_and_colorspace + // will not change probed_colorspace, and even if a user sets + // a colorspace it will not get applied. Thus warn user in case a + // value was specifically set and finish. + if (requested_csp != -1) { + mp_warn(log, "User selected a D3D11 color space %s (%d), " + "but configuration of color spaces is only supported" + "from Windows 10! The default configuration has been " + "left as-is.\n", + csp_name, selected_colorspace); + } + + return true; + } + + if (!mp_csp_mapped) { + mp_warn(log, "Color space %s (%d) does not have an mpv color space " + "mapping! Overriding to standard sRGB!\n", + csp_name, selected_colorspace); + selected_colorspace = DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P709; + d3d11_get_mp_csp(selected_colorspace, &mp_csp); + } + + mp_verbose(log, "Selected swapchain color space %s (%d), attempting to " + "utilize it.\n", + csp_name, selected_colorspace); + + if (!update_swapchain_color_space(log, swapchain, selected_colorspace)) { + return false; + } + + if (configured_csp) { + *configured_csp = mp_csp; + } + + return true; +} + +// Create a Direct3D 11 swapchain +bool mp_d3d11_create_swapchain(ID3D11Device *dev, struct mp_log *log, + struct d3d11_swapchain_opts *opts, + IDXGISwapChain **swapchain_out) +{ + IDXGIDevice1 *dxgi_dev = NULL; + IDXGIAdapter1 *adapter = NULL; + IDXGIFactory1 *factory = NULL; + IDXGIFactory2 *factory2 = NULL; + IDXGISwapChain *swapchain = NULL; + bool success = false; + HRESULT hr; + + hr = ID3D11Device_QueryInterface(dev, &IID_IDXGIDevice1, (void**)&dxgi_dev); + if (FAILED(hr)) { + mp_fatal(log, "Failed to get DXGI device\n"); + goto done; + } + hr = IDXGIDevice1_GetParent(dxgi_dev, &IID_IDXGIAdapter1, (void**)&adapter); + if (FAILED(hr)) { + mp_fatal(log, "Failed to get DXGI adapter\n"); + goto done; + } + hr = IDXGIAdapter1_GetParent(adapter, &IID_IDXGIFactory1, (void**)&factory); + if (FAILED(hr)) { + mp_fatal(log, "Failed to get DXGI factory\n"); + goto done; + } + hr = IDXGIFactory1_QueryInterface(factory, &IID_IDXGIFactory2, + (void**)&factory2); + if (FAILED(hr)) + factory2 = NULL; + + bool flip = factory2 && opts->flip; + + // Return here to retry creating the swapchain + do { + if (factory2) { + // Create a DXGI 1.2+ (Windows 8+) swap chain if possible + hr = create_swapchain_1_2(dev, factory2, log, opts, flip, + DXGI_FORMAT_R8G8B8A8_UNORM, &swapchain); + } else { + // Fall back to DXGI 1.1 (Windows 7) + hr = create_swapchain_1_1(dev, factory, log, opts, + DXGI_FORMAT_R8G8B8A8_UNORM, &swapchain); + } + if (SUCCEEDED(hr)) + break; + + if (flip) { + mp_dbg(log, "Failed to create flip-model swapchain, trying bitblt\n"); + flip = false; + continue; + } + + mp_fatal(log, "Failed to create swapchain: %s\n", mp_HRESULT_to_str(hr)); + goto done; + } while (true); + + // Prevent DXGI from making changes to the VO window, otherwise it will + // hook the Alt+Enter keystroke and make it trigger an ugly transition to + // exclusive fullscreen mode instead of running the user-set command. + IDXGIFactory_MakeWindowAssociation(factory, opts->window, + DXGI_MWA_NO_WINDOW_CHANGES | DXGI_MWA_NO_ALT_ENTER | + DXGI_MWA_NO_PRINT_SCREEN); + + if (factory2) { + mp_verbose(log, "Using DXGI 1.2+\n"); + } else { + mp_verbose(log, "Using DXGI 1.1\n"); + } + + configure_created_swapchain(log, swapchain, opts->format, + opts->color_space, + opts->configured_csp); + + DXGI_SWAP_CHAIN_DESC scd = {0}; + IDXGISwapChain_GetDesc(swapchain, &scd); + if (scd.SwapEffect == DXGI_SWAP_EFFECT_FLIP_SEQUENTIAL || + scd.SwapEffect == DXGI_SWAP_EFFECT_FLIP_DISCARD) + { + mp_verbose(log, "Using flip-model presentation\n"); + } else { + mp_verbose(log, "Using bitblt-model presentation\n"); + } + + *swapchain_out = swapchain; + swapchain = NULL; + success = true; + +done: + SAFE_RELEASE(swapchain); + SAFE_RELEASE(factory2); + SAFE_RELEASE(factory); + SAFE_RELEASE(adapter); + SAFE_RELEASE(dxgi_dev); + return success; +} diff --git a/video/out/gpu/d3d11_helpers.h b/video/out/gpu/d3d11_helpers.h new file mode 100644 index 0000000..c115d33 --- /dev/null +++ b/video/out/gpu/d3d11_helpers.h @@ -0,0 +1,103 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#ifndef MP_D3D11_HELPERS_H_ +#define MP_D3D11_HELPERS_H_ + +#include <stdbool.h> +#include <windows.h> +#include <d3d11.h> +#include <dxgi1_2.h> + +#include "video/mp_image.h" + +#define D3D_FEATURE_LEVEL_12_0 (0xc000) +#define D3D_FEATURE_LEVEL_12_1 (0xc100) + +#define DXGI_COLOR_SPACE_RGB_STUDIO_G24_NONE_P709 ((DXGI_COLOR_SPACE_TYPE)20) +#define DXGI_COLOR_SPACE_RGB_STUDIO_G24_NONE_P2020 ((DXGI_COLOR_SPACE_TYPE)21) +#define DXGI_COLOR_SPACE_YCBCR_STUDIO_G24_LEFT_P709 ((DXGI_COLOR_SPACE_TYPE)22) +#define DXGI_COLOR_SPACE_YCBCR_STUDIO_G24_LEFT_P2020 ((DXGI_COLOR_SPACE_TYPE)23) +#define DXGI_COLOR_SPACE_YCBCR_STUDIO_G24_TOPLEFT_P2020 ((DXGI_COLOR_SPACE_TYPE)24) + +struct d3d11_device_opts { + // Enable the debug layer (D3D11_CREATE_DEVICE_DEBUG) + bool debug; + + // Allow a software (WARP) adapter. Note, sometimes a software adapter will + // be used even when allow_warp is false. This is because, on Windows 8 and + // up, if there are no hardware adapters, Windows will pretend the WARP + // adapter is the primary hardware adapter. + bool allow_warp; + + // Always use a WARP adapter. This is mainly for testing purposes. + bool force_warp; + + // The maximum number of pending frames allowed to be queued to a swapchain + int max_frame_latency; + + // The maximum Direct3D 11 feature level to attempt to create + // If unset, defaults to D3D_FEATURE_LEVEL_11_0 + int max_feature_level; + + // The minimum Direct3D 11 feature level to attempt to create. If this is + // not supported, device creation will fail. + // If unset, defaults to D3D_FEATURE_LEVEL_9_1 + int min_feature_level; + + // The adapter name to utilize if a specific adapter is required + // If unset, the default adapter will be utilized when creating + // a device. + char *adapter_name; +}; + +bool mp_d3d11_list_or_verify_adapters(struct mp_log *log, + bstr adapter_name, + bstr *listing); + +bool mp_d3d11_create_present_device(struct mp_log *log, + struct d3d11_device_opts *opts, + ID3D11Device **dev_out); + +struct d3d11_swapchain_opts { + HWND window; + int width; + int height; + DXGI_FORMAT format; + DXGI_COLOR_SPACE_TYPE color_space; + + // mp_colorspace mapping of the configured swapchain colorspace + // shall be written into this memory location if configuration + // succeeds. Will be ignored if NULL. + struct mp_colorspace *configured_csp; + + // Use DXGI_SWAP_EFFECT_FLIP_SEQUENTIAL if possible + bool flip; + + // Number of surfaces in the swapchain + int length; + + // The BufferUsage value for swapchain surfaces. This should probably + // contain DXGI_USAGE_RENDER_TARGET_OUTPUT. + DXGI_USAGE usage; +}; + +bool mp_d3d11_create_swapchain(ID3D11Device *dev, struct mp_log *log, + struct d3d11_swapchain_opts *opts, + IDXGISwapChain **swapchain_out); + +#endif diff --git a/video/out/gpu/error_diffusion.c b/video/out/gpu/error_diffusion.c new file mode 100644 index 0000000..c1ea542 --- /dev/null +++ b/video/out/gpu/error_diffusion.c @@ -0,0 +1,316 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <stdlib.h> + +#include "error_diffusion.h" + +#include "common/common.h" + +#define GLSL(...) gl_sc_addf(sc, __VA_ARGS__) +#define GLSLH(...) gl_sc_haddf(sc, __VA_ARGS__) + +// After a (y, x) -> (y, x + y * shift) mapping, find the right most column that +// will be affected by the current column. +static int compute_rightmost_shifted_column(const struct error_diffusion_kernel *k) +{ + int ret = 0; + for (int y = 0; y <= EF_MAX_DELTA_Y; y++) { + for (int x = EF_MIN_DELTA_X; x <= EF_MAX_DELTA_X; x++) { + if (k->pattern[y][x - EF_MIN_DELTA_X] != 0) { + int shifted_x = x + y * k->shift; + + // The shift mapping guarantees current column (or left of it) + // won't be affected by error diffusion. + assert(shifted_x > 0); + + ret = MPMAX(ret, shifted_x); + } + } + } + return ret; +} + +const struct error_diffusion_kernel *mp_find_error_diffusion_kernel(const char *name) +{ + if (!name) + return NULL; + for (const struct error_diffusion_kernel *k = mp_error_diffusion_kernels; + k->name; + k++) { + if (strcmp(k->name, name) == 0) + return k; + } + return NULL; +} + +int mp_ef_compute_shared_memory_size(const struct error_diffusion_kernel *k, + int height) +{ + // We add EF_MAX_DELTA_Y empty lines on the bottom to handle errors + // propagated out from bottom side. + int rows = height + EF_MAX_DELTA_Y; + int shifted_columns = compute_rightmost_shifted_column(k) + 1; + + // The shared memory is an array of size rows*shifted_columns. Each element + // is a single uint for three RGB component. + return rows * shifted_columns * 4; +} + +void pass_error_diffusion(struct gl_shader_cache *sc, + const struct error_diffusion_kernel *k, + int tex, int width, int height, int depth, int block_size) +{ + assert(block_size <= height); + + // The parallel error diffusion works by applying the shift mapping first. + // Taking the Floyd and Steinberg algorithm for example. After applying + // the (y, x) -> (y, x + y * shift) mapping (with shift=2), all errors are + // propagated into the next few columns, which makes parallel processing on + // the same column possible. + // + // X 7/16 X 7/16 + // 3/16 5/16 1/16 ==> 0 0 3/16 5/16 1/16 + + // Figuring out the size of rectangle containing all shifted pixels. + // The rectangle height is not changed. + int shifted_width = width + (height - 1) * k->shift; + + // We process all pixels from the shifted rectangles column by column, with + // a single global work group of size |block_size|. + // Figuring out how many block are required to process all pixels. We need + // this explicitly to make the number of barrier() calls match. + int blocks = (height * shifted_width + block_size - 1) / block_size; + + // If we figure out how many of the next columns will be affected while the + // current columns is being processed. We can store errors of only a few + // columns in the shared memory. Using a ring buffer will further save the + // cost while iterating to next column. + int ring_buffer_rows = height + EF_MAX_DELTA_Y; + int ring_buffer_columns = compute_rightmost_shifted_column(k) + 1; + int ring_buffer_size = ring_buffer_rows * ring_buffer_columns; + + // Defines the ring buffer in shared memory. + GLSLH("shared uint err_rgb8[%d];\n", ring_buffer_size); + + // Initialize the ring buffer. + GLSL("for (int i = int(gl_LocalInvocationIndex); i < %d; i += %d) ", + ring_buffer_size, block_size); + GLSL("err_rgb8[i] = 0u;\n"); + + GLSL("for (int block_id = 0; block_id < %d; ++block_id) {\n", blocks); + + // Add barrier here to have previous block all processed before starting + // the processing of the next. + GLSL("groupMemoryBarrier();\n"); + GLSL("barrier();\n"); + + // Compute the coordinate of the pixel we are currently processing, both + // before and after the shift mapping. + GLSL("int id = int(gl_LocalInvocationIndex) + block_id * %d;\n", block_size); + GLSL("int y = id %% %d, x_shifted = id / %d;\n", height, height); + GLSL("int x = x_shifted - y * %d;\n", k->shift); + + // Proceed only if we are processing a valid pixel. + GLSL("if (0 <= x && x < %d) {\n", width); + + // The index that the current pixel have on the ring buffer. + GLSL("int idx = (x_shifted * %d + y) %% %d;\n", ring_buffer_rows, ring_buffer_size); + + // Fetch the current pixel. + GLSL("vec3 pix = texelFetch(texture%d, ivec2(x, y), 0).rgb;\n", tex); + + // The dithering will quantize pixel value into multiples of 1/dither_quant. + int dither_quant = (1 << depth) - 1; + + // We encode errors in RGB components into a single 32-bit unsigned integer. + // The error we propagate from the current pixel is in range of + // [-0.5 / dither_quant, 0.5 / dither_quant]. While not quite obvious, the + // sum of all errors been propagated into a pixel is also in the same range. + // It's possible to map errors in this range into [-127, 127], and use an + // unsigned 8-bit integer to store it (using standard two's complement). + // The three 8-bit unsigned integers can then be encoded into a single + // 32-bit unsigned integer, with two 4-bit padding to prevent addition + // operation overflows affecting other component. There are at most 12 + // addition operations on each pixel, so 4-bit padding should be enough. + // The overflow from R component will be discarded. + // + // The following figure is how the encoding looks like. + // + // +------------------------------------+ + // |RRRRRRRR|0000|GGGGGGGG|0000|BBBBBBBB| + // +------------------------------------+ + // + + // The bitshift position for R and G component. + int bitshift_r = 24, bitshift_g = 12; + // The multiplier we use to map [-0.5, 0.5] to [-127, 127]. + int uint8_mul = 127 * 2; + + // Adding the error previously propagated into current pixel, and clear it + // in the buffer. + GLSL("uint err_u32 = err_rgb8[idx] + %uu;\n", + (128u << bitshift_r) | (128u << bitshift_g) | 128u); + GLSL("pix = pix * %d.0 + vec3(" + "int((err_u32 >> %d) & 255u) - 128," + "int((err_u32 >> %d) & 255u) - 128," + "int( err_u32 & 255u) - 128" + ") / %d.0;\n", dither_quant, bitshift_r, bitshift_g, uint8_mul); + GLSL("err_rgb8[idx] = 0u;\n"); + + // Write the dithered pixel. + GLSL("vec3 dithered = round(pix);\n"); + GLSL("imageStore(out_image, ivec2(x, y), vec4(dithered / %d.0, 0.0));\n", + dither_quant); + + GLSL("vec3 err_divided = (pix - dithered) * %d.0 / %d.0;\n", + uint8_mul, k->divisor); + GLSL("ivec3 tmp;\n"); + + // Group error propagation with same weight factor together, in order to + // reduce the number of annoying error encoding. + for (int dividend = 1; dividend <= k->divisor; dividend++) { + bool err_assigned = false; + + for (int y = 0; y <= EF_MAX_DELTA_Y; y++) { + for (int x = EF_MIN_DELTA_X; x <= EF_MAX_DELTA_X; x++) { + if (k->pattern[y][x - EF_MIN_DELTA_X] != dividend) + continue; + + if (!err_assigned) { + err_assigned = true; + + GLSL("tmp = ivec3(round(err_divided * %d.0));\n", dividend); + + GLSL("err_u32 = " + "(uint(tmp.r & 255) << %d)|" + "(uint(tmp.g & 255) << %d)|" + " uint(tmp.b & 255);\n", + bitshift_r, bitshift_g); + } + + int shifted_x = x + y * k->shift; + + // Unlike the right border, errors propagated out from left + // border will remain in the ring buffer. This will produce + // visible artifacts near the left border, especially for + // shift=3 kernels. + if (x < 0) + GLSL("if (x >= %d) ", -x); + + // Calculate the new position in the ring buffer to propagate + // the error into. + int ring_buffer_delta = shifted_x * ring_buffer_rows + y; + GLSL("atomicAdd(err_rgb8[(idx + %d) %% %d], err_u32);\n", + ring_buffer_delta, ring_buffer_size); + } + } + } + + GLSL("}\n"); // if (0 <= x && x < width) + + GLSL("}\n"); // block_id +} + +// Different kernels for error diffusion. +// Patterns are from http://www.efg2.com/Lab/Library/ImageProcessing/DHALF.TXT +const struct error_diffusion_kernel mp_error_diffusion_kernels[] = { + { + .name = "simple", + .shift = 1, + .pattern = {{0, 0, 0, 1, 0}, + {0, 0, 1, 0, 0}, + {0, 0, 0, 0, 0}}, + .divisor = 2 + }, + { + // The "false" Floyd-Steinberg kernel + .name = "false-fs", + .shift = 1, + .pattern = {{0, 0, 0, 3, 0}, + {0, 0, 3, 2, 0}, + {0, 0, 0, 0, 0}}, + .divisor = 8 + }, + { + .name = "sierra-lite", + .shift = 2, + .pattern = {{0, 0, 0, 2, 0}, + {0, 1, 1, 0, 0}, + {0, 0, 0, 0, 0}}, + .divisor = 4 + }, + { + .name = "floyd-steinberg", + .shift = 2, + .pattern = {{0, 0, 0, 7, 0}, + {0, 3, 5, 1, 0}, + {0, 0, 0, 0, 0}}, + .divisor = 16 + }, + { + .name = "atkinson", + .shift = 2, + .pattern = {{0, 0, 0, 1, 1}, + {0, 1, 1, 1, 0}, + {0, 0, 1, 0, 0}}, + .divisor = 8 + }, + // All kernels below have shift value of 3, and probably are too heavy for + // low end GPU. + { + .name = "jarvis-judice-ninke", + .shift = 3, + .pattern = {{0, 0, 0, 7, 5}, + {3, 5, 7, 5, 3}, + {1, 3, 5, 3, 1}}, + .divisor = 48 + }, + { + .name = "stucki", + .shift = 3, + .pattern = {{0, 0, 0, 8, 4}, + {2, 4, 8, 4, 2}, + {1, 2, 4, 2, 1}}, + .divisor = 42 + }, + { + .name = "burkes", + .shift = 3, + .pattern = {{0, 0, 0, 8, 4}, + {2, 4, 8, 4, 2}, + {0, 0, 0, 0, 0}}, + .divisor = 32 + }, + { + .name = "sierra-3", + .shift = 3, + .pattern = {{0, 0, 0, 5, 3}, + {2, 4, 5, 4, 2}, + {0, 2, 3, 2, 0}}, + .divisor = 32 + }, + { + .name = "sierra-2", + .shift = 3, + .pattern = {{0, 0, 0, 4, 3}, + {1, 2, 3, 2, 1}, + {0, 0, 0, 0, 0}}, + .divisor = 16 + }, + {0} +}; diff --git a/video/out/gpu/error_diffusion.h b/video/out/gpu/error_diffusion.h new file mode 100644 index 0000000..6bdcea1 --- /dev/null +++ b/video/out/gpu/error_diffusion.h @@ -0,0 +1,48 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#ifndef MP_GL_ERROR_DIFFUSION +#define MP_GL_ERROR_DIFFUSION + +#include "shader_cache.h" + +// defines the border of all error diffusion kernels +#define EF_MIN_DELTA_X (-2) +#define EF_MAX_DELTA_X (2) +#define EF_MAX_DELTA_Y (2) + +struct error_diffusion_kernel { + const char *name; + + // The minimum value such that a (y, x) -> (y, x + y * shift) mapping will + // make all error pushing operations affect next column (and after it) only. + int shift; + + // The diffusion factor for (y, x) is pattern[y][x - EF_MIN_DELTA_X] / divisor. + int pattern[EF_MAX_DELTA_Y + 1][EF_MAX_DELTA_X - EF_MIN_DELTA_X + 1]; + int divisor; +}; + +extern const struct error_diffusion_kernel mp_error_diffusion_kernels[]; + +const struct error_diffusion_kernel *mp_find_error_diffusion_kernel(const char *name); +int mp_ef_compute_shared_memory_size(const struct error_diffusion_kernel *k, int height); +void pass_error_diffusion(struct gl_shader_cache *sc, + const struct error_diffusion_kernel *k, + int tex, int width, int height, int depth, int block_size); + +#endif /* MP_GL_ERROR_DIFFUSION */ diff --git a/video/out/gpu/hwdec.c b/video/out/gpu/hwdec.c new file mode 100644 index 0000000..c8098f3 --- /dev/null +++ b/video/out/gpu/hwdec.c @@ -0,0 +1,358 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <stddef.h> +#include <string.h> + +#include "config.h" + +#include "common/common.h" +#include "common/msg.h" +#include "options/m_config.h" +#include "hwdec.h" + +extern const struct ra_hwdec_driver ra_hwdec_vaapi; +extern const struct ra_hwdec_driver ra_hwdec_videotoolbox; +extern const struct ra_hwdec_driver ra_hwdec_vdpau; +extern const struct ra_hwdec_driver ra_hwdec_dxva2egl; +extern const struct ra_hwdec_driver ra_hwdec_d3d11egl; +extern const struct ra_hwdec_driver ra_hwdec_dxva2gldx; +extern const struct ra_hwdec_driver ra_hwdec_d3d11va; +extern const struct ra_hwdec_driver ra_hwdec_dxva2dxgi; +extern const struct ra_hwdec_driver ra_hwdec_cuda; +extern const struct ra_hwdec_driver ra_hwdec_rpi_overlay; +extern const struct ra_hwdec_driver ra_hwdec_drmprime; +extern const struct ra_hwdec_driver ra_hwdec_drmprime_overlay; +extern const struct ra_hwdec_driver ra_hwdec_aimagereader; +extern const struct ra_hwdec_driver ra_hwdec_vulkan; + +const struct ra_hwdec_driver *const ra_hwdec_drivers[] = { +#if HAVE_VAAPI + &ra_hwdec_vaapi, +#endif +#if HAVE_VIDEOTOOLBOX_GL || HAVE_IOS_GL || HAVE_VIDEOTOOLBOX_PL + &ra_hwdec_videotoolbox, +#endif +#if HAVE_D3D_HWACCEL + #if HAVE_EGL_ANGLE + &ra_hwdec_d3d11egl, + #if HAVE_D3D9_HWACCEL + &ra_hwdec_dxva2egl, + #endif + #endif + #if HAVE_D3D11 + &ra_hwdec_d3d11va, + #if HAVE_D3D9_HWACCEL + &ra_hwdec_dxva2dxgi, + #endif + #endif +#endif +#if HAVE_GL_DXINTEROP_D3D9 + &ra_hwdec_dxva2gldx, +#endif +#if HAVE_CUDA_INTEROP + &ra_hwdec_cuda, +#endif +#if HAVE_VDPAU_GL_X11 + &ra_hwdec_vdpau, +#endif +#if HAVE_RPI_MMAL + &ra_hwdec_rpi_overlay, +#endif +#if HAVE_DRM + &ra_hwdec_drmprime, + &ra_hwdec_drmprime_overlay, +#endif +#if HAVE_ANDROID_MEDIA_NDK + &ra_hwdec_aimagereader, +#endif +#if HAVE_VULKAN_INTEROP + &ra_hwdec_vulkan, +#endif + + NULL +}; + +struct ra_hwdec *ra_hwdec_load_driver(struct ra_ctx *ra_ctx, + struct mp_log *log, + struct mpv_global *global, + struct mp_hwdec_devices *devs, + const struct ra_hwdec_driver *drv, + bool is_auto) +{ + struct ra_hwdec *hwdec = talloc(NULL, struct ra_hwdec); + *hwdec = (struct ra_hwdec) { + .driver = drv, + .log = mp_log_new(hwdec, log, drv->name), + .global = global, + .ra_ctx = ra_ctx, + .devs = devs, + .probing = is_auto, + .priv = talloc_zero_size(hwdec, drv->priv_size), + }; + mp_verbose(log, "Loading hwdec driver '%s'\n", drv->name); + if (hwdec->driver->init(hwdec) < 0) { + ra_hwdec_uninit(hwdec); + mp_verbose(log, "Loading failed.\n"); + return NULL; + } + return hwdec; +} + +void ra_hwdec_uninit(struct ra_hwdec *hwdec) +{ + if (hwdec) + hwdec->driver->uninit(hwdec); + talloc_free(hwdec); +} + +bool ra_hwdec_test_format(struct ra_hwdec *hwdec, int imgfmt) +{ + for (int n = 0; hwdec->driver->imgfmts[n]; n++) { + if (hwdec->driver->imgfmts[n] == imgfmt) + return true; + } + return false; +} + +struct ra_hwdec_mapper *ra_hwdec_mapper_create(struct ra_hwdec *hwdec, + const struct mp_image_params *params) +{ + assert(ra_hwdec_test_format(hwdec, params->imgfmt)); + + struct ra_hwdec_mapper *mapper = talloc_ptrtype(NULL, mapper); + *mapper = (struct ra_hwdec_mapper){ + .owner = hwdec, + .driver = hwdec->driver->mapper, + .log = hwdec->log, + .ra = hwdec->ra_ctx->ra, + .priv = talloc_zero_size(mapper, hwdec->driver->mapper->priv_size), + .src_params = *params, + .dst_params = *params, + }; + if (mapper->driver->init(mapper) < 0) + ra_hwdec_mapper_free(&mapper); + return mapper; +} + +void ra_hwdec_mapper_free(struct ra_hwdec_mapper **mapper) +{ + struct ra_hwdec_mapper *p = *mapper; + if (p) { + ra_hwdec_mapper_unmap(p); + p->driver->uninit(p); + talloc_free(p); + } + *mapper = NULL; +} + +void ra_hwdec_mapper_unmap(struct ra_hwdec_mapper *mapper) +{ + if (mapper->driver->unmap) + mapper->driver->unmap(mapper); + + // Clean up after the image if the mapper didn't already + mp_image_unrefp(&mapper->src); +} + +int ra_hwdec_mapper_map(struct ra_hwdec_mapper *mapper, struct mp_image *img) +{ + ra_hwdec_mapper_unmap(mapper); + mp_image_setrefp(&mapper->src, img); + if (mapper->driver->map(mapper) < 0) { + ra_hwdec_mapper_unmap(mapper); + return -1; + } + return 0; +} + +static int ra_hwdec_validate_opt_full(struct mp_log *log, bool include_modes, + const m_option_t *opt, + struct bstr name, const char **value) +{ + struct bstr param = bstr0(*value); + bool help = bstr_equals0(param, "help"); + if (help) + mp_info(log, "Available hwdecs:\n"); + for (int n = 0; ra_hwdec_drivers[n]; n++) { + const struct ra_hwdec_driver *drv = ra_hwdec_drivers[n]; + if (help) { + mp_info(log, " %s\n", drv->name); + } else if (bstr_equals0(param, drv->name)) { + return 1; + } + } + if (help) { + if (include_modes) { + mp_info(log, " auto (behavior depends on context)\n" + " all (load all hwdecs)\n" + " no (do not load any and block loading on demand)\n"); + } + return M_OPT_EXIT; + } + if (!param.len) + return 1; // "" is treated specially + if (include_modes && + (bstr_equals0(param, "all") || bstr_equals0(param, "auto") || + bstr_equals0(param, "no"))) + return 1; + mp_fatal(log, "No hwdec backend named '%.*s' found!\n", BSTR_P(param)); + return M_OPT_INVALID; +} + +int ra_hwdec_validate_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value) +{ + return ra_hwdec_validate_opt_full(log, true, opt, name, value); +} + +int ra_hwdec_validate_drivers_only_opt(struct mp_log *log, + const m_option_t *opt, + struct bstr name, const char **value) +{ + return ra_hwdec_validate_opt_full(log, false, opt, name, value); +} + +static void load_add_hwdec(struct ra_hwdec_ctx *ctx, struct mp_hwdec_devices *devs, + const struct ra_hwdec_driver *drv, bool is_auto) +{ + // Don't load duplicate hwdecs + for (int j = 0; j < ctx->num_hwdecs; j++) { + if (ctx->hwdecs[j]->driver == drv) + return; + } + + struct ra_hwdec *hwdec = + ra_hwdec_load_driver(ctx->ra_ctx, ctx->log, ctx->global, devs, drv, is_auto); + if (hwdec) + MP_TARRAY_APPEND(NULL, ctx->hwdecs, ctx->num_hwdecs, hwdec); +} + +static void load_hwdecs_all(struct ra_hwdec_ctx *ctx, struct mp_hwdec_devices *devs) +{ + if (!ctx->loading_done) { + for (int n = 0; ra_hwdec_drivers[n]; n++) + load_add_hwdec(ctx, devs, ra_hwdec_drivers[n], true); + ctx->loading_done = true; + } +} + +void ra_hwdec_ctx_init(struct ra_hwdec_ctx *ctx, struct mp_hwdec_devices *devs, + const char *type, bool load_all_by_default) +{ + assert(ctx->ra_ctx); + + /* + * By default, or if the option value is "auto", we will not pre-emptively + * load any interops, and instead allow them to be loaded on-demand. + * + * If the option value is "no", then no interops will be loaded now, and + * no interops will be loaded, even if requested later. + * + * If the option value is "all", then all interops will be loaded now, and + * obviously no interops will need to be loaded later. + * + * Finally, if a specific interop is requested, it will be loaded now, and + * other interops can be loaded, if requested later. + */ + if (!type || !type[0] || strcmp(type, "auto") == 0) { + if (!load_all_by_default) + return; + type = "all"; + } + if (strcmp(type, "no") == 0) { + // do nothing, just block further loading + } else if (strcmp(type, "all") == 0) { + load_hwdecs_all(ctx, devs); + } else { + for (int n = 0; ra_hwdec_drivers[n]; n++) { + const struct ra_hwdec_driver *drv = ra_hwdec_drivers[n]; + if (strcmp(type, drv->name) == 0) { + load_add_hwdec(ctx, devs, drv, false); + break; + } + } + } + ctx->loading_done = true; +} + +void ra_hwdec_ctx_uninit(struct ra_hwdec_ctx *ctx) +{ + for (int n = 0; n < ctx->num_hwdecs; n++) + ra_hwdec_uninit(ctx->hwdecs[n]); + + talloc_free(ctx->hwdecs); + memset(ctx, 0, sizeof(*ctx)); +} + +void ra_hwdec_ctx_load_fmt(struct ra_hwdec_ctx *ctx, struct mp_hwdec_devices *devs, + struct hwdec_imgfmt_request *params) +{ + int imgfmt = params->imgfmt; + if (ctx->loading_done) { + /* + * If we previously marked interop loading as done (for reasons + * discussed above), then do not load any other interops regardless + * of imgfmt. + */ + return; + } + + if (imgfmt == IMGFMT_NONE) { + MP_VERBOSE(ctx, "Loading hwdec drivers for all formats\n"); + load_hwdecs_all(ctx, devs); + return; + } + + MP_VERBOSE(ctx, "Loading hwdec drivers for format: '%s'\n", + mp_imgfmt_to_name(imgfmt)); + for (int i = 0; ra_hwdec_drivers[i]; i++) { + bool matched_fmt = false; + const struct ra_hwdec_driver *drv = ra_hwdec_drivers[i]; + for (int j = 0; drv->imgfmts[j]; j++) { + if (imgfmt == drv->imgfmts[j]) { + matched_fmt = true; + break; + } + } + if (!matched_fmt) { + continue; + } + + load_add_hwdec(ctx, devs, drv, params->probing); + } +} + +struct ra_hwdec *ra_hwdec_get(struct ra_hwdec_ctx *ctx, int imgfmt) +{ + for (int n = 0; n < ctx->num_hwdecs; n++) { + if (ra_hwdec_test_format(ctx->hwdecs[n], imgfmt)) + return ctx->hwdecs[n]; + } + + return NULL; +} + +int ra_hwdec_driver_get_imgfmt_for_name(const char *name) +{ + for (int i = 0; ra_hwdec_drivers[i]; i++) { + if (!strcmp(ra_hwdec_drivers[i]->name, name)) { + return ra_hwdec_drivers[i]->imgfmts[0]; + } + } + return IMGFMT_NONE; +} diff --git a/video/out/gpu/hwdec.h b/video/out/gpu/hwdec.h new file mode 100644 index 0000000..7766073 --- /dev/null +++ b/video/out/gpu/hwdec.h @@ -0,0 +1,156 @@ +#ifndef MPGL_HWDEC_H_ +#define MPGL_HWDEC_H_ + +#include "video/mp_image.h" +#include "context.h" +#include "ra.h" +#include "video/hwdec.h" + +// Helper to organize/load hwdecs dynamically +struct ra_hwdec_ctx { + // Set these before calling `ra_hwdec_ctx_init` + struct mp_log *log; + struct mpv_global *global; + struct ra_ctx *ra_ctx; + + bool loading_done; + struct ra_hwdec **hwdecs; + int num_hwdecs; +}; + +int ra_hwdec_validate_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value); + +int ra_hwdec_validate_drivers_only_opt(struct mp_log *log, + const m_option_t *opt, + struct bstr name, const char **value); + +void ra_hwdec_ctx_init(struct ra_hwdec_ctx *ctx, struct mp_hwdec_devices *devs, + const char *opt, bool load_all_by_default); +void ra_hwdec_ctx_uninit(struct ra_hwdec_ctx *ctx); + +void ra_hwdec_ctx_load_fmt(struct ra_hwdec_ctx *ctx, struct mp_hwdec_devices *devs, + struct hwdec_imgfmt_request *params); + +// Gets the right `ra_hwdec` for a format, if any +struct ra_hwdec *ra_hwdec_get(struct ra_hwdec_ctx *ctx, int imgfmt); + +struct ra_hwdec { + const struct ra_hwdec_driver *driver; + struct mp_log *log; + struct mpv_global *global; + struct ra_ctx *ra_ctx; + struct mp_hwdec_devices *devs; + // GLSL extensions required to sample textures from this. + const char **glsl_extensions; + // For free use by hwdec driver + void *priv; + // For working around the vdpau vs. vaapi mess. + bool probing; + // Used in overlay mode only. + float overlay_colorkey[4]; +}; + +struct ra_hwdec_mapper { + const struct ra_hwdec_mapper_driver *driver; + struct mp_log *log; + struct ra *ra; + void *priv; + struct ra_hwdec *owner; + // Input frame parameters. (Set before init(), immutable.) + struct mp_image_params src_params; + // Output frame parameters (represents the format the textures return). Must + // be set by init(), immutable afterwards, + struct mp_image_params dst_params; + + // The currently mapped source image (or the image about to be mapped in + // ->map()). NULL if unmapped. The mapper can also clear this reference if + // the mapped textures contain a full copy. + struct mp_image *src; + + // The mapped textures and metadata about them. These fields change if a + // new frame is mapped (or unmapped), but otherwise remain constant. + // The common code won't mess with these, so you can e.g. set them in the + // .init() callback. + struct ra_tex *tex[4]; +}; + +// This can be used to map frames of a specific hw format as GL textures. +struct ra_hwdec_mapper_driver { + // Used to create ra_hwdec_mapper.priv. + size_t priv_size; + + // Init the mapper implementation. At this point, the field src_params, + // fns, devs, priv are initialized. + int (*init)(struct ra_hwdec_mapper *mapper); + // Destroy the mapper. unmap is called before this. + void (*uninit)(struct ra_hwdec_mapper *mapper); + + // Map mapper->src as texture, and set mapper->frame to textures using it. + // It is expected that the textures remain valid until the next unmap + // or uninit call. + // The function is allowed to unref mapper->src if it's not needed (i.e. + // this function creates a copy). + // The underlying format can change, so you might need to do some form + // of change detection. You also must reject unsupported formats with an + // error. + // On error, returns negative value on error and remains unmapped. + int (*map)(struct ra_hwdec_mapper *mapper); + // Unmap the frame. Does nothing if already unmapped. Optional. + void (*unmap)(struct ra_hwdec_mapper *mapper); +}; + +struct ra_hwdec_driver { + // Name of the interop backend. This is used for informational purposes and + // for use with debugging options. + const char *name; + // Used to create ra_hwdec.priv. + size_t priv_size; + // One of the hardware surface IMGFMT_ that must be passed to map_image later. + // Terminated with a 0 entry. (Extend the array size as needed.) + const int imgfmts[3]; + + // Create the hwdec device. It must add it to hw->devs, if applicable. + int (*init)(struct ra_hwdec *hw); + void (*uninit)(struct ra_hwdec *hw); + + // This will be used to create a ra_hwdec_mapper from ra_hwdec. + const struct ra_hwdec_mapper_driver *mapper; + + // The following function provides an alternative API. Each ra_hwdec_driver + // must have either provide a mapper or overlay_frame (not both or none), and + // if overlay_frame is set, it operates in overlay mode. In this mode, + // OSD etc. is rendered via OpenGL, but the video is rendered as a separate + // layer below it. + // Non-overlay mode is strictly preferred, so try not to use overlay mode. + // Set the given frame as overlay, replacing the previous one. This can also + // just change the position of the overlay. + // hw_image==src==dst==NULL is passed to clear the overlay. + int (*overlay_frame)(struct ra_hwdec *hw, struct mp_image *hw_image, + struct mp_rect *src, struct mp_rect *dst, bool newframe); +}; + +extern const struct ra_hwdec_driver *const ra_hwdec_drivers[]; + +struct ra_hwdec *ra_hwdec_load_driver(struct ra_ctx *ra_ctx, + struct mp_log *log, + struct mpv_global *global, + struct mp_hwdec_devices *devs, + const struct ra_hwdec_driver *drv, + bool is_auto); + +void ra_hwdec_uninit(struct ra_hwdec *hwdec); + +bool ra_hwdec_test_format(struct ra_hwdec *hwdec, int imgfmt); + +struct ra_hwdec_mapper *ra_hwdec_mapper_create(struct ra_hwdec *hwdec, + const struct mp_image_params *params); +void ra_hwdec_mapper_free(struct ra_hwdec_mapper **mapper); +void ra_hwdec_mapper_unmap(struct ra_hwdec_mapper *mapper); +int ra_hwdec_mapper_map(struct ra_hwdec_mapper *mapper, struct mp_image *img); + +// Get the primary image format for the given driver name. +// Returns IMGFMT_NONE if the name doesn't get matched. +int ra_hwdec_driver_get_imgfmt_for_name(const char *name); + +#endif diff --git a/video/out/gpu/lcms.c b/video/out/gpu/lcms.c new file mode 100644 index 0000000..7006a96 --- /dev/null +++ b/video/out/gpu/lcms.c @@ -0,0 +1,526 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <string.h> +#include <math.h> + +#include "mpv_talloc.h" + +#include "config.h" + +#include "stream/stream.h" +#include "common/common.h" +#include "misc/bstr.h" +#include "common/msg.h" +#include "options/m_option.h" +#include "options/path.h" +#include "video/csputils.h" +#include "lcms.h" + +#include "osdep/io.h" + +#if HAVE_LCMS2 + +#include <lcms2.h> +#include <libavutil/sha.h> +#include <libavutil/mem.h> + +struct gl_lcms { + void *icc_data; + size_t icc_size; + struct AVBufferRef *vid_profile; + char *current_profile; + bool using_memory_profile; + bool changed; + enum mp_csp_prim current_prim; + enum mp_csp_trc current_trc; + + struct mp_log *log; + struct mpv_global *global; + struct mp_icc_opts *opts; +}; + +static void lcms2_error_handler(cmsContext ctx, cmsUInt32Number code, + const char *msg) +{ + struct gl_lcms *p = cmsGetContextUserData(ctx); + MP_ERR(p, "lcms2: %s\n", msg); +} + +static void load_profile(struct gl_lcms *p) +{ + talloc_free(p->icc_data); + p->icc_data = NULL; + p->icc_size = 0; + p->using_memory_profile = false; + talloc_free(p->current_profile); + p->current_profile = NULL; + + if (!p->opts->profile || !p->opts->profile[0]) + return; + + char *fname = mp_get_user_path(NULL, p->global, p->opts->profile); + MP_VERBOSE(p, "Opening ICC profile '%s'\n", fname); + struct bstr iccdata = stream_read_file(fname, p, p->global, + 100000000); // 100 MB + talloc_free(fname); + if (!iccdata.len) + return; + + talloc_free(p->icc_data); + + p->icc_data = iccdata.start; + p->icc_size = iccdata.len; + p->current_profile = talloc_strdup(p, p->opts->profile); +} + +static void gl_lcms_destructor(void *ptr) +{ + struct gl_lcms *p = ptr; + av_buffer_unref(&p->vid_profile); +} + +struct gl_lcms *gl_lcms_init(void *talloc_ctx, struct mp_log *log, + struct mpv_global *global, + struct mp_icc_opts *opts) +{ + struct gl_lcms *p = talloc_ptrtype(talloc_ctx, p); + talloc_set_destructor(p, gl_lcms_destructor); + *p = (struct gl_lcms) { + .global = global, + .log = log, + .opts = opts, + }; + gl_lcms_update_options(p); + return p; +} + +void gl_lcms_update_options(struct gl_lcms *p) +{ + if ((p->using_memory_profile && !p->opts->profile_auto) || + !bstr_equals(bstr0(p->opts->profile), bstr0(p->current_profile))) + { + load_profile(p); + } + + p->changed = true; // probably +} + +// Warning: profile.start must point to a ta allocation, and the function +// takes over ownership. +// Returns whether the internal profile was changed. +bool gl_lcms_set_memory_profile(struct gl_lcms *p, bstr profile) +{ + if (!p->opts->profile_auto || (p->opts->profile && p->opts->profile[0])) { + talloc_free(profile.start); + return false; + } + + if (p->using_memory_profile && + p->icc_data && profile.start && + profile.len == p->icc_size && + memcmp(profile.start, p->icc_data, p->icc_size) == 0) + { + talloc_free(profile.start); + return false; + } + + p->changed = true; + p->using_memory_profile = true; + + talloc_free(p->icc_data); + + p->icc_data = talloc_steal(p, profile.start); + p->icc_size = profile.len; + + return true; +} + +// Guards against NULL and uses bstr_equals to short-circuit some special cases +static bool vid_profile_eq(struct AVBufferRef *a, struct AVBufferRef *b) +{ + if (!a || !b) + return a == b; + + return bstr_equals((struct bstr){ a->data, a->size }, + (struct bstr){ b->data, b->size }); +} + +// Return whether the profile or config has changed since the last time it was +// retrieved. If it has changed, gl_lcms_get_lut3d() should be called. +bool gl_lcms_has_changed(struct gl_lcms *p, enum mp_csp_prim prim, + enum mp_csp_trc trc, struct AVBufferRef *vid_profile) +{ + if (p->changed || p->current_prim != prim || p->current_trc != trc) + return true; + + return !vid_profile_eq(p->vid_profile, vid_profile); +} + +// Whether a profile is set. (gl_lcms_get_lut3d() is expected to return a lut, +// but it could still fail due to runtime errors, such as invalid icc data.) +bool gl_lcms_has_profile(struct gl_lcms *p) +{ + return p->icc_size > 0; +} + +static cmsHPROFILE get_vid_profile(struct gl_lcms *p, cmsContext cms, + cmsHPROFILE disp_profile, + enum mp_csp_prim prim, enum mp_csp_trc trc) +{ + if (p->opts->use_embedded && p->vid_profile) { + // Try using the embedded ICC profile + cmsHPROFILE prof = cmsOpenProfileFromMemTHR(cms, p->vid_profile->data, + p->vid_profile->size); + if (prof) { + MP_VERBOSE(p, "Successfully opened embedded ICC profile\n"); + return prof; + } + + // Otherwise, warn the user and generate the profile as usual + MP_WARN(p, "Video contained an invalid ICC profile! Ignoring...\n"); + } + + // The input profile for the transformation is dependent on the video + // primaries and transfer characteristics + struct mp_csp_primaries csp = mp_get_csp_primaries(prim); + cmsCIExyY wp_xyY = {csp.white.x, csp.white.y, 1.0}; + cmsCIExyYTRIPLE prim_xyY = { + .Red = {csp.red.x, csp.red.y, 1.0}, + .Green = {csp.green.x, csp.green.y, 1.0}, + .Blue = {csp.blue.x, csp.blue.y, 1.0}, + }; + + cmsToneCurve *tonecurve[3] = {0}; + switch (trc) { + case MP_CSP_TRC_LINEAR: tonecurve[0] = cmsBuildGamma(cms, 1.0); break; + case MP_CSP_TRC_GAMMA18: tonecurve[0] = cmsBuildGamma(cms, 1.8); break; + case MP_CSP_TRC_GAMMA20: tonecurve[0] = cmsBuildGamma(cms, 2.0); break; + case MP_CSP_TRC_GAMMA22: tonecurve[0] = cmsBuildGamma(cms, 2.2); break; + case MP_CSP_TRC_GAMMA24: tonecurve[0] = cmsBuildGamma(cms, 2.4); break; + case MP_CSP_TRC_GAMMA26: tonecurve[0] = cmsBuildGamma(cms, 2.6); break; + case MP_CSP_TRC_GAMMA28: tonecurve[0] = cmsBuildGamma(cms, 2.8); break; + + case MP_CSP_TRC_SRGB: + // Values copied from Little-CMS + tonecurve[0] = cmsBuildParametricToneCurve(cms, 4, + (double[5]){2.40, 1/1.055, 0.055/1.055, 1/12.92, 0.04045}); + break; + + case MP_CSP_TRC_PRO_PHOTO: + tonecurve[0] = cmsBuildParametricToneCurve(cms, 4, + (double[5]){1.8, 1.0, 0.0, 1/16.0, 0.03125}); + break; + + case MP_CSP_TRC_BT_1886: { + double src_black[3]; + if (p->opts->contrast < 0) { + // User requested infinite contrast, return 2.4 profile + tonecurve[0] = cmsBuildGamma(cms, 2.4); + break; + } else if (p->opts->contrast > 0) { + MP_VERBOSE(p, "Using specified contrast: %d\n", p->opts->contrast); + for (int i = 0; i < 3; i++) + src_black[i] = 1.0 / p->opts->contrast; + } else { + // To build an appropriate BT.1886 transformation we need access to + // the display's black point, so we use LittleCMS' detection + // function. Relative colorimetric is used since we want to + // approximate the BT.1886 to the target device's actual black + // point even in e.g. perceptual mode + const int intent = MP_INTENT_RELATIVE_COLORIMETRIC; + cmsCIEXYZ bp_XYZ; + if (!cmsDetectBlackPoint(&bp_XYZ, disp_profile, intent, 0)) + return false; + + // Map this XYZ value back into the (linear) source space + cmsHPROFILE rev_profile; + cmsToneCurve *linear = cmsBuildGamma(cms, 1.0); + rev_profile = cmsCreateRGBProfileTHR(cms, &wp_xyY, &prim_xyY, + (cmsToneCurve*[3]){linear, linear, linear}); + cmsHPROFILE xyz_profile = cmsCreateXYZProfile(); + cmsHTRANSFORM xyz2src = cmsCreateTransformTHR(cms, + xyz_profile, TYPE_XYZ_DBL, rev_profile, TYPE_RGB_DBL, + intent, cmsFLAGS_NOCACHE | cmsFLAGS_NOOPTIMIZE); + cmsFreeToneCurve(linear); + cmsCloseProfile(rev_profile); + cmsCloseProfile(xyz_profile); + if (!xyz2src) + return false; + + cmsDoTransform(xyz2src, &bp_XYZ, src_black, 1); + cmsDeleteTransform(xyz2src); + + double contrast = 3.0 / (src_black[0] + src_black[1] + src_black[2]); + MP_VERBOSE(p, "Detected ICC profile contrast: %f\n", contrast); + } + + // Build the parametric BT.1886 transfer curve, one per channel + for (int i = 0; i < 3; i++) { + const double gamma = 2.40; + double binv = pow(src_black[i], 1.0/gamma); + tonecurve[i] = cmsBuildParametricToneCurve(cms, 6, + (double[4]){gamma, 1.0 - binv, binv, 0.0}); + } + break; + } + + default: + abort(); + } + + if (!tonecurve[0]) + return false; + + if (!tonecurve[1]) tonecurve[1] = tonecurve[0]; + if (!tonecurve[2]) tonecurve[2] = tonecurve[0]; + + cmsHPROFILE *vid_profile = cmsCreateRGBProfileTHR(cms, &wp_xyY, &prim_xyY, + tonecurve); + + if (tonecurve[2] != tonecurve[0]) cmsFreeToneCurve(tonecurve[2]); + if (tonecurve[1] != tonecurve[0]) cmsFreeToneCurve(tonecurve[1]); + cmsFreeToneCurve(tonecurve[0]); + + return vid_profile; +} + +bool gl_lcms_get_lut3d(struct gl_lcms *p, struct lut3d **result_lut3d, + enum mp_csp_prim prim, enum mp_csp_trc trc, + struct AVBufferRef *vid_profile) +{ + int s_r, s_g, s_b; + bool result = false; + + p->changed = false; + p->current_prim = prim; + p->current_trc = trc; + + // We need to hold on to a reference to the video's ICC profile for as long + // as we still need to perform equality checking, so generate a new + // reference here + av_buffer_unref(&p->vid_profile); + if (vid_profile) { + MP_VERBOSE(p, "Got an embedded ICC profile.\n"); + p->vid_profile = av_buffer_ref(vid_profile); + MP_HANDLE_OOM(p->vid_profile); + } + + if (!gl_parse_3dlut_size(p->opts->size_str, &s_r, &s_g, &s_b)) + return false; + + if (!gl_lcms_has_profile(p)) + return false; + + // For simplicity, default to 65x65x65, which is large enough to cover + // typical profiles with good accuracy while not being too wasteful + s_r = s_r ? s_r : 65; + s_g = s_g ? s_g : 65; + s_b = s_b ? s_b : 65; + + void *tmp = talloc_new(NULL); + uint16_t *output = talloc_array(tmp, uint16_t, s_r * s_g * s_b * 4); + struct lut3d *lut = NULL; + cmsContext cms = NULL; + + char *cache_file = NULL; + if (p->opts->cache) { + // Gamma is included in the header to help uniquely identify it, + // because we may change the parameter in the future or make it + // customizable, same for the primaries. + char *cache_info = talloc_asprintf(tmp, + "ver=1.4, intent=%d, size=%dx%dx%d, prim=%d, trc=%d, " + "contrast=%d\n", + p->opts->intent, s_r, s_g, s_b, prim, trc, p->opts->contrast); + + uint8_t hash[32]; + struct AVSHA *sha = av_sha_alloc(); + MP_HANDLE_OOM(sha); + av_sha_init(sha, 256); + av_sha_update(sha, cache_info, strlen(cache_info)); + if (vid_profile) + av_sha_update(sha, vid_profile->data, vid_profile->size); + av_sha_update(sha, p->icc_data, p->icc_size); + av_sha_final(sha, hash); + av_free(sha); + + char *cache_dir = p->opts->cache_dir; + if (cache_dir && cache_dir[0]) { + cache_dir = mp_get_user_path(tmp, p->global, cache_dir); + } else { + cache_dir = mp_find_user_file(tmp, p->global, "cache", ""); + } + + if (cache_dir && cache_dir[0]) { + cache_file = talloc_strdup(tmp, ""); + for (int i = 0; i < sizeof(hash); i++) + cache_file = talloc_asprintf_append(cache_file, "%02X", hash[i]); + cache_file = mp_path_join(tmp, cache_dir, cache_file); + mp_mkdirp(cache_dir); + } + } + + // check cache + if (cache_file && stat(cache_file, &(struct stat){0}) == 0) { + MP_VERBOSE(p, "Opening 3D LUT cache in file '%s'.\n", cache_file); + struct bstr cachedata = stream_read_file(cache_file, tmp, p->global, + 1000000000); // 1 GB + if (cachedata.len == talloc_get_size(output)) { + memcpy(output, cachedata.start, cachedata.len); + goto done; + } else { + MP_WARN(p, "3D LUT cache invalid!\n"); + } + } + + cms = cmsCreateContext(NULL, p); + if (!cms) + goto error_exit; + cmsSetLogErrorHandlerTHR(cms, lcms2_error_handler); + + cmsHPROFILE profile = + cmsOpenProfileFromMemTHR(cms, p->icc_data, p->icc_size); + if (!profile) + goto error_exit; + + cmsHPROFILE vid_hprofile = get_vid_profile(p, cms, profile, prim, trc); + if (!vid_hprofile) { + cmsCloseProfile(profile); + goto error_exit; + } + + cmsHTRANSFORM trafo = cmsCreateTransformTHR(cms, vid_hprofile, TYPE_RGB_16, + profile, TYPE_RGBA_16, + p->opts->intent, + cmsFLAGS_NOCACHE | + cmsFLAGS_NOOPTIMIZE | + cmsFLAGS_BLACKPOINTCOMPENSATION); + cmsCloseProfile(profile); + cmsCloseProfile(vid_hprofile); + + if (!trafo) + goto error_exit; + + // transform a (s_r)x(s_g)x(s_b) cube, with 3 components per channel + uint16_t *input = talloc_array(tmp, uint16_t, s_r * 3); + for (int b = 0; b < s_b; b++) { + for (int g = 0; g < s_g; g++) { + for (int r = 0; r < s_r; r++) { + input[r * 3 + 0] = r * 65535 / (s_r - 1); + input[r * 3 + 1] = g * 65535 / (s_g - 1); + input[r * 3 + 2] = b * 65535 / (s_b - 1); + } + size_t base = (b * s_r * s_g + g * s_r) * 4; + cmsDoTransform(trafo, input, output + base, s_r); + } + } + + cmsDeleteTransform(trafo); + + if (cache_file) { + FILE *out = fopen(cache_file, "wb"); + if (out) { + fwrite(output, talloc_get_size(output), 1, out); + fclose(out); + } + } + +done: ; + + lut = talloc_ptrtype(NULL, lut); + *lut = (struct lut3d) { + .data = talloc_steal(lut, output), + .size = {s_r, s_g, s_b}, + }; + + *result_lut3d = lut; + result = true; + +error_exit: + + if (cms) + cmsDeleteContext(cms); + + if (!lut) + MP_FATAL(p, "Error loading ICC profile.\n"); + + talloc_free(tmp); + return result; +} + +#else /* HAVE_LCMS2 */ + +struct gl_lcms *gl_lcms_init(void *talloc_ctx, struct mp_log *log, + struct mpv_global *global, + struct mp_icc_opts *opts) +{ + return (struct gl_lcms *) talloc_new(talloc_ctx); +} + +void gl_lcms_update_options(struct gl_lcms *p) { } +bool gl_lcms_set_memory_profile(struct gl_lcms *p, bstr profile) {return false;} + +bool gl_lcms_has_changed(struct gl_lcms *p, enum mp_csp_prim prim, + enum mp_csp_trc trc, struct AVBufferRef *vid_profile) +{ + return false; +} + +bool gl_lcms_has_profile(struct gl_lcms *p) +{ + return false; +} + +bool gl_lcms_get_lut3d(struct gl_lcms *p, struct lut3d **result_lut3d, + enum mp_csp_prim prim, enum mp_csp_trc trc, + struct AVBufferRef *vid_profile) +{ + return false; +} + +#endif + +static int validate_3dlut_size_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value) +{ + int p1, p2, p3; + return gl_parse_3dlut_size(*value, &p1, &p2, &p3) ? 0 : M_OPT_INVALID; +} + +#define OPT_BASE_STRUCT struct mp_icc_opts +const struct m_sub_options mp_icc_conf = { + .opts = (const m_option_t[]) { + {"use-embedded-icc-profile", OPT_BOOL(use_embedded)}, + {"icc-profile", OPT_STRING(profile), .flags = M_OPT_FILE}, + {"icc-profile-auto", OPT_BOOL(profile_auto)}, + {"icc-cache", OPT_BOOL(cache)}, + {"icc-cache-dir", OPT_STRING(cache_dir), .flags = M_OPT_FILE}, + {"icc-intent", OPT_INT(intent)}, + {"icc-force-contrast", OPT_CHOICE(contrast, {"no", 0}, {"inf", -1}), + M_RANGE(0, 1000000)}, + {"icc-3dlut-size", OPT_STRING_VALIDATE(size_str, validate_3dlut_size_opt)}, + {"icc-use-luma", OPT_BOOL(icc_use_luma)}, + {0} + }, + .size = sizeof(struct mp_icc_opts), + .defaults = &(const struct mp_icc_opts) { + .size_str = "auto", + .intent = MP_INTENT_RELATIVE_COLORIMETRIC, + .use_embedded = true, + .cache = true, + }, +}; diff --git a/video/out/gpu/lcms.h b/video/out/gpu/lcms.h new file mode 100644 index 0000000..607353a --- /dev/null +++ b/video/out/gpu/lcms.h @@ -0,0 +1,61 @@ +#ifndef MP_GL_LCMS_H +#define MP_GL_LCMS_H + +#include <stddef.h> +#include <stdbool.h> +#include "misc/bstr.h" +#include "video/csputils.h" +#include <libavutil/buffer.h> + +extern const struct m_sub_options mp_icc_conf; + +struct mp_icc_opts { + bool use_embedded; + char *profile; + bool profile_auto; + bool cache; + char *cache_dir; + char *size_str; + int intent; + int contrast; + bool icc_use_luma; +}; + +struct lut3d { + uint16_t *data; + int size[3]; +}; + +struct mp_log; +struct mpv_global; +struct gl_lcms; + +struct gl_lcms *gl_lcms_init(void *talloc_ctx, struct mp_log *log, + struct mpv_global *global, + struct mp_icc_opts *opts); +void gl_lcms_update_options(struct gl_lcms *p); +bool gl_lcms_set_memory_profile(struct gl_lcms *p, bstr profile); +bool gl_lcms_has_profile(struct gl_lcms *p); +bool gl_lcms_get_lut3d(struct gl_lcms *p, struct lut3d **, + enum mp_csp_prim prim, enum mp_csp_trc trc, + struct AVBufferRef *vid_profile); +bool gl_lcms_has_changed(struct gl_lcms *p, enum mp_csp_prim prim, + enum mp_csp_trc trc, struct AVBufferRef *vid_profile); + +static inline bool gl_parse_3dlut_size(const char *arg, int *p1, int *p2, int *p3) +{ + if (!strcmp(arg, "auto")) { + *p1 = *p2 = *p3 = 0; + return true; + } + if (sscanf(arg, "%dx%dx%d", p1, p2, p3) != 3) + return false; + for (int n = 0; n < 3; n++) { + int s = ((int[]) { *p1, *p2, *p3 })[n]; + if (s < 2 || s > 512) + return false; + } + return true; +} + +#endif diff --git a/video/out/gpu/libmpv_gpu.c b/video/out/gpu/libmpv_gpu.c new file mode 100644 index 0000000..aae1d18 --- /dev/null +++ b/video/out/gpu/libmpv_gpu.c @@ -0,0 +1,248 @@ +#include "config.h" +#include "hwdec.h" +#include "libmpv_gpu.h" +#include "libmpv/render_gl.h" +#include "video.h" +#include "video/out/libmpv.h" + +static const struct libmpv_gpu_context_fns *context_backends[] = { +#if HAVE_GL + &libmpv_gpu_context_gl, +#endif + NULL +}; + +struct priv { + struct libmpv_gpu_context *context; + + struct gl_video *renderer; +}; + +struct native_resource_entry { + const char *name; // ra_add_native_resource() internal name argument + size_t size; // size of struct pointed to (0 for no copy) +}; + +static const struct native_resource_entry native_resource_map[] = { + [MPV_RENDER_PARAM_X11_DISPLAY] = { + .name = "x11", + .size = 0, + }, + [MPV_RENDER_PARAM_WL_DISPLAY] = { + .name = "wl", + .size = 0, + }, + [MPV_RENDER_PARAM_DRM_DRAW_SURFACE_SIZE] = { + .name = "drm_draw_surface_size", + .size = sizeof (mpv_opengl_drm_draw_surface_size), + }, + [MPV_RENDER_PARAM_DRM_DISPLAY_V2] = { + .name = "drm_params_v2", + .size = sizeof (mpv_opengl_drm_params_v2), + }, +}; + +static int init(struct render_backend *ctx, mpv_render_param *params) +{ + ctx->priv = talloc_zero(NULL, struct priv); + struct priv *p = ctx->priv; + + char *api = get_mpv_render_param(params, MPV_RENDER_PARAM_API_TYPE, NULL); + if (!api) + return MPV_ERROR_INVALID_PARAMETER; + + for (int n = 0; context_backends[n]; n++) { + const struct libmpv_gpu_context_fns *backend = context_backends[n]; + if (strcmp(backend->api_name, api) == 0) { + p->context = talloc_zero(NULL, struct libmpv_gpu_context); + *p->context = (struct libmpv_gpu_context){ + .global = ctx->global, + .log = ctx->log, + .fns = backend, + }; + break; + } + } + + if (!p->context) + return MPV_ERROR_NOT_IMPLEMENTED; + + int err = p->context->fns->init(p->context, params); + if (err < 0) + return err; + + for (int n = 0; params && params[n].type; n++) { + if (params[n].type > 0 && + params[n].type < MP_ARRAY_SIZE(native_resource_map) && + native_resource_map[params[n].type].name) + { + const struct native_resource_entry *entry = + &native_resource_map[params[n].type]; + void *data = params[n].data; + if (entry->size) + data = talloc_memdup(p, data, entry->size); + ra_add_native_resource(p->context->ra_ctx->ra, entry->name, data); + } + } + + p->renderer = gl_video_init(p->context->ra_ctx->ra, ctx->log, ctx->global); + + ctx->hwdec_devs = hwdec_devices_create(); + gl_video_init_hwdecs(p->renderer, p->context->ra_ctx, ctx->hwdec_devs, true); + ctx->driver_caps = VO_CAP_ROTATE90; + return 0; +} + +static bool check_format(struct render_backend *ctx, int imgfmt) +{ + struct priv *p = ctx->priv; + + return gl_video_check_format(p->renderer, imgfmt); +} + +static int set_parameter(struct render_backend *ctx, mpv_render_param param) +{ + struct priv *p = ctx->priv; + + switch (param.type) { + case MPV_RENDER_PARAM_ICC_PROFILE: { + mpv_byte_array *data = param.data; + gl_video_set_icc_profile(p->renderer, (bstr){data->data, data->size}); + return 0; + } + case MPV_RENDER_PARAM_AMBIENT_LIGHT: { + int lux = *(int *)param.data; + gl_video_set_ambient_lux(p->renderer, lux); + return 0; + } + default: + return MPV_ERROR_NOT_IMPLEMENTED; + } +} + +static void reconfig(struct render_backend *ctx, struct mp_image_params *params) +{ + struct priv *p = ctx->priv; + + gl_video_config(p->renderer, params); +} + +static void reset(struct render_backend *ctx) +{ + struct priv *p = ctx->priv; + + gl_video_reset(p->renderer); +} + +static void update_external(struct render_backend *ctx, struct vo *vo) +{ + struct priv *p = ctx->priv; + + gl_video_set_osd_source(p->renderer, vo ? vo->osd : NULL); + if (vo) + gl_video_configure_queue(p->renderer, vo); +} + +static void resize(struct render_backend *ctx, struct mp_rect *src, + struct mp_rect *dst, struct mp_osd_res *osd) +{ + struct priv *p = ctx->priv; + + gl_video_resize(p->renderer, src, dst, osd); +} + +static int get_target_size(struct render_backend *ctx, mpv_render_param *params, + int *out_w, int *out_h) +{ + struct priv *p = ctx->priv; + + // Mapping the surface is cheap, better than adding new backend entrypoints. + struct ra_tex *tex; + int err = p->context->fns->wrap_fbo(p->context, params, &tex); + if (err < 0) + return err; + *out_w = tex->params.w; + *out_h = tex->params.h; + return 0; +} + +static int render(struct render_backend *ctx, mpv_render_param *params, + struct vo_frame *frame) +{ + struct priv *p = ctx->priv; + + // Mapping the surface is cheap, better than adding new backend entrypoints. + struct ra_tex *tex; + int err = p->context->fns->wrap_fbo(p->context, params, &tex); + if (err < 0) + return err; + + int depth = *(int *)get_mpv_render_param(params, MPV_RENDER_PARAM_DEPTH, + &(int){0}); + gl_video_set_fb_depth(p->renderer, depth); + + bool flip = *(int *)get_mpv_render_param(params, MPV_RENDER_PARAM_FLIP_Y, + &(int){0}); + + struct ra_fbo target = {.tex = tex, .flip = flip}; + gl_video_render_frame(p->renderer, frame, target, RENDER_FRAME_DEF); + p->context->fns->done_frame(p->context, frame->display_synced); + + return 0; +} + +static struct mp_image *get_image(struct render_backend *ctx, int imgfmt, + int w, int h, int stride_align, int flags) +{ + struct priv *p = ctx->priv; + + return gl_video_get_image(p->renderer, imgfmt, w, h, stride_align, flags); +} + +static void screenshot(struct render_backend *ctx, struct vo_frame *frame, + struct voctrl_screenshot *args) +{ + struct priv *p = ctx->priv; + + gl_video_screenshot(p->renderer, frame, args); +} + +static void perfdata(struct render_backend *ctx, + struct voctrl_performance_data *out) +{ + struct priv *p = ctx->priv; + + gl_video_perfdata(p->renderer, out); +} + +static void destroy(struct render_backend *ctx) +{ + struct priv *p = ctx->priv; + + if (p->renderer) + gl_video_uninit(p->renderer); + + hwdec_devices_destroy(ctx->hwdec_devs); + + if (p->context) { + p->context->fns->destroy(p->context); + talloc_free(p->context->priv); + talloc_free(p->context); + } +} + +const struct render_backend_fns render_backend_gpu = { + .init = init, + .check_format = check_format, + .set_parameter = set_parameter, + .reconfig = reconfig, + .reset = reset, + .update_external = update_external, + .resize = resize, + .get_target_size = get_target_size, + .render = render, + .get_image = get_image, + .screenshot = screenshot, + .perfdata = perfdata, + .destroy = destroy, +}; diff --git a/video/out/gpu/libmpv_gpu.h b/video/out/gpu/libmpv_gpu.h new file mode 100644 index 0000000..497dcc3 --- /dev/null +++ b/video/out/gpu/libmpv_gpu.h @@ -0,0 +1,40 @@ +#pragma once + +#include "video/out/libmpv.h" + +struct ra_tex; + +struct libmpv_gpu_context { + struct mpv_global *global; + struct mp_log *log; + const struct libmpv_gpu_context_fns *fns; + + struct ra_ctx *ra_ctx; + void *priv; +}; + +// Manage backend specific interaction between libmpv and ra backend, that can't +// be managed by ra itself (initialization and passing FBOs). +struct libmpv_gpu_context_fns { + // The libmpv API type name, see MPV_RENDER_PARAM_API_TYPE. + const char *api_name; + // Pretty much works like render_backend_fns.init, except that the + // API type is already checked by the caller. + // Successful init must set ctx->ra. + int (*init)(struct libmpv_gpu_context *ctx, mpv_render_param *params); + // Wrap the surface passed to mpv_render_context_render() (via the params + // array) into a ra_tex and return it. Returns a libmpv error code, and sets + // *out to a temporary object on success. The returned object is valid until + // another wrap_fbo() or done_frame() is called. + // This does not need to care about generic attributes, like flipping. + int (*wrap_fbo)(struct libmpv_gpu_context *ctx, mpv_render_param *params, + struct ra_tex **out); + // Signal that the ra_tex object obtained with wrap_fbo is no longer used. + // For certain backends, this might also be used to signal the end of + // rendering (like OpenGL doing weird crap). + void (*done_frame)(struct libmpv_gpu_context *ctx, bool ds); + // Free all data in ctx->priv. + void (*destroy)(struct libmpv_gpu_context *ctx); +}; + +extern const struct libmpv_gpu_context_fns libmpv_gpu_context_gl; diff --git a/video/out/gpu/osd.c b/video/out/gpu/osd.c new file mode 100644 index 0000000..91505a9 --- /dev/null +++ b/video/out/gpu/osd.c @@ -0,0 +1,363 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <stdlib.h> +#include <assert.h> +#include <limits.h> + +#include "common/common.h" +#include "common/msg.h" +#include "video/csputils.h" +#include "video/mp_image.h" +#include "osd.h" + +#define GLSL(x) gl_sc_add(sc, #x "\n"); + +// glBlendFuncSeparate() arguments +static const int blend_factors[SUBBITMAP_COUNT][4] = { + [SUBBITMAP_LIBASS] = {RA_BLEND_SRC_ALPHA, RA_BLEND_ONE_MINUS_SRC_ALPHA, + RA_BLEND_ONE, RA_BLEND_ONE_MINUS_SRC_ALPHA}, + [SUBBITMAP_BGRA] = {RA_BLEND_ONE, RA_BLEND_ONE_MINUS_SRC_ALPHA, + RA_BLEND_ONE, RA_BLEND_ONE_MINUS_SRC_ALPHA}, +}; + +struct vertex { + float position[2]; + float texcoord[2]; + uint8_t ass_color[4]; +}; + +static const struct ra_renderpass_input vertex_vao[] = { + {"position", RA_VARTYPE_FLOAT, 2, 1, offsetof(struct vertex, position)}, + {"texcoord" , RA_VARTYPE_FLOAT, 2, 1, offsetof(struct vertex, texcoord)}, + {"ass_color", RA_VARTYPE_BYTE_UNORM, 4, 1, offsetof(struct vertex, ass_color)}, +}; + +struct mpgl_osd_part { + enum sub_bitmap_format format; + int change_id; + struct ra_tex *texture; + int w, h; + int num_subparts; + int prev_num_subparts; + struct sub_bitmap *subparts; + int num_vertices; + struct vertex *vertices; +}; + +struct mpgl_osd { + struct mp_log *log; + struct osd_state *osd; + struct ra *ra; + struct mpgl_osd_part *parts[MAX_OSD_PARTS]; + const struct ra_format *fmt_table[SUBBITMAP_COUNT]; + bool formats[SUBBITMAP_COUNT]; + bool change_flag; // for reporting to API user only + // temporary + int stereo_mode; + struct mp_osd_res osd_res; + void *scratch; +}; + +struct mpgl_osd *mpgl_osd_init(struct ra *ra, struct mp_log *log, + struct osd_state *osd) +{ + struct mpgl_osd *ctx = talloc_ptrtype(NULL, ctx); + *ctx = (struct mpgl_osd) { + .log = log, + .osd = osd, + .ra = ra, + .change_flag = true, + .scratch = talloc_zero_size(ctx, 1), + }; + + ctx->fmt_table[SUBBITMAP_LIBASS] = ra_find_unorm_format(ra, 1, 1); + ctx->fmt_table[SUBBITMAP_BGRA] = ra_find_unorm_format(ra, 1, 4); + + for (int n = 0; n < MAX_OSD_PARTS; n++) + ctx->parts[n] = talloc_zero(ctx, struct mpgl_osd_part); + + for (int n = 0; n < SUBBITMAP_COUNT; n++) + ctx->formats[n] = !!ctx->fmt_table[n]; + + return ctx; +} + +void mpgl_osd_destroy(struct mpgl_osd *ctx) +{ + if (!ctx) + return; + + for (int n = 0; n < MAX_OSD_PARTS; n++) { + struct mpgl_osd_part *p = ctx->parts[n]; + ra_tex_free(ctx->ra, &p->texture); + } + talloc_free(ctx); +} + +static int next_pow2(int v) +{ + for (int x = 0; x < 30; x++) { + if ((1 << x) >= v) + return 1 << x; + } + return INT_MAX; +} + +static bool upload_osd(struct mpgl_osd *ctx, struct mpgl_osd_part *osd, + struct sub_bitmaps *imgs) +{ + struct ra *ra = ctx->ra; + bool ok = false; + + assert(imgs->packed); + + int req_w = next_pow2(imgs->packed_w); + int req_h = next_pow2(imgs->packed_h); + + const struct ra_format *fmt = ctx->fmt_table[imgs->format]; + assert(fmt); + + if (!osd->texture || req_w > osd->w || req_h > osd->h || + osd->format != imgs->format) + { + ra_tex_free(ra, &osd->texture); + + osd->format = imgs->format; + osd->w = MPMAX(32, req_w); + osd->h = MPMAX(32, req_h); + + MP_VERBOSE(ctx, "Reallocating OSD texture to %dx%d.\n", osd->w, osd->h); + + if (osd->w > ra->max_texture_wh || osd->h > ra->max_texture_wh) { + MP_ERR(ctx, "OSD bitmaps do not fit on a surface with the maximum " + "supported size %dx%d.\n", ra->max_texture_wh, + ra->max_texture_wh); + goto done; + } + + struct ra_tex_params params = { + .dimensions = 2, + .w = osd->w, + .h = osd->h, + .d = 1, + .format = fmt, + .render_src = true, + .src_linear = true, + .host_mutable = true, + }; + osd->texture = ra_tex_create(ra, ¶ms); + if (!osd->texture) + goto done; + } + + struct ra_tex_upload_params params = { + .tex = osd->texture, + .src = imgs->packed->planes[0], + .invalidate = true, + .rc = &(struct mp_rect){0, 0, imgs->packed_w, imgs->packed_h}, + .stride = imgs->packed->stride[0], + }; + + ok = ra->fns->tex_upload(ra, ¶ms); + +done: + return ok; +} + +static void gen_osd_cb(void *pctx, struct sub_bitmaps *imgs) +{ + struct mpgl_osd *ctx = pctx; + + if (imgs->num_parts == 0 || !ctx->formats[imgs->format]) + return; + + struct mpgl_osd_part *osd = ctx->parts[imgs->render_index]; + + bool ok = true; + if (imgs->change_id != osd->change_id) { + if (!upload_osd(ctx, osd, imgs)) + ok = false; + + osd->change_id = imgs->change_id; + ctx->change_flag = true; + } + osd->num_subparts = ok ? imgs->num_parts : 0; + + MP_TARRAY_GROW(osd, osd->subparts, osd->num_subparts); + memcpy(osd->subparts, imgs->parts, + osd->num_subparts * sizeof(osd->subparts[0])); +} + +bool mpgl_osd_draw_prepare(struct mpgl_osd *ctx, int index, + struct gl_shader_cache *sc) +{ + assert(index >= 0 && index < MAX_OSD_PARTS); + struct mpgl_osd_part *part = ctx->parts[index]; + + enum sub_bitmap_format fmt = part->format; + if (!fmt || !part->num_subparts || !part->texture) + return false; + + gl_sc_uniform_texture(sc, "osdtex", part->texture); + switch (fmt) { + case SUBBITMAP_BGRA: { + GLSL(color = texture(osdtex, texcoord).bgra;) + break; + } + case SUBBITMAP_LIBASS: { + GLSL(color = + vec4(ass_color.rgb, ass_color.a * texture(osdtex, texcoord).r);) + break; + } + default: + MP_ASSERT_UNREACHABLE(); + } + + return true; +} + +static void write_quad(struct vertex *va, struct gl_transform t, + float x0, float y0, float x1, float y1, + float tx0, float ty0, float tx1, float ty1, + float tex_w, float tex_h, const uint8_t color[4]) +{ + gl_transform_vec(t, &x0, &y0); + gl_transform_vec(t, &x1, &y1); + +#define COLOR_INIT {color[0], color[1], color[2], color[3]} + va[0] = (struct vertex){ {x0, y0}, {tx0 / tex_w, ty0 / tex_h}, COLOR_INIT }; + va[1] = (struct vertex){ {x0, y1}, {tx0 / tex_w, ty1 / tex_h}, COLOR_INIT }; + va[2] = (struct vertex){ {x1, y0}, {tx1 / tex_w, ty0 / tex_h}, COLOR_INIT }; + va[3] = (struct vertex){ {x1, y1}, {tx1 / tex_w, ty1 / tex_h}, COLOR_INIT }; + va[4] = va[2]; + va[5] = va[1]; +#undef COLOR_INIT +} + +static void generate_verts(struct mpgl_osd_part *part, struct gl_transform t) +{ + MP_TARRAY_GROW(part, part->vertices, + part->num_vertices + part->num_subparts * 6); + + for (int n = 0; n < part->num_subparts; n++) { + struct sub_bitmap *b = &part->subparts[n]; + struct vertex *va = &part->vertices[part->num_vertices]; + + // NOTE: the blend color is used with SUBBITMAP_LIBASS only, so it + // doesn't matter that we upload garbage for the other formats + uint32_t c = b->libass.color; + uint8_t color[4] = { c >> 24, (c >> 16) & 0xff, + (c >> 8) & 0xff, 255 - (c & 0xff) }; + + write_quad(va, t, + b->x, b->y, b->x + b->dw, b->y + b->dh, + b->src_x, b->src_y, b->src_x + b->w, b->src_y + b->h, + part->w, part->h, color); + + part->num_vertices += 6; + } +} + +// number of screen divisions per axis (x=0, y=1) for the current 3D mode +static void get_3d_side_by_side(int stereo_mode, int div[2]) +{ + div[0] = div[1] = 1; + switch (stereo_mode) { + case MP_STEREO3D_SBS2L: + case MP_STEREO3D_SBS2R: div[0] = 2; break; + case MP_STEREO3D_AB2R: + case MP_STEREO3D_AB2L: div[1] = 2; break; + } +} + +void mpgl_osd_draw_finish(struct mpgl_osd *ctx, int index, + struct gl_shader_cache *sc, struct ra_fbo fbo) +{ + struct mpgl_osd_part *part = ctx->parts[index]; + + int div[2]; + get_3d_side_by_side(ctx->stereo_mode, div); + + part->num_vertices = 0; + + for (int x = 0; x < div[0]; x++) { + for (int y = 0; y < div[1]; y++) { + struct gl_transform t; + gl_transform_ortho_fbo(&t, fbo); + + float a_x = ctx->osd_res.w * x; + float a_y = ctx->osd_res.h * y; + t.t[0] += a_x * t.m[0][0] + a_y * t.m[1][0]; + t.t[1] += a_x * t.m[0][1] + a_y * t.m[1][1]; + + generate_verts(part, t); + } + } + + const int *factors = &blend_factors[part->format][0]; + gl_sc_blend(sc, factors[0], factors[1], factors[2], factors[3]); + + gl_sc_dispatch_draw(sc, fbo.tex, false, vertex_vao, MP_ARRAY_SIZE(vertex_vao), + sizeof(struct vertex), part->vertices, part->num_vertices); +} + +static void set_res(struct mpgl_osd *ctx, struct mp_osd_res res, int stereo_mode) +{ + int div[2]; + get_3d_side_by_side(stereo_mode, div); + + res.w /= div[0]; + res.h /= div[1]; + ctx->osd_res = res; +} + +void mpgl_osd_generate(struct mpgl_osd *ctx, struct mp_osd_res res, double pts, + int stereo_mode, int draw_flags) +{ + for (int n = 0; n < MAX_OSD_PARTS; n++) + ctx->parts[n]->num_subparts = 0; + + set_res(ctx, res, stereo_mode); + + osd_draw(ctx->osd, ctx->osd_res, pts, draw_flags, ctx->formats, gen_osd_cb, ctx); + ctx->stereo_mode = stereo_mode; + + // Parts going away does not necessarily result in gen_osd_cb() being called + // (not even with num_parts==0), so check this separately. + for (int n = 0; n < MAX_OSD_PARTS; n++) { + struct mpgl_osd_part *part = ctx->parts[n]; + if (part->num_subparts != part->prev_num_subparts) + ctx->change_flag = true; + part->prev_num_subparts = part->num_subparts; + } +} + +// See osd_resize() for remarks. This function is an optional optimization too. +void mpgl_osd_resize(struct mpgl_osd *ctx, struct mp_osd_res res, int stereo_mode) +{ + set_res(ctx, res, stereo_mode); + osd_resize(ctx->osd, ctx->osd_res); +} + +bool mpgl_osd_check_change(struct mpgl_osd *ctx, struct mp_osd_res *res, + double pts) +{ + ctx->change_flag = false; + mpgl_osd_generate(ctx, *res, pts, 0, 0); + return ctx->change_flag; +} diff --git a/video/out/gpu/osd.h b/video/out/gpu/osd.h new file mode 100644 index 0000000..00fbc49 --- /dev/null +++ b/video/out/gpu/osd.h @@ -0,0 +1,25 @@ +#ifndef MPLAYER_GL_OSD_H +#define MPLAYER_GL_OSD_H + +#include <stdbool.h> +#include <inttypes.h> + +#include "utils.h" +#include "shader_cache.h" +#include "sub/osd.h" + +struct mpgl_osd *mpgl_osd_init(struct ra *ra, struct mp_log *log, + struct osd_state *osd); +void mpgl_osd_destroy(struct mpgl_osd *ctx); + +void mpgl_osd_generate(struct mpgl_osd *ctx, struct mp_osd_res res, double pts, + int stereo_mode, int draw_flags); +void mpgl_osd_resize(struct mpgl_osd *ctx, struct mp_osd_res res, int stereo_mode); +bool mpgl_osd_draw_prepare(struct mpgl_osd *ctx, int index, + struct gl_shader_cache *sc); +void mpgl_osd_draw_finish(struct mpgl_osd *ctx, int index, + struct gl_shader_cache *sc, struct ra_fbo fbo); +bool mpgl_osd_check_change(struct mpgl_osd *ctx, struct mp_osd_res *res, + double pts); + +#endif diff --git a/video/out/gpu/ra.c b/video/out/gpu/ra.c new file mode 100644 index 0000000..855f9b6 --- /dev/null +++ b/video/out/gpu/ra.c @@ -0,0 +1,424 @@ +#include "common/common.h" +#include "common/msg.h" +#include "video/img_format.h" + +#include "ra.h" + +void ra_add_native_resource(struct ra *ra, const char *name, void *data) +{ + struct ra_native_resource r = { + .name = name, + .data = data, + }; + MP_TARRAY_APPEND(ra, ra->native_resources, ra->num_native_resources, r); +} + +void *ra_get_native_resource(struct ra *ra, const char *name) +{ + for (int n = 0; n < ra->num_native_resources; n++) { + struct ra_native_resource *r = &ra->native_resources[n]; + if (strcmp(r->name, name) == 0) + return r->data; + } + + return NULL; +} + +struct ra_tex *ra_tex_create(struct ra *ra, const struct ra_tex_params *params) +{ + switch (params->dimensions) { + case 1: + assert(params->h == 1 && params->d == 1); + break; + case 2: + assert(params->d == 1); + break; + default: + assert(params->dimensions >= 1 && params->dimensions <= 3); + } + return ra->fns->tex_create(ra, params); +} + +void ra_tex_free(struct ra *ra, struct ra_tex **tex) +{ + if (*tex) + ra->fns->tex_destroy(ra, *tex); + *tex = NULL; +} + +struct ra_buf *ra_buf_create(struct ra *ra, const struct ra_buf_params *params) +{ + return ra->fns->buf_create(ra, params); +} + +void ra_buf_free(struct ra *ra, struct ra_buf **buf) +{ + if (*buf) + ra->fns->buf_destroy(ra, *buf); + *buf = NULL; +} + +void ra_free(struct ra **ra) +{ + if (*ra) + (*ra)->fns->destroy(*ra); + talloc_free(*ra); + *ra = NULL; +} + +size_t ra_vartype_size(enum ra_vartype type) +{ + switch (type) { + case RA_VARTYPE_INT: return sizeof(int); + case RA_VARTYPE_FLOAT: return sizeof(float); + case RA_VARTYPE_BYTE_UNORM: return 1; + default: return 0; + } +} + +struct ra_layout ra_renderpass_input_layout(struct ra_renderpass_input *input) +{ + size_t el_size = ra_vartype_size(input->type); + if (!el_size) + return (struct ra_layout){0}; + + // host data is always tightly packed + return (struct ra_layout) { + .align = 1, + .stride = el_size * input->dim_v, + .size = el_size * input->dim_v * input->dim_m, + }; +} + +static struct ra_renderpass_input *dup_inputs(void *ta_parent, + const struct ra_renderpass_input *inputs, int num_inputs) +{ + struct ra_renderpass_input *res = + talloc_memdup(ta_parent, (void *)inputs, num_inputs * sizeof(inputs[0])); + for (int n = 0; n < num_inputs; n++) + res[n].name = talloc_strdup(res, res[n].name); + return res; +} + +// Return a newly allocated deep-copy of params. +struct ra_renderpass_params *ra_renderpass_params_copy(void *ta_parent, + const struct ra_renderpass_params *params) +{ + struct ra_renderpass_params *res = talloc_ptrtype(ta_parent, res); + *res = *params; + res->inputs = dup_inputs(res, res->inputs, res->num_inputs); + res->vertex_attribs = + dup_inputs(res, res->vertex_attribs, res->num_vertex_attribs); + res->cached_program = bstrdup(res, res->cached_program); + res->vertex_shader = talloc_strdup(res, res->vertex_shader); + res->frag_shader = talloc_strdup(res, res->frag_shader); + res->compute_shader = talloc_strdup(res, res->compute_shader); + return res; +} + +struct glsl_fmt { + enum ra_ctype ctype; + int num_components; + int component_depth[4]; + const char *glsl_format; +}; + +// List taken from the GLSL specification, sans snorm and sint formats +static const struct glsl_fmt ra_glsl_fmts[] = { + {RA_CTYPE_FLOAT, 1, {16}, "r16f"}, + {RA_CTYPE_FLOAT, 1, {32}, "r32f"}, + {RA_CTYPE_FLOAT, 2, {16, 16}, "rg16f"}, + {RA_CTYPE_FLOAT, 2, {32, 32}, "rg32f"}, + {RA_CTYPE_FLOAT, 4, {16, 16, 16, 16}, "rgba16f"}, + {RA_CTYPE_FLOAT, 4, {32, 32, 32, 32}, "rgba32f"}, + {RA_CTYPE_FLOAT, 3, {11, 11, 10}, "r11f_g11f_b10f"}, + + {RA_CTYPE_UNORM, 1, {8}, "r8"}, + {RA_CTYPE_UNORM, 1, {16}, "r16"}, + {RA_CTYPE_UNORM, 2, {8, 8}, "rg8"}, + {RA_CTYPE_UNORM, 2, {16, 16}, "rg16"}, + {RA_CTYPE_UNORM, 4, {8, 8, 8, 8}, "rgba8"}, + {RA_CTYPE_UNORM, 4, {16, 16, 16, 16}, "rgba16"}, + {RA_CTYPE_UNORM, 4, {10, 10, 10, 2}, "rgb10_a2"}, + + {RA_CTYPE_UINT, 1, {8}, "r8ui"}, + {RA_CTYPE_UINT, 1, {16}, "r16ui"}, + {RA_CTYPE_UINT, 1, {32}, "r32ui"}, + {RA_CTYPE_UINT, 2, {8, 8}, "rg8ui"}, + {RA_CTYPE_UINT, 2, {16, 16}, "rg16ui"}, + {RA_CTYPE_UINT, 2, {32, 32}, "rg32ui"}, + {RA_CTYPE_UINT, 4, {8, 8, 8, 8}, "rgba8ui"}, + {RA_CTYPE_UINT, 4, {16, 16, 16, 16}, "rgba16ui"}, + {RA_CTYPE_UINT, 4, {32, 32, 32, 32}, "rgba32ui"}, + {RA_CTYPE_UINT, 4, {10, 10, 10, 2}, "rgb10_a2ui"}, +}; + +const char *ra_fmt_glsl_format(const struct ra_format *fmt) +{ + for (int n = 0; n < MP_ARRAY_SIZE(ra_glsl_fmts); n++) { + const struct glsl_fmt *gfmt = &ra_glsl_fmts[n]; + + if (fmt->ctype != gfmt->ctype) + continue; + if (fmt->num_components != gfmt->num_components) + continue; + + for (int i = 0; i < fmt->num_components; i++) { + if (fmt->component_depth[i] != gfmt->component_depth[i]) + goto next_fmt; + } + + return gfmt->glsl_format; + +next_fmt: ; // equivalent to `continue` + } + + return NULL; +} + +// Return whether this is a tightly packed format with no external padding and +// with the same bit size/depth in all components, and the shader returns +// components in the same order as in memory. +static bool ra_format_is_regular(const struct ra_format *fmt) +{ + if (!fmt->pixel_size || !fmt->num_components || !fmt->ordered) + return false; + for (int n = 1; n < fmt->num_components; n++) { + if (fmt->component_size[n] != fmt->component_size[0] || + fmt->component_depth[n] != fmt->component_depth[0]) + return false; + } + if (fmt->component_size[0] * fmt->num_components != fmt->pixel_size * 8) + return false; + return true; +} + +// Return a regular filterable format using RA_CTYPE_UNORM. +const struct ra_format *ra_find_unorm_format(struct ra *ra, + int bytes_per_component, + int n_components) +{ + for (int n = 0; n < ra->num_formats; n++) { + const struct ra_format *fmt = ra->formats[n]; + if (fmt->ctype == RA_CTYPE_UNORM && fmt->num_components == n_components && + fmt->pixel_size == bytes_per_component * n_components && + fmt->component_depth[0] == bytes_per_component * 8 && + fmt->linear_filter && ra_format_is_regular(fmt)) + return fmt; + } + return NULL; +} + +// Return a regular format using RA_CTYPE_UINT. +const struct ra_format *ra_find_uint_format(struct ra *ra, + int bytes_per_component, + int n_components) +{ + for (int n = 0; n < ra->num_formats; n++) { + const struct ra_format *fmt = ra->formats[n]; + if (fmt->ctype == RA_CTYPE_UINT && fmt->num_components == n_components && + fmt->pixel_size == bytes_per_component * n_components && + fmt->component_depth[0] == bytes_per_component * 8 && + ra_format_is_regular(fmt)) + return fmt; + } + return NULL; +} + +// Find a float format of any precision that matches the C type of the same +// size for upload. +// May drop bits from the mantissa (such as selecting float16 even if +// bytes_per_component == 32); prefers possibly faster formats first. +static const struct ra_format *ra_find_float_format(struct ra *ra, + int bytes_per_component, + int n_components) +{ + // Assumes ra_format are ordered by performance. + // The >=16 check is to avoid catching fringe formats. + for (int n = 0; n < ra->num_formats; n++) { + const struct ra_format *fmt = ra->formats[n]; + if (fmt->ctype == RA_CTYPE_FLOAT && fmt->num_components == n_components && + fmt->pixel_size == bytes_per_component * n_components && + fmt->component_depth[0] >= 16 && + fmt->linear_filter && ra_format_is_regular(fmt)) + return fmt; + } + return NULL; +} + +// Return a filterable regular format that uses at least float16 internally, and +// uses a normal C float for transfer on the CPU side. (This is just so we don't +// need 32->16 bit conversion on CPU, which would be messy.) +const struct ra_format *ra_find_float16_format(struct ra *ra, int n_components) +{ + return ra_find_float_format(ra, sizeof(float), n_components); +} + +const struct ra_format *ra_find_named_format(struct ra *ra, const char *name) +{ + for (int n = 0; n < ra->num_formats; n++) { + const struct ra_format *fmt = ra->formats[n]; + if (strcmp(fmt->name, name) == 0) + return fmt; + } + return NULL; +} + +// Like ra_find_unorm_format(), but if no fixed point format is available, +// return an unsigned integer format. +static const struct ra_format *find_plane_format(struct ra *ra, int bytes, + int n_channels, + enum mp_component_type ctype) +{ + switch (ctype) { + case MP_COMPONENT_TYPE_UINT: { + const struct ra_format *f = ra_find_unorm_format(ra, bytes, n_channels); + if (f) + return f; + return ra_find_uint_format(ra, bytes, n_channels); + } + case MP_COMPONENT_TYPE_FLOAT: + return ra_find_float_format(ra, bytes, n_channels); + default: return NULL; + } +} + +// Put a mapping of imgfmt to texture formats into *out. Basically it selects +// the correct texture formats needed to represent an imgfmt in a shader, with +// textures using the same memory organization as on the CPU. +// Each plane is represented by a texture, and each texture has a RGBA +// component order. out->components describes the meaning of them. +// May return integer formats for >8 bit formats, if the driver has no +// normalized 16 bit formats. +// Returns false (and *out is not touched) if no format found. +bool ra_get_imgfmt_desc(struct ra *ra, int imgfmt, struct ra_imgfmt_desc *out) +{ + struct ra_imgfmt_desc res = {.component_type = RA_CTYPE_UNKNOWN}; + + struct mp_regular_imgfmt regfmt; + if (mp_get_regular_imgfmt(®fmt, imgfmt)) { + res.num_planes = regfmt.num_planes; + res.component_bits = regfmt.component_size * 8; + res.component_pad = regfmt.component_pad; + for (int n = 0; n < regfmt.num_planes; n++) { + struct mp_regular_imgfmt_plane *plane = ®fmt.planes[n]; + res.planes[n] = find_plane_format(ra, regfmt.component_size, + plane->num_components, + regfmt.component_type); + if (!res.planes[n]) + return false; + for (int i = 0; i < plane->num_components; i++) + res.components[n][i] = plane->components[i]; + // Dropping LSBs when shifting will lead to dropped MSBs. + if (res.component_bits > res.planes[n]->component_depth[0] && + res.component_pad < 0) + return false; + // Renderer restriction, but actually an unwanted corner case. + if (res.component_type != RA_CTYPE_UNKNOWN && + res.component_type != res.planes[n]->ctype) + return false; + res.component_type = res.planes[n]->ctype; + } + res.chroma_w = 1 << regfmt.chroma_xs; + res.chroma_h = 1 << regfmt.chroma_ys; + goto supported; + } + + for (int n = 0; n < ra->num_formats; n++) { + if (imgfmt && ra->formats[n]->special_imgfmt == imgfmt) { + res = *ra->formats[n]->special_imgfmt_desc; + goto supported; + } + } + + // Unsupported format + return false; + +supported: + + *out = res; + return true; +} + +static const char *ctype_to_str(enum ra_ctype ctype) +{ + switch (ctype) { + case RA_CTYPE_UNORM: return "unorm"; + case RA_CTYPE_UINT: return "uint "; + case RA_CTYPE_FLOAT: return "float"; + default: return "unknown"; + } +} + +void ra_dump_tex_formats(struct ra *ra, int msgl) +{ + if (!mp_msg_test(ra->log, msgl)) + return; + MP_MSG(ra, msgl, "Texture formats:\n"); + MP_MSG(ra, msgl, " NAME COMP*TYPE SIZE DEPTH PER COMP.\n"); + for (int n = 0; n < ra->num_formats; n++) { + const struct ra_format *fmt = ra->formats[n]; + const char *ctype = ctype_to_str(fmt->ctype); + char cl[40] = ""; + for (int i = 0; i < fmt->num_components; i++) { + mp_snprintf_cat(cl, sizeof(cl), "%s%d", i ? " " : "", + fmt->component_size[i]); + if (fmt->component_size[i] != fmt->component_depth[i]) + mp_snprintf_cat(cl, sizeof(cl), "/%d", fmt->component_depth[i]); + } + MP_MSG(ra, msgl, " %-10s %d*%s %3dB %s %s %s %s {%s}\n", fmt->name, + fmt->num_components, ctype, fmt->pixel_size, + fmt->luminance_alpha ? "LA" : " ", + fmt->linear_filter ? "LF" : " ", + fmt->renderable ? "CR" : " ", + fmt->storable ? "ST" : " ", cl); + } + MP_MSG(ra, msgl, " LA = LUMINANCE_ALPHA hack format\n"); + MP_MSG(ra, msgl, " LF = linear filterable\n"); + MP_MSG(ra, msgl, " CR = can be used for render targets\n"); + MP_MSG(ra, msgl, " ST = can be used for storable images\n"); +} + +void ra_dump_imgfmt_desc(struct ra *ra, const struct ra_imgfmt_desc *desc, + int msgl) +{ + char pl[80] = ""; + char pf[80] = ""; + for (int n = 0; n < desc->num_planes; n++) { + if (n > 0) { + mp_snprintf_cat(pl, sizeof(pl), "/"); + mp_snprintf_cat(pf, sizeof(pf), "/"); + } + char t[5] = {0}; + for (int i = 0; i < 4; i++) + t[i] = "_rgba"[desc->components[n][i]]; + for (int i = 3; i > 0 && t[i] == '_'; i--) + t[i] = '\0'; + mp_snprintf_cat(pl, sizeof(pl), "%s", t); + mp_snprintf_cat(pf, sizeof(pf), "%s", desc->planes[n]->name); + } + MP_MSG(ra, msgl, "%d planes %dx%d %d/%d [%s] (%s) [%s]\n", + desc->num_planes, desc->chroma_w, desc->chroma_h, + desc->component_bits, desc->component_pad, pf, pl, + ctype_to_str(desc->component_type)); +} + +void ra_dump_img_formats(struct ra *ra, int msgl) +{ + if (!mp_msg_test(ra->log, msgl)) + return; + MP_MSG(ra, msgl, "Image formats:\n"); + for (int imgfmt = IMGFMT_START; imgfmt < IMGFMT_END; imgfmt++) { + const char *name = mp_imgfmt_to_name(imgfmt); + if (strcmp(name, "unknown") == 0) + continue; + MP_MSG(ra, msgl, " %s", name); + struct ra_imgfmt_desc desc; + if (ra_get_imgfmt_desc(ra, imgfmt, &desc)) { + MP_MSG(ra, msgl, " => "); + ra_dump_imgfmt_desc(ra, &desc, msgl); + } else { + MP_MSG(ra, msgl, "\n"); + } + } +} diff --git a/video/out/gpu/ra.h b/video/out/gpu/ra.h new file mode 100644 index 0000000..5f229f8 --- /dev/null +++ b/video/out/gpu/ra.h @@ -0,0 +1,559 @@ +#pragma once + +#include "common/common.h" +#include "misc/bstr.h" + +// Handle for a rendering API backend. +struct ra { + struct ra_fns *fns; + void *priv; + + int glsl_version; // GLSL version (e.g. 300 => 3.0) + bool glsl_es; // use ES dialect + bool glsl_vulkan; // use vulkan dialect + + struct mp_log *log; + + // RA_CAP_* bit field. The RA backend must set supported features at init + // time. + uint64_t caps; + + // Maximum supported width and height of a 2D texture. Set by the RA backend + // at init time. + int max_texture_wh; + + // Maximum shared memory for compute shaders. Set by the RA backend at init + // time. + size_t max_shmem; + + // Maximum number of threads in a compute work group. Set by the RA backend + // at init time. + size_t max_compute_group_threads; + + // Maximum push constant size. Set by the RA backend at init time. + size_t max_pushc_size; + + // Set of supported texture formats. Must be added by RA backend at init time. + // If there are equivalent formats with different caveats, the preferred + // formats should have a lower index. (E.g. GLES3 should put rg8 before la.) + struct ra_format **formats; + int num_formats; + + // Accelerate texture uploads via an extra PBO even when + // RA_CAP_DIRECT_UPLOAD is supported. This is basically only relevant for + // OpenGL. Set by the RA user. + bool use_pbo; + + // Array of native resources. For the most part an "escape" mechanism, and + // usually does not contain parameters required for basic functionality. + struct ra_native_resource *native_resources; + int num_native_resources; +}; + +// For passing through windowing system specific parameters and such. The +// names are always internal (the libmpv render API uses mpv_render_param_type +// and maps them to names internally). +// For example, a name="x11" entry has a X11 display as (Display*)data. +struct ra_native_resource { + const char *name; + void *data; +}; + +// Add a ra_native_resource entry. Both name and data pointers must stay valid +// until ra termination. +void ra_add_native_resource(struct ra *ra, const char *name, void *data); + +// Search ra->native_resources, returns NULL on failure. +void *ra_get_native_resource(struct ra *ra, const char *name); + +enum { + RA_CAP_TEX_1D = 1 << 0, // supports 1D textures (as shader inputs) + RA_CAP_TEX_3D = 1 << 1, // supports 3D textures (as shader inputs) + RA_CAP_BLIT = 1 << 2, // supports ra_fns.blit + RA_CAP_COMPUTE = 1 << 3, // supports compute shaders + RA_CAP_DIRECT_UPLOAD = 1 << 4, // supports tex_upload without ra_buf + RA_CAP_BUF_RO = 1 << 5, // supports RA_VARTYPE_BUF_RO + RA_CAP_BUF_RW = 1 << 6, // supports RA_VARTYPE_BUF_RW + RA_CAP_NESTED_ARRAY = 1 << 7, // supports nested arrays + RA_CAP_GLOBAL_UNIFORM = 1 << 8, // supports using "naked" uniforms (not UBO) + RA_CAP_GATHER = 1 << 9, // supports textureGather in GLSL + RA_CAP_FRAGCOORD = 1 << 10, // supports reading from gl_FragCoord + RA_CAP_PARALLEL_COMPUTE = 1 << 11, // supports parallel compute shaders + RA_CAP_NUM_GROUPS = 1 << 12, // supports gl_NumWorkGroups + RA_CAP_SLOW_DR = 1 << 13, // direct rendering is assumed to be slow +}; + +enum ra_ctype { + RA_CTYPE_UNKNOWN = 0, // also used for inconsistent multi-component formats + RA_CTYPE_UNORM, // unsigned normalized integer (fixed point) formats + RA_CTYPE_UINT, // full integer formats + RA_CTYPE_FLOAT, // float formats (signed, any bit size) +}; + +// All formats must be useable as texture formats. All formats must be byte +// aligned (all pixels start and end on a byte boundary), at least as far CPU +// transfers are concerned. +struct ra_format { + // All fields are read-only after creation. + const char *name; // symbolic name for user interaction/debugging + void *priv; + enum ra_ctype ctype; // data type of each component + bool ordered; // components are sequential in memory, and returned + // by the shader in memory order (the shader can + // return arbitrary values for unused components) + int num_components; // component count, 0 if not applicable, max. 4 + int component_size[4]; // in bits, all entries 0 if not applicable + int component_depth[4]; // bits in use for each component, 0 if not applicable + // (_must_ be set if component_size[] includes padding, + // and the real procession as seen by shader is lower) + int pixel_size; // in bytes, total pixel size (0 if opaque) + bool luminance_alpha; // pre-GL_ARB_texture_rg hack for 2 component textures + // if this is set, shader must use .ra instead of .rg + // only applies to 2-component textures + bool linear_filter; // linear filtering available from shader + bool renderable; // can be used for render targets + bool storable; // can be used for storage images + bool dummy_format; // is not a real ra_format but a fake one (e.g. FBO). + // dummy formats cannot be used to create textures + + // If not 0, the format represents some sort of packed fringe format, whose + // shader representation is given by the special_imgfmt_desc pointer. + int special_imgfmt; + const struct ra_imgfmt_desc *special_imgfmt_desc; + + // This gives the GLSL image format corresponding to the format, if any. + // (e.g. rgba16ui) + const char *glsl_format; +}; + +struct ra_tex_params { + int dimensions; // 1-3 for 1D-3D textures + // Size of the texture. 1D textures require h=d=1, 2D textures require d=1. + int w, h, d; + const struct ra_format *format; + bool render_src; // must be useable as source texture in a shader + bool render_dst; // must be useable as target texture in a shader + bool storage_dst; // must be usable as a storage image (RA_VARTYPE_IMG_W) + bool blit_src; // must be usable as a blit source + bool blit_dst; // must be usable as a blit destination + bool host_mutable; // texture may be updated with tex_upload + bool downloadable; // texture can be read with tex_download + // When used as render source texture. + bool src_linear; // if false, use nearest sampling (whether this can + // be true depends on ra_format.linear_filter) + bool src_repeat; // if false, clamp texture coordinates to edge + // if true, repeat texture coordinates + bool non_normalized; // hack for GL_TEXTURE_RECTANGLE OSX idiocy + // always set to false, except in OSX code + bool external_oes; // hack for GL_TEXTURE_EXTERNAL_OES idiocy + // If non-NULL, the texture will be created with these contents. Using + // this does *not* require setting host_mutable. Otherwise, the initial + // data is undefined. + void *initial_data; +}; + +// Conflates the following typical GPU API concepts: +// - texture itself +// - sampler state +// - staging buffers for texture upload +// - framebuffer objects +// - wrappers for swapchain framebuffers +// - synchronization needed for upload/rendering/etc. +struct ra_tex { + // All fields are read-only after creation. + struct ra_tex_params params; + void *priv; +}; + +struct ra_tex_upload_params { + struct ra_tex *tex; // Texture to upload to + bool invalidate; // Discard pre-existing data not in the region uploaded + // Uploading from buffer: + struct ra_buf *buf; // Buffer to upload from (mutually exclusive with `src`) + size_t buf_offset; // Start of data within buffer (bytes) + // Uploading directly: (Note: If RA_CAP_DIRECT_UPLOAD is not set, then this + // will be internally translated to a tex_upload buffer by the RA) + const void *src; // Address of data + // For 2D textures only: + struct mp_rect *rc; // Region to upload. NULL means entire image + ptrdiff_t stride; // The size of a horizontal line in bytes (*not* texels!) +}; + +struct ra_tex_download_params { + struct ra_tex *tex; // Texture to download from + // Downloading directly (set by caller, data written to by callee): + void *dst; // Address of data (packed with no alignment) + ptrdiff_t stride; // The size of a horizontal line in bytes (*not* texels!) +}; + +// Buffer usage type. This restricts what types of operations may be performed +// on a buffer. +enum ra_buf_type { + RA_BUF_TYPE_INVALID, + RA_BUF_TYPE_TEX_UPLOAD, // texture upload buffer (pixel buffer object) + RA_BUF_TYPE_SHADER_STORAGE, // shader buffer (SSBO), for RA_VARTYPE_BUF_RW + RA_BUF_TYPE_UNIFORM, // uniform buffer (UBO), for RA_VARTYPE_BUF_RO + RA_BUF_TYPE_VERTEX, // not publicly usable (RA-internal usage) + RA_BUF_TYPE_SHARED_MEMORY, // device memory for sharing with external API +}; + +struct ra_buf_params { + enum ra_buf_type type; + size_t size; + bool host_mapped; // create a read-writable persistent mapping (ra_buf.data) + bool host_mutable; // contents may be updated via buf_update() + // If non-NULL, the buffer will be created with these contents. Otherwise, + // the initial data is undefined. + void *initial_data; +}; + +// A generic buffer, which can be used for many purposes (texture upload, +// storage buffer, uniform buffer, etc.) +struct ra_buf { + // All fields are read-only after creation. + struct ra_buf_params params; + void *data; // for persistently mapped buffers, points to the first byte + void *priv; +}; + +// Type of a shader uniform variable, or a vertex attribute. In all cases, +// vectors are matrices are done by having more than 1 value. +enum ra_vartype { + RA_VARTYPE_INVALID, + RA_VARTYPE_INT, // C: int, GLSL: int, ivec* + RA_VARTYPE_FLOAT, // C: float, GLSL: float, vec*, mat* + RA_VARTYPE_TEX, // C: ra_tex*, GLSL: various sampler types + // ra_tex.params.render_src must be true + RA_VARTYPE_IMG_W, // C: ra_tex*, GLSL: various image types + // write-only (W) image for compute shaders + // ra_tex.params.storage_dst must be true + RA_VARTYPE_BYTE_UNORM, // C: uint8_t, GLSL: int, vec* (vertex data only) + RA_VARTYPE_BUF_RO, // C: ra_buf*, GLSL: uniform buffer block + // buf type must be RA_BUF_TYPE_UNIFORM + RA_VARTYPE_BUF_RW, // C: ra_buf*, GLSL: shader storage buffer block + // buf type must be RA_BUF_TYPE_SHADER_STORAGE + RA_VARTYPE_COUNT +}; + +// Returns the host size of a ra_vartype, or 0 for abstract vartypes (e.g. tex) +size_t ra_vartype_size(enum ra_vartype type); + +// Represents a uniform, texture input parameter, and similar things. +struct ra_renderpass_input { + const char *name; // name as used in the shader + enum ra_vartype type; + // The total number of values is given by dim_v * dim_m. + int dim_v; // vector dimension (1 for non-vector and non-matrix) + int dim_m; // additional matrix dimension (dim_v x dim_m) + // Vertex data: byte offset of the attribute into the vertex struct + size_t offset; + // RA_VARTYPE_TEX: texture unit + // RA_VARTYPE_IMG_W: image unit + // RA_VARTYPE_BUF_* buffer binding point + // Other uniforms: unused + // Bindings must be unique within each namespace, as specified by + // desc_namespace() + int binding; +}; + +// Represents the layout requirements of an input value +struct ra_layout { + size_t align; // the alignment requirements (always a power of two) + size_t stride; // the delta between two rows of an array/matrix + size_t size; // the total size of the input +}; + +// Returns the host layout of a render pass input. Returns {0} for renderpass +// inputs without a corresponding host representation (e.g. textures/buffers) +struct ra_layout ra_renderpass_input_layout(struct ra_renderpass_input *input); + +enum ra_blend { + RA_BLEND_ZERO, + RA_BLEND_ONE, + RA_BLEND_SRC_ALPHA, + RA_BLEND_ONE_MINUS_SRC_ALPHA, +}; + +enum ra_renderpass_type { + RA_RENDERPASS_TYPE_INVALID, + RA_RENDERPASS_TYPE_RASTER, // vertex+fragment shader + RA_RENDERPASS_TYPE_COMPUTE, // compute shader +}; + +// Static part of a rendering pass. It conflates the following: +// - compiled shader and its list of uniforms +// - vertex attributes and its shader mappings +// - blending parameters +// (For Vulkan, this would be shader module + pipeline state.) +// Upon creation, the values of dynamic values such as uniform contents (whose +// initial values are not provided here) are required to be 0. +struct ra_renderpass_params { + enum ra_renderpass_type type; + + // Uniforms, including texture/sampler inputs. + struct ra_renderpass_input *inputs; + int num_inputs; + size_t push_constants_size; // must be <= ra.max_pushc_size and a multiple of 4 + + // Highly implementation-specific byte array storing a compiled version + // of the program. Can be used to speed up shader compilation. A backend + // xan read this in renderpass_create, or set this on the newly created + // ra_renderpass params field. + bstr cached_program; + + // --- type==RA_RENDERPASS_TYPE_RASTER only + + // Describes the format of the vertex data. When using ra.glsl_vulkan, + // the order of this array must match the vertex attribute locations. + struct ra_renderpass_input *vertex_attribs; + int num_vertex_attribs; + int vertex_stride; + + // Format of the target texture + const struct ra_format *target_format; + + // Shader text, in GLSL. (Yes, you need a GLSL compiler.) + // These are complete shaders, including prelude and declarations. + const char *vertex_shader; + const char *frag_shader; + + // Target blending mode. If enable_blend is false, the blend_ fields can + // be ignored. + bool enable_blend; + enum ra_blend blend_src_rgb; + enum ra_blend blend_dst_rgb; + enum ra_blend blend_src_alpha; + enum ra_blend blend_dst_alpha; + + // If true, the contents of `target` not written to will become undefined + bool invalidate_target; + + // --- type==RA_RENDERPASS_TYPE_COMPUTE only + + // Shader text, like vertex_shader/frag_shader. + const char *compute_shader; +}; + +struct ra_renderpass_params *ra_renderpass_params_copy(void *ta_parent, + const struct ra_renderpass_params *params); + +// Conflates the following typical GPU API concepts: +// - various kinds of shaders +// - rendering pipelines +// - descriptor sets, uniforms, other bindings +// - all synchronization necessary +// - the current values of all uniforms (this one makes it relatively stateful +// from an API perspective) +struct ra_renderpass { + // All fields are read-only after creation. + struct ra_renderpass_params params; + void *priv; +}; + +// An input value (see ra_renderpass_input). +struct ra_renderpass_input_val { + int index; // index into ra_renderpass_params.inputs[] + void *data; // pointer to data according to ra_renderpass_input + // (e.g. type==RA_VARTYPE_FLOAT+dim_v=3,dim_m=3 => float[9]) +}; + +// Parameters for performing a rendering pass (basically the dynamic params). +// These change potentially every time. +struct ra_renderpass_run_params { + struct ra_renderpass *pass; + + // Generally this lists parameters only which changed since the last + // invocation and need to be updated. The ra_renderpass instance is + // supposed to keep unchanged values from the previous run. + // For non-primitive types like textures, these entries are always added, + // even if they do not change. + struct ra_renderpass_input_val *values; + int num_values; + void *push_constants; // must be set if params.push_constants_size > 0 + + // --- pass->params.type==RA_RENDERPASS_TYPE_RASTER only + + // target->params.render_dst must be true, and target->params.format must + // match pass->params.target_format. + struct ra_tex *target; + struct mp_rect viewport; + struct mp_rect scissors; + + // (The primitive type is always a triangle list.) + void *vertex_data; + int vertex_count; // number of vertex elements, not bytes + + // --- pass->params.type==RA_RENDERPASS_TYPE_COMPUTE only + + // Number of work groups to be run in X/Y/Z dimensions. + int compute_groups[3]; +}; + +// This is an opaque type provided by the implementation, but we want to at +// least give it a saner name than void* for code readability purposes. +typedef void ra_timer; + +// Rendering API entrypoints. (Note: there are some additional hidden features +// you need to take care of. For example, hwdec mapping will be provided +// separately from ra, but might need to call into ra private code.) +struct ra_fns { + void (*destroy)(struct ra *ra); + + // Create a texture (with undefined contents). Return NULL on failure. + // This is a rare operation, and normally textures and even FBOs for + // temporary rendering intermediate data are cached. + struct ra_tex *(*tex_create)(struct ra *ra, + const struct ra_tex_params *params); + + void (*tex_destroy)(struct ra *ra, struct ra_tex *tex); + + // Upload data to a texture. This is an extremely common operation. When + // using a buffer, the contents of the buffer must exactly match the image + // - conversions between bit depth etc. are not supported. The buffer *may* + // be marked as "in use" while this operation is going on, and the contents + // must not be touched again by the API user until buf_poll returns true. + // Returns whether successful. + bool (*tex_upload)(struct ra *ra, const struct ra_tex_upload_params *params); + + // Copy data from the texture to memory. ra_tex_params.downloadable must + // have been set to true on texture creation. + bool (*tex_download)(struct ra *ra, struct ra_tex_download_params *params); + + // Create a buffer. This can be used as a persistently mapped buffer, + // a uniform buffer, a shader storage buffer or possibly others. + // Not all usage types must be supported; may return NULL if unavailable. + struct ra_buf *(*buf_create)(struct ra *ra, + const struct ra_buf_params *params); + + void (*buf_destroy)(struct ra *ra, struct ra_buf *buf); + + // Update the contents of a buffer, starting at a given offset (*must* be a + // multiple of 4) and up to a given size, with the contents of *data. This + // is an extremely common operation. Calling this while the buffer is + // considered "in use" is an error. (See: buf_poll) + void (*buf_update)(struct ra *ra, struct ra_buf *buf, ptrdiff_t offset, + const void *data, size_t size); + + // Returns if a buffer is currently "in use" or not. Updating the contents + // of a buffer (via buf_update or writing to buf->data) while it is still + // in use is an error and may result in graphical corruption. Optional, if + // NULL then all buffers are always usable. + bool (*buf_poll)(struct ra *ra, struct ra_buf *buf); + + // Returns the layout requirements of a uniform buffer element. Optional, + // but must be implemented if RA_CAP_BUF_RO is supported. + struct ra_layout (*uniform_layout)(struct ra_renderpass_input *inp); + + // Returns the layout requirements of a push constant element. Optional, + // but must be implemented if ra.max_pushc_size > 0. + struct ra_layout (*push_constant_layout)(struct ra_renderpass_input *inp); + + // Returns an abstract namespace index for a given renderpass input type. + // This will always be a value >= 0 and < RA_VARTYPE_COUNT. This is used to + // figure out which inputs may share the same value of `binding`. + int (*desc_namespace)(struct ra *ra, enum ra_vartype type); + + // Clear the dst with the given color (rgba) and within the given scissor. + // dst must have dst->params.render_dst==true. Content outside of the + // scissor is preserved. + void (*clear)(struct ra *ra, struct ra_tex *dst, float color[4], + struct mp_rect *scissor); + + // Copy a sub-rectangle from one texture to another. The source/dest region + // is always within the texture bounds. Areas outside the dest region are + // preserved. The formats of the textures must be loosely compatible. The + // dst texture can be a swapchain framebuffer, but src can not. Only 2D + // textures are supported. + // The textures must have blit_src and blit_dst set, respectively. + // Rectangles with negative width/height lead to flipping, different src/dst + // sizes lead to point scaling. Coordinates are always in pixels. + // Optional. Only available if RA_CAP_BLIT is set (if it's not set, it must + // not be called, even if it's non-NULL). + void (*blit)(struct ra *ra, struct ra_tex *dst, struct ra_tex *src, + struct mp_rect *dst_rc, struct mp_rect *src_rc); + + // Compile a shader and create a pipeline. This is a rare operation. + // The params pointer and anything it points to must stay valid until + // renderpass_destroy. + struct ra_renderpass *(*renderpass_create)(struct ra *ra, + const struct ra_renderpass_params *params); + + void (*renderpass_destroy)(struct ra *ra, struct ra_renderpass *pass); + + // Perform a render pass, basically drawing a list of triangles to a FBO. + // This is an extremely common operation. + void (*renderpass_run)(struct ra *ra, + const struct ra_renderpass_run_params *params); + + // Create a timer object. Returns NULL on failure, or if timers are + // unavailable for some reason. Optional. + ra_timer *(*timer_create)(struct ra *ra); + + void (*timer_destroy)(struct ra *ra, ra_timer *timer); + + // Start recording a timer. Note that valid usage requires you to pair + // every start with a stop. Trying to start a timer twice, or trying to + // stop a timer before having started it, consistutes invalid usage. + void (*timer_start)(struct ra *ra, ra_timer *timer); + + // Stop recording a timer. This also returns any results that have been + // measured since the last usage of this ra_timer. It's important to note + // that GPU timer measurement are asynchronous, so this function does not + // always produce a value - and the values it does produce are typically + // delayed by a few frames. When no value is available, this returns 0. + uint64_t (*timer_stop)(struct ra *ra, ra_timer *timer); + + // Associates a marker with any past error messages, for debugging + // purposes. Optional. + void (*debug_marker)(struct ra *ra, const char *msg); +}; + +struct ra_tex *ra_tex_create(struct ra *ra, const struct ra_tex_params *params); +void ra_tex_free(struct ra *ra, struct ra_tex **tex); + +struct ra_buf *ra_buf_create(struct ra *ra, const struct ra_buf_params *params); +void ra_buf_free(struct ra *ra, struct ra_buf **buf); + +void ra_free(struct ra **ra); + +const struct ra_format *ra_find_unorm_format(struct ra *ra, + int bytes_per_component, + int n_components); +const struct ra_format *ra_find_uint_format(struct ra *ra, + int bytes_per_component, + int n_components); +const struct ra_format *ra_find_float16_format(struct ra *ra, int n_components); +const struct ra_format *ra_find_named_format(struct ra *ra, const char *name); + +struct ra_imgfmt_desc { + int num_planes; + const struct ra_format *planes[4]; + // Chroma pixel size (1x1 is 4:4:4) + uint8_t chroma_w, chroma_h; + // Component storage size in bits (possibly padded). For formats with + // different sizes per component, this is arbitrary. For padded formats + // like P010 or YUV420P10, padding is included. + int component_bits; + // Like mp_regular_imgfmt.component_pad. + int component_pad; + // == planes[n].ctype (RA_CTYPE_UNKNOWN if not applicable) + enum ra_ctype component_type; + // For each texture and each texture output (rgba order) describe what + // component it returns. + // The values are like the values in mp_regular_imgfmt_plane.components[]. + // Access as components[plane_nr][component_index]. Set unused items to 0. + // For ra_format.luminance_alpha, this returns 1/2 ("rg") instead of 1/4 + // ("ra"). the logic is that the texture format has 2 channels, thus the + // data must be returned in the first two components. The renderer fixes + // this later. + uint8_t components[4][4]; +}; + +const char *ra_fmt_glsl_format(const struct ra_format *fmt); + +bool ra_get_imgfmt_desc(struct ra *ra, int imgfmt, struct ra_imgfmt_desc *out); + +void ra_dump_tex_formats(struct ra *ra, int msgl); +void ra_dump_imgfmt_desc(struct ra *ra, const struct ra_imgfmt_desc *desc, + int msgl); +void ra_dump_img_formats(struct ra *ra, int msgl); diff --git a/video/out/gpu/shader_cache.c b/video/out/gpu/shader_cache.c new file mode 100644 index 0000000..3e05173 --- /dev/null +++ b/video/out/gpu/shader_cache.c @@ -0,0 +1,1056 @@ +#include <stddef.h> +#include <stdint.h> +#include <stdlib.h> +#include <string.h> +#include <stdarg.h> +#include <assert.h> + +#include <libavutil/sha.h> +#include <libavutil/mem.h> + +#include "osdep/io.h" + +#include "common/common.h" +#include "options/path.h" +#include "stream/stream.h" +#include "shader_cache.h" +#include "utils.h" + +// Force cache flush if more than this number of shaders is created. +#define SC_MAX_ENTRIES 256 + +union uniform_val { + float f[9]; // RA_VARTYPE_FLOAT + int i[4]; // RA_VARTYPE_INT + struct ra_tex *tex; // RA_VARTYPE_TEX, RA_VARTYPE_IMG_* + struct ra_buf *buf; // RA_VARTYPE_BUF_* +}; + +enum sc_uniform_type { + SC_UNIFORM_TYPE_GLOBAL = 0, // global uniform (RA_CAP_GLOBAL_UNIFORM) + SC_UNIFORM_TYPE_UBO = 1, // uniform buffer (RA_CAP_BUF_RO) + SC_UNIFORM_TYPE_PUSHC = 2, // push constant (ra.max_pushc_size) +}; + +struct sc_uniform { + enum sc_uniform_type type; + struct ra_renderpass_input input; + const char *glsl_type; + union uniform_val v; + char *buffer_format; + // for SC_UNIFORM_TYPE_UBO/PUSHC: + struct ra_layout layout; + size_t offset; // byte offset within the buffer +}; + +struct sc_cached_uniform { + union uniform_val v; + int index; // for ra_renderpass_input_val + bool set; // whether the uniform has ever been set +}; + +struct sc_entry { + struct ra_renderpass *pass; + struct sc_cached_uniform *cached_uniforms; + int num_cached_uniforms; + bstr total; + struct timer_pool *timer; + struct ra_buf *ubo; + int ubo_index; // for ra_renderpass_input_val.index + void *pushc; +}; + +struct gl_shader_cache { + struct ra *ra; + struct mp_log *log; + + // permanent + char **exts; + int num_exts; + + // this is modified during use (gl_sc_add() etc.) and reset for each shader + bstr prelude_text; + bstr header_text; + bstr text; + + // Next binding point (texture unit, image unit, buffer binding, etc.) + // In OpenGL these are separate for each input type + int next_binding[RA_VARTYPE_COUNT]; + bool next_uniform_dynamic; + + struct ra_renderpass_params params; + + struct sc_entry **entries; + int num_entries; + + struct sc_entry *current_shader; // set by gl_sc_generate() + + struct sc_uniform *uniforms; + int num_uniforms; + + int ubo_binding; + size_t ubo_size; + size_t pushc_size; + + struct ra_renderpass_input_val *values; + int num_values; + + // For checking that the user is calling gl_sc_reset() properly. + bool needs_reset; + + bool error_state; // true if an error occurred + + // temporary buffers (avoids frequent reallocations) + bstr tmp[6]; + + // For the disk-cache. + char *cache_dir; + struct mpv_global *global; // can be NULL +}; + +struct gl_shader_cache *gl_sc_create(struct ra *ra, struct mpv_global *global, + struct mp_log *log) +{ + struct gl_shader_cache *sc = talloc_ptrtype(NULL, sc); + *sc = (struct gl_shader_cache){ + .ra = ra, + .global = global, + .log = log, + }; + gl_sc_reset(sc); + return sc; +} + +// Reset the previous pass. This must be called after gl_sc_generate and before +// starting a new shader. It may also be called on errors. +void gl_sc_reset(struct gl_shader_cache *sc) +{ + sc->prelude_text.len = 0; + sc->header_text.len = 0; + sc->text.len = 0; + for (int n = 0; n < sc->num_uniforms; n++) + talloc_free((void *)sc->uniforms[n].input.name); + sc->num_uniforms = 0; + sc->ubo_binding = 0; + sc->ubo_size = 0; + sc->pushc_size = 0; + for (int i = 0; i < RA_VARTYPE_COUNT; i++) + sc->next_binding[i] = 0; + sc->next_uniform_dynamic = false; + sc->current_shader = NULL; + sc->params = (struct ra_renderpass_params){0}; + sc->needs_reset = false; +} + +static void sc_flush_cache(struct gl_shader_cache *sc) +{ + MP_DBG(sc, "flushing shader cache\n"); + + for (int n = 0; n < sc->num_entries; n++) { + struct sc_entry *e = sc->entries[n]; + ra_buf_free(sc->ra, &e->ubo); + if (e->pass) + sc->ra->fns->renderpass_destroy(sc->ra, e->pass); + timer_pool_destroy(e->timer); + talloc_free(e); + } + sc->num_entries = 0; +} + +void gl_sc_destroy(struct gl_shader_cache *sc) +{ + if (!sc) + return; + gl_sc_reset(sc); + sc_flush_cache(sc); + talloc_free(sc); +} + +bool gl_sc_error_state(struct gl_shader_cache *sc) +{ + return sc->error_state; +} + +void gl_sc_reset_error(struct gl_shader_cache *sc) +{ + sc->error_state = false; +} + +void gl_sc_enable_extension(struct gl_shader_cache *sc, char *name) +{ + for (int n = 0; n < sc->num_exts; n++) { + if (strcmp(sc->exts[n], name) == 0) + return; + } + MP_TARRAY_APPEND(sc, sc->exts, sc->num_exts, talloc_strdup(sc, name)); +} + +#define bstr_xappend0(sc, b, s) bstr_xappend(sc, b, bstr0(s)) + +void gl_sc_add(struct gl_shader_cache *sc, const char *text) +{ + bstr_xappend0(sc, &sc->text, text); +} + +void gl_sc_addf(struct gl_shader_cache *sc, const char *textf, ...) +{ + va_list ap; + va_start(ap, textf); + bstr_xappend_vasprintf(sc, &sc->text, textf, ap); + va_end(ap); +} + +void gl_sc_hadd(struct gl_shader_cache *sc, const char *text) +{ + bstr_xappend0(sc, &sc->header_text, text); +} + +void gl_sc_haddf(struct gl_shader_cache *sc, const char *textf, ...) +{ + va_list ap; + va_start(ap, textf); + bstr_xappend_vasprintf(sc, &sc->header_text, textf, ap); + va_end(ap); +} + +void gl_sc_hadd_bstr(struct gl_shader_cache *sc, struct bstr text) +{ + bstr_xappend(sc, &sc->header_text, text); +} + +void gl_sc_paddf(struct gl_shader_cache *sc, const char *textf, ...) +{ + va_list ap; + va_start(ap, textf); + bstr_xappend_vasprintf(sc, &sc->prelude_text, textf, ap); + va_end(ap); +} + +static struct sc_uniform *find_uniform(struct gl_shader_cache *sc, + const char *name) +{ + struct sc_uniform new = { + .input = { + .dim_v = 1, + .dim_m = 1, + }, + }; + + for (int n = 0; n < sc->num_uniforms; n++) { + struct sc_uniform *u = &sc->uniforms[n]; + if (strcmp(u->input.name, name) == 0) { + const char *allocname = u->input.name; + *u = new; + u->input.name = allocname; + return u; + } + } + + // not found -> add it + new.input.name = talloc_strdup(NULL, name); + MP_TARRAY_APPEND(sc, sc->uniforms, sc->num_uniforms, new); + return &sc->uniforms[sc->num_uniforms - 1]; +} + +static int gl_sc_next_binding(struct gl_shader_cache *sc, enum ra_vartype type) +{ + return sc->next_binding[sc->ra->fns->desc_namespace(sc->ra, type)]++; +} + +void gl_sc_uniform_dynamic(struct gl_shader_cache *sc) +{ + sc->next_uniform_dynamic = true; +} + +// Updates the metadata for the given sc_uniform. Assumes sc_uniform->input +// and glsl_type/buffer_format are already set. +static void update_uniform_params(struct gl_shader_cache *sc, struct sc_uniform *u) +{ + bool dynamic = sc->next_uniform_dynamic; + sc->next_uniform_dynamic = false; + + // Try not using push constants for "large" values like matrices, since + // this is likely to both exceed the VGPR budget as well as the pushc size + // budget + bool try_pushc = u->input.dim_m == 1 || dynamic; + + // Attempt using push constants first + if (try_pushc && sc->ra->glsl_vulkan && sc->ra->max_pushc_size) { + struct ra_layout layout = sc->ra->fns->push_constant_layout(&u->input); + size_t offset = MP_ALIGN_UP(sc->pushc_size, layout.align); + // Push constants have limited size, so make sure we don't exceed this + size_t new_size = offset + layout.size; + if (new_size <= sc->ra->max_pushc_size) { + u->type = SC_UNIFORM_TYPE_PUSHC; + u->layout = layout; + u->offset = offset; + sc->pushc_size = new_size; + return; + } + } + + // Attempt using uniform buffer next. The GLSL version 440 check is due + // to explicit offsets on UBO entries. In theory we could leave away + // the offsets and support UBOs for older GL as well, but this is a nice + // safety net for driver bugs (and also rules out potentially buggy drivers) + // Also avoid UBOs for highly dynamic stuff since that requires synchronizing + // the UBO writes every frame + bool try_ubo = !(sc->ra->caps & RA_CAP_GLOBAL_UNIFORM) || !dynamic; + if (try_ubo && sc->ra->glsl_version >= 440 && (sc->ra->caps & RA_CAP_BUF_RO)) { + u->type = SC_UNIFORM_TYPE_UBO; + u->layout = sc->ra->fns->uniform_layout(&u->input); + u->offset = MP_ALIGN_UP(sc->ubo_size, u->layout.align); + sc->ubo_size = u->offset + u->layout.size; + return; + } + + // If all else fails, use global uniforms + assert(sc->ra->caps & RA_CAP_GLOBAL_UNIFORM); + u->type = SC_UNIFORM_TYPE_GLOBAL; +} + +void gl_sc_uniform_texture(struct gl_shader_cache *sc, char *name, + struct ra_tex *tex) +{ + const char *glsl_type = "sampler2D"; + if (tex->params.dimensions == 1) { + glsl_type = "sampler1D"; + } else if (tex->params.dimensions == 3) { + glsl_type = "sampler3D"; + } else if (tex->params.non_normalized) { + glsl_type = "sampler2DRect"; + } else if (tex->params.external_oes) { + glsl_type = "samplerExternalOES"; + } else if (tex->params.format->ctype == RA_CTYPE_UINT) { + glsl_type = sc->ra->glsl_es ? "highp usampler2D" : "usampler2D"; + } + + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_TEX; + u->glsl_type = glsl_type; + u->input.binding = gl_sc_next_binding(sc, u->input.type); + u->v.tex = tex; +} + +void gl_sc_uniform_image2D_wo(struct gl_shader_cache *sc, const char *name, + struct ra_tex *tex) +{ + gl_sc_enable_extension(sc, "GL_ARB_shader_image_load_store"); + + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_IMG_W; + u->glsl_type = sc->ra->glsl_es ? "writeonly highp image2D" : "writeonly image2D"; + u->input.binding = gl_sc_next_binding(sc, u->input.type); + u->v.tex = tex; +} + +void gl_sc_ssbo(struct gl_shader_cache *sc, char *name, struct ra_buf *buf, + char *format, ...) +{ + assert(sc->ra->caps & RA_CAP_BUF_RW); + gl_sc_enable_extension(sc, "GL_ARB_shader_storage_buffer_object"); + + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_BUF_RW; + u->glsl_type = ""; + u->input.binding = gl_sc_next_binding(sc, u->input.type); + u->v.buf = buf; + + va_list ap; + va_start(ap, format); + u->buffer_format = ta_vasprintf(sc, format, ap); + va_end(ap); +} + +void gl_sc_uniform_f(struct gl_shader_cache *sc, char *name, float f) +{ + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_FLOAT; + u->glsl_type = "float"; + update_uniform_params(sc, u); + u->v.f[0] = f; +} + +void gl_sc_uniform_i(struct gl_shader_cache *sc, char *name, int i) +{ + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_INT; + u->glsl_type = "int"; + update_uniform_params(sc, u); + u->v.i[0] = i; +} + +void gl_sc_uniform_vec2(struct gl_shader_cache *sc, char *name, float f[2]) +{ + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_FLOAT; + u->input.dim_v = 2; + u->glsl_type = "vec2"; + update_uniform_params(sc, u); + u->v.f[0] = f[0]; + u->v.f[1] = f[1]; +} + +void gl_sc_uniform_vec3(struct gl_shader_cache *sc, char *name, float f[3]) +{ + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_FLOAT; + u->input.dim_v = 3; + u->glsl_type = "vec3"; + update_uniform_params(sc, u); + u->v.f[0] = f[0]; + u->v.f[1] = f[1]; + u->v.f[2] = f[2]; +} + +static void transpose2x2(float r[2 * 2]) +{ + MPSWAP(float, r[0+2*1], r[1+2*0]); +} + +void gl_sc_uniform_mat2(struct gl_shader_cache *sc, char *name, + bool transpose, float *v) +{ + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_FLOAT; + u->input.dim_v = 2; + u->input.dim_m = 2; + u->glsl_type = "mat2"; + update_uniform_params(sc, u); + for (int n = 0; n < 4; n++) + u->v.f[n] = v[n]; + if (transpose) + transpose2x2(&u->v.f[0]); +} + +static void transpose3x3(float r[3 * 3]) +{ + MPSWAP(float, r[0+3*1], r[1+3*0]); + MPSWAP(float, r[0+3*2], r[2+3*0]); + MPSWAP(float, r[1+3*2], r[2+3*1]); +} + +void gl_sc_uniform_mat3(struct gl_shader_cache *sc, char *name, + bool transpose, float *v) +{ + struct sc_uniform *u = find_uniform(sc, name); + u->input.type = RA_VARTYPE_FLOAT; + u->input.dim_v = 3; + u->input.dim_m = 3; + u->glsl_type = "mat3"; + update_uniform_params(sc, u); + for (int n = 0; n < 9; n++) + u->v.f[n] = v[n]; + if (transpose) + transpose3x3(&u->v.f[0]); +} + +void gl_sc_blend(struct gl_shader_cache *sc, + enum ra_blend blend_src_rgb, + enum ra_blend blend_dst_rgb, + enum ra_blend blend_src_alpha, + enum ra_blend blend_dst_alpha) +{ + sc->params.enable_blend = true; + sc->params.blend_src_rgb = blend_src_rgb; + sc->params.blend_dst_rgb = blend_dst_rgb; + sc->params.blend_src_alpha = blend_src_alpha; + sc->params.blend_dst_alpha = blend_dst_alpha; +} + +const char *gl_sc_bvec(struct gl_shader_cache *sc, int dims) +{ + static const char *bvecs[] = { + [1] = "bool", + [2] = "bvec2", + [3] = "bvec3", + [4] = "bvec4", + }; + + static const char *vecs[] = { + [1] = "float", + [2] = "vec2", + [3] = "vec3", + [4] = "vec4", + }; + + assert(dims > 0 && dims < MP_ARRAY_SIZE(bvecs)); + return sc->ra->glsl_version >= 130 ? bvecs[dims] : vecs[dims]; +} + +static const char *vao_glsl_type(const struct ra_renderpass_input *e) +{ + // pretty dumb... too dumb, but works for us + switch (e->dim_v) { + case 1: return "float"; + case 2: return "vec2"; + case 3: return "vec3"; + case 4: return "vec4"; + default: MP_ASSERT_UNREACHABLE(); + } +} + +static void update_ubo(struct ra *ra, struct ra_buf *ubo, struct sc_uniform *u) +{ + uintptr_t src = (uintptr_t) &u->v; + size_t dst = u->offset; + struct ra_layout src_layout = ra_renderpass_input_layout(&u->input); + struct ra_layout dst_layout = u->layout; + + for (int i = 0; i < u->input.dim_m; i++) { + ra->fns->buf_update(ra, ubo, dst, (void *)src, src_layout.stride); + src += src_layout.stride; + dst += dst_layout.stride; + } +} + +static void update_pushc(struct ra *ra, void *pushc, struct sc_uniform *u) +{ + uintptr_t src = (uintptr_t) &u->v; + uintptr_t dst = (uintptr_t) pushc + (ptrdiff_t) u->offset; + struct ra_layout src_layout = ra_renderpass_input_layout(&u->input); + struct ra_layout dst_layout = u->layout; + + for (int i = 0; i < u->input.dim_m; i++) { + memcpy((void *)dst, (void *)src, src_layout.stride); + src += src_layout.stride; + dst += dst_layout.stride; + } +} + +static void update_uniform(struct gl_shader_cache *sc, struct sc_entry *e, + struct sc_uniform *u, int n) +{ + struct sc_cached_uniform *un = &e->cached_uniforms[n]; + struct ra_layout layout = ra_renderpass_input_layout(&u->input); + if (layout.size > 0 && un->set && memcmp(&un->v, &u->v, layout.size) == 0) + return; + + un->v = u->v; + un->set = true; + + static const char *desc[] = { + [SC_UNIFORM_TYPE_UBO] = "UBO", + [SC_UNIFORM_TYPE_PUSHC] = "PC", + [SC_UNIFORM_TYPE_GLOBAL] = "global", + }; + MP_TRACE(sc, "Updating %s uniform '%s'\n", desc[u->type], u->input.name); + + switch (u->type) { + case SC_UNIFORM_TYPE_GLOBAL: { + struct ra_renderpass_input_val value = { + .index = un->index, + .data = &un->v, + }; + MP_TARRAY_APPEND(sc, sc->values, sc->num_values, value); + break; + } + case SC_UNIFORM_TYPE_UBO: + assert(e->ubo); + update_ubo(sc->ra, e->ubo, u); + break; + case SC_UNIFORM_TYPE_PUSHC: + assert(e->pushc); + update_pushc(sc->ra, e->pushc, u); + break; + default: MP_ASSERT_UNREACHABLE(); + } +} + +void gl_sc_set_cache_dir(struct gl_shader_cache *sc, char *dir) +{ + talloc_free(sc->cache_dir); + if (dir && dir[0]) { + dir = mp_get_user_path(NULL, sc->global, dir); + } else { + dir = mp_find_user_file(NULL, sc->global, "cache", ""); + } + sc->cache_dir = talloc_strdup(sc, dir); + talloc_free(dir); +} + +static bool create_pass(struct gl_shader_cache *sc, struct sc_entry *entry) +{ + bool ret = false; + + void *tmp = talloc_new(NULL); + struct ra_renderpass_params params = sc->params; + + const char *cache_header = "mpv shader cache v1\n"; + char *cache_filename = NULL; + char *cache_dir = NULL; + + if (sc->cache_dir && sc->cache_dir[0]) { + // Try to load it from a disk cache. + cache_dir = mp_get_user_path(tmp, sc->global, sc->cache_dir); + + struct AVSHA *sha = av_sha_alloc(); + MP_HANDLE_OOM(sha); + av_sha_init(sha, 256); + av_sha_update(sha, entry->total.start, entry->total.len); + + uint8_t hash[256 / 8]; + av_sha_final(sha, hash); + av_free(sha); + + char hashstr[256 / 8 * 2 + 1]; + for (int n = 0; n < 256 / 8; n++) + snprintf(hashstr + n * 2, sizeof(hashstr) - n * 2, "%02X", hash[n]); + + cache_filename = mp_path_join(tmp, cache_dir, hashstr); + if (stat(cache_filename, &(struct stat){0}) == 0) { + MP_DBG(sc, "Trying to load shader from disk...\n"); + struct bstr cachedata = + stream_read_file(cache_filename, tmp, sc->global, 1000000000); + if (bstr_eatstart0(&cachedata, cache_header)) + params.cached_program = cachedata; + } + } + + // If using a UBO, also make sure to add it as an input value so the RA + // can see it + if (sc->ubo_size) { + entry->ubo_index = sc->params.num_inputs; + struct ra_renderpass_input ubo_input = { + .name = "UBO", + .type = RA_VARTYPE_BUF_RO, + .dim_v = 1, + .dim_m = 1, + .binding = sc->ubo_binding, + }; + MP_TARRAY_APPEND(sc, params.inputs, params.num_inputs, ubo_input); + } + + if (sc->pushc_size) { + params.push_constants_size = MP_ALIGN_UP(sc->pushc_size, 4); + entry->pushc = talloc_zero_size(entry, params.push_constants_size); + } + + if (sc->ubo_size) { + struct ra_buf_params ubo_params = { + .type = RA_BUF_TYPE_UNIFORM, + .size = sc->ubo_size, + .host_mutable = true, + }; + + entry->ubo = ra_buf_create(sc->ra, &ubo_params); + if (!entry->ubo) { + MP_ERR(sc, "Failed creating uniform buffer!\n"); + goto error; + } + } + + entry->pass = sc->ra->fns->renderpass_create(sc->ra, ¶ms); + if (!entry->pass) + goto error; + + if (entry->pass && cache_filename) { + bstr nc = entry->pass->params.cached_program; + if (nc.len && !bstr_equals(params.cached_program, nc)) { + mp_mkdirp(cache_dir); + + MP_DBG(sc, "Writing shader cache file: %s\n", cache_filename); + FILE *out = fopen(cache_filename, "wb"); + if (out) { + fwrite(cache_header, strlen(cache_header), 1, out); + fwrite(nc.start, nc.len, 1, out); + fclose(out); + } + } + } + + ret = true; + +error: + talloc_free(tmp); + return ret; +} + +#define ADD(x, ...) bstr_xappend_asprintf(sc, (x), __VA_ARGS__) +#define ADD_BSTR(x, s) bstr_xappend(sc, (x), (s)) + +static void add_uniforms(struct gl_shader_cache *sc, bstr *dst) +{ + // Add all of the UBO entries separately as members of their own buffer + if (sc->ubo_size > 0) { + ADD(dst, "layout(std140, binding=%d) uniform UBO {\n", sc->ubo_binding); + for (int n = 0; n < sc->num_uniforms; n++) { + struct sc_uniform *u = &sc->uniforms[n]; + if (u->type != SC_UNIFORM_TYPE_UBO) + continue; + ADD(dst, "layout(offset=%zu) %s %s;\n", u->offset, u->glsl_type, + u->input.name); + } + ADD(dst, "};\n"); + } + + // Ditto for push constants + if (sc->pushc_size > 0) { + ADD(dst, "layout(std430, push_constant) uniform PushC {\n"); + for (int n = 0; n < sc->num_uniforms; n++) { + struct sc_uniform *u = &sc->uniforms[n]; + if (u->type != SC_UNIFORM_TYPE_PUSHC) + continue; + ADD(dst, "layout(offset=%zu) %s %s;\n", u->offset, u->glsl_type, + u->input.name); + } + ADD(dst, "};\n"); + } + + for (int n = 0; n < sc->num_uniforms; n++) { + struct sc_uniform *u = &sc->uniforms[n]; + if (u->type != SC_UNIFORM_TYPE_GLOBAL) + continue; + switch (u->input.type) { + case RA_VARTYPE_INT: + case RA_VARTYPE_FLOAT: + assert(sc->ra->caps & RA_CAP_GLOBAL_UNIFORM); + MP_FALLTHROUGH; + case RA_VARTYPE_TEX: + // Vulkan requires explicitly assigning the bindings in the shader + // source. For OpenGL it's optional, but requires higher GL version + // so we don't do it (and instead have ra_gl update the bindings + // after program creation). + if (sc->ra->glsl_vulkan) + ADD(dst, "layout(binding=%d) ", u->input.binding); + ADD(dst, "uniform %s %s;\n", u->glsl_type, u->input.name); + break; + case RA_VARTYPE_BUF_RO: + ADD(dst, "layout(std140, binding=%d) uniform %s { %s };\n", + u->input.binding, u->input.name, u->buffer_format); + break; + case RA_VARTYPE_BUF_RW: + ADD(dst, "layout(std430, binding=%d) restrict coherent buffer %s { %s };\n", + u->input.binding, u->input.name, u->buffer_format); + break; + case RA_VARTYPE_IMG_W: { + // For better compatibility, we have to explicitly label the + // type of data we will be reading/writing to this image. + const char *fmt = u->v.tex->params.format->glsl_format; + + if (sc->ra->glsl_vulkan) { + if (fmt) { + ADD(dst, "layout(binding=%d, %s) ", u->input.binding, fmt); + } else { + ADD(dst, "layout(binding=%d) ", u->input.binding); + } + } else if (fmt) { + ADD(dst, "layout(%s) ", fmt); + } + ADD(dst, "uniform restrict %s %s;\n", u->glsl_type, u->input.name); + } + } + } +} + +// 1. Generate vertex and fragment shaders from the fragment shader text added +// with gl_sc_add(). The generated shader program is cached (based on the +// text), so actual compilation happens only the first time. +// 2. Update the uniforms and textures set with gl_sc_uniform_*. +// 3. Make the new shader program current (glUseProgram()). +// After that, you render, and then you call gc_sc_reset(), which does: +// 1. Unbind the program and all textures. +// 2. Reset the sc state and prepare for a new shader program. (All uniforms +// and fragment operations needed for the next program have to be re-added.) +static void gl_sc_generate(struct gl_shader_cache *sc, + enum ra_renderpass_type type, + const struct ra_format *target_format, + const struct ra_renderpass_input *vao, + int vao_len, size_t vertex_stride) +{ + int glsl_version = sc->ra->glsl_version; + int glsl_es = sc->ra->glsl_es ? glsl_version : 0; + + sc->params.type = type; + + // gl_sc_reset() must be called after ending the previous render process, + // and before starting a new one. + assert(!sc->needs_reset); + sc->needs_reset = true; + + // If using a UBO, pick a binding (needed for shader generation) + if (sc->ubo_size) + sc->ubo_binding = gl_sc_next_binding(sc, RA_VARTYPE_BUF_RO); + + for (int n = 0; n < MP_ARRAY_SIZE(sc->tmp); n++) + sc->tmp[n].len = 0; + + // set up shader text (header + uniforms + body) + bstr *header = &sc->tmp[0]; + ADD(header, "#version %d%s\n", glsl_version, glsl_es >= 300 ? " es" : ""); + if (type == RA_RENDERPASS_TYPE_COMPUTE) { + // This extension cannot be enabled in fragment shader. Enable it as + // an exception for compute shader. + ADD(header, "#extension GL_ARB_compute_shader : enable\n"); + } + for (int n = 0; n < sc->num_exts; n++) + ADD(header, "#extension %s : enable\n", sc->exts[n]); + if (glsl_es) { + ADD(header, "#ifdef GL_FRAGMENT_PRECISION_HIGH\n"); + ADD(header, "precision highp float;\n"); + ADD(header, "#else\n"); + ADD(header, "precision mediump float;\n"); + ADD(header, "#endif\n"); + + ADD(header, "precision mediump sampler2D;\n"); + if (sc->ra->caps & RA_CAP_TEX_3D) + ADD(header, "precision mediump sampler3D;\n"); + } + + if (glsl_version >= 130) { + ADD(header, "#define tex1D texture\n"); + ADD(header, "#define tex3D texture\n"); + } else { + ADD(header, "#define tex1D texture1D\n"); + ADD(header, "#define tex3D texture3D\n"); + ADD(header, "#define texture texture2D\n"); + } + + // Additional helpers. + ADD(header, "#define LUT_POS(x, lut_size)" + " mix(0.5 / (lut_size), 1.0 - 0.5 / (lut_size), (x))\n"); + + char *vert_in = glsl_version >= 130 ? "in" : "attribute"; + char *vert_out = glsl_version >= 130 ? "out" : "varying"; + char *frag_in = glsl_version >= 130 ? "in" : "varying"; + + struct bstr *vert = NULL, *frag = NULL, *comp = NULL; + + if (type == RA_RENDERPASS_TYPE_RASTER) { + // vertex shader: we don't use the vertex shader, so just setup a + // dummy, which passes through the vertex array attributes. + bstr *vert_head = &sc->tmp[1]; + ADD_BSTR(vert_head, *header); + bstr *vert_body = &sc->tmp[2]; + ADD(vert_body, "void main() {\n"); + bstr *frag_vaos = &sc->tmp[3]; + for (int n = 0; n < vao_len; n++) { + const struct ra_renderpass_input *e = &vao[n]; + const char *glsl_type = vao_glsl_type(e); + char loc[32] = {0}; + if (sc->ra->glsl_vulkan) + snprintf(loc, sizeof(loc), "layout(location=%d) ", n); + if (strcmp(e->name, "position") == 0) { + // setting raster pos. requires setting gl_Position magic variable + assert(e->dim_v == 2 && e->type == RA_VARTYPE_FLOAT); + ADD(vert_head, "%s%s vec2 vertex_position;\n", loc, vert_in); + ADD(vert_body, "gl_Position = vec4(vertex_position, 1.0, 1.0);\n"); + } else { + ADD(vert_head, "%s%s %s vertex_%s;\n", loc, vert_in, glsl_type, e->name); + ADD(vert_head, "%s%s %s %s;\n", loc, vert_out, glsl_type, e->name); + ADD(vert_body, "%s = vertex_%s;\n", e->name, e->name); + ADD(frag_vaos, "%s%s %s %s;\n", loc, frag_in, glsl_type, e->name); + } + } + ADD(vert_body, "}\n"); + vert = vert_head; + ADD_BSTR(vert, *vert_body); + + // fragment shader; still requires adding used uniforms and VAO elements + frag = &sc->tmp[4]; + ADD_BSTR(frag, *header); + if (glsl_version >= 130) { + ADD(frag, "%sout vec4 out_color;\n", + sc->ra->glsl_vulkan ? "layout(location=0) " : ""); + } + ADD_BSTR(frag, *frag_vaos); + add_uniforms(sc, frag); + + ADD_BSTR(frag, sc->prelude_text); + ADD_BSTR(frag, sc->header_text); + + ADD(frag, "void main() {\n"); + // we require _all_ frag shaders to write to a "vec4 color" + ADD(frag, "vec4 color = vec4(0.0, 0.0, 0.0, 1.0);\n"); + ADD_BSTR(frag, sc->text); + if (glsl_version >= 130) { + ADD(frag, "out_color = color;\n"); + } else { + ADD(frag, "gl_FragColor = color;\n"); + } + ADD(frag, "}\n"); + + // We need to fix the format of the render dst at renderpass creation + // time + assert(target_format); + sc->params.target_format = target_format; + } + + if (type == RA_RENDERPASS_TYPE_COMPUTE) { + comp = &sc->tmp[4]; + ADD_BSTR(comp, *header); + + add_uniforms(sc, comp); + + ADD_BSTR(comp, sc->prelude_text); + ADD_BSTR(comp, sc->header_text); + + ADD(comp, "void main() {\n"); + ADD(comp, "vec4 color = vec4(0.0, 0.0, 0.0, 1.0);\n"); // convenience + ADD_BSTR(comp, sc->text); + ADD(comp, "}\n"); + } + + bstr *hash_total = &sc->tmp[5]; + + ADD(hash_total, "type %d\n", sc->params.type); + + if (frag) { + ADD_BSTR(hash_total, *frag); + sc->params.frag_shader = frag->start; + } + ADD(hash_total, "\n"); + if (vert) { + ADD_BSTR(hash_total, *vert); + sc->params.vertex_shader = vert->start; + } + ADD(hash_total, "\n"); + if (comp) { + ADD_BSTR(hash_total, *comp); + sc->params.compute_shader = comp->start; + } + ADD(hash_total, "\n"); + + if (sc->params.enable_blend) { + ADD(hash_total, "blend %d %d %d %d\n", + sc->params.blend_src_rgb, sc->params.blend_dst_rgb, + sc->params.blend_src_alpha, sc->params.blend_dst_alpha); + } + + if (sc->params.target_format) + ADD(hash_total, "format %s\n", sc->params.target_format->name); + + struct sc_entry *entry = NULL; + for (int n = 0; n < sc->num_entries; n++) { + struct sc_entry *cur = sc->entries[n]; + if (bstr_equals(cur->total, *hash_total)) { + entry = cur; + break; + } + } + if (!entry) { + if (sc->num_entries == SC_MAX_ENTRIES) + sc_flush_cache(sc); + entry = talloc_ptrtype(NULL, entry); + *entry = (struct sc_entry){ + .total = bstrdup(entry, *hash_total), + .timer = timer_pool_create(sc->ra), + }; + + // The vertex shader uses mangled names for the vertex attributes, so + // that the fragment shader can use the "real" names. But the shader is + // expecting the vertex attribute names (at least with older GLSL + // targets for GL). + sc->params.vertex_stride = vertex_stride; + for (int n = 0; n < vao_len; n++) { + struct ra_renderpass_input attrib = vao[n]; + attrib.name = talloc_asprintf(entry, "vertex_%s", attrib.name); + MP_TARRAY_APPEND(sc, sc->params.vertex_attribs, + sc->params.num_vertex_attribs, attrib); + } + + for (int n = 0; n < sc->num_uniforms; n++) { + struct sc_cached_uniform u = {0}; + if (sc->uniforms[n].type == SC_UNIFORM_TYPE_GLOBAL) { + // global uniforms need to be made visible to the ra_renderpass + u.index = sc->params.num_inputs; + MP_TARRAY_APPEND(sc, sc->params.inputs, sc->params.num_inputs, + sc->uniforms[n].input); + } + MP_TARRAY_APPEND(entry, entry->cached_uniforms, + entry->num_cached_uniforms, u); + } + if (!create_pass(sc, entry)) + sc->error_state = true; + MP_TARRAY_APPEND(sc, sc->entries, sc->num_entries, entry); + } + + if (!entry->pass) { + sc->current_shader = NULL; + return; + } + + assert(sc->num_uniforms == entry->num_cached_uniforms); + + sc->num_values = 0; + for (int n = 0; n < sc->num_uniforms; n++) + update_uniform(sc, entry, &sc->uniforms[n], n); + + // If we're using a UBO, make sure to bind it as well + if (sc->ubo_size) { + struct ra_renderpass_input_val ubo_val = { + .index = entry->ubo_index, + .data = &entry->ubo, + }; + MP_TARRAY_APPEND(sc, sc->values, sc->num_values, ubo_val); + } + + sc->current_shader = entry; +} + +struct mp_pass_perf gl_sc_dispatch_draw(struct gl_shader_cache *sc, + struct ra_tex *target, bool discard, + const struct ra_renderpass_input *vao, + int vao_len, size_t vertex_stride, + void *vertices, size_t num_vertices) +{ + struct timer_pool *timer = NULL; + + sc->params.invalidate_target = discard; + gl_sc_generate(sc, RA_RENDERPASS_TYPE_RASTER, target->params.format, + vao, vao_len, vertex_stride); + if (!sc->current_shader) + goto error; + + timer = sc->current_shader->timer; + + struct mp_rect full_rc = {0, 0, target->params.w, target->params.h}; + + struct ra_renderpass_run_params run = { + .pass = sc->current_shader->pass, + .values = sc->values, + .num_values = sc->num_values, + .push_constants = sc->current_shader->pushc, + .target = target, + .vertex_data = vertices, + .vertex_count = num_vertices, + .viewport = full_rc, + .scissors = full_rc, + }; + + timer_pool_start(timer); + sc->ra->fns->renderpass_run(sc->ra, &run); + timer_pool_stop(timer); + +error: + gl_sc_reset(sc); + return timer_pool_measure(timer); +} + +struct mp_pass_perf gl_sc_dispatch_compute(struct gl_shader_cache *sc, + int w, int h, int d) +{ + struct timer_pool *timer = NULL; + + gl_sc_generate(sc, RA_RENDERPASS_TYPE_COMPUTE, NULL, NULL, 0, 0); + if (!sc->current_shader) + goto error; + + timer = sc->current_shader->timer; + + struct ra_renderpass_run_params run = { + .pass = sc->current_shader->pass, + .values = sc->values, + .num_values = sc->num_values, + .push_constants = sc->current_shader->pushc, + .compute_groups = {w, h, d}, + }; + + timer_pool_start(timer); + sc->ra->fns->renderpass_run(sc->ra, &run); + timer_pool_stop(timer); + +error: + gl_sc_reset(sc); + return timer_pool_measure(timer); +} diff --git a/video/out/gpu/shader_cache.h b/video/out/gpu/shader_cache.h new file mode 100644 index 0000000..7c51c7a --- /dev/null +++ b/video/out/gpu/shader_cache.h @@ -0,0 +1,66 @@ +#pragma once + +#include "common/common.h" +#include "misc/bstr.h" +#include "ra.h" + +// For mp_pass_perf +#include "video/out/vo.h" + +struct mp_log; +struct mpv_global; +struct gl_shader_cache; + +struct gl_shader_cache *gl_sc_create(struct ra *ra, struct mpv_global *global, + struct mp_log *log); +void gl_sc_destroy(struct gl_shader_cache *sc); +bool gl_sc_error_state(struct gl_shader_cache *sc); +void gl_sc_reset_error(struct gl_shader_cache *sc); +void gl_sc_add(struct gl_shader_cache *sc, const char *text); +void gl_sc_addf(struct gl_shader_cache *sc, const char *textf, ...) + PRINTF_ATTRIBUTE(2, 3); +void gl_sc_hadd(struct gl_shader_cache *sc, const char *text); +void gl_sc_haddf(struct gl_shader_cache *sc, const char *textf, ...) + PRINTF_ATTRIBUTE(2, 3); +void gl_sc_hadd_bstr(struct gl_shader_cache *sc, struct bstr text); +void gl_sc_paddf(struct gl_shader_cache *sc, const char *textf, ...) + PRINTF_ATTRIBUTE(2, 3); + +// A hint that the next data-type (i.e. non-binding) uniform is expected to +// change frequently. This refers to the _f, _i, _vecN etc. uniform types. +void gl_sc_uniform_dynamic(struct gl_shader_cache *sc); +void gl_sc_uniform_texture(struct gl_shader_cache *sc, char *name, + struct ra_tex *tex); +void gl_sc_uniform_image2D_wo(struct gl_shader_cache *sc, const char *name, + struct ra_tex *tex); +void gl_sc_ssbo(struct gl_shader_cache *sc, char *name, struct ra_buf *buf, + char *format, ...) PRINTF_ATTRIBUTE(4, 5); +void gl_sc_uniform_f(struct gl_shader_cache *sc, char *name, float f); +void gl_sc_uniform_i(struct gl_shader_cache *sc, char *name, int f); +void gl_sc_uniform_vec2(struct gl_shader_cache *sc, char *name, float f[2]); +void gl_sc_uniform_vec3(struct gl_shader_cache *sc, char *name, float f[3]); +void gl_sc_uniform_mat2(struct gl_shader_cache *sc, char *name, + bool transpose, float *v); +void gl_sc_uniform_mat3(struct gl_shader_cache *sc, char *name, + bool transpose, float *v); + +// Return the correct bvecN() variant for using mix() in this GLSL version +const char *gl_sc_bvec(struct gl_shader_cache *sc, int dims); + +void gl_sc_blend(struct gl_shader_cache *sc, + enum ra_blend blend_src_rgb, + enum ra_blend blend_dst_rgb, + enum ra_blend blend_src_alpha, + enum ra_blend blend_dst_alpha); +void gl_sc_enable_extension(struct gl_shader_cache *sc, char *name); +struct mp_pass_perf gl_sc_dispatch_draw(struct gl_shader_cache *sc, + struct ra_tex *target, bool discard, + const struct ra_renderpass_input *vao, + int vao_len, size_t vertex_stride, + void *ptr, size_t num); +struct mp_pass_perf gl_sc_dispatch_compute(struct gl_shader_cache *sc, + int w, int h, int d); +// The application can call this on errors, to reset the current shader. This +// is normally done implicitly by gl_sc_dispatch_* +void gl_sc_reset(struct gl_shader_cache *sc); +void gl_sc_set_cache_dir(struct gl_shader_cache *sc, char *dir); diff --git a/video/out/gpu/spirv.c b/video/out/gpu/spirv.c new file mode 100644 index 0000000..67088bc --- /dev/null +++ b/video/out/gpu/spirv.c @@ -0,0 +1,70 @@ +#include "common/msg.h" +#include "options/m_config.h" + +#include "spirv.h" +#include "config.h" + +extern const struct spirv_compiler_fns spirv_shaderc; + +// in probe-order +enum { + SPIRV_AUTO = 0, + SPIRV_SHADERC, // generally preferred, but not packaged everywhere +}; + +static const struct spirv_compiler_fns *compilers[] = { +#if HAVE_SHADERC + [SPIRV_SHADERC] = &spirv_shaderc, +#endif +}; + +static const struct m_opt_choice_alternatives compiler_choices[] = { + {"auto", SPIRV_AUTO}, +#if HAVE_SHADERC + {"shaderc", SPIRV_SHADERC}, +#endif + {0} +}; + +struct spirv_opts { + int compiler; +}; + +#define OPT_BASE_STRUCT struct spirv_opts +const struct m_sub_options spirv_conf = { + .opts = (const struct m_option[]) { + {"spirv-compiler", OPT_CHOICE_C(compiler, compiler_choices)}, + {0} + }, + .size = sizeof(struct spirv_opts), +}; + +bool spirv_compiler_init(struct ra_ctx *ctx) +{ + void *tmp = talloc_new(NULL); + struct spirv_opts *opts = mp_get_config_group(tmp, ctx->global, &spirv_conf); + int compiler = opts->compiler; + talloc_free(tmp); + + for (int i = SPIRV_AUTO+1; i < MP_ARRAY_SIZE(compilers); i++) { + if (compiler != SPIRV_AUTO && i != compiler) + continue; + if (!compilers[i]) + continue; + + ctx->spirv = talloc_zero(ctx, struct spirv_compiler); + ctx->spirv->log = ctx->log, + ctx->spirv->fns = compilers[i]; + + const char *name = m_opt_choice_str(compiler_choices, i); + strncpy(ctx->spirv->name, name, sizeof(ctx->spirv->name) - 1); + MP_VERBOSE(ctx, "Initializing SPIR-V compiler '%s'\n", name); + if (ctx->spirv->fns->init(ctx)) + return true; + talloc_free(ctx->spirv); + ctx->spirv = NULL; + } + + MP_ERR(ctx, "Failed initializing SPIR-V compiler!\n"); + return false; +} diff --git a/video/out/gpu/spirv.h b/video/out/gpu/spirv.h new file mode 100644 index 0000000..e3dbd4f --- /dev/null +++ b/video/out/gpu/spirv.h @@ -0,0 +1,41 @@ +#pragma once + +#include "common/msg.h" +#include "common/common.h" +#include "context.h" + +enum glsl_shader { + GLSL_SHADER_VERTEX, + GLSL_SHADER_FRAGMENT, + GLSL_SHADER_COMPUTE, +}; + +#define SPIRV_NAME_MAX_LEN 32 + +struct spirv_compiler { + char name[SPIRV_NAME_MAX_LEN]; + const struct spirv_compiler_fns *fns; + struct mp_log *log; + void *priv; + + const char *required_ext; // or NULL + int glsl_version; // GLSL version supported + int compiler_version; // for cache invalidation, may be left as 0 + int ra_caps; // RA_CAP_* provided by this implementation, if any +}; + +struct spirv_compiler_fns { + // Compile GLSL to SPIR-V, under GL_KHR_vulkan_glsl semantics. + bool (*compile_glsl)(struct spirv_compiler *spirv, void *tactx, + enum glsl_shader type, const char *glsl, + struct bstr *out_spirv); + + // Called by spirv_compiler_init / ra_ctx_destroy. These don't need to + // allocate/free ctx->spirv, that is done by the caller + bool (*init)(struct ra_ctx *ctx); + void (*uninit)(struct ra_ctx *ctx); // optional +}; + +// Initializes ctx->spirv to a valid SPIR-V compiler, or returns false on +// failure. Cleanup will be handled by ra_ctx_destroy. +bool spirv_compiler_init(struct ra_ctx *ctx); diff --git a/video/out/gpu/spirv_shaderc.c b/video/out/gpu/spirv_shaderc.c new file mode 100644 index 0000000..f285631 --- /dev/null +++ b/video/out/gpu/spirv_shaderc.c @@ -0,0 +1,125 @@ +#include "common/msg.h" + +#include "context.h" +#include "spirv.h" + +#include <shaderc/shaderc.h> + +struct priv { + shaderc_compiler_t compiler; + shaderc_compile_options_t opts; +}; + +static void shaderc_uninit(struct ra_ctx *ctx) +{ + struct priv *p = ctx->spirv->priv; + if (!p) + return; + + shaderc_compile_options_release(p->opts); + shaderc_compiler_release(p->compiler); +} + +static bool shaderc_init(struct ra_ctx *ctx) +{ + struct priv *p = ctx->spirv->priv = talloc_zero(ctx->spirv, struct priv); + + p->compiler = shaderc_compiler_initialize(); + if (!p->compiler) + goto error; + p->opts = shaderc_compile_options_initialize(); + if (!p->opts) + goto error; + + shaderc_compile_options_set_optimization_level(p->opts, + shaderc_optimization_level_performance); + if (ctx->opts.debug) + shaderc_compile_options_set_generate_debug_info(p->opts); + + int ver, rev; + shaderc_get_spv_version(&ver, &rev); + ctx->spirv->compiler_version = ver * 100 + rev; // forwards compatibility + ctx->spirv->glsl_version = 450; // impossible to query? + return true; + +error: + shaderc_uninit(ctx); + return false; +} + +static shaderc_compilation_result_t compile(struct priv *p, + enum glsl_shader type, + const char *glsl, bool debug) +{ + static const shaderc_shader_kind kinds[] = { + [GLSL_SHADER_VERTEX] = shaderc_glsl_vertex_shader, + [GLSL_SHADER_FRAGMENT] = shaderc_glsl_fragment_shader, + [GLSL_SHADER_COMPUTE] = shaderc_glsl_compute_shader, + }; + + if (debug) { + return shaderc_compile_into_spv_assembly(p->compiler, glsl, strlen(glsl), + kinds[type], "input", "main", p->opts); + } else { + return shaderc_compile_into_spv(p->compiler, glsl, strlen(glsl), + kinds[type], "input", "main", p->opts); + } +} + +static bool shaderc_compile(struct spirv_compiler *spirv, void *tactx, + enum glsl_shader type, const char *glsl, + struct bstr *out_spirv) +{ + struct priv *p = spirv->priv; + + shaderc_compilation_result_t res = compile(p, type, glsl, false); + int errs = shaderc_result_get_num_errors(res), + warn = shaderc_result_get_num_warnings(res), + msgl = errs ? MSGL_ERR : warn ? MSGL_WARN : MSGL_V; + + const char *msg = shaderc_result_get_error_message(res); + if (msg[0]) + MP_MSG(spirv, msgl, "shaderc output:\n%s", msg); + + int s = shaderc_result_get_compilation_status(res); + bool success = s == shaderc_compilation_status_success; + + static const char *results[] = { + [shaderc_compilation_status_success] = "success", + [shaderc_compilation_status_invalid_stage] = "invalid stage", + [shaderc_compilation_status_compilation_error] = "error", + [shaderc_compilation_status_internal_error] = "internal error", + [shaderc_compilation_status_null_result_object] = "no result", + [shaderc_compilation_status_invalid_assembly] = "invalid assembly", + }; + + const char *status = s < MP_ARRAY_SIZE(results) ? results[s] : "unknown"; + MP_MSG(spirv, msgl, "shaderc compile status '%s' (%d errors, %d warnings)\n", + status, errs, warn); + + if (success) { + void *bytes = (void *) shaderc_result_get_bytes(res); + out_spirv->len = shaderc_result_get_length(res); + out_spirv->start = talloc_memdup(tactx, bytes, out_spirv->len); + } + + // Also print SPIR-V disassembly for debugging purposes. Unfortunately + // there doesn't seem to be a way to get this except compiling the shader + // a second time.. + if (mp_msg_test(spirv->log, MSGL_TRACE)) { + shaderc_compilation_result_t dis = compile(p, type, glsl, true); + MP_TRACE(spirv, "Generated SPIR-V:\n%.*s", + (int)shaderc_result_get_length(dis), + shaderc_result_get_bytes(dis)); + shaderc_result_release(dis); + } + + shaderc_result_release(res); + return success; +} + +const struct spirv_compiler_fns spirv_shaderc = { + .compile_glsl = shaderc_compile, + .init = shaderc_init, + .uninit = shaderc_uninit, +}; diff --git a/video/out/gpu/user_shaders.c b/video/out/gpu/user_shaders.c new file mode 100644 index 0000000..708de87 --- /dev/null +++ b/video/out/gpu/user_shaders.c @@ -0,0 +1,463 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <assert.h> +#include <math.h> + +#include "common/msg.h" +#include "misc/ctype.h" +#include "user_shaders.h" + +static bool parse_rpn_szexpr(struct bstr line, struct szexp out[MAX_SZEXP_SIZE]) +{ + int pos = 0; + + while (line.len > 0) { + struct bstr word = bstr_strip(bstr_splitchar(line, &line, ' ')); + if (word.len == 0) + continue; + + if (pos >= MAX_SZEXP_SIZE) + return false; + + struct szexp *exp = &out[pos++]; + + if (bstr_eatend0(&word, ".w") || bstr_eatend0(&word, ".width")) { + exp->tag = SZEXP_VAR_W; + exp->val.varname = word; + continue; + } + + if (bstr_eatend0(&word, ".h") || bstr_eatend0(&word, ".height")) { + exp->tag = SZEXP_VAR_H; + exp->val.varname = word; + continue; + } + + switch (word.start[0]) { + case '+': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_ADD; continue; + case '-': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_SUB; continue; + case '*': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_MUL; continue; + case '/': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_DIV; continue; + case '%': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_MOD; continue; + case '!': exp->tag = SZEXP_OP1; exp->val.op = SZEXP_OP_NOT; continue; + case '>': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_GT; continue; + case '<': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_LT; continue; + case '=': exp->tag = SZEXP_OP2; exp->val.op = SZEXP_OP_EQ; continue; + } + + if (mp_isdigit(word.start[0])) { + exp->tag = SZEXP_CONST; + if (bstr_sscanf(word, "%f", &exp->val.cval) != 1) + return false; + continue; + } + + // Some sort of illegal expression + return false; + } + + return true; +} + +// Returns whether successful. 'result' is left untouched on failure +bool eval_szexpr(struct mp_log *log, void *priv, + bool (*lookup)(void *priv, struct bstr var, float size[2]), + struct szexp expr[MAX_SZEXP_SIZE], float *result) +{ + float stack[MAX_SZEXP_SIZE] = {0}; + int idx = 0; // points to next element to push + + for (int i = 0; i < MAX_SZEXP_SIZE; i++) { + switch (expr[i].tag) { + case SZEXP_END: + goto done; + + case SZEXP_CONST: + // Since our SZEXPs are bound by MAX_SZEXP_SIZE, it should be + // impossible to overflow the stack + assert(idx < MAX_SZEXP_SIZE); + stack[idx++] = expr[i].val.cval; + continue; + + case SZEXP_OP1: + if (idx < 1) { + mp_warn(log, "Stack underflow in RPN expression!\n"); + return false; + } + + switch (expr[i].val.op) { + case SZEXP_OP_NOT: stack[idx-1] = !stack[idx-1]; break; + default: MP_ASSERT_UNREACHABLE(); + } + continue; + + case SZEXP_OP2: + if (idx < 2) { + mp_warn(log, "Stack underflow in RPN expression!\n"); + return false; + } + + // Pop the operands in reverse order + float op2 = stack[--idx]; + float op1 = stack[--idx]; + float res = 0.0; + switch (expr[i].val.op) { + case SZEXP_OP_ADD: res = op1 + op2; break; + case SZEXP_OP_SUB: res = op1 - op2; break; + case SZEXP_OP_MUL: res = op1 * op2; break; + case SZEXP_OP_DIV: res = op1 / op2; break; + case SZEXP_OP_MOD: res = fmodf(op1, op2); break; + case SZEXP_OP_GT: res = op1 > op2; break; + case SZEXP_OP_LT: res = op1 < op2; break; + case SZEXP_OP_EQ: res = op1 == op2; break; + default: MP_ASSERT_UNREACHABLE(); + } + + if (!isfinite(res)) { + mp_warn(log, "Illegal operation in RPN expression!\n"); + return false; + } + + stack[idx++] = res; + continue; + + case SZEXP_VAR_W: + case SZEXP_VAR_H: { + struct bstr name = expr[i].val.varname; + float size[2]; + + if (!lookup(priv, name, size)) { + mp_warn(log, "Variable %.*s not found in RPN expression!\n", + BSTR_P(name)); + return false; + } + + stack[idx++] = (expr[i].tag == SZEXP_VAR_W) ? size[0] : size[1]; + continue; + } + } + } + +done: + // Return the single stack element + if (idx != 1) { + mp_warn(log, "Malformed stack after RPN expression!\n"); + return false; + } + + *result = stack[0]; + return true; +} + +static bool parse_hook(struct mp_log *log, struct bstr *body, + struct gl_user_shader_hook *out) +{ + *out = (struct gl_user_shader_hook){ + .pass_desc = bstr0("(unknown)"), + .offset = identity_trans, + .align_offset = false, + .width = {{ SZEXP_VAR_W, { .varname = bstr0("HOOKED") }}}, + .height = {{ SZEXP_VAR_H, { .varname = bstr0("HOOKED") }}}, + .cond = {{ SZEXP_CONST, { .cval = 1.0 }}}, + }; + + int hook_idx = 0; + int bind_idx = 0; + + // Parse all headers + while (true) { + struct bstr rest; + struct bstr line = bstr_strip(bstr_getline(*body, &rest)); + + // Check for the presence of the magic line beginning + if (!bstr_eatstart0(&line, "//!")) + break; + + *body = rest; + + // Parse the supported commands + if (bstr_eatstart0(&line, "HOOK")) { + if (hook_idx == SHADER_MAX_HOOKS) { + mp_err(log, "Passes may only hook up to %d textures!\n", + SHADER_MAX_HOOKS); + return false; + } + out->hook_tex[hook_idx++] = bstr_strip(line); + continue; + } + + if (bstr_eatstart0(&line, "BIND")) { + if (bind_idx == SHADER_MAX_BINDS) { + mp_err(log, "Passes may only bind up to %d textures!\n", + SHADER_MAX_BINDS); + return false; + } + out->bind_tex[bind_idx++] = bstr_strip(line); + continue; + } + + if (bstr_eatstart0(&line, "SAVE")) { + out->save_tex = bstr_strip(line); + continue; + } + + if (bstr_eatstart0(&line, "DESC")) { + out->pass_desc = bstr_strip(line); + continue; + } + + if (bstr_eatstart0(&line, "OFFSET")) { + line = bstr_strip(line); + if (bstr_equals0(line, "ALIGN")) { + out->align_offset = true; + } else { + float ox, oy; + if (bstr_sscanf(line, "%f %f", &ox, &oy) != 2) { + mp_err(log, "Error while parsing OFFSET!\n"); + return false; + } + out->offset.t[0] = ox; + out->offset.t[1] = oy; + } + continue; + } + + if (bstr_eatstart0(&line, "WIDTH")) { + if (!parse_rpn_szexpr(line, out->width)) { + mp_err(log, "Error while parsing WIDTH!\n"); + return false; + } + continue; + } + + if (bstr_eatstart0(&line, "HEIGHT")) { + if (!parse_rpn_szexpr(line, out->height)) { + mp_err(log, "Error while parsing HEIGHT!\n"); + return false; + } + continue; + } + + if (bstr_eatstart0(&line, "WHEN")) { + if (!parse_rpn_szexpr(line, out->cond)) { + mp_err(log, "Error while parsing WHEN!\n"); + return false; + } + continue; + } + + if (bstr_eatstart0(&line, "COMPONENTS")) { + if (bstr_sscanf(line, "%d", &out->components) != 1) { + mp_err(log, "Error while parsing COMPONENTS!\n"); + return false; + } + continue; + } + + if (bstr_eatstart0(&line, "COMPUTE")) { + struct compute_info *ci = &out->compute; + int num = bstr_sscanf(line, "%d %d %d %d", &ci->block_w, &ci->block_h, + &ci->threads_w, &ci->threads_h); + + if (num == 2 || num == 4) { + ci->active = true; + ci->directly_writes = true; + } else { + mp_err(log, "Error while parsing COMPUTE!\n"); + return false; + } + continue; + } + + // Unknown command type + mp_err(log, "Unrecognized command '%.*s'!\n", BSTR_P(line)); + return false; + } + + // The rest of the file up until the next magic line beginning (if any) + // shall be the shader body + if (bstr_split_tok(*body, "//!", &out->pass_body, body)) { + // Make sure the magic line is part of the rest + body->start -= 3; + body->len += 3; + } + + // Sanity checking + if (hook_idx == 0) + mp_warn(log, "Pass has no hooked textures (will be ignored)!\n"); + + return true; +} + +static bool parse_tex(struct mp_log *log, struct ra *ra, struct bstr *body, + struct gl_user_shader_tex *out) +{ + *out = (struct gl_user_shader_tex){ + .name = bstr0("USER_TEX"), + .params = { + .dimensions = 2, + .w = 1, .h = 1, .d = 1, + .render_src = true, + .src_linear = true, + }, + }; + struct ra_tex_params *p = &out->params; + + while (true) { + struct bstr rest; + struct bstr line = bstr_strip(bstr_getline(*body, &rest)); + + if (!bstr_eatstart0(&line, "//!")) + break; + + *body = rest; + + if (bstr_eatstart0(&line, "TEXTURE")) { + out->name = bstr_strip(line); + continue; + } + + if (bstr_eatstart0(&line, "SIZE")) { + p->dimensions = bstr_sscanf(line, "%d %d %d", &p->w, &p->h, &p->d); + if (p->dimensions < 1 || p->dimensions > 3 || + p->w < 1 || p->h < 1 || p->d < 1) + { + mp_err(log, "Error while parsing SIZE!\n"); + return false; + } + continue; + } + + if (bstr_eatstart0(&line, "FORMAT ")) { + p->format = NULL; + for (int n = 0; n < ra->num_formats; n++) { + const struct ra_format *fmt = ra->formats[n]; + if (bstr_equals0(line, fmt->name)) { + p->format = fmt; + break; + } + } + // (pixel_size==0 is for opaque formats) + if (!p->format || !p->format->pixel_size) { + mp_err(log, "Unrecognized/unavailable FORMAT name: '%.*s'!\n", + BSTR_P(line)); + return false; + } + continue; + } + + if (bstr_eatstart0(&line, "FILTER")) { + line = bstr_strip(line); + if (bstr_equals0(line, "LINEAR")) { + p->src_linear = true; + } else if (bstr_equals0(line, "NEAREST")) { + p->src_linear = false; + } else { + mp_err(log, "Unrecognized FILTER: '%.*s'!\n", BSTR_P(line)); + return false; + } + continue; + } + + if (bstr_eatstart0(&line, "BORDER")) { + line = bstr_strip(line); + if (bstr_equals0(line, "CLAMP")) { + p->src_repeat = false; + } else if (bstr_equals0(line, "REPEAT")) { + p->src_repeat = true; + } else { + mp_err(log, "Unrecognized BORDER: '%.*s'!\n", BSTR_P(line)); + return false; + } + continue; + } + + mp_err(log, "Unrecognized command '%.*s'!\n", BSTR_P(line)); + return false; + } + + if (!p->format) { + mp_err(log, "No FORMAT specified.\n"); + return false; + } + + if (p->src_linear && !p->format->linear_filter) { + mp_err(log, "The specified texture format cannot be filtered!\n"); + return false; + } + + // Decode the rest of the section (up to the next //! marker) as raw hex + // data for the texture + struct bstr hexdata; + if (bstr_split_tok(*body, "//!", &hexdata, body)) { + // Make sure the magic line is part of the rest + body->start -= 3; + body->len += 3; + } + + struct bstr tex; + if (!bstr_decode_hex(NULL, bstr_strip(hexdata), &tex)) { + mp_err(log, "Error while parsing TEXTURE body: must be a valid " + "hexadecimal sequence, on a single line!\n"); + return false; + } + + int expected_len = p->w * p->h * p->d * p->format->pixel_size; + if (tex.len != expected_len) { + mp_err(log, "Shader TEXTURE size mismatch: got %zd bytes, expected %d!\n", + tex.len, expected_len); + talloc_free(tex.start); + return false; + } + + p->initial_data = tex.start; + return true; +} + +void parse_user_shader(struct mp_log *log, struct ra *ra, struct bstr shader, + void *priv, + bool (*dohook)(void *p, struct gl_user_shader_hook hook), + bool (*dotex)(void *p, struct gl_user_shader_tex tex)) +{ + if (!dohook || !dotex || !shader.len) + return; + + // Skip all garbage (e.g. comments) before the first header + int pos = bstr_find(shader, bstr0("//!")); + if (pos < 0) { + mp_warn(log, "Shader appears to contain no headers!\n"); + return; + } + shader = bstr_cut(shader, pos); + + // Loop over the file + while (shader.len > 0) + { + // Peek at the first header to dispatch the right type + if (bstr_startswith0(shader, "//!TEXTURE")) { + struct gl_user_shader_tex t; + if (!parse_tex(log, ra, &shader, &t) || !dotex(priv, t)) + return; + continue; + } + + struct gl_user_shader_hook h; + if (!parse_hook(log, &shader, &h) || !dohook(priv, h)) + return; + } +} diff --git a/video/out/gpu/user_shaders.h b/video/out/gpu/user_shaders.h new file mode 100644 index 0000000..4bb7c22 --- /dev/null +++ b/video/out/gpu/user_shaders.h @@ -0,0 +1,99 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#ifndef MP_GL_USER_SHADERS_H +#define MP_GL_USER_SHADERS_H + +#include "utils.h" +#include "ra.h" + +#define SHADER_MAX_HOOKS 16 +#define SHADER_MAX_BINDS 16 +#define MAX_SZEXP_SIZE 32 + +enum szexp_op { + SZEXP_OP_ADD, + SZEXP_OP_SUB, + SZEXP_OP_MUL, + SZEXP_OP_DIV, + SZEXP_OP_MOD, + SZEXP_OP_NOT, + SZEXP_OP_GT, + SZEXP_OP_LT, + SZEXP_OP_EQ, +}; + +enum szexp_tag { + SZEXP_END = 0, // End of an RPN expression + SZEXP_CONST, // Push a constant value onto the stack + SZEXP_VAR_W, // Get the width/height of a named texture (variable) + SZEXP_VAR_H, + SZEXP_OP2, // Pop two elements and push the result of a dyadic operation + SZEXP_OP1, // Pop one element and push the result of a monadic operation +}; + +struct szexp { + enum szexp_tag tag; + union { + float cval; + struct bstr varname; + enum szexp_op op; + } val; +}; + +struct compute_info { + bool active; + int block_w, block_h; // Block size (each block corresponds to one WG) + int threads_w, threads_h; // How many threads form a working group + bool directly_writes; // If true, shader is assumed to imageStore(out_image) +}; + +struct gl_user_shader_hook { + struct bstr pass_desc; + struct bstr hook_tex[SHADER_MAX_HOOKS]; + struct bstr bind_tex[SHADER_MAX_BINDS]; + struct bstr save_tex; + struct bstr pass_body; + struct gl_transform offset; + bool align_offset; + struct szexp width[MAX_SZEXP_SIZE]; + struct szexp height[MAX_SZEXP_SIZE]; + struct szexp cond[MAX_SZEXP_SIZE]; + int components; + struct compute_info compute; +}; + +struct gl_user_shader_tex { + struct bstr name; + struct ra_tex_params params; + // for video.c + struct ra_tex *tex; +}; + +// Parse the next shader block from `body`. The callbacks are invoked on every +// valid shader block parsed. +void parse_user_shader(struct mp_log *log, struct ra *ra, struct bstr shader, + void *priv, + bool (*dohook)(void *p, struct gl_user_shader_hook hook), + bool (*dotex)(void *p, struct gl_user_shader_tex tex)); + +// Evaluate a szexp, given a lookup function for named textures +bool eval_szexpr(struct mp_log *log, void *priv, + bool (*lookup)(void *priv, struct bstr var, float size[2]), + struct szexp expr[MAX_SZEXP_SIZE], float *result); + +#endif diff --git a/video/out/gpu/utils.c b/video/out/gpu/utils.c new file mode 100644 index 0000000..8a1aacf --- /dev/null +++ b/video/out/gpu/utils.c @@ -0,0 +1,349 @@ +#include "common/msg.h" +#include "video/out/vo.h" +#include "utils.h" + +// Standard parallel 2D projection, except y1 < y0 means that the coordinate +// system is flipped, not the projection. +void gl_transform_ortho(struct gl_transform *t, float x0, float x1, + float y0, float y1) +{ + if (y1 < y0) { + float tmp = y0; + y0 = tmp - y1; + y1 = tmp; + } + + t->m[0][0] = 2.0f / (x1 - x0); + t->m[0][1] = 0.0f; + t->m[1][0] = 0.0f; + t->m[1][1] = 2.0f / (y1 - y0); + t->t[0] = -(x1 + x0) / (x1 - x0); + t->t[1] = -(y1 + y0) / (y1 - y0); +} + +// Apply the effects of one transformation to another, transforming it in the +// process. In other words: post-composes t onto x +void gl_transform_trans(struct gl_transform t, struct gl_transform *x) +{ + struct gl_transform xt = *x; + x->m[0][0] = t.m[0][0] * xt.m[0][0] + t.m[0][1] * xt.m[1][0]; + x->m[1][0] = t.m[1][0] * xt.m[0][0] + t.m[1][1] * xt.m[1][0]; + x->m[0][1] = t.m[0][0] * xt.m[0][1] + t.m[0][1] * xt.m[1][1]; + x->m[1][1] = t.m[1][0] * xt.m[0][1] + t.m[1][1] * xt.m[1][1]; + gl_transform_vec(t, &x->t[0], &x->t[1]); +} + +void gl_transform_ortho_fbo(struct gl_transform *t, struct ra_fbo fbo) +{ + int y_dir = fbo.flip ? -1 : 1; + gl_transform_ortho(t, 0, fbo.tex->params.w, 0, fbo.tex->params.h * y_dir); +} + +float gl_video_scale_ambient_lux(float lmin, float lmax, + float rmin, float rmax, float lux) +{ + assert(lmax > lmin); + + float num = (rmax - rmin) * (log10(lux) - log10(lmin)); + float den = log10(lmax) - log10(lmin); + float result = num / den + rmin; + + // clamp the result + float max = MPMAX(rmax, rmin); + float min = MPMIN(rmax, rmin); + return MPMAX(MPMIN(result, max), min); +} + +void ra_buf_pool_uninit(struct ra *ra, struct ra_buf_pool *pool) +{ + for (int i = 0; i < pool->num_buffers; i++) + ra_buf_free(ra, &pool->buffers[i]); + + talloc_free(pool->buffers); + *pool = (struct ra_buf_pool){0}; +} + +static bool ra_buf_params_compatible(const struct ra_buf_params *new, + const struct ra_buf_params *old) +{ + return new->type == old->type && + new->size <= old->size && + new->host_mapped == old->host_mapped && + new->host_mutable == old->host_mutable; +} + +static bool ra_buf_pool_grow(struct ra *ra, struct ra_buf_pool *pool) +{ + struct ra_buf *buf = ra_buf_create(ra, &pool->current_params); + if (!buf) + return false; + + MP_TARRAY_INSERT_AT(NULL, pool->buffers, pool->num_buffers, pool->index, buf); + MP_VERBOSE(ra, "Resized buffer pool of type %u to size %d\n", + pool->current_params.type, pool->num_buffers); + return true; +} + +struct ra_buf *ra_buf_pool_get(struct ra *ra, struct ra_buf_pool *pool, + const struct ra_buf_params *params) +{ + assert(!params->initial_data); + + if (!ra_buf_params_compatible(params, &pool->current_params)) { + ra_buf_pool_uninit(ra, pool); + pool->current_params = *params; + } + + // Make sure we have at least one buffer available + if (!pool->buffers && !ra_buf_pool_grow(ra, pool)) + return NULL; + + // Make sure the next buffer is available for use + if (!ra->fns->buf_poll(ra, pool->buffers[pool->index]) && + !ra_buf_pool_grow(ra, pool)) + { + return NULL; + } + + struct ra_buf *buf = pool->buffers[pool->index++]; + pool->index %= pool->num_buffers; + + return buf; +} + +bool ra_tex_upload_pbo(struct ra *ra, struct ra_buf_pool *pbo, + const struct ra_tex_upload_params *params) +{ + if (params->buf) + return ra->fns->tex_upload(ra, params); + + struct ra_tex *tex = params->tex; + size_t row_size = tex->params.dimensions == 2 ? params->stride : + tex->params.w * tex->params.format->pixel_size; + + int height = tex->params.h; + if (tex->params.dimensions == 2 && params->rc) + height = mp_rect_h(*params->rc); + + struct ra_buf_params bufparams = { + .type = RA_BUF_TYPE_TEX_UPLOAD, + .size = row_size * height * tex->params.d, + .host_mutable = true, + }; + + struct ra_buf *buf = ra_buf_pool_get(ra, pbo, &bufparams); + if (!buf) + return false; + + ra->fns->buf_update(ra, buf, 0, params->src, bufparams.size); + + struct ra_tex_upload_params newparams = *params; + newparams.buf = buf; + newparams.src = NULL; + + return ra->fns->tex_upload(ra, &newparams); +} + +struct ra_layout std140_layout(struct ra_renderpass_input *inp) +{ + size_t el_size = ra_vartype_size(inp->type); + + // std140 packing rules: + // 1. The alignment of generic values is their size in bytes + // 2. The alignment of vectors is the vector length * the base count, with + // the exception of vec3 which is always aligned like vec4 + // 3. The alignment of arrays is that of the element size rounded up to + // the nearest multiple of vec4 + // 4. Matrices are treated like arrays of vectors + // 5. Arrays/matrices are laid out with a stride equal to the alignment + size_t stride = el_size * inp->dim_v; + size_t align = stride; + if (inp->dim_v == 3) + align += el_size; + if (inp->dim_m > 1) + stride = align = MP_ALIGN_UP(stride, sizeof(float[4])); + + return (struct ra_layout) { + .align = align, + .stride = stride, + .size = stride * inp->dim_m, + }; +} + +struct ra_layout std430_layout(struct ra_renderpass_input *inp) +{ + size_t el_size = ra_vartype_size(inp->type); + + // std430 packing rules: like std140, except arrays/matrices are always + // "tightly" packed, even arrays/matrices of vec3s + size_t stride = el_size * inp->dim_v; + size_t align = stride; + if (inp->dim_v == 3 && inp->dim_m == 1) + align += el_size; + + return (struct ra_layout) { + .align = align, + .stride = stride, + .size = stride * inp->dim_m, + }; +} + +// Resize a texture to a new desired size and format if necessary +bool ra_tex_resize(struct ra *ra, struct mp_log *log, struct ra_tex **tex, + int w, int h, const struct ra_format *fmt) +{ + if (*tex) { + struct ra_tex_params cur_params = (*tex)->params; + if (cur_params.w == w && cur_params.h == h && cur_params.format == fmt) + return true; + } + + mp_dbg(log, "Resizing texture: %dx%d\n", w, h); + + if (!fmt || !fmt->renderable || !fmt->linear_filter) { + mp_err(log, "Format %s not supported.\n", fmt ? fmt->name : "(unset)"); + return false; + } + + ra_tex_free(ra, tex); + struct ra_tex_params params = { + .dimensions = 2, + .w = w, + .h = h, + .d = 1, + .format = fmt, + .src_linear = true, + .render_src = true, + .render_dst = true, + .storage_dst = fmt->storable, + .blit_src = true, + }; + + *tex = ra_tex_create(ra, ¶ms); + if (!*tex) + mp_err(log, "Error: texture could not be created.\n"); + + return *tex; +} + +struct timer_pool { + struct ra *ra; + ra_timer *timer; + bool running; // detect invalid usage + + uint64_t samples[VO_PERF_SAMPLE_COUNT]; + int sample_idx; + int sample_count; + + uint64_t sum; + uint64_t peak; +}; + +struct timer_pool *timer_pool_create(struct ra *ra) +{ + if (!ra->fns->timer_create) + return NULL; + + ra_timer *timer = ra->fns->timer_create(ra); + if (!timer) + return NULL; + + struct timer_pool *pool = talloc(NULL, struct timer_pool); + if (!pool) { + ra->fns->timer_destroy(ra, timer); + return NULL; + } + + *pool = (struct timer_pool){ .ra = ra, .timer = timer }; + return pool; +} + +void timer_pool_destroy(struct timer_pool *pool) +{ + if (!pool) + return; + + pool->ra->fns->timer_destroy(pool->ra, pool->timer); + talloc_free(pool); +} + +void timer_pool_start(struct timer_pool *pool) +{ + if (!pool) + return; + + assert(!pool->running); + pool->ra->fns->timer_start(pool->ra, pool->timer); + pool->running = true; +} + +void timer_pool_stop(struct timer_pool *pool) +{ + if (!pool) + return; + + assert(pool->running); + uint64_t res = pool->ra->fns->timer_stop(pool->ra, pool->timer); + pool->running = false; + + if (res) { + // Input res into the buffer and grab the previous value + uint64_t old = pool->samples[pool->sample_idx]; + pool->sample_count = MPMIN(pool->sample_count + 1, VO_PERF_SAMPLE_COUNT); + pool->samples[pool->sample_idx++] = res; + pool->sample_idx %= VO_PERF_SAMPLE_COUNT; + pool->sum = pool->sum + res - old; + + // Update peak if necessary + if (res >= pool->peak) { + pool->peak = res; + } else if (pool->peak == old) { + // It's possible that the last peak was the value we just removed, + // if so we need to scan for the new peak + uint64_t peak = res; + for (int i = 0; i < VO_PERF_SAMPLE_COUNT; i++) + peak = MPMAX(peak, pool->samples[i]); + pool->peak = peak; + } + } +} + +struct mp_pass_perf timer_pool_measure(struct timer_pool *pool) +{ + if (!pool) + return (struct mp_pass_perf){0}; + + struct mp_pass_perf res = { + .peak = pool->peak, + .count = pool->sample_count, + }; + + int idx = pool->sample_idx - pool->sample_count + VO_PERF_SAMPLE_COUNT; + for (int i = 0; i < res.count; i++) { + idx %= VO_PERF_SAMPLE_COUNT; + res.samples[i] = pool->samples[idx++]; + } + + if (res.count > 0) { + res.last = res.samples[res.count - 1]; + res.avg = pool->sum / res.count; + } + + return res; +} + +void mp_log_source(struct mp_log *log, int lev, const char *src) +{ + int line = 1; + if (!src) + return; + while (*src) { + const char *end = strchr(src, '\n'); + const char *next = end + 1; + if (!end) + next = end = src + strlen(src); + mp_msg(log, lev, "[%3d] %.*s\n", line, (int)(end - src), src); + line++; + src = next; + } +} diff --git a/video/out/gpu/utils.h b/video/out/gpu/utils.h new file mode 100644 index 0000000..215873e --- /dev/null +++ b/video/out/gpu/utils.h @@ -0,0 +1,108 @@ +#pragma once + +#include <stdbool.h> +#include <math.h> + +#include "ra.h" +#include "context.h" + +// A 3x2 matrix, with the translation part separate. +struct gl_transform { + // row-major, e.g. in mathematical notation: + // | m[0][0] m[0][1] | + // | m[1][0] m[1][1] | + float m[2][2]; + float t[2]; +}; + +static const struct gl_transform identity_trans = { + .m = {{1.0, 0.0}, {0.0, 1.0}}, + .t = {0.0, 0.0}, +}; + +void gl_transform_ortho(struct gl_transform *t, float x0, float x1, + float y0, float y1); + +// This treats m as an affine transformation, in other words m[2][n] gets +// added to the output. +static inline void gl_transform_vec(struct gl_transform t, float *x, float *y) +{ + float vx = *x, vy = *y; + *x = vx * t.m[0][0] + vy * t.m[0][1] + t.t[0]; + *y = vx * t.m[1][0] + vy * t.m[1][1] + t.t[1]; +} + +struct mp_rect_f { + float x0, y0, x1, y1; +}; + +// Semantic equality (fuzzy comparison) +static inline bool mp_rect_f_seq(struct mp_rect_f a, struct mp_rect_f b) +{ + return fabs(a.x0 - b.x0) < 1e-6 && fabs(a.x1 - b.x1) < 1e-6 && + fabs(a.y0 - b.y0) < 1e-6 && fabs(a.y1 - b.y1) < 1e-6; +} + +static inline void gl_transform_rect(struct gl_transform t, struct mp_rect_f *r) +{ + gl_transform_vec(t, &r->x0, &r->y0); + gl_transform_vec(t, &r->x1, &r->y1); +} + +static inline bool gl_transform_eq(struct gl_transform a, struct gl_transform b) +{ + for (int x = 0; x < 2; x++) { + for (int y = 0; y < 2; y++) { + if (a.m[x][y] != b.m[x][y]) + return false; + } + } + + return a.t[0] == b.t[0] && a.t[1] == b.t[1]; +} + +void gl_transform_trans(struct gl_transform t, struct gl_transform *x); + +void gl_transform_ortho_fbo(struct gl_transform *t, struct ra_fbo fbo); + +float gl_video_scale_ambient_lux(float lmin, float lmax, + float rmin, float rmax, float lux); + +// A pool of buffers, which can grow as needed +struct ra_buf_pool { + struct ra_buf_params current_params; + struct ra_buf **buffers; + int num_buffers; + int index; +}; + +void ra_buf_pool_uninit(struct ra *ra, struct ra_buf_pool *pool); + +// Note: params->initial_data is *not* supported +struct ra_buf *ra_buf_pool_get(struct ra *ra, struct ra_buf_pool *pool, + const struct ra_buf_params *params); + +// Helper that wraps ra_tex_upload using texture upload buffers to ensure that +// params->buf is always set. This is intended for RA-internal usage. +bool ra_tex_upload_pbo(struct ra *ra, struct ra_buf_pool *pbo, + const struct ra_tex_upload_params *params); + +// Layout rules for GLSL's packing modes +struct ra_layout std140_layout(struct ra_renderpass_input *inp); +struct ra_layout std430_layout(struct ra_renderpass_input *inp); + +bool ra_tex_resize(struct ra *ra, struct mp_log *log, struct ra_tex **tex, + int w, int h, const struct ra_format *fmt); + +// A wrapper around ra_timer that does result pooling, averaging etc. +struct timer_pool; + +struct timer_pool *timer_pool_create(struct ra *ra); +void timer_pool_destroy(struct timer_pool *pool); +void timer_pool_start(struct timer_pool *pool); +void timer_pool_stop(struct timer_pool *pool); +struct mp_pass_perf timer_pool_measure(struct timer_pool *pool); + +// print a multi line string with line numbers (e.g. for shader sources) +// log, lev: module and log level, as in mp_msg() +void mp_log_source(struct mp_log *log, int lev, const char *src); diff --git a/video/out/gpu/video.c b/video/out/gpu/video.c new file mode 100644 index 0000000..852ee78 --- /dev/null +++ b/video/out/gpu/video.c @@ -0,0 +1,4364 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <assert.h> +#include <float.h> +#include <math.h> +#include <stdarg.h> +#include <stdbool.h> +#include <string.h> + +#include <libavutil/common.h> +#include <libavutil/lfg.h> + +#include "video.h" + +#include "misc/bstr.h" +#include "options/m_config.h" +#include "options/path.h" +#include "common/global.h" +#include "options/options.h" +#include "utils.h" +#include "hwdec.h" +#include "osd.h" +#include "ra.h" +#include "stream/stream.h" +#include "video_shaders.h" +#include "user_shaders.h" +#include "error_diffusion.h" +#include "video/out/filter_kernels.h" +#include "video/out/aspect.h" +#include "video/out/dither.h" +#include "video/out/vo.h" + +// scale/cscale arguments that map directly to shader filter routines. +// Note that the convolution filters are not included in this list. +static const char *const fixed_scale_filters[] = { + "bilinear", + "bicubic_fast", + "oversample", + NULL +}; +static const char *const fixed_tscale_filters[] = { + "oversample", + "linear", + NULL +}; + +// must be sorted, and terminated with 0 +int filter_sizes[] = + {2, 4, 6, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60, 64, 0}; +int tscale_sizes[] = {2, 4, 6, 8, 0}; + +struct vertex_pt { + float x, y; +}; + +struct texplane { + struct ra_tex *tex; + int w, h; + bool flipped; +}; + +struct video_image { + struct texplane planes[4]; + struct mp_image *mpi; // original input image + uint64_t id; // unique ID identifying mpi contents + bool hwdec_mapped; +}; + +enum plane_type { + PLANE_NONE = 0, + PLANE_RGB, + PLANE_LUMA, + PLANE_CHROMA, + PLANE_ALPHA, + PLANE_XYZ, +}; + +static const char *plane_names[] = { + [PLANE_NONE] = "unknown", + [PLANE_RGB] = "rgb", + [PLANE_LUMA] = "luma", + [PLANE_CHROMA] = "chroma", + [PLANE_ALPHA] = "alpha", + [PLANE_XYZ] = "xyz", +}; + +// A self-contained description of a source image which can be bound to a +// texture unit and sampled from. Contains metadata about how it's to be used +struct image { + enum plane_type type; // must be set to something non-zero + int components; // number of relevant coordinates + float multiplier; // multiplier to be used when sampling + struct ra_tex *tex; + int w, h; // logical size (after transformation) + struct gl_transform transform; // rendering transformation + int padding; // number of leading padding components (e.g. 2 = rg is padding) +}; + +// A named image, for user scripting purposes +struct saved_img { + const char *name; + struct image img; +}; + +// A texture hook. This is some operation that transforms a named texture as +// soon as it's generated +struct tex_hook { + const char *save_tex; + const char *hook_tex[SHADER_MAX_HOOKS]; + const char *bind_tex[SHADER_MAX_BINDS]; + int components; // how many components are relevant (0 = same as input) + bool align_offset; // whether to align hooked tex with reference. + void *priv; // this gets talloc_freed when the tex_hook is removed + void (*hook)(struct gl_video *p, struct image img, // generates GLSL + struct gl_transform *trans, void *priv); + bool (*cond)(struct gl_video *p, struct image img, void *priv); +}; + +struct surface { + struct ra_tex *tex; + uint64_t id; + double pts; +}; + +#define SURFACES_MAX 10 + +struct cached_file { + char *path; + struct bstr body; +}; + +struct pass_info { + struct bstr desc; + struct mp_pass_perf perf; +}; + +struct dr_buffer { + struct ra_buf *buf; + // The mpi reference will keep the data from being recycled (or from other + // references gaining write access) while the GPU is accessing the buffer. + struct mp_image *mpi; +}; + +struct gl_video { + struct ra *ra; + + struct mpv_global *global; + struct mp_log *log; + struct gl_video_opts opts; + struct m_config_cache *opts_cache; + struct gl_lcms *cms; + + int fb_depth; // actual bits available in GL main framebuffer + struct m_color clear_color; + bool force_clear_color; + + struct gl_shader_cache *sc; + + struct osd_state *osd_state; + struct mpgl_osd *osd; + double osd_pts; + + struct ra_tex *lut_3d_texture; + bool use_lut_3d; + int lut_3d_size[3]; + + struct ra_tex *dither_texture; + + struct mp_image_params real_image_params; // configured format + struct mp_image_params image_params; // texture format (mind hwdec case) + struct ra_imgfmt_desc ra_format; // texture format + int plane_count; + + bool is_gray; + bool has_alpha; + char color_swizzle[5]; + bool use_integer_conversion; + + struct video_image image; + + struct dr_buffer *dr_buffers; + int num_dr_buffers; + + bool using_dr_path; + + bool dumb_mode; + bool forced_dumb_mode; + + // Cached vertex array, to avoid re-allocation per frame. For simplicity, + // our vertex format is simply a list of `vertex_pt`s, since this greatly + // simplifies offset calculation at the cost of (unneeded) flexibility. + struct vertex_pt *tmp_vertex; + struct ra_renderpass_input *vao; + int vao_len; + + const struct ra_format *fbo_format; + struct ra_tex *merge_tex[4]; + struct ra_tex *scale_tex[4]; + struct ra_tex *integer_tex[4]; + struct ra_tex *indirect_tex; + struct ra_tex *blend_subs_tex; + struct ra_tex *error_diffusion_tex[2]; + struct ra_tex *screen_tex; + struct ra_tex *output_tex; + struct ra_tex **hook_textures; + int num_hook_textures; + int idx_hook_textures; + + struct ra_buf *hdr_peak_ssbo; + struct surface surfaces[SURFACES_MAX]; + + // user pass descriptions and textures + struct tex_hook *tex_hooks; + int num_tex_hooks; + struct gl_user_shader_tex *user_textures; + int num_user_textures; + + int surface_idx; + int surface_now; + int frames_drawn; + bool is_interpolated; + bool output_tex_valid; + + // state for configured scalers + struct scaler scaler[SCALER_COUNT]; + + struct mp_csp_equalizer_state *video_eq; + + struct mp_rect src_rect; // displayed part of the source video + struct mp_rect dst_rect; // video rectangle on output window + struct mp_osd_res osd_rect; // OSD size/margins + + // temporary during rendering + struct compute_info pass_compute; // compute shader metadata for this pass + struct image *pass_imgs; // bound images for this pass + int num_pass_imgs; + struct saved_img *saved_imgs; // saved (named) images for this frame + int num_saved_imgs; + + // effective current texture metadata - this will essentially affect the + // next render pass target, as well as implicitly tracking what needs to + // be done with the image + int texture_w, texture_h; + struct gl_transform texture_offset; // texture transform without rotation + int components; + bool use_linear; + float user_gamma; + + // pass info / metrics + struct pass_info pass_fresh[VO_PASS_PERF_MAX]; + struct pass_info pass_redraw[VO_PASS_PERF_MAX]; + struct pass_info *pass; + int pass_idx; + struct timer_pool *upload_timer; + struct timer_pool *blit_timer; + struct timer_pool *osd_timer; + + int frames_uploaded; + int frames_rendered; + AVLFG lfg; + + // Cached because computing it can take relatively long + int last_dither_matrix_size; + float *last_dither_matrix; + + struct cached_file *files; + int num_files; + + struct ra_hwdec_ctx hwdec_ctx; + struct ra_hwdec_mapper *hwdec_mapper; + struct ra_hwdec *hwdec_overlay; + bool hwdec_active; + + bool dsi_warned; + bool broken_frame; // temporary error state + + bool colorspace_override_warned; + bool correct_downscaling_warned; +}; + +static const struct gl_video_opts gl_video_opts_def = { + .dither_algo = DITHER_FRUIT, + .dither_size = 6, + .temporal_dither_period = 1, + .error_diffusion = "sierra-lite", + .fbo_format = "auto", + .sigmoid_center = 0.75, + .sigmoid_slope = 6.5, + .scaler = { + {{"lanczos", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // scale + {{"hermite", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // dscale + {{NULL, .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // cscale + {{"oversample", .params={NAN, NAN}}, {.params = {NAN, NAN}}}, // tscale + }, + .scaler_resizes_only = true, + .correct_downscaling = true, + .linear_downscaling = true, + .sigmoid_upscaling = true, + .interpolation_threshold = 0.01, + .alpha_mode = ALPHA_BLEND_TILES, + .background = {0, 0, 0, 255}, + .gamma = 1.0f, + .tone_map = { + .curve = TONE_MAPPING_AUTO, + .curve_param = NAN, + .max_boost = 1.0, + .decay_rate = 20.0, + .scene_threshold_low = 1.0, + .scene_threshold_high = 3.0, + .contrast_smoothness = 3.5, + }, + .early_flush = -1, + .shader_cache = true, + .hwdec_interop = "auto", +}; + +static int validate_scaler_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value); + +static int validate_window_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value); + +static int validate_error_diffusion_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value); + +#define OPT_BASE_STRUCT struct gl_video_opts + +// Use for options which use NAN for defaults. +#define OPT_FLOATDEF(field) \ + OPT_FLOAT(field), \ + .flags = M_OPT_DEFAULT_NAN + +#define SCALER_OPTS(n, i) \ + {n, OPT_STRING_VALIDATE(scaler[i].kernel.name, validate_scaler_opt)}, \ + {n"-param1", OPT_FLOATDEF(scaler[i].kernel.params[0])}, \ + {n"-param2", OPT_FLOATDEF(scaler[i].kernel.params[1])}, \ + {n"-blur", OPT_FLOAT(scaler[i].kernel.blur)}, \ + {n"-cutoff", OPT_REMOVED("Hard-coded as 0.001")}, \ + {n"-taper", OPT_FLOAT(scaler[i].kernel.taper), M_RANGE(0.0, 1.0)}, \ + {n"-wparam", OPT_FLOATDEF(scaler[i].window.params[0])}, \ + {n"-wblur", OPT_REMOVED("Just adjust filter radius directly")}, \ + {n"-wtaper", OPT_FLOAT(scaler[i].window.taper), M_RANGE(0.0, 1.0)}, \ + {n"-clamp", OPT_FLOAT(scaler[i].clamp), M_RANGE(0.0, 1.0)}, \ + {n"-radius", OPT_FLOAT(scaler[i].radius), M_RANGE(0.5, 16.0)}, \ + {n"-antiring", OPT_FLOAT(scaler[i].antiring), M_RANGE(0.0, 1.0)}, \ + {n"-window", OPT_STRING_VALIDATE(scaler[i].window.name, validate_window_opt)} + +const struct m_sub_options gl_video_conf = { + .opts = (const m_option_t[]) { + {"gpu-dumb-mode", OPT_CHOICE(dumb_mode, + {"auto", 0}, {"yes", 1}, {"no", -1})}, + {"gamma-factor", OPT_FLOAT(gamma), M_RANGE(0.1, 2.0), + .deprecation_message = "no replacement"}, + {"gamma-auto", OPT_BOOL(gamma_auto), + .deprecation_message = "no replacement"}, + {"target-prim", OPT_CHOICE_C(target_prim, mp_csp_prim_names)}, + {"target-trc", OPT_CHOICE_C(target_trc, mp_csp_trc_names)}, + {"target-peak", OPT_CHOICE(target_peak, {"auto", 0}), + M_RANGE(10, 10000)}, + {"target-contrast", OPT_CHOICE(target_contrast, {"auto", 0}, {"inf", -1}), + M_RANGE(10, 1000000)}, + {"target-gamut", OPT_CHOICE_C(target_gamut, mp_csp_prim_names)}, + {"tone-mapping", OPT_CHOICE(tone_map.curve, + {"auto", TONE_MAPPING_AUTO}, + {"clip", TONE_MAPPING_CLIP}, + {"mobius", TONE_MAPPING_MOBIUS}, + {"reinhard", TONE_MAPPING_REINHARD}, + {"hable", TONE_MAPPING_HABLE}, + {"gamma", TONE_MAPPING_GAMMA}, + {"linear", TONE_MAPPING_LINEAR}, + {"spline", TONE_MAPPING_SPLINE}, + {"bt.2390", TONE_MAPPING_BT_2390}, + {"bt.2446a", TONE_MAPPING_BT_2446A}, + {"st2094-40", TONE_MAPPING_ST2094_40}, + {"st2094-10", TONE_MAPPING_ST2094_10})}, + {"tone-mapping-param", OPT_FLOATDEF(tone_map.curve_param)}, + {"inverse-tone-mapping", OPT_BOOL(tone_map.inverse)}, + {"tone-mapping-max-boost", OPT_FLOAT(tone_map.max_boost), + M_RANGE(1.0, 10.0)}, + {"tone-mapping-visualize", OPT_BOOL(tone_map.visualize)}, + {"gamut-mapping-mode", OPT_CHOICE(tone_map.gamut_mode, + {"auto", GAMUT_AUTO}, + {"clip", GAMUT_CLIP}, + {"perceptual", GAMUT_PERCEPTUAL}, + {"relative", GAMUT_RELATIVE}, + {"saturation", GAMUT_SATURATION}, + {"absolute", GAMUT_ABSOLUTE}, + {"desaturate", GAMUT_DESATURATE}, + {"darken", GAMUT_DARKEN}, + {"warn", GAMUT_WARN}, + {"linear", GAMUT_LINEAR})}, + {"hdr-compute-peak", OPT_CHOICE(tone_map.compute_peak, + {"auto", 0}, + {"yes", 1}, + {"no", -1})}, + {"hdr-peak-percentile", OPT_FLOAT(tone_map.peak_percentile), + M_RANGE(0.0, 100.0)}, + {"hdr-peak-decay-rate", OPT_FLOAT(tone_map.decay_rate), + M_RANGE(0.0, 1000.0)}, + {"hdr-scene-threshold-low", OPT_FLOAT(tone_map.scene_threshold_low), + M_RANGE(0, 20.0)}, + {"hdr-scene-threshold-high", OPT_FLOAT(tone_map.scene_threshold_high), + M_RANGE(0, 20.0)}, + {"hdr-contrast-recovery", OPT_FLOAT(tone_map.contrast_recovery), + M_RANGE(0, 2.0)}, + {"hdr-contrast-smoothness", OPT_FLOAT(tone_map.contrast_smoothness), + M_RANGE(1.0, 100.0)}, + {"opengl-pbo", OPT_BOOL(pbo)}, + SCALER_OPTS("scale", SCALER_SCALE), + SCALER_OPTS("dscale", SCALER_DSCALE), + SCALER_OPTS("cscale", SCALER_CSCALE), + SCALER_OPTS("tscale", SCALER_TSCALE), + {"scaler-lut-size", OPT_REMOVED("hard-coded as 8")}, + {"scaler-resizes-only", OPT_BOOL(scaler_resizes_only)}, + {"correct-downscaling", OPT_BOOL(correct_downscaling)}, + {"linear-downscaling", OPT_BOOL(linear_downscaling)}, + {"linear-upscaling", OPT_BOOL(linear_upscaling)}, + {"sigmoid-upscaling", OPT_BOOL(sigmoid_upscaling)}, + {"sigmoid-center", OPT_FLOAT(sigmoid_center), M_RANGE(0.0, 1.0)}, + {"sigmoid-slope", OPT_FLOAT(sigmoid_slope), M_RANGE(1.0, 20.0)}, + {"fbo-format", OPT_STRING(fbo_format)}, + {"dither-depth", OPT_CHOICE(dither_depth, {"no", -1}, {"auto", 0}), + M_RANGE(-1, 16)}, + {"dither", OPT_CHOICE(dither_algo, + {"fruit", DITHER_FRUIT}, + {"ordered", DITHER_ORDERED}, + {"error-diffusion", DITHER_ERROR_DIFFUSION}, + {"no", DITHER_NONE})}, + {"dither-size-fruit", OPT_INT(dither_size), M_RANGE(2, 8)}, + {"temporal-dither", OPT_BOOL(temporal_dither)}, + {"temporal-dither-period", OPT_INT(temporal_dither_period), + M_RANGE(1, 128)}, + {"error-diffusion", + OPT_STRING_VALIDATE(error_diffusion, validate_error_diffusion_opt)}, + {"alpha", OPT_CHOICE(alpha_mode, + {"no", ALPHA_NO}, + {"yes", ALPHA_YES}, + {"blend", ALPHA_BLEND}, + {"blend-tiles", ALPHA_BLEND_TILES})}, + {"opengl-rectangle-textures", OPT_BOOL(use_rectangle)}, + {"background", OPT_COLOR(background)}, + {"interpolation", OPT_BOOL(interpolation)}, + {"interpolation-threshold", OPT_FLOAT(interpolation_threshold)}, + {"blend-subtitles", OPT_CHOICE(blend_subs, + {"no", BLEND_SUBS_NO}, + {"yes", BLEND_SUBS_YES}, + {"video", BLEND_SUBS_VIDEO})}, + {"glsl-shaders", OPT_PATHLIST(user_shaders), .flags = M_OPT_FILE}, + {"glsl-shader", OPT_CLI_ALIAS("glsl-shaders-append")}, + {"glsl-shader-opts", OPT_KEYVALUELIST(user_shader_opts)}, + {"deband", OPT_BOOL(deband)}, + {"deband", OPT_SUBSTRUCT(deband_opts, deband_conf)}, + {"sharpen", OPT_FLOAT(unsharp)}, + {"gpu-tex-pad-x", OPT_INT(tex_pad_x), M_RANGE(0, 4096)}, + {"gpu-tex-pad-y", OPT_INT(tex_pad_y), M_RANGE(0, 4096)}, + {"", OPT_SUBSTRUCT(icc_opts, mp_icc_conf)}, + {"gpu-shader-cache", OPT_BOOL(shader_cache)}, + {"gpu-shader-cache-dir", OPT_STRING(shader_cache_dir), .flags = M_OPT_FILE}, + {"gpu-hwdec-interop", + OPT_STRING_VALIDATE(hwdec_interop, ra_hwdec_validate_opt)}, + {"gamut-warning", OPT_REMOVED("Replaced by --gamut-mapping-mode=warn")}, + {"gamut-clipping", OPT_REMOVED("Replaced by --gamut-mapping-mode=desaturate")}, + {"tone-mapping-desaturate", OPT_REMOVED("Replaced by --tone-mapping-mode")}, + {"tone-mapping-desaturate-exponent", OPT_REMOVED("Replaced by --tone-mapping-mode")}, + {"tone-mapping-crosstalk", OPT_REMOVED("Hard-coded as 0.04")}, + {"tone-mapping-mode", OPT_REMOVED("no replacement")}, + {0} + }, + .size = sizeof(struct gl_video_opts), + .defaults = &gl_video_opts_def, +}; + +static void uninit_rendering(struct gl_video *p); +static void uninit_scaler(struct gl_video *p, struct scaler *scaler); +static void check_gl_features(struct gl_video *p); +static bool pass_upload_image(struct gl_video *p, struct mp_image *mpi, uint64_t id); +static const char *handle_scaler_opt(const char *name, bool tscale); +static void reinit_from_options(struct gl_video *p); +static void get_scale_factors(struct gl_video *p, bool transpose_rot, double xy[2]); +static void gl_video_setup_hooks(struct gl_video *p); +static void gl_video_update_options(struct gl_video *p); + +#define GLSL(x) gl_sc_add(p->sc, #x "\n"); +#define GLSLF(...) gl_sc_addf(p->sc, __VA_ARGS__) +#define GLSLHF(...) gl_sc_haddf(p->sc, __VA_ARGS__) +#define PRELUDE(...) gl_sc_paddf(p->sc, __VA_ARGS__) + +static struct bstr load_cached_file(struct gl_video *p, const char *path) +{ + if (!path || !path[0]) + return (struct bstr){0}; + for (int n = 0; n < p->num_files; n++) { + if (strcmp(p->files[n].path, path) == 0) + return p->files[n].body; + } + // not found -> load it + char *fname = mp_get_user_path(NULL, p->global, path); + struct bstr s = stream_read_file(fname, p, p->global, 1000000000); // 1GB + talloc_free(fname); + if (s.len) { + struct cached_file new = { + .path = talloc_strdup(p, path), + .body = s, + }; + MP_TARRAY_APPEND(p, p->files, p->num_files, new); + return new.body; + } + return (struct bstr){0}; +} + +static void debug_check_gl(struct gl_video *p, const char *msg) +{ + if (p->ra->fns->debug_marker) + p->ra->fns->debug_marker(p->ra, msg); +} + +static void gl_video_reset_surfaces(struct gl_video *p) +{ + for (int i = 0; i < SURFACES_MAX; i++) { + p->surfaces[i].id = 0; + p->surfaces[i].pts = MP_NOPTS_VALUE; + } + p->surface_idx = 0; + p->surface_now = 0; + p->frames_drawn = 0; + p->output_tex_valid = false; +} + +static void gl_video_reset_hooks(struct gl_video *p) +{ + for (int i = 0; i < p->num_tex_hooks; i++) + talloc_free(p->tex_hooks[i].priv); + + for (int i = 0; i < p->num_user_textures; i++) + ra_tex_free(p->ra, &p->user_textures[i].tex); + + p->num_tex_hooks = 0; + p->num_user_textures = 0; +} + +static inline int surface_wrap(int id) +{ + id = id % SURFACES_MAX; + return id < 0 ? id + SURFACES_MAX : id; +} + +static void reinit_osd(struct gl_video *p) +{ + mpgl_osd_destroy(p->osd); + p->osd = NULL; + if (p->osd_state) + p->osd = mpgl_osd_init(p->ra, p->log, p->osd_state); +} + +static void uninit_rendering(struct gl_video *p) +{ + for (int n = 0; n < SCALER_COUNT; n++) + uninit_scaler(p, &p->scaler[n]); + + ra_tex_free(p->ra, &p->dither_texture); + + for (int n = 0; n < 4; n++) { + ra_tex_free(p->ra, &p->merge_tex[n]); + ra_tex_free(p->ra, &p->scale_tex[n]); + ra_tex_free(p->ra, &p->integer_tex[n]); + } + + ra_tex_free(p->ra, &p->indirect_tex); + ra_tex_free(p->ra, &p->blend_subs_tex); + ra_tex_free(p->ra, &p->screen_tex); + ra_tex_free(p->ra, &p->output_tex); + + for (int n = 0; n < 2; n++) + ra_tex_free(p->ra, &p->error_diffusion_tex[n]); + + for (int n = 0; n < SURFACES_MAX; n++) + ra_tex_free(p->ra, &p->surfaces[n].tex); + + for (int n = 0; n < p->num_hook_textures; n++) + ra_tex_free(p->ra, &p->hook_textures[n]); + + gl_video_reset_surfaces(p); + gl_video_reset_hooks(p); + + gl_sc_reset_error(p->sc); +} + +bool gl_video_gamma_auto_enabled(struct gl_video *p) +{ + return p->opts.gamma_auto; +} + +struct mp_colorspace gl_video_get_output_colorspace(struct gl_video *p) +{ + return (struct mp_colorspace) { + .primaries = p->opts.target_prim, + .gamma = p->opts.target_trc, + .hdr.max_luma = p->opts.target_peak, + }; +} + +// Warning: profile.start must point to a ta allocation, and the function +// takes over ownership. +void gl_video_set_icc_profile(struct gl_video *p, bstr icc_data) +{ + if (gl_lcms_set_memory_profile(p->cms, icc_data)) + reinit_from_options(p); +} + +bool gl_video_icc_auto_enabled(struct gl_video *p) +{ + return p->opts.icc_opts ? p->opts.icc_opts->profile_auto : false; +} + +static bool gl_video_get_lut3d(struct gl_video *p, enum mp_csp_prim prim, + enum mp_csp_trc trc) +{ + if (!p->use_lut_3d) + return false; + + struct AVBufferRef *icc = NULL; + if (p->image.mpi) + icc = p->image.mpi->icc_profile; + + if (p->lut_3d_texture && !gl_lcms_has_changed(p->cms, prim, trc, icc)) + return true; + + // GLES3 doesn't provide filtered 16 bit integer textures + // GLES2 doesn't even provide 3D textures + const struct ra_format *fmt = ra_find_unorm_format(p->ra, 2, 4); + if (!fmt || !(p->ra->caps & RA_CAP_TEX_3D)) { + p->use_lut_3d = false; + MP_WARN(p, "Disabling color management (no RGBA16 3D textures).\n"); + return false; + } + + struct lut3d *lut3d = NULL; + if (!fmt || !gl_lcms_get_lut3d(p->cms, &lut3d, prim, trc, icc) || !lut3d) { + p->use_lut_3d = false; + return false; + } + + ra_tex_free(p->ra, &p->lut_3d_texture); + + struct ra_tex_params params = { + .dimensions = 3, + .w = lut3d->size[0], + .h = lut3d->size[1], + .d = lut3d->size[2], + .format = fmt, + .render_src = true, + .src_linear = true, + .initial_data = lut3d->data, + }; + p->lut_3d_texture = ra_tex_create(p->ra, ¶ms); + + debug_check_gl(p, "after 3d lut creation"); + + for (int i = 0; i < 3; i++) + p->lut_3d_size[i] = lut3d->size[i]; + + talloc_free(lut3d); + + if (!p->lut_3d_texture) { + p->use_lut_3d = false; + return false; + } + + return true; +} + +// Fill an image struct from a ra_tex + some metadata +static struct image image_wrap(struct ra_tex *tex, enum plane_type type, + int components) +{ + assert(type != PLANE_NONE); + return (struct image){ + .type = type, + .tex = tex, + .multiplier = 1.0, + .w = tex ? tex->params.w : 1, + .h = tex ? tex->params.h : 1, + .transform = identity_trans, + .components = components, + }; +} + +// Bind an image to a free texture unit and return its ID. +static int pass_bind(struct gl_video *p, struct image img) +{ + int idx = p->num_pass_imgs; + MP_TARRAY_APPEND(p, p->pass_imgs, p->num_pass_imgs, img); + return idx; +} + +// Rotation by 90° and flipping. +// w/h is used for recentering. +static void get_transform(float w, float h, int rotate, bool flip, + struct gl_transform *out_tr) +{ + int a = rotate % 90 ? 0 : rotate / 90; + int sin90[4] = {0, 1, 0, -1}; // just to avoid rounding issues etc. + int cos90[4] = {1, 0, -1, 0}; + struct gl_transform tr = {{{ cos90[a], sin90[a]}, + {-sin90[a], cos90[a]}}}; + + // basically, recenter to keep the whole image in view + float b[2] = {1, 1}; + gl_transform_vec(tr, &b[0], &b[1]); + tr.t[0] += b[0] < 0 ? w : 0; + tr.t[1] += b[1] < 0 ? h : 0; + + if (flip) { + struct gl_transform fliptr = {{{1, 0}, {0, -1}}, {0, h}}; + gl_transform_trans(fliptr, &tr); + } + + *out_tr = tr; +} + +// Return the chroma plane upscaled to luma size, but with additional padding +// for image sizes not aligned to subsampling. +static int chroma_upsize(int size, int pixel) +{ + return (size + pixel - 1) / pixel * pixel; +} + +// If a and b are on the same plane, return what plane type should be used. +// If a or b are none, the other type always wins. +// Usually: LUMA/RGB/XYZ > CHROMA > ALPHA +static enum plane_type merge_plane_types(enum plane_type a, enum plane_type b) +{ + if (a == PLANE_NONE) + return b; + if (b == PLANE_LUMA || b == PLANE_RGB || b == PLANE_XYZ) + return b; + if (b != PLANE_NONE && a == PLANE_ALPHA) + return b; + return a; +} + +// Places a video_image's image textures + associated metadata into img[]. The +// number of textures is equal to p->plane_count. Any necessary plane offsets +// are stored in off. (e.g. chroma position) +static void pass_get_images(struct gl_video *p, struct video_image *vimg, + struct image img[4], struct gl_transform off[4]) +{ + assert(vimg->mpi); + + int w = p->image_params.w; + int h = p->image_params.h; + + // Determine the chroma offset + float ls_w = 1.0 / p->ra_format.chroma_w; + float ls_h = 1.0 / p->ra_format.chroma_h; + + struct gl_transform chroma = {{{ls_w, 0.0}, {0.0, ls_h}}}; + + if (p->image_params.chroma_location != MP_CHROMA_CENTER) { + int cx, cy; + mp_get_chroma_location(p->image_params.chroma_location, &cx, &cy); + // By default texture coordinates are such that chroma is centered with + // any chroma subsampling. If a specific direction is given, make it + // so that the luma and chroma sample line up exactly. + // For 4:4:4, setting chroma location should have no effect at all. + // luma sample size (in chroma coord. space) + chroma.t[0] = ls_w < 1 ? ls_w * -cx / 2 : 0; + chroma.t[1] = ls_h < 1 ? ls_h * -cy / 2 : 0; + } + + memset(img, 0, 4 * sizeof(img[0])); + for (int n = 0; n < p->plane_count; n++) { + struct texplane *t = &vimg->planes[n]; + + enum plane_type type = PLANE_NONE; + int padding = 0; + for (int i = 0; i < 4; i++) { + int c = p->ra_format.components[n][i]; + enum plane_type ctype; + if (c == 0) { + ctype = PLANE_NONE; + } else if (c == 4) { + ctype = PLANE_ALPHA; + } else if (p->image_params.color.space == MP_CSP_RGB) { + ctype = PLANE_RGB; + } else if (p->image_params.color.space == MP_CSP_XYZ) { + ctype = PLANE_XYZ; + } else { + ctype = c == 1 ? PLANE_LUMA : PLANE_CHROMA; + } + type = merge_plane_types(type, ctype); + if (!c && padding == i) + padding = i + 1; + } + + int msb_valid_bits = + p->ra_format.component_bits + MPMIN(p->ra_format.component_pad, 0); + int csp = type == PLANE_ALPHA ? MP_CSP_RGB : p->image_params.color.space; + float tex_mul = + 1.0 / mp_get_csp_mul(csp, msb_valid_bits, p->ra_format.component_bits); + if (p->ra_format.component_type == RA_CTYPE_FLOAT) + tex_mul = 1.0; + + img[n] = (struct image){ + .type = type, + .tex = t->tex, + .multiplier = tex_mul, + .w = t->w, + .h = t->h, + .padding = padding, + }; + + for (int i = 0; i < 4; i++) + img[n].components += !!p->ra_format.components[n][i]; + + get_transform(t->w, t->h, p->image_params.rotate, t->flipped, + &img[n].transform); + if (p->image_params.rotate % 180 == 90) + MPSWAP(int, img[n].w, img[n].h); + + off[n] = identity_trans; + + if (type == PLANE_CHROMA) { + struct gl_transform rot; + get_transform(0, 0, p->image_params.rotate, true, &rot); + + struct gl_transform tr = chroma; + gl_transform_vec(rot, &tr.t[0], &tr.t[1]); + + float dx = (chroma_upsize(w, p->ra_format.chroma_w) - w) * ls_w; + float dy = (chroma_upsize(h, p->ra_format.chroma_h) - h) * ls_h; + + // Adjust the chroma offset if the real chroma size is fractional + // due image sizes not aligned to chroma subsampling. + struct gl_transform rot2; + get_transform(0, 0, p->image_params.rotate, t->flipped, &rot2); + if (rot2.m[0][0] < 0) + tr.t[0] += dx; + if (rot2.m[1][0] < 0) + tr.t[0] += dy; + if (rot2.m[0][1] < 0) + tr.t[1] += dx; + if (rot2.m[1][1] < 0) + tr.t[1] += dy; + + off[n] = tr; + } + } +} + +// Return the index of the given component (assuming all non-padding components +// of all planes are concatenated into a linear list). +static int find_comp(struct ra_imgfmt_desc *desc, int component) +{ + int cur = 0; + for (int n = 0; n < desc->num_planes; n++) { + for (int i = 0; i < 4; i++) { + if (desc->components[n][i]) { + if (desc->components[n][i] == component) + return cur; + cur++; + } + } + } + return -1; +} + +static void init_video(struct gl_video *p) +{ + p->use_integer_conversion = false; + + struct ra_hwdec *hwdec = ra_hwdec_get(&p->hwdec_ctx, p->image_params.imgfmt); + if (hwdec) { + if (hwdec->driver->overlay_frame) { + MP_WARN(p, "Using HW-overlay mode. No GL filtering is performed " + "on the video!\n"); + p->hwdec_overlay = hwdec; + } else { + p->hwdec_mapper = ra_hwdec_mapper_create(hwdec, &p->image_params); + if (!p->hwdec_mapper) + MP_ERR(p, "Initializing texture for hardware decoding failed.\n"); + } + if (p->hwdec_mapper) + p->image_params = p->hwdec_mapper->dst_params; + const char **exts = hwdec->glsl_extensions; + for (int n = 0; exts && exts[n]; n++) + gl_sc_enable_extension(p->sc, (char *)exts[n]); + p->hwdec_active = true; + } + + p->ra_format = (struct ra_imgfmt_desc){0}; + ra_get_imgfmt_desc(p->ra, p->image_params.imgfmt, &p->ra_format); + + p->plane_count = p->ra_format.num_planes; + + p->has_alpha = false; + p->is_gray = true; + + for (int n = 0; n < p->ra_format.num_planes; n++) { + for (int i = 0; i < 4; i++) { + if (p->ra_format.components[n][i]) { + p->has_alpha |= p->ra_format.components[n][i] == 4; + p->is_gray &= p->ra_format.components[n][i] == 1 || + p->ra_format.components[n][i] == 4; + } + } + } + + for (int c = 0; c < 4; c++) { + int loc = find_comp(&p->ra_format, c + 1); + p->color_swizzle[c] = "rgba"[loc >= 0 && loc < 4 ? loc : 0]; + } + p->color_swizzle[4] = '\0'; + + mp_image_params_guess_csp(&p->image_params); + + av_lfg_init(&p->lfg, 1); + + debug_check_gl(p, "before video texture creation"); + + if (!p->hwdec_active) { + struct video_image *vimg = &p->image; + + struct mp_image layout = {0}; + mp_image_set_params(&layout, &p->image_params); + + for (int n = 0; n < p->plane_count; n++) { + struct texplane *plane = &vimg->planes[n]; + const struct ra_format *format = p->ra_format.planes[n]; + + plane->w = mp_image_plane_w(&layout, n); + plane->h = mp_image_plane_h(&layout, n); + + struct ra_tex_params params = { + .dimensions = 2, + .w = plane->w + p->opts.tex_pad_x, + .h = plane->h + p->opts.tex_pad_y, + .d = 1, + .format = format, + .render_src = true, + .src_linear = format->linear_filter, + .non_normalized = p->opts.use_rectangle, + .host_mutable = true, + }; + + MP_VERBOSE(p, "Texture for plane %d: %dx%d\n", n, + params.w, params.h); + + plane->tex = ra_tex_create(p->ra, ¶ms); + p->use_integer_conversion |= format->ctype == RA_CTYPE_UINT; + } + } + + debug_check_gl(p, "after video texture creation"); + + // Format-dependent checks. + check_gl_features(p); + + gl_video_setup_hooks(p); +} + +static struct dr_buffer *gl_find_dr_buffer(struct gl_video *p, uint8_t *ptr) +{ + for (int i = 0; i < p->num_dr_buffers; i++) { + struct dr_buffer *buffer = &p->dr_buffers[i]; + uint8_t *bufptr = buffer->buf->data; + size_t size = buffer->buf->params.size; + if (ptr >= bufptr && ptr < bufptr + size) + return buffer; + } + + return NULL; +} + +static void gc_pending_dr_fences(struct gl_video *p, bool force) +{ +again:; + for (int n = 0; n < p->num_dr_buffers; n++) { + struct dr_buffer *buffer = &p->dr_buffers[n]; + if (!buffer->mpi) + continue; + + bool res = p->ra->fns->buf_poll(p->ra, buffer->buf); + if (res || force) { + // Unreferencing the image could cause gl_video_dr_free_buffer() + // to be called by the talloc destructor (if it was the last + // reference). This will implicitly invalidate the buffer pointer + // and change the p->dr_buffers array. To make it worse, it could + // free multiple dr_buffers due to weird theoretical corner cases. + // This is also why we use the goto to iterate again from the + // start, because everything gets fucked up. Hail satan! + struct mp_image *ref = buffer->mpi; + buffer->mpi = NULL; + talloc_free(ref); + goto again; + } + } +} + +static void unref_current_image(struct gl_video *p) +{ + struct video_image *vimg = &p->image; + + if (vimg->hwdec_mapped) { + assert(p->hwdec_active && p->hwdec_mapper); + ra_hwdec_mapper_unmap(p->hwdec_mapper); + memset(vimg->planes, 0, sizeof(vimg->planes)); + vimg->hwdec_mapped = false; + } + + vimg->id = 0; + + mp_image_unrefp(&vimg->mpi); + + // While we're at it, also garbage collect pending fences in here to + // get it out of the way. + gc_pending_dr_fences(p, false); +} + +// If overlay mode is used, make sure to remove the overlay. +// Be careful with this. Removing the overlay and adding another one will +// lead to flickering artifacts. +static void unmap_overlay(struct gl_video *p) +{ + if (p->hwdec_overlay) + p->hwdec_overlay->driver->overlay_frame(p->hwdec_overlay, NULL, NULL, NULL, true); +} + +static void uninit_video(struct gl_video *p) +{ + uninit_rendering(p); + + struct video_image *vimg = &p->image; + + unmap_overlay(p); + unref_current_image(p); + + for (int n = 0; n < p->plane_count; n++) { + struct texplane *plane = &vimg->planes[n]; + ra_tex_free(p->ra, &plane->tex); + } + *vimg = (struct video_image){0}; + + // Invalidate image_params to ensure that gl_video_config() will call + // init_video() on uninitialized gl_video. + p->real_image_params = (struct mp_image_params){0}; + p->image_params = p->real_image_params; + p->hwdec_active = false; + p->hwdec_overlay = NULL; + ra_hwdec_mapper_free(&p->hwdec_mapper); +} + +static void pass_record(struct gl_video *p, struct mp_pass_perf perf) +{ + if (!p->pass || p->pass_idx == VO_PASS_PERF_MAX) + return; + + struct pass_info *pass = &p->pass[p->pass_idx]; + pass->perf = perf; + + if (pass->desc.len == 0) + bstr_xappend(p, &pass->desc, bstr0("(unknown)")); + + p->pass_idx++; +} + +PRINTF_ATTRIBUTE(2, 3) +static void pass_describe(struct gl_video *p, const char *textf, ...) +{ + if (!p->pass || p->pass_idx == VO_PASS_PERF_MAX) + return; + + struct pass_info *pass = &p->pass[p->pass_idx]; + + if (pass->desc.len > 0) + bstr_xappend(p, &pass->desc, bstr0(" + ")); + + va_list ap; + va_start(ap, textf); + bstr_xappend_vasprintf(p, &pass->desc, textf, ap); + va_end(ap); +} + +static void pass_info_reset(struct gl_video *p, bool is_redraw) +{ + p->pass = is_redraw ? p->pass_redraw : p->pass_fresh; + p->pass_idx = 0; + + for (int i = 0; i < VO_PASS_PERF_MAX; i++) { + p->pass[i].desc.len = 0; + p->pass[i].perf = (struct mp_pass_perf){0}; + } +} + +static void pass_report_performance(struct gl_video *p) +{ + if (!p->pass) + return; + + for (int i = 0; i < VO_PASS_PERF_MAX; i++) { + struct pass_info *pass = &p->pass[i]; + if (pass->desc.len) { + MP_TRACE(p, "pass '%.*s': last %dus avg %dus peak %dus\n", + BSTR_P(pass->desc), + (int)pass->perf.last/1000, + (int)pass->perf.avg/1000, + (int)pass->perf.peak/1000); + } + } +} + +static void pass_prepare_src_tex(struct gl_video *p) +{ + struct gl_shader_cache *sc = p->sc; + + for (int n = 0; n < p->num_pass_imgs; n++) { + struct image *s = &p->pass_imgs[n]; + if (!s->tex) + continue; + + char *texture_name = mp_tprintf(32, "texture%d", n); + char *texture_size = mp_tprintf(32, "texture_size%d", n); + char *texture_rot = mp_tprintf(32, "texture_rot%d", n); + char *texture_off = mp_tprintf(32, "texture_off%d", n); + char *pixel_size = mp_tprintf(32, "pixel_size%d", n); + + gl_sc_uniform_texture(sc, texture_name, s->tex); + float f[2] = {1, 1}; + if (!s->tex->params.non_normalized) { + f[0] = s->tex->params.w; + f[1] = s->tex->params.h; + } + gl_sc_uniform_vec2(sc, texture_size, f); + gl_sc_uniform_mat2(sc, texture_rot, true, (float *)s->transform.m); + gl_sc_uniform_vec2(sc, texture_off, (float *)s->transform.t); + gl_sc_uniform_vec2(sc, pixel_size, (float[]){1.0f / f[0], + 1.0f / f[1]}); + } +} + +static void cleanup_binds(struct gl_video *p) +{ + p->num_pass_imgs = 0; +} + +// Sets the appropriate compute shader metadata for an implicit compute pass +// bw/bh: block size +static void pass_is_compute(struct gl_video *p, int bw, int bh, bool flexible) +{ + if (p->pass_compute.active && flexible) { + // Avoid overwriting existing block sizes when using a flexible pass + bw = p->pass_compute.block_w; + bh = p->pass_compute.block_h; + } + + p->pass_compute = (struct compute_info){ + .active = true, + .block_w = bw, + .block_h = bh, + }; +} + +// w/h: the width/height of the compute shader's operating domain (e.g. the +// target target that needs to be written, or the source texture that needs to +// be reduced) +static void dispatch_compute(struct gl_video *p, int w, int h, + struct compute_info info) +{ + PRELUDE("layout (local_size_x = %d, local_size_y = %d) in;\n", + info.threads_w > 0 ? info.threads_w : info.block_w, + info.threads_h > 0 ? info.threads_h : info.block_h); + + pass_prepare_src_tex(p); + + // Since we don't actually have vertices, we pretend for convenience + // reasons that we do and calculate the right texture coordinates based on + // the output sample ID + gl_sc_uniform_vec2(p->sc, "out_scale", (float[2]){ 1.0 / w, 1.0 / h }); + PRELUDE("#define outcoord(id) (out_scale * (vec2(id) + vec2(0.5)))\n"); + + for (int n = 0; n < p->num_pass_imgs; n++) { + struct image *s = &p->pass_imgs[n]; + if (!s->tex) + continue; + + PRELUDE("#define texmap%d(id) (texture_rot%d * outcoord(id) + " + "pixel_size%d * texture_off%d)\n", n, n, n, n); + PRELUDE("#define texcoord%d texmap%d(gl_GlobalInvocationID)\n", n, n); + } + + // always round up when dividing to make sure we don't leave off a part of + // the image + int num_x = info.block_w > 0 ? (w + info.block_w - 1) / info.block_w : 1, + num_y = info.block_h > 0 ? (h + info.block_h - 1) / info.block_h : 1; + + if (!(p->ra->caps & RA_CAP_NUM_GROUPS)) + PRELUDE("#define gl_NumWorkGroups uvec3(%d, %d, 1)\n", num_x, num_y); + + pass_record(p, gl_sc_dispatch_compute(p->sc, num_x, num_y, 1)); + cleanup_binds(p); +} + +static struct mp_pass_perf render_pass_quad(struct gl_video *p, + struct ra_fbo fbo, bool discard, + const struct mp_rect *dst) +{ + // The first element is reserved for `vec2 position` + int num_vertex_attribs = 1 + p->num_pass_imgs; + size_t vertex_stride = num_vertex_attribs * sizeof(struct vertex_pt); + + // Expand the VAO if necessary + while (p->vao_len < num_vertex_attribs) { + MP_TARRAY_APPEND(p, p->vao, p->vao_len, (struct ra_renderpass_input) { + .name = talloc_asprintf(p, "texcoord%d", p->vao_len - 1), + .type = RA_VARTYPE_FLOAT, + .dim_v = 2, + .dim_m = 1, + .offset = p->vao_len * sizeof(struct vertex_pt), + }); + } + + int num_vertices = 6; // quad as triangle list + int num_attribs_total = num_vertices * num_vertex_attribs; + MP_TARRAY_GROW(p, p->tmp_vertex, num_attribs_total); + + struct gl_transform t; + gl_transform_ortho_fbo(&t, fbo); + + float x[2] = {dst->x0, dst->x1}; + float y[2] = {dst->y0, dst->y1}; + gl_transform_vec(t, &x[0], &y[0]); + gl_transform_vec(t, &x[1], &y[1]); + + for (int n = 0; n < 4; n++) { + struct vertex_pt *vs = &p->tmp_vertex[num_vertex_attribs * n]; + // vec2 position in idx 0 + vs[0].x = x[n / 2]; + vs[0].y = y[n % 2]; + for (int i = 0; i < p->num_pass_imgs; i++) { + struct image *s = &p->pass_imgs[i]; + if (!s->tex) + continue; + struct gl_transform tr = s->transform; + float tx = (n / 2) * s->w; + float ty = (n % 2) * s->h; + gl_transform_vec(tr, &tx, &ty); + bool rect = s->tex->params.non_normalized; + // vec2 texcoordN in idx N+1 + vs[i + 1].x = tx / (rect ? 1 : s->tex->params.w); + vs[i + 1].y = ty / (rect ? 1 : s->tex->params.h); + } + } + + memmove(&p->tmp_vertex[num_vertex_attribs * 4], + &p->tmp_vertex[num_vertex_attribs * 2], + vertex_stride); + + memmove(&p->tmp_vertex[num_vertex_attribs * 5], + &p->tmp_vertex[num_vertex_attribs * 1], + vertex_stride); + + return gl_sc_dispatch_draw(p->sc, fbo.tex, discard, p->vao, num_vertex_attribs, + vertex_stride, p->tmp_vertex, num_vertices); +} + +static void finish_pass_fbo(struct gl_video *p, struct ra_fbo fbo, + bool discard, const struct mp_rect *dst) +{ + pass_prepare_src_tex(p); + pass_record(p, render_pass_quad(p, fbo, discard, dst)); + debug_check_gl(p, "after rendering"); + cleanup_binds(p); +} + +// dst_fbo: this will be used for rendering; possibly reallocating the whole +// FBO, if the required parameters have changed +// w, h: required FBO target dimension, and also defines the target rectangle +// used for rasterization +static void finish_pass_tex(struct gl_video *p, struct ra_tex **dst_tex, + int w, int h) +{ + if (!ra_tex_resize(p->ra, p->log, dst_tex, w, h, p->fbo_format)) { + cleanup_binds(p); + gl_sc_reset(p->sc); + return; + } + + // If RA_CAP_PARALLEL_COMPUTE is set, try to prefer compute shaders + // over fragment shaders wherever possible. + if (!p->pass_compute.active && (p->ra->caps & RA_CAP_PARALLEL_COMPUTE) && + (*dst_tex)->params.storage_dst) + { + pass_is_compute(p, 16, 16, true); + } + + if (p->pass_compute.active) { + gl_sc_uniform_image2D_wo(p->sc, "out_image", *dst_tex); + if (!p->pass_compute.directly_writes) + GLSL(imageStore(out_image, ivec2(gl_GlobalInvocationID), color);) + + dispatch_compute(p, w, h, p->pass_compute); + p->pass_compute = (struct compute_info){0}; + + debug_check_gl(p, "after dispatching compute shader"); + } else { + struct ra_fbo fbo = { .tex = *dst_tex, }; + finish_pass_fbo(p, fbo, true, &(struct mp_rect){0, 0, w, h}); + } +} + +static const char *get_tex_swizzle(struct image *img) +{ + if (!img->tex) + return "rgba"; + return img->tex->params.format->luminance_alpha ? "raaa" : "rgba"; +} + +// Copy a texture to the vec4 color, while increasing offset. Also applies +// the texture multiplier to the sampled color +static void copy_image(struct gl_video *p, unsigned int *offset, struct image img) +{ + const unsigned int count = img.components; + char src[5] = {0}; + char dst[5] = {0}; + + assert(*offset + count < sizeof(dst)); + assert(img.padding + count < sizeof(src)); + + int id = pass_bind(p, img); + + const char *tex_fmt = get_tex_swizzle(&img); + const char *dst_fmt = "rgba"; + for (unsigned int i = 0; i < count; i++) { + src[i] = tex_fmt[img.padding + i]; + dst[i] = dst_fmt[*offset + i]; + } + + if (img.tex && img.tex->params.format->ctype == RA_CTYPE_UINT) { + uint64_t tex_max = 1ull << p->ra_format.component_bits; + img.multiplier *= 1.0 / (tex_max - 1); + } + + GLSLF("color.%s = %f * vec4(texture(texture%d, texcoord%d)).%s;\n", + dst, img.multiplier, id, id, src); + + *offset += count; +} + +static void skip_unused(struct gl_video *p, int num_components) +{ + for (int i = num_components; i < 4; i++) + GLSLF("color.%c = %f;\n", "rgba"[i], i < 3 ? 0.0 : 1.0); +} + +static void uninit_scaler(struct gl_video *p, struct scaler *scaler) +{ + ra_tex_free(p->ra, &scaler->sep_fbo); + ra_tex_free(p->ra, &scaler->lut); + scaler->kernel = NULL; + scaler->initialized = false; +} + +static void hook_prelude(struct gl_video *p, const char *name, int id, + struct image img) +{ + GLSLHF("#define %s_raw texture%d\n", name, id); + GLSLHF("#define %s_pos texcoord%d\n", name, id); + GLSLHF("#define %s_size texture_size%d\n", name, id); + GLSLHF("#define %s_rot texture_rot%d\n", name, id); + GLSLHF("#define %s_off texture_off%d\n", name, id); + GLSLHF("#define %s_pt pixel_size%d\n", name, id); + GLSLHF("#define %s_map texmap%d\n", name, id); + GLSLHF("#define %s_mul %f\n", name, img.multiplier); + + char crap[5] = ""; + snprintf(crap, sizeof(crap), "%s", get_tex_swizzle(&img)); + + // Remove leading padding by rotating the swizzle mask. + int len = strlen(crap); + for (int n = 0; n < img.padding; n++) { + if (len) { + char f = crap[0]; + memmove(crap, crap + 1, len - 1); + crap[len - 1] = f; + } + } + + // Set up the sampling functions + GLSLHF("#define %s_tex(pos) (%s_mul * vec4(texture(%s_raw, pos)).%s)\n", + name, name, name, crap); + + if (p->ra->caps & RA_CAP_GATHER) { + GLSLHF("#define %s_gather(pos, c) (%s_mul * vec4(" + "textureGather(%s_raw, pos, c)))\n", name, name, name); + } + + // Since the extra matrix multiplication impacts performance, + // skip it unless the texture was actually rotated + if (gl_transform_eq(img.transform, identity_trans)) { + GLSLHF("#define %s_texOff(off) %s_tex(%s_pos + %s_pt * vec2(off))\n", + name, name, name, name); + } else { + GLSLHF("#define %s_texOff(off) " + "%s_tex(%s_pos + %s_rot * vec2(off)/%s_size)\n", + name, name, name, name, name); + } +} + +static bool saved_img_find(struct gl_video *p, const char *name, + struct image *out) +{ + if (!name || !out) + return false; + + for (int i = 0; i < p->num_saved_imgs; i++) { + if (strcmp(p->saved_imgs[i].name, name) == 0) { + *out = p->saved_imgs[i].img; + return true; + } + } + + return false; +} + +static void saved_img_store(struct gl_video *p, const char *name, + struct image img) +{ + assert(name); + + for (int i = 0; i < p->num_saved_imgs; i++) { + if (strcmp(p->saved_imgs[i].name, name) == 0) { + p->saved_imgs[i].img = img; + return; + } + } + + MP_TARRAY_APPEND(p, p->saved_imgs, p->num_saved_imgs, (struct saved_img) { + .name = name, + .img = img + }); +} + +static bool pass_hook_setup_binds(struct gl_video *p, const char *name, + struct image img, struct tex_hook *hook) +{ + for (int t = 0; t < SHADER_MAX_BINDS; t++) { + char *bind_name = (char *)hook->bind_tex[t]; + + if (!bind_name) + continue; + + // This is a special name that means "currently hooked texture" + if (strcmp(bind_name, "HOOKED") == 0) { + int id = pass_bind(p, img); + hook_prelude(p, "HOOKED", id, img); + hook_prelude(p, name, id, img); + continue; + } + + // BIND can also be used to load user-defined textures, in which + // case we will directly load them as a uniform instead of + // generating the hook_prelude boilerplate + for (int u = 0; u < p->num_user_textures; u++) { + struct gl_user_shader_tex *utex = &p->user_textures[u]; + if (bstr_equals0(utex->name, bind_name)) { + gl_sc_uniform_texture(p->sc, bind_name, utex->tex); + goto next_bind; + } + } + + struct image bind_img; + if (!saved_img_find(p, bind_name, &bind_img)) { + // Clean up texture bindings and move on to the next hook + MP_TRACE(p, "Skipping hook on %s due to no texture named %s.\n", + name, bind_name); + p->num_pass_imgs -= t; + return false; + } + + hook_prelude(p, bind_name, pass_bind(p, bind_img), bind_img); + +next_bind: ; + } + + return true; +} + +static struct ra_tex **next_hook_tex(struct gl_video *p) +{ + if (p->idx_hook_textures == p->num_hook_textures) + MP_TARRAY_APPEND(p, p->hook_textures, p->num_hook_textures, NULL); + + return &p->hook_textures[p->idx_hook_textures++]; +} + +// Process hooks for a plane, saving the result and returning a new image +// If 'trans' is NULL, the shader is forbidden from transforming img +static struct image pass_hook(struct gl_video *p, const char *name, + struct image img, struct gl_transform *trans) +{ + if (!name) + return img; + + saved_img_store(p, name, img); + + MP_TRACE(p, "Running hooks for %s\n", name); + for (int i = 0; i < p->num_tex_hooks; i++) { + struct tex_hook *hook = &p->tex_hooks[i]; + + // Figure out if this pass hooks this texture + for (int h = 0; h < SHADER_MAX_HOOKS; h++) { + if (hook->hook_tex[h] && strcmp(hook->hook_tex[h], name) == 0) + goto found; + } + + continue; + +found: + // Check the hook's condition + if (hook->cond && !hook->cond(p, img, hook->priv)) { + MP_TRACE(p, "Skipping hook on %s due to condition.\n", name); + continue; + } + + const char *store_name = hook->save_tex ? hook->save_tex : name; + bool is_overwrite = strcmp(store_name, name) == 0; + + // If user shader is set to align HOOKED with reference and fix its + // offset, it requires HOOKED to be resizable and overwrited. + if (is_overwrite && hook->align_offset) { + if (!trans) { + MP_ERR(p, "Hook tried to align unresizable texture %s!\n", + name); + return img; + } + + struct gl_transform align_off = identity_trans; + align_off.t[0] = trans->t[0]; + align_off.t[1] = trans->t[1]; + + gl_transform_trans(align_off, &img.transform); + } + + if (!pass_hook_setup_binds(p, name, img, hook)) + continue; + + // Run the actual hook. This generates a series of GLSL shader + // instructions sufficient for drawing the hook's output + struct gl_transform hook_off = identity_trans; + hook->hook(p, img, &hook_off, hook->priv); + + int comps = hook->components ? hook->components : img.components; + skip_unused(p, comps); + + // Compute the updated FBO dimensions and store the result + struct mp_rect_f sz = {0, 0, img.w, img.h}; + gl_transform_rect(hook_off, &sz); + int w = lroundf(fabs(sz.x1 - sz.x0)); + int h = lroundf(fabs(sz.y1 - sz.y0)); + + struct ra_tex **tex = next_hook_tex(p); + finish_pass_tex(p, tex, w, h); + struct image saved_img = image_wrap(*tex, img.type, comps); + + // If the texture we're saving overwrites the "current" texture, also + // update the tex parameter so that the future loop cycles will use the + // updated values, and export the offset + if (is_overwrite) { + if (!trans && !gl_transform_eq(hook_off, identity_trans)) { + MP_ERR(p, "Hook tried changing size of unscalable texture %s!\n", + name); + return img; + } + + img = saved_img; + if (trans) { + gl_transform_trans(hook_off, trans); + + // If user shader is set to align HOOKED, the offset it produces + // is dynamic (with static resizing factor though). + // Align it with reference manually to get offset fixed. + if (hook->align_offset) { + trans->t[0] = 0.0; + trans->t[1] = 0.0; + } + } + } + + saved_img_store(p, store_name, saved_img); + } + + return img; +} + +// This can be used at any time in the middle of rendering to specify an +// optional hook point, which if triggered will render out to a new FBO and +// load the result back into vec4 color. Offsets applied by the hooks are +// accumulated in tex_trans, and the FBO is dimensioned according +// to p->texture_w/h +static void pass_opt_hook_point(struct gl_video *p, const char *name, + struct gl_transform *tex_trans) +{ + if (!name) + return; + + for (int i = 0; i < p->num_tex_hooks; i++) { + struct tex_hook *hook = &p->tex_hooks[i]; + + for (int h = 0; h < SHADER_MAX_HOOKS; h++) { + if (hook->hook_tex[h] && strcmp(hook->hook_tex[h], name) == 0) + goto found; + } + + for (int b = 0; b < SHADER_MAX_BINDS; b++) { + if (hook->bind_tex[b] && strcmp(hook->bind_tex[b], name) == 0) + goto found; + } + } + + // Nothing uses this texture, don't bother storing it + return; + +found: ; + struct ra_tex **tex = next_hook_tex(p); + finish_pass_tex(p, tex, p->texture_w, p->texture_h); + struct image img = image_wrap(*tex, PLANE_RGB, p->components); + img = pass_hook(p, name, img, tex_trans); + copy_image(p, &(int){0}, img); + p->texture_w = img.w; + p->texture_h = img.h; + p->components = img.components; + pass_describe(p, "(remainder pass)"); +} + +static void load_shader(struct gl_video *p, struct bstr body) +{ + gl_sc_hadd_bstr(p->sc, body); + gl_sc_uniform_dynamic(p->sc); + gl_sc_uniform_f(p->sc, "random", (double)av_lfg_get(&p->lfg) / UINT32_MAX); + gl_sc_uniform_dynamic(p->sc); + gl_sc_uniform_i(p->sc, "frame", p->frames_uploaded); + gl_sc_uniform_vec2(p->sc, "input_size", + (float[]){(p->src_rect.x1 - p->src_rect.x0) * + p->texture_offset.m[0][0], + (p->src_rect.y1 - p->src_rect.y0) * + p->texture_offset.m[1][1]}); + gl_sc_uniform_vec2(p->sc, "target_size", + (float[]){p->dst_rect.x1 - p->dst_rect.x0, + p->dst_rect.y1 - p->dst_rect.y0}); + gl_sc_uniform_vec2(p->sc, "tex_offset", + (float[]){p->src_rect.x0 * p->texture_offset.m[0][0] + + p->texture_offset.t[0], + p->src_rect.y0 * p->texture_offset.m[1][1] + + p->texture_offset.t[1]}); +} + +// Semantic equality +static bool double_seq(double a, double b) +{ + return (isnan(a) && isnan(b)) || a == b; +} + +static bool scaler_fun_eq(struct scaler_fun a, struct scaler_fun b) +{ + if ((a.name && !b.name) || (b.name && !a.name)) + return false; + + return ((!a.name && !b.name) || strcmp(a.name, b.name) == 0) && + double_seq(a.params[0], b.params[0]) && + double_seq(a.params[1], b.params[1]) && + a.blur == b.blur && + a.taper == b.taper; +} + +static bool scaler_conf_eq(struct scaler_config a, struct scaler_config b) +{ + // Note: antiring isn't compared because it doesn't affect LUT + // generation + return scaler_fun_eq(a.kernel, b.kernel) && + scaler_fun_eq(a.window, b.window) && + a.radius == b.radius && + a.clamp == b.clamp; +} + +static void reinit_scaler(struct gl_video *p, struct scaler *scaler, + const struct scaler_config *conf, + double scale_factor, + int sizes[]) +{ + assert(conf); + if (scaler_conf_eq(scaler->conf, *conf) && + scaler->scale_factor == scale_factor && + scaler->initialized) + return; + + uninit_scaler(p, scaler); + + if (scaler->index == SCALER_DSCALE && (!conf->kernel.name || + !conf->kernel.name[0])) + { + conf = &p->opts.scaler[SCALER_SCALE]; + } + + if (scaler->index == SCALER_CSCALE && (!conf->kernel.name || + !conf->kernel.name[0])) + { + conf = &p->opts.scaler[SCALER_SCALE]; + } + + struct filter_kernel bare_window; + const struct filter_kernel *t_kernel = mp_find_filter_kernel(conf->kernel.name); + const struct filter_window *t_window = mp_find_filter_window(conf->window.name); + bool is_tscale = scaler->index == SCALER_TSCALE; + if (!t_kernel) { + const struct filter_window *window = mp_find_filter_window(conf->kernel.name); + if (window) { + bare_window = (struct filter_kernel) { .f = *window }; + t_kernel = &bare_window; + } + } + + scaler->conf = *conf; + scaler->conf.kernel.name = (char *)handle_scaler_opt(conf->kernel.name, is_tscale); + scaler->conf.window.name = t_window ? (char *)t_window->name : NULL; + scaler->scale_factor = scale_factor; + scaler->insufficient = false; + scaler->initialized = true; + if (!t_kernel) + return; + + scaler->kernel_storage = *t_kernel; + scaler->kernel = &scaler->kernel_storage; + + if (!t_window) { + // fall back to the scaler's default window if available + t_window = mp_find_filter_window(t_kernel->window); + } + if (t_window) + scaler->kernel->w = *t_window; + + for (int n = 0; n < 2; n++) { + if (!isnan(conf->kernel.params[n])) + scaler->kernel->f.params[n] = conf->kernel.params[n]; + if (!isnan(conf->window.params[n])) + scaler->kernel->w.params[n] = conf->window.params[n]; + } + + if (conf->kernel.blur > 0.0) + scaler->kernel->f.blur = conf->kernel.blur; + if (conf->window.blur > 0.0) + scaler->kernel->w.blur = conf->window.blur; + + if (conf->kernel.taper > 0.0) + scaler->kernel->f.taper = conf->kernel.taper; + if (conf->window.taper > 0.0) + scaler->kernel->w.taper = conf->window.taper; + + if (scaler->kernel->f.resizable && conf->radius > 0.0) + scaler->kernel->f.radius = conf->radius; + + scaler->kernel->clamp = conf->clamp; + scaler->insufficient = !mp_init_filter(scaler->kernel, sizes, scale_factor); + + int size = scaler->kernel->size; + int num_components = size > 2 ? 4 : size; + const struct ra_format *fmt = ra_find_float16_format(p->ra, num_components); + assert(fmt); + + int width = (size + num_components - 1) / num_components; // round up + int stride = width * num_components; + assert(size <= stride); + + static const int lut_size = 256; + float *weights = talloc_array(NULL, float, lut_size * stride); + mp_compute_lut(scaler->kernel, lut_size, stride, weights); + + bool use_1d = scaler->kernel->polar && (p->ra->caps & RA_CAP_TEX_1D); + + struct ra_tex_params lut_params = { + .dimensions = use_1d ? 1 : 2, + .w = use_1d ? lut_size : width, + .h = use_1d ? 1 : lut_size, + .d = 1, + .format = fmt, + .render_src = true, + .src_linear = true, + .initial_data = weights, + }; + scaler->lut = ra_tex_create(p->ra, &lut_params); + + talloc_free(weights); + + debug_check_gl(p, "after initializing scaler"); +} + +// Special helper for sampling from two separated stages +static void pass_sample_separated(struct gl_video *p, struct image src, + struct scaler *scaler, int w, int h) +{ + // Separate the transformation into x and y components, per pass + struct gl_transform t_x = { + .m = {{src.transform.m[0][0], 0.0}, {src.transform.m[1][0], 1.0}}, + .t = {src.transform.t[0], 0.0}, + }; + struct gl_transform t_y = { + .m = {{1.0, src.transform.m[0][1]}, {0.0, src.transform.m[1][1]}}, + .t = {0.0, src.transform.t[1]}, + }; + + // First pass (scale only in the y dir) + src.transform = t_y; + sampler_prelude(p->sc, pass_bind(p, src)); + GLSLF("// first pass\n"); + pass_sample_separated_gen(p->sc, scaler, 0, 1); + GLSLF("color *= %f;\n", src.multiplier); + finish_pass_tex(p, &scaler->sep_fbo, src.w, h); + + // Second pass (scale only in the x dir) + src = image_wrap(scaler->sep_fbo, src.type, src.components); + src.transform = t_x; + pass_describe(p, "%s second pass", scaler->conf.kernel.name); + sampler_prelude(p->sc, pass_bind(p, src)); + pass_sample_separated_gen(p->sc, scaler, 1, 0); +} + +// Picks either the compute shader version or the regular sampler version +// depending on hardware support +static void pass_dispatch_sample_polar(struct gl_video *p, struct scaler *scaler, + struct image img, int w, int h) +{ + uint64_t reqs = RA_CAP_COMPUTE; + if ((p->ra->caps & reqs) != reqs) + goto fallback; + + int bound = ceil(scaler->kernel->radius_cutoff); + int offset = bound - 1; // padding top/left + int padding = offset + bound; // total padding + + float ratiox = (float)w / img.w, + ratioy = (float)h / img.h; + + // For performance we want to load at least as many pixels + // horizontally as there are threads in a warp (32 for nvidia), as + // well as enough to take advantage of shmem parallelism + const int warp_size = 32, threads = 256; + int bw = warp_size; + int bh = threads / bw; + + // We need to sample everything from base_min to base_max, so make sure + // we have enough room in shmem + int iw = (int)ceil(bw / ratiox) + padding + 1, + ih = (int)ceil(bh / ratioy) + padding + 1; + + int shmem_req = iw * ih * img.components * sizeof(float); + if (shmem_req > p->ra->max_shmem) + goto fallback; + + pass_is_compute(p, bw, bh, false); + pass_compute_polar(p->sc, scaler, img.components, bw, bh, iw, ih); + return; + +fallback: + // Fall back to regular polar shader when compute shaders are unsupported + // or the kernel is too big for shmem + pass_sample_polar(p->sc, scaler, img.components, + p->ra->caps & RA_CAP_GATHER); +} + +// Sample from image, with the src rectangle given by it. +// The dst rectangle is implicit by what the caller will do next, but w and h +// must still be what is going to be used (to dimension FBOs correctly). +// This will write the scaled contents to the vec4 "color". +// The scaler unit is initialized by this function; in order to avoid cache +// thrashing, the scaler unit should usually use the same parameters. +static void pass_sample(struct gl_video *p, struct image img, + struct scaler *scaler, const struct scaler_config *conf, + double scale_factor, int w, int h) +{ + reinit_scaler(p, scaler, conf, scale_factor, filter_sizes); + + // Describe scaler + const char *scaler_opt[] = { + [SCALER_SCALE] = "scale", + [SCALER_DSCALE] = "dscale", + [SCALER_CSCALE] = "cscale", + [SCALER_TSCALE] = "tscale", + }; + + pass_describe(p, "%s=%s (%s)", scaler_opt[scaler->index], + scaler->conf.kernel.name, plane_names[img.type]); + + bool is_separated = scaler->kernel && !scaler->kernel->polar; + + // Set up the transformation+prelude and bind the texture, for everything + // other than separated scaling (which does this in the subfunction) + if (!is_separated) + sampler_prelude(p->sc, pass_bind(p, img)); + + // Dispatch the scaler. They're all wildly different. + const char *name = scaler->conf.kernel.name; + if (strcmp(name, "bilinear") == 0) { + GLSL(color = texture(tex, pos);) + } else if (strcmp(name, "bicubic_fast") == 0) { + pass_sample_bicubic_fast(p->sc); + } else if (strcmp(name, "oversample") == 0) { + pass_sample_oversample(p->sc, scaler, w, h); + } else if (scaler->kernel && scaler->kernel->polar) { + pass_dispatch_sample_polar(p, scaler, img, w, h); + } else if (scaler->kernel) { + pass_sample_separated(p, img, scaler, w, h); + } else { + MP_ASSERT_UNREACHABLE(); // should never happen + } + + // Apply any required multipliers. Separated scaling already does this in + // its first stage + if (!is_separated) + GLSLF("color *= %f;\n", img.multiplier); + + // Micro-optimization: Avoid scaling unneeded channels + skip_unused(p, img.components); +} + +// Returns true if two images are semantically equivalent (same metadata) +static bool image_equiv(struct image a, struct image b) +{ + return a.type == b.type && + a.components == b.components && + a.multiplier == b.multiplier && + a.tex->params.format == b.tex->params.format && + a.tex->params.w == b.tex->params.w && + a.tex->params.h == b.tex->params.h && + a.w == b.w && + a.h == b.h && + gl_transform_eq(a.transform, b.transform); +} + +static void deband_hook(struct gl_video *p, struct image img, + struct gl_transform *trans, void *priv) +{ + pass_describe(p, "debanding (%s)", plane_names[img.type]); + pass_sample_deband(p->sc, p->opts.deband_opts, &p->lfg, + p->image_params.color.gamma); +} + +static void unsharp_hook(struct gl_video *p, struct image img, + struct gl_transform *trans, void *priv) +{ + pass_describe(p, "unsharp masking"); + pass_sample_unsharp(p->sc, p->opts.unsharp); +} + +struct szexp_ctx { + struct gl_video *p; + struct image img; +}; + +static bool szexp_lookup(void *priv, struct bstr var, float size[2]) +{ + struct szexp_ctx *ctx = priv; + struct gl_video *p = ctx->p; + + if (bstr_equals0(var, "NATIVE_CROPPED")) { + size[0] = (p->src_rect.x1 - p->src_rect.x0) * p->texture_offset.m[0][0]; + size[1] = (p->src_rect.y1 - p->src_rect.y0) * p->texture_offset.m[1][1]; + return true; + } + + // The size of OUTPUT is determined. It could be useful for certain + // user shaders to skip passes. + if (bstr_equals0(var, "OUTPUT")) { + size[0] = p->dst_rect.x1 - p->dst_rect.x0; + size[1] = p->dst_rect.y1 - p->dst_rect.y0; + return true; + } + + // HOOKED is a special case + if (bstr_equals0(var, "HOOKED")) { + size[0] = ctx->img.w; + size[1] = ctx->img.h; + return true; + } + + for (int o = 0; o < p->num_saved_imgs; o++) { + if (bstr_equals0(var, p->saved_imgs[o].name)) { + size[0] = p->saved_imgs[o].img.w; + size[1] = p->saved_imgs[o].img.h; + return true; + } + } + + return false; +} + +static bool user_hook_cond(struct gl_video *p, struct image img, void *priv) +{ + struct gl_user_shader_hook *shader = priv; + assert(shader); + + float res = false; + struct szexp_ctx ctx = {p, img}; + eval_szexpr(p->log, &ctx, szexp_lookup, shader->cond, &res); + return res; +} + +static void user_hook(struct gl_video *p, struct image img, + struct gl_transform *trans, void *priv) +{ + struct gl_user_shader_hook *shader = priv; + assert(shader); + load_shader(p, shader->pass_body); + + pass_describe(p, "user shader: %.*s (%s)", BSTR_P(shader->pass_desc), + plane_names[img.type]); + + if (shader->compute.active) { + p->pass_compute = shader->compute; + GLSLF("hook();\n"); + } else { + GLSLF("color = hook();\n"); + } + + // Make sure we at least create a legal FBO on failure, since it's better + // to do this and display an error message than just crash OpenGL + float w = 1.0, h = 1.0; + + eval_szexpr(p->log, &(struct szexp_ctx){p, img}, szexp_lookup, shader->width, &w); + eval_szexpr(p->log, &(struct szexp_ctx){p, img}, szexp_lookup, shader->height, &h); + + *trans = (struct gl_transform){{{w / img.w, 0}, {0, h / img.h}}}; + gl_transform_trans(shader->offset, trans); +} + +static bool add_user_hook(void *priv, struct gl_user_shader_hook hook) +{ + struct gl_video *p = priv; + struct gl_user_shader_hook *copy = talloc_ptrtype(p, copy); + *copy = hook; + + struct tex_hook texhook = { + .save_tex = bstrdup0(copy, hook.save_tex), + .components = hook.components, + .align_offset = hook.align_offset, + .hook = user_hook, + .cond = user_hook_cond, + .priv = copy, + }; + + for (int h = 0; h < SHADER_MAX_HOOKS; h++) + texhook.hook_tex[h] = bstrdup0(copy, hook.hook_tex[h]); + for (int h = 0; h < SHADER_MAX_BINDS; h++) + texhook.bind_tex[h] = bstrdup0(copy, hook.bind_tex[h]); + + MP_TARRAY_APPEND(p, p->tex_hooks, p->num_tex_hooks, texhook); + return true; +} + +static bool add_user_tex(void *priv, struct gl_user_shader_tex tex) +{ + struct gl_video *p = priv; + + tex.tex = ra_tex_create(p->ra, &tex.params); + TA_FREEP(&tex.params.initial_data); + + if (!tex.tex) + return false; + + MP_TARRAY_APPEND(p, p->user_textures, p->num_user_textures, tex); + return true; +} + +static void load_user_shaders(struct gl_video *p, char **shaders) +{ + if (!shaders) + return; + + for (int n = 0; shaders[n] != NULL; n++) { + struct bstr file = load_cached_file(p, shaders[n]); + parse_user_shader(p->log, p->ra, file, p, add_user_hook, add_user_tex); + } +} + +static void gl_video_setup_hooks(struct gl_video *p) +{ + gl_video_reset_hooks(p); + + if (p->opts.deband) { + MP_TARRAY_APPEND(p, p->tex_hooks, p->num_tex_hooks, (struct tex_hook) { + .hook_tex = {"LUMA", "CHROMA", "RGB", "XYZ"}, + .bind_tex = {"HOOKED"}, + .hook = deband_hook, + }); + } + + if (p->opts.unsharp != 0.0) { + MP_TARRAY_APPEND(p, p->tex_hooks, p->num_tex_hooks, (struct tex_hook) { + .hook_tex = {"MAIN"}, + .bind_tex = {"HOOKED"}, + .hook = unsharp_hook, + }); + } + + load_user_shaders(p, p->opts.user_shaders); +} + +// sample from video textures, set "color" variable to yuv value +static void pass_read_video(struct gl_video *p) +{ + struct image img[4]; + struct gl_transform offsets[4]; + pass_get_images(p, &p->image, img, offsets); + + // To keep the code as simple as possibly, we currently run all shader + // stages even if they would be unnecessary (e.g. no hooks for a texture). + // In the future, deferred image should optimize this away. + + // Merge semantically identical textures. This loop is done from back + // to front so that merged textures end up in the right order while + // simultaneously allowing us to skip unnecessary merges + for (int n = 3; n >= 0; n--) { + if (img[n].type == PLANE_NONE) + continue; + + int first = n; + int num = 0; + + for (int i = 0; i < n; i++) { + if (image_equiv(img[n], img[i]) && + gl_transform_eq(offsets[n], offsets[i])) + { + GLSLF("// merging plane %d ...\n", i); + copy_image(p, &num, img[i]); + first = MPMIN(first, i); + img[i] = (struct image){0}; + } + } + + if (num > 0) { + GLSLF("// merging plane %d ... into %d\n", n, first); + copy_image(p, &num, img[n]); + pass_describe(p, "merging planes"); + finish_pass_tex(p, &p->merge_tex[n], img[n].w, img[n].h); + img[first] = image_wrap(p->merge_tex[n], img[n].type, num); + img[n] = (struct image){0}; + } + } + + // If any textures are still in integer format by this point, we need + // to introduce an explicit conversion pass to avoid breaking hooks/scaling + for (int n = 0; n < 4; n++) { + if (img[n].tex && img[n].tex->params.format->ctype == RA_CTYPE_UINT) { + GLSLF("// use_integer fix for plane %d\n", n); + copy_image(p, &(int){0}, img[n]); + pass_describe(p, "use_integer fix"); + finish_pass_tex(p, &p->integer_tex[n], img[n].w, img[n].h); + img[n] = image_wrap(p->integer_tex[n], img[n].type, + img[n].components); + } + } + + // The basic idea is we assume the rgb/luma texture is the "reference" and + // scale everything else to match, after all planes are finalized. + // We find the reference texture first, in order to maintain texture offset + // between hooks on different type of planes. + int reference_tex_num = 0; + for (int n = 0; n < 4; n++) { + switch (img[n].type) { + case PLANE_RGB: + case PLANE_XYZ: + case PLANE_LUMA: break; + default: continue; + } + + reference_tex_num = n; + break; + } + + // Dispatch the hooks for all of these textures, saving and perhaps + // modifying them in the process + for (int n = 0; n < 4; n++) { + const char *name; + switch (img[n].type) { + case PLANE_RGB: name = "RGB"; break; + case PLANE_LUMA: name = "LUMA"; break; + case PLANE_CHROMA: name = "CHROMA"; break; + case PLANE_ALPHA: name = "ALPHA"; break; + case PLANE_XYZ: name = "XYZ"; break; + default: continue; + } + + img[n] = pass_hook(p, name, img[n], &offsets[n]); + + if (reference_tex_num == n) { + // The reference texture is finalized now. + p->texture_w = img[n].w; + p->texture_h = img[n].h; + p->texture_offset = offsets[n]; + } + } + + // At this point all planes are finalized but they may not be at the + // required size yet. Furthermore, they may have texture offsets that + // require realignment. + + // Compute the reference rect + struct mp_rect_f src = {0.0, 0.0, p->image_params.w, p->image_params.h}; + struct mp_rect_f ref = src; + gl_transform_rect(p->texture_offset, &ref); + + // Explicitly scale all of the textures that don't match + for (int n = 0; n < 4; n++) { + if (img[n].type == PLANE_NONE) + continue; + + // If the planes are aligned identically, we will end up with the + // exact same source rectangle. + struct mp_rect_f rect = src; + gl_transform_rect(offsets[n], &rect); + if (mp_rect_f_seq(ref, rect)) + continue; + + // If the rectangles differ, then our planes have a different + // alignment and/or size. First of all, we have to compute the + // corrections required to meet the target rectangle + struct gl_transform fix = { + .m = {{(ref.x1 - ref.x0) / (rect.x1 - rect.x0), 0.0}, + {0.0, (ref.y1 - ref.y0) / (rect.y1 - rect.y0)}}, + .t = {ref.x0, ref.y0}, + }; + + // Since the scale in texture space is different from the scale in + // absolute terms, we have to scale the coefficients down to be + // relative to the texture's physical dimensions and local offset + struct gl_transform scale = { + .m = {{(float)img[n].w / p->texture_w, 0.0}, + {0.0, (float)img[n].h / p->texture_h}}, + .t = {-rect.x0, -rect.y0}, + }; + if (p->image_params.rotate % 180 == 90) + MPSWAP(double, scale.m[0][0], scale.m[1][1]); + + gl_transform_trans(scale, &fix); + + // Since the texture transform is a function of the texture coordinates + // to texture space, rather than the other way around, we have to + // actually apply the *inverse* of this. Fortunately, calculating + // the inverse is relatively easy here. + fix.m[0][0] = 1.0 / fix.m[0][0]; + fix.m[1][1] = 1.0 / fix.m[1][1]; + fix.t[0] = fix.m[0][0] * -fix.t[0]; + fix.t[1] = fix.m[1][1] * -fix.t[1]; + gl_transform_trans(fix, &img[n].transform); + + int scaler_id = -1; + const char *name = NULL; + switch (img[n].type) { + case PLANE_RGB: + case PLANE_LUMA: + case PLANE_XYZ: + scaler_id = SCALER_SCALE; + // these aren't worth hooking, fringe hypothetical cases only + break; + case PLANE_CHROMA: + scaler_id = SCALER_CSCALE; + name = "CHROMA_SCALED"; + break; + case PLANE_ALPHA: + // alpha always uses bilinear + name = "ALPHA_SCALED"; + } + + if (scaler_id < 0) + continue; + + const struct scaler_config *conf = &p->opts.scaler[scaler_id]; + + if (scaler_id == SCALER_CSCALE && (!conf->kernel.name || + !conf->kernel.name[0])) + { + conf = &p->opts.scaler[SCALER_SCALE]; + } + + struct scaler *scaler = &p->scaler[scaler_id]; + + // bilinear scaling is a free no-op thanks to GPU sampling + if (strcmp(conf->kernel.name, "bilinear") != 0) { + GLSLF("// upscaling plane %d\n", n); + pass_sample(p, img[n], scaler, conf, 1.0, p->texture_w, p->texture_h); + finish_pass_tex(p, &p->scale_tex[n], p->texture_w, p->texture_h); + img[n] = image_wrap(p->scale_tex[n], img[n].type, img[n].components); + } + + // Run any post-scaling hooks + img[n] = pass_hook(p, name, img[n], NULL); + } + + // All planes are of the same size and properly aligned at this point + pass_describe(p, "combining planes"); + int coord = 0; + for (int i = 0; i < 4; i++) { + if (img[i].type != PLANE_NONE) + copy_image(p, &coord, img[i]); + } + p->components = coord; +} + +// Utility function that simply binds a texture and reads from it, without any +// transformations. +static void pass_read_tex(struct gl_video *p, struct ra_tex *tex) +{ + struct image img = image_wrap(tex, PLANE_RGB, p->components); + copy_image(p, &(int){0}, img); +} + +// yuv conversion, and any other conversions before main up/down-scaling +static void pass_convert_yuv(struct gl_video *p) +{ + struct gl_shader_cache *sc = p->sc; + + struct mp_csp_params cparams = MP_CSP_PARAMS_DEFAULTS; + cparams.gray = p->is_gray; + cparams.is_float = p->ra_format.component_type == RA_CTYPE_FLOAT; + mp_csp_set_image_params(&cparams, &p->image_params); + mp_csp_equalizer_state_get(p->video_eq, &cparams); + p->user_gamma = 1.0 / (cparams.gamma * p->opts.gamma); + + pass_describe(p, "color conversion"); + + if (p->color_swizzle[0]) + GLSLF("color = color.%s;\n", p->color_swizzle); + + // Pre-colormatrix input gamma correction + if (cparams.color.space == MP_CSP_XYZ) + pass_linearize(p->sc, p->image_params.color.gamma); + + // We always explicitly normalize the range in pass_read_video + cparams.input_bits = cparams.texture_bits = 0; + + // Conversion to RGB. For RGB itself, this still applies e.g. brightness + // and contrast controls, or expansion of e.g. LSB-packed 10 bit data. + struct mp_cmat m = {{{0}}}; + mp_get_csp_matrix(&cparams, &m); + gl_sc_uniform_mat3(sc, "colormatrix", true, &m.m[0][0]); + gl_sc_uniform_vec3(sc, "colormatrix_c", m.c); + + GLSL(color.rgb = mat3(colormatrix) * color.rgb + colormatrix_c;) + + if (cparams.color.space == MP_CSP_XYZ) { + pass_delinearize(p->sc, p->image_params.color.gamma); + // mp_get_csp_matrix implicitly converts XYZ to DCI-P3 + p->image_params.color.space = MP_CSP_RGB; + p->image_params.color.primaries = MP_CSP_PRIM_DCI_P3; + } + + if (p->image_params.color.space == MP_CSP_BT_2020_C) { + // Conversion for C'rcY'cC'bc via the BT.2020 CL system: + // C'bc = (B'-Y'c) / 1.9404 | C'bc <= 0 + // = (B'-Y'c) / 1.5816 | C'bc > 0 + // + // C'rc = (R'-Y'c) / 1.7184 | C'rc <= 0 + // = (R'-Y'c) / 0.9936 | C'rc > 0 + // + // as per the BT.2020 specification, table 4. This is a non-linear + // transformation because (constant) luminance receives non-equal + // contributions from the three different channels. + GLSLF("// constant luminance conversion \n" + "color.br = color.br * mix(vec2(1.5816, 0.9936), \n" + " vec2(1.9404, 1.7184), \n" + " %s(lessThanEqual(color.br, vec2(0))))\n" + " + color.gg; \n", + gl_sc_bvec(p->sc, 2)); + // Expand channels to camera-linear light. This shader currently just + // assumes everything uses the BT.2020 12-bit gamma function, since the + // difference between 10 and 12-bit is negligible for anything other + // than 12-bit content. + GLSLF("color.rgb = mix(color.rgb * vec3(1.0/4.5), \n" + " pow((color.rgb + vec3(0.0993))*vec3(1.0/1.0993), \n" + " vec3(1.0/0.45)), \n" + " %s(lessThanEqual(vec3(0.08145), color.rgb))); \n", + gl_sc_bvec(p->sc, 3)); + // Calculate the green channel from the expanded RYcB + // The BT.2020 specification says Yc = 0.2627*R + 0.6780*G + 0.0593*B + GLSL(color.g = (color.g - 0.2627*color.r - 0.0593*color.b)*1.0/0.6780;) + // Recompress to receive the R'G'B' result, same as other systems + GLSLF("color.rgb = mix(color.rgb * vec3(4.5), \n" + " vec3(1.0993) * pow(color.rgb, vec3(0.45)) - vec3(0.0993), \n" + " %s(lessThanEqual(vec3(0.0181), color.rgb))); \n", + gl_sc_bvec(p->sc, 3)); + } + + p->components = 3; + if (!p->has_alpha || p->opts.alpha_mode == ALPHA_NO) { + GLSL(color.a = 1.0;) + } else if (p->image_params.alpha == MP_ALPHA_PREMUL) { + p->components = 4; + } else { + p->components = 4; + GLSL(color = vec4(color.rgb * color.a, color.a);) // straight -> premul + } +} + +static void get_scale_factors(struct gl_video *p, bool transpose_rot, double xy[2]) +{ + double target_w = p->src_rect.x1 - p->src_rect.x0; + double target_h = p->src_rect.y1 - p->src_rect.y0; + if (transpose_rot && p->image_params.rotate % 180 == 90) + MPSWAP(double, target_w, target_h); + xy[0] = (p->dst_rect.x1 - p->dst_rect.x0) / target_w; + xy[1] = (p->dst_rect.y1 - p->dst_rect.y0) / target_h; +} + +// Cropping. +static void compute_src_transform(struct gl_video *p, struct gl_transform *tr) +{ + float sx = (p->src_rect.x1 - p->src_rect.x0) / (float)p->texture_w, + sy = (p->src_rect.y1 - p->src_rect.y0) / (float)p->texture_h, + ox = p->src_rect.x0, + oy = p->src_rect.y0; + struct gl_transform transform = {{{sx, 0}, {0, sy}}, {ox, oy}}; + + gl_transform_trans(p->texture_offset, &transform); + + *tr = transform; +} + +// Takes care of the main scaling and pre/post-conversions +static void pass_scale_main(struct gl_video *p) +{ + // Figure out the main scaler. + double xy[2]; + get_scale_factors(p, true, xy); + + // actual scale factor should be divided by the scale factor of prescaling. + xy[0] /= p->texture_offset.m[0][0]; + xy[1] /= p->texture_offset.m[1][1]; + + // The calculation of scale factor involves 32-bit float(from gl_transform), + // use non-strict equality test to tolerate precision loss. + bool downscaling = xy[0] < 1.0 - FLT_EPSILON || xy[1] < 1.0 - FLT_EPSILON; + bool upscaling = !downscaling && (xy[0] > 1.0 + FLT_EPSILON || + xy[1] > 1.0 + FLT_EPSILON); + double scale_factor = 1.0; + + struct scaler *scaler = &p->scaler[SCALER_SCALE]; + struct scaler_config scaler_conf = p->opts.scaler[SCALER_SCALE]; + if (p->opts.scaler_resizes_only && !downscaling && !upscaling) { + scaler_conf.kernel.name = "bilinear"; + // For scaler-resizes-only, we round the texture offset to + // the nearest round value in order to prevent ugly blurriness + // (in exchange for slightly shifting the image by up to half a + // subpixel) + p->texture_offset.t[0] = roundf(p->texture_offset.t[0]); + p->texture_offset.t[1] = roundf(p->texture_offset.t[1]); + } + if (downscaling && p->opts.scaler[SCALER_DSCALE].kernel.name) { + scaler_conf = p->opts.scaler[SCALER_DSCALE]; + scaler = &p->scaler[SCALER_DSCALE]; + } + + // When requesting correct-downscaling and the clip is anamorphic, and + // because only a single scale factor is used for both axes, enable it only + // when both axes are downscaled, and use the milder of the factors to not + // end up with too much blur on one axis (even if we end up with sub-optimal + // scale factor on the other axis). This is better than not respecting + // correct scaling at all for anamorphic clips. + double f = MPMAX(xy[0], xy[1]); + if (p->opts.correct_downscaling && f < 1.0) + scale_factor = 1.0 / f; + + // Pre-conversion, like linear light/sigmoidization + GLSLF("// scaler pre-conversion\n"); + bool use_linear = false; + if (downscaling) { + use_linear = p->opts.linear_downscaling; + + // Linear light downscaling results in nasty artifacts for HDR curves + // due to the potentially extreme brightness differences severely + // compounding any ringing. So just scale in gamma light instead. + if (mp_trc_is_hdr(p->image_params.color.gamma)) + use_linear = false; + } else if (upscaling) { + use_linear = p->opts.linear_upscaling || p->opts.sigmoid_upscaling; + } + + if (use_linear) { + p->use_linear = true; + pass_linearize(p->sc, p->image_params.color.gamma); + pass_opt_hook_point(p, "LINEAR", NULL); + } + + bool use_sigmoid = use_linear && p->opts.sigmoid_upscaling && upscaling; + float sig_center, sig_slope, sig_offset, sig_scale; + if (use_sigmoid) { + // Coefficients for the sigmoidal transform are taken from the + // formula here: http://www.imagemagick.org/Usage/color_mods/#sigmoidal + sig_center = p->opts.sigmoid_center; + sig_slope = p->opts.sigmoid_slope; + // This function needs to go through (0,0) and (1,1) so we compute the + // values at 1 and 0, and then scale/shift them, respectively. + sig_offset = 1.0/(1+expf(sig_slope * sig_center)); + sig_scale = 1.0/(1+expf(sig_slope * (sig_center-1))) - sig_offset; + GLSL(color.rgb = clamp(color.rgb, 0.0, 1.0);) + GLSLF("color.rgb = %f - log(1.0/(color.rgb * %f + %f) - 1.0) * 1.0/%f;\n", + sig_center, sig_scale, sig_offset, sig_slope); + pass_opt_hook_point(p, "SIGMOID", NULL); + } + + pass_opt_hook_point(p, "PREKERNEL", NULL); + + int vp_w = p->dst_rect.x1 - p->dst_rect.x0; + int vp_h = p->dst_rect.y1 - p->dst_rect.y0; + struct gl_transform transform; + compute_src_transform(p, &transform); + + GLSLF("// main scaling\n"); + finish_pass_tex(p, &p->indirect_tex, p->texture_w, p->texture_h); + struct image src = image_wrap(p->indirect_tex, PLANE_RGB, p->components); + gl_transform_trans(transform, &src.transform); + pass_sample(p, src, scaler, &scaler_conf, scale_factor, vp_w, vp_h); + + // Changes the texture size to display size after main scaler. + p->texture_w = vp_w; + p->texture_h = vp_h; + + pass_opt_hook_point(p, "POSTKERNEL", NULL); + + GLSLF("// scaler post-conversion\n"); + if (use_sigmoid) { + // Inverse of the transformation above + GLSL(color.rgb = clamp(color.rgb, 0.0, 1.0);) + GLSLF("color.rgb = (1.0/(1.0 + exp(%f * (%f - color.rgb))) - %f) * 1.0/%f;\n", + sig_slope, sig_center, sig_offset, sig_scale); + } +} + +// Adapts the colors to the right output color space. (Final pass during +// rendering) +// If OSD is true, ignore any changes that may have been made to the video +// by previous passes (i.e. linear scaling) +static void pass_colormanage(struct gl_video *p, struct mp_colorspace src, + struct mp_colorspace fbo_csp, int flags, bool osd) +{ + struct ra *ra = p->ra; + + // Configure the destination according to the FBO color space, + // unless specific transfer function, primaries or target peak + // is set. If values are set to _AUTO, the most likely intended + // values are guesstimated later in this function. + struct mp_colorspace dst = { + .gamma = p->opts.target_trc == MP_CSP_TRC_AUTO ? + fbo_csp.gamma : p->opts.target_trc, + .primaries = p->opts.target_prim == MP_CSP_PRIM_AUTO ? + fbo_csp.primaries : p->opts.target_prim, + .light = MP_CSP_LIGHT_DISPLAY, + .hdr.max_luma = !p->opts.target_peak ? + fbo_csp.hdr.max_luma : p->opts.target_peak, + }; + + if (!p->colorspace_override_warned && + ((fbo_csp.gamma && dst.gamma != fbo_csp.gamma) || + (fbo_csp.primaries && dst.primaries != fbo_csp.primaries))) + { + MP_WARN(p, "One or more colorspace value is being overridden " + "by user while the FBO provides colorspace information: " + "transfer function: (dst: %s, fbo: %s), " + "primaries: (dst: %s, fbo: %s). " + "Rendering can lead to incorrect results!\n", + m_opt_choice_str(mp_csp_trc_names, dst.gamma), + m_opt_choice_str(mp_csp_trc_names, fbo_csp.gamma), + m_opt_choice_str(mp_csp_prim_names, dst.primaries), + m_opt_choice_str(mp_csp_prim_names, fbo_csp.primaries)); + p->colorspace_override_warned = true; + } + + if (dst.gamma == MP_CSP_TRC_HLG) + dst.light = MP_CSP_LIGHT_SCENE_HLG; + + if (p->use_lut_3d && (flags & RENDER_SCREEN_COLOR)) { + // The 3DLUT is always generated against the video's original source + // space, *not* the reference space. (To avoid having to regenerate + // the 3DLUT for the OSD on every frame) + enum mp_csp_prim prim_orig = p->image_params.color.primaries; + enum mp_csp_trc trc_orig = p->image_params.color.gamma; + + // One exception: HDR is not implemented by LittleCMS for technical + // limitation reasons, so we use a gamma 2.2 input curve here instead. + // We could pick any value we want here, the difference is just coding + // efficiency. + if (mp_trc_is_hdr(trc_orig)) + trc_orig = MP_CSP_TRC_GAMMA22; + + if (gl_video_get_lut3d(p, prim_orig, trc_orig)) { + dst.primaries = prim_orig; + dst.gamma = trc_orig; + assert(dst.primaries && dst.gamma); + } + } + + if (dst.primaries == MP_CSP_PRIM_AUTO) { + // The vast majority of people are on sRGB or BT.709 displays, so pick + // this as the default output color space. + dst.primaries = MP_CSP_PRIM_BT_709; + + if (src.primaries == MP_CSP_PRIM_BT_601_525 || + src.primaries == MP_CSP_PRIM_BT_601_625) + { + // Since we auto-pick BT.601 and BT.709 based on the dimensions, + // combined with the fact that they're very similar to begin with, + // and to avoid confusing the average user, just don't adapt BT.601 + // content automatically at all. + dst.primaries = src.primaries; + } + } + + if (dst.gamma == MP_CSP_TRC_AUTO) { + // Most people seem to complain when the image is darker or brighter + // than what they're "used to", so just avoid changing the gamma + // altogether by default. The only exceptions to this rule apply to + // very unusual TRCs, which even hardcode technoluddites would probably + // not enjoy viewing unaltered. + dst.gamma = src.gamma; + + // Avoid outputting linear light or HDR content "by default". For these + // just pick gamma 2.2 as a default, since it's a good estimate for + // the response of typical displays + if (dst.gamma == MP_CSP_TRC_LINEAR || mp_trc_is_hdr(dst.gamma)) + dst.gamma = MP_CSP_TRC_GAMMA22; + } + + // If there's no specific signal peak known for the output display, infer + // it from the chosen transfer function. Also normalize the src peak, in + // case it was unknown + if (!dst.hdr.max_luma) + dst.hdr.max_luma = mp_trc_nom_peak(dst.gamma) * MP_REF_WHITE; + if (!src.hdr.max_luma) + src.hdr.max_luma = mp_trc_nom_peak(src.gamma) * MP_REF_WHITE; + + // Whitelist supported modes + switch (p->opts.tone_map.curve) { + case TONE_MAPPING_AUTO: + case TONE_MAPPING_CLIP: + case TONE_MAPPING_MOBIUS: + case TONE_MAPPING_REINHARD: + case TONE_MAPPING_HABLE: + case TONE_MAPPING_GAMMA: + case TONE_MAPPING_LINEAR: + case TONE_MAPPING_BT_2390: + break; + default: + MP_WARN(p, "Tone mapping curve unsupported by vo_gpu, falling back.\n"); + p->opts.tone_map.curve = TONE_MAPPING_AUTO; + break; + } + + switch (p->opts.tone_map.gamut_mode) { + case GAMUT_AUTO: + case GAMUT_WARN: + case GAMUT_CLIP: + case GAMUT_DESATURATE: + break; + default: + MP_WARN(p, "Gamut mapping mode unsupported by vo_gpu, falling back.\n"); + p->opts.tone_map.gamut_mode = GAMUT_AUTO; + break; + } + + struct gl_tone_map_opts tone_map = p->opts.tone_map; + bool detect_peak = tone_map.compute_peak >= 0 && mp_trc_is_hdr(src.gamma) + && src.hdr.max_luma > dst.hdr.max_luma; + + if (detect_peak && !p->hdr_peak_ssbo) { + struct { + float average[2]; + int32_t frame_sum; + uint32_t frame_max; + uint32_t counter; + } peak_ssbo = {0}; + + struct ra_buf_params params = { + .type = RA_BUF_TYPE_SHADER_STORAGE, + .size = sizeof(peak_ssbo), + .initial_data = &peak_ssbo, + }; + + p->hdr_peak_ssbo = ra_buf_create(ra, ¶ms); + if (!p->hdr_peak_ssbo) { + MP_WARN(p, "Failed to create HDR peak detection SSBO, disabling.\n"); + tone_map.compute_peak = p->opts.tone_map.compute_peak = -1; + detect_peak = false; + } + } + + if (detect_peak) { + pass_describe(p, "detect HDR peak"); + pass_is_compute(p, 8, 8, true); // 8x8 is good for performance + gl_sc_ssbo(p->sc, "PeakDetect", p->hdr_peak_ssbo, + "vec2 average;" + "int frame_sum;" + "uint frame_max;" + "uint counter;" + ); + } else { + tone_map.compute_peak = -1; + } + + // Adapt from src to dst as necessary + pass_color_map(p->sc, p->use_linear && !osd, src, dst, &tone_map); + + if (p->use_lut_3d && (flags & RENDER_SCREEN_COLOR)) { + gl_sc_uniform_texture(p->sc, "lut_3d", p->lut_3d_texture); + GLSL(vec3 cpos;) + for (int i = 0; i < 3; i++) + GLSLF("cpos[%d] = LUT_POS(color[%d], %d.0);\n", i, i, p->lut_3d_size[i]); + GLSL(color.rgb = tex3D(lut_3d, cpos).rgb;) + } +} + +void gl_video_set_fb_depth(struct gl_video *p, int fb_depth) +{ + p->fb_depth = fb_depth; +} + +static void pass_dither(struct gl_video *p) +{ + // Assume 8 bits per component if unknown. + int dst_depth = p->fb_depth > 0 ? p->fb_depth : 8; + if (p->opts.dither_depth > 0) + dst_depth = p->opts.dither_depth; + + if (p->opts.dither_depth < 0 || p->opts.dither_algo == DITHER_NONE) + return; + + if (p->opts.dither_algo == DITHER_ERROR_DIFFUSION) { + const struct error_diffusion_kernel *kernel = + mp_find_error_diffusion_kernel(p->opts.error_diffusion); + int o_w = p->dst_rect.x1 - p->dst_rect.x0, + o_h = p->dst_rect.y1 - p->dst_rect.y0; + + int shmem_req = mp_ef_compute_shared_memory_size(kernel, o_h); + if (shmem_req > p->ra->max_shmem) { + MP_WARN(p, "Fallback to dither=fruit because there is no enough " + "shared memory (%d/%d).\n", + shmem_req, (int)p->ra->max_shmem); + p->opts.dither_algo = DITHER_FRUIT; + } else { + finish_pass_tex(p, &p->error_diffusion_tex[0], o_w, o_h); + + struct image img = image_wrap(p->error_diffusion_tex[0], PLANE_RGB, p->components); + + // Ensure the block size doesn't exceed the maximum of the + // implementation. + int block_size = MPMIN(p->ra->max_compute_group_threads, o_h); + + pass_describe(p, "dither=error-diffusion (kernel=%s, depth=%d)", + kernel->name, dst_depth); + + p->pass_compute = (struct compute_info) { + .active = true, + .threads_w = block_size, + .threads_h = 1, + .directly_writes = true + }; + + int tex_id = pass_bind(p, img); + + pass_error_diffusion(p->sc, kernel, tex_id, o_w, o_h, + dst_depth, block_size); + + finish_pass_tex(p, &p->error_diffusion_tex[1], o_w, o_h); + + img = image_wrap(p->error_diffusion_tex[1], PLANE_RGB, p->components); + copy_image(p, &(int){0}, img); + + return; + } + } + + if (!p->dither_texture) { + MP_VERBOSE(p, "Dither to %d.\n", dst_depth); + + int tex_size = 0; + void *tex_data = NULL; + const struct ra_format *fmt = NULL; + void *temp = NULL; + + if (p->opts.dither_algo == DITHER_FRUIT) { + int sizeb = p->opts.dither_size; + int size = 1 << sizeb; + + if (p->last_dither_matrix_size != size) { + p->last_dither_matrix = talloc_realloc(p, p->last_dither_matrix, + float, size * size); + mp_make_fruit_dither_matrix(p->last_dither_matrix, sizeb); + p->last_dither_matrix_size = size; + } + + // Prefer R16 texture since they provide higher precision. + fmt = ra_find_unorm_format(p->ra, 2, 1); + if (!fmt) + fmt = ra_find_float16_format(p->ra, 1); + if (fmt) { + tex_size = size; + tex_data = p->last_dither_matrix; + if (fmt->ctype == RA_CTYPE_UNORM) { + uint16_t *t = temp = talloc_array(NULL, uint16_t, size * size); + for (int n = 0; n < size * size; n++) + t[n] = p->last_dither_matrix[n] * UINT16_MAX; + tex_data = t; + } + } else { + MP_VERBOSE(p, "GL too old. Falling back to ordered dither.\n"); + p->opts.dither_algo = DITHER_ORDERED; + } + } + + if (p->opts.dither_algo == DITHER_ORDERED) { + temp = talloc_array(NULL, char, 8 * 8); + mp_make_ordered_dither_matrix(temp, 8); + + fmt = ra_find_unorm_format(p->ra, 1, 1); + tex_size = 8; + tex_data = temp; + } + + struct ra_tex_params params = { + .dimensions = 2, + .w = tex_size, + .h = tex_size, + .d = 1, + .format = fmt, + .render_src = true, + .src_repeat = true, + .initial_data = tex_data, + }; + p->dither_texture = ra_tex_create(p->ra, ¶ms); + + debug_check_gl(p, "dither setup"); + + talloc_free(temp); + + if (!p->dither_texture) + return; + } + + GLSLF("// dithering\n"); + + // This defines how many bits are considered significant for output on + // screen. The superfluous bits will be used for rounding according to the + // dither matrix. The precision of the source implicitly decides how many + // dither patterns can be visible. + int dither_quantization = (1 << dst_depth) - 1; + int dither_size = p->dither_texture->params.w; + + gl_sc_uniform_texture(p->sc, "dither", p->dither_texture); + + GLSLF("vec2 dither_pos = gl_FragCoord.xy * 1.0/%d.0;\n", dither_size); + + if (p->opts.temporal_dither) { + int phase = (p->frames_rendered / p->opts.temporal_dither_period) % 8u; + float r = phase * (M_PI / 2); // rotate + float m = phase < 4 ? 1 : -1; // mirror + + float matrix[2][2] = {{cos(r), -sin(r) }, + {sin(r) * m, cos(r) * m}}; + gl_sc_uniform_dynamic(p->sc); + gl_sc_uniform_mat2(p->sc, "dither_trafo", true, &matrix[0][0]); + + GLSL(dither_pos = dither_trafo * dither_pos;) + } + + GLSL(float dither_value = texture(dither, dither_pos).r;) + GLSLF("color = floor(color * %d.0 + dither_value + 0.5 / %d.0) * 1.0/%d.0;\n", + dither_quantization, dither_size * dither_size, dither_quantization); +} + +// Draws the OSD, in scene-referred colors.. If cms is true, subtitles are +// instead adapted to the display's gamut. +static void pass_draw_osd(struct gl_video *p, int osd_flags, int frame_flags, + double pts, struct mp_osd_res rect, struct ra_fbo fbo, + bool cms) +{ + if (frame_flags & RENDER_FRAME_VF_SUBS) + osd_flags |= OSD_DRAW_SUB_FILTER; + + if ((osd_flags & OSD_DRAW_SUB_ONLY) && (osd_flags & OSD_DRAW_OSD_ONLY)) + return; + + mpgl_osd_generate(p->osd, rect, pts, p->image_params.stereo3d, osd_flags); + + timer_pool_start(p->osd_timer); + for (int n = 0; n < MAX_OSD_PARTS; n++) { + // (This returns false if this part is empty with nothing to draw.) + if (!mpgl_osd_draw_prepare(p->osd, n, p->sc)) + continue; + // When subtitles need to be color managed, assume they're in sRGB + // (for lack of anything saner to do) + if (cms) { + static const struct mp_colorspace csp_srgb = { + .primaries = MP_CSP_PRIM_BT_709, + .gamma = MP_CSP_TRC_SRGB, + .light = MP_CSP_LIGHT_DISPLAY, + }; + + pass_colormanage(p, csp_srgb, fbo.color_space, frame_flags, true); + } + mpgl_osd_draw_finish(p->osd, n, p->sc, fbo); + } + + timer_pool_stop(p->osd_timer); + pass_describe(p, "drawing osd"); + pass_record(p, timer_pool_measure(p->osd_timer)); +} + +static float chroma_realign(int size, int pixel) +{ + return size / (float)chroma_upsize(size, pixel); +} + +// Minimal rendering code path, for GLES or OpenGL 2.1 without proper FBOs. +static void pass_render_frame_dumb(struct gl_video *p) +{ + struct image img[4]; + struct gl_transform off[4]; + pass_get_images(p, &p->image, img, off); + + struct gl_transform transform; + compute_src_transform(p, &transform); + + int index = 0; + for (int i = 0; i < p->plane_count; i++) { + int cw = img[i].type == PLANE_CHROMA ? p->ra_format.chroma_w : 1; + int ch = img[i].type == PLANE_CHROMA ? p->ra_format.chroma_h : 1; + if (p->image_params.rotate % 180 == 90) + MPSWAP(int, cw, ch); + + struct gl_transform t = transform; + t.m[0][0] *= chroma_realign(p->texture_w, cw); + t.m[1][1] *= chroma_realign(p->texture_h, ch); + + t.t[0] /= cw; + t.t[1] /= ch; + + t.t[0] += off[i].t[0]; + t.t[1] += off[i].t[1]; + + gl_transform_trans(img[i].transform, &t); + img[i].transform = t; + + copy_image(p, &index, img[i]); + } + + pass_convert_yuv(p); +} + +// The main rendering function, takes care of everything up to and including +// upscaling. p->image is rendered. +// flags: bit set of RENDER_FRAME_* flags +static bool pass_render_frame(struct gl_video *p, struct mp_image *mpi, + uint64_t id, int flags) +{ + // initialize the texture parameters and temporary variables + p->texture_w = p->image_params.w; + p->texture_h = p->image_params.h; + p->texture_offset = identity_trans; + p->components = 0; + p->num_saved_imgs = 0; + p->idx_hook_textures = 0; + p->use_linear = false; + + // try uploading the frame + if (!pass_upload_image(p, mpi, id)) + return false; + + if (p->image_params.rotate % 180 == 90) + MPSWAP(int, p->texture_w, p->texture_h); + + if (p->dumb_mode) + return true; + + pass_read_video(p); + pass_opt_hook_point(p, "NATIVE", &p->texture_offset); + pass_convert_yuv(p); + pass_opt_hook_point(p, "MAINPRESUB", &p->texture_offset); + + // For subtitles + double vpts = p->image.mpi->pts; + if (vpts == MP_NOPTS_VALUE) + vpts = p->osd_pts; + + if (p->osd && p->opts.blend_subs == BLEND_SUBS_VIDEO && + (flags & RENDER_FRAME_SUBS)) + { + double scale[2]; + get_scale_factors(p, false, scale); + struct mp_osd_res rect = { + .w = p->texture_w, .h = p->texture_h, + .display_par = scale[1] / scale[0], // counter compensate scaling + }; + finish_pass_tex(p, &p->blend_subs_tex, rect.w, rect.h); + struct ra_fbo fbo = { p->blend_subs_tex }; + pass_draw_osd(p, OSD_DRAW_SUB_ONLY, flags, vpts, rect, fbo, false); + pass_read_tex(p, p->blend_subs_tex); + pass_describe(p, "blend subs video"); + } + pass_opt_hook_point(p, "MAIN", &p->texture_offset); + + pass_scale_main(p); + + int vp_w = p->dst_rect.x1 - p->dst_rect.x0, + vp_h = p->dst_rect.y1 - p->dst_rect.y0; + if (p->osd && p->opts.blend_subs == BLEND_SUBS_YES && + (flags & RENDER_FRAME_SUBS)) + { + // Recreate the real video size from the src/dst rects + struct mp_osd_res rect = { + .w = vp_w, .h = vp_h, + .ml = -p->src_rect.x0, .mr = p->src_rect.x1 - p->image_params.w, + .mt = -p->src_rect.y0, .mb = p->src_rect.y1 - p->image_params.h, + .display_par = 1.0, + }; + // Adjust margins for scale + double scale[2]; + get_scale_factors(p, true, scale); + rect.ml *= scale[0]; rect.mr *= scale[0]; + rect.mt *= scale[1]; rect.mb *= scale[1]; + // We should always blend subtitles in non-linear light + if (p->use_linear) { + pass_delinearize(p->sc, p->image_params.color.gamma); + p->use_linear = false; + } + finish_pass_tex(p, &p->blend_subs_tex, p->texture_w, p->texture_h); + struct ra_fbo fbo = { p->blend_subs_tex }; + pass_draw_osd(p, OSD_DRAW_SUB_ONLY, flags, vpts, rect, fbo, false); + pass_read_tex(p, p->blend_subs_tex); + pass_describe(p, "blend subs"); + } + + pass_opt_hook_point(p, "SCALED", NULL); + + return true; +} + +static void pass_draw_to_screen(struct gl_video *p, struct ra_fbo fbo, int flags) +{ + if (p->dumb_mode) + pass_render_frame_dumb(p); + + // Adjust the overall gamma before drawing to screen + if (p->user_gamma != 1) { + gl_sc_uniform_f(p->sc, "user_gamma", p->user_gamma); + GLSL(color.rgb = clamp(color.rgb, 0.0, 1.0);) + GLSL(color.rgb = pow(color.rgb, vec3(user_gamma));) + } + + pass_colormanage(p, p->image_params.color, fbo.color_space, flags, false); + + // Since finish_pass_fbo doesn't work with compute shaders, and neither + // does the checkerboard/dither code, we may need an indirection via + // p->screen_tex here. + if (p->pass_compute.active) { + int o_w = p->dst_rect.x1 - p->dst_rect.x0, + o_h = p->dst_rect.y1 - p->dst_rect.y0; + finish_pass_tex(p, &p->screen_tex, o_w, o_h); + struct image tmp = image_wrap(p->screen_tex, PLANE_RGB, p->components); + copy_image(p, &(int){0}, tmp); + } + + if (p->has_alpha){ + if (p->opts.alpha_mode == ALPHA_BLEND_TILES) { + // Draw checkerboard pattern to indicate transparency + GLSLF("// transparency checkerboard\n"); + GLSL(bvec2 tile = lessThan(fract(gl_FragCoord.xy * 1.0/32.0), vec2(0.5));) + GLSL(vec3 background = vec3(tile.x == tile.y ? 0.93 : 0.87);) + GLSL(color.rgb += background.rgb * (1.0 - color.a);) + GLSL(color.a = 1.0;) + } else if (p->opts.alpha_mode == ALPHA_BLEND) { + // Blend into background color (usually black) + struct m_color c = p->opts.background; + GLSLF("vec4 background = vec4(%f, %f, %f, %f);\n", + c.r / 255.0, c.g / 255.0, c.b / 255.0, c.a / 255.0); + GLSL(color.rgb += background.rgb * (1.0 - color.a);) + GLSL(color.a = background.a;) + } + } + + pass_opt_hook_point(p, "OUTPUT", NULL); + + if (flags & RENDER_SCREEN_COLOR) + pass_dither(p); + pass_describe(p, "output to screen"); + finish_pass_fbo(p, fbo, false, &p->dst_rect); +} + +// flags: bit set of RENDER_FRAME_* flags +static bool update_surface(struct gl_video *p, struct mp_image *mpi, + uint64_t id, struct surface *surf, int flags) +{ + int vp_w = p->dst_rect.x1 - p->dst_rect.x0, + vp_h = p->dst_rect.y1 - p->dst_rect.y0; + + pass_info_reset(p, false); + if (!pass_render_frame(p, mpi, id, flags)) + return false; + + // Frame blending should always be done in linear light to preserve the + // overall brightness, otherwise this will result in flashing dark frames + // because mixing in compressed light artificially darkens the results + if (!p->use_linear) { + p->use_linear = true; + pass_linearize(p->sc, p->image_params.color.gamma); + } + + finish_pass_tex(p, &surf->tex, vp_w, vp_h); + surf->id = id; + surf->pts = mpi->pts; + return true; +} + +// Draws an interpolate frame to fbo, based on the frame timing in t +// flags: bit set of RENDER_FRAME_* flags +static void gl_video_interpolate_frame(struct gl_video *p, struct vo_frame *t, + struct ra_fbo fbo, int flags) +{ + bool is_new = false; + + // Reset the queue completely if this is a still image, to avoid any + // interpolation artifacts from surrounding frames when unpausing or + // framestepping + if (t->still) + gl_video_reset_surfaces(p); + + // First of all, figure out if we have a frame available at all, and draw + // it manually + reset the queue if not + if (p->surfaces[p->surface_now].id == 0) { + struct surface *now = &p->surfaces[p->surface_now]; + if (!update_surface(p, t->current, t->frame_id, now, flags)) + return; + p->surface_idx = p->surface_now; + is_new = true; + } + + // Find the right frame for this instant + if (t->current) { + int next = surface_wrap(p->surface_now + 1); + while (p->surfaces[next].id && + p->surfaces[next].id > p->surfaces[p->surface_now].id && + p->surfaces[p->surface_now].id < t->frame_id) + { + p->surface_now = next; + next = surface_wrap(next + 1); + } + } + + // Figure out the queue size. For illustration, a filter radius of 2 would + // look like this: _ A [B] C D _ + // A is surface_bse, B is surface_now, C is surface_now+1 and D is + // surface_end. + struct scaler *tscale = &p->scaler[SCALER_TSCALE]; + reinit_scaler(p, tscale, &p->opts.scaler[SCALER_TSCALE], 1, tscale_sizes); + bool oversample = strcmp(tscale->conf.kernel.name, "oversample") == 0; + bool linear = strcmp(tscale->conf.kernel.name, "linear") == 0; + int size; + + if (oversample || linear) { + size = 2; + } else { + assert(tscale->kernel && !tscale->kernel->polar); + size = ceil(tscale->kernel->size); + } + + int radius = size/2; + int surface_now = p->surface_now; + int surface_bse = surface_wrap(surface_now - (radius-1)); + int surface_end = surface_wrap(surface_now + radius); + assert(surface_wrap(surface_bse + size-1) == surface_end); + + // Render new frames while there's room in the queue. Note that technically, + // this should be done before the step where we find the right frame, but + // it only barely matters at the very beginning of playback, and this way + // makes the code much more linear. + int surface_dst = surface_wrap(p->surface_idx + 1); + for (int i = 0; i < t->num_frames; i++) { + // Avoid overwriting data we might still need + if (surface_dst == surface_bse - 1) + break; + + struct mp_image *f = t->frames[i]; + uint64_t f_id = t->frame_id + i; + if (!mp_image_params_equal(&f->params, &p->real_image_params)) + continue; + + if (f_id > p->surfaces[p->surface_idx].id) { + struct surface *dst = &p->surfaces[surface_dst]; + if (!update_surface(p, f, f_id, dst, flags)) + return; + p->surface_idx = surface_dst; + surface_dst = surface_wrap(surface_dst + 1); + is_new = true; + } + } + + // Figure out whether the queue is "valid". A queue is invalid if the + // frames' PTS is not monotonically increasing. Anything else is invalid, + // so avoid blending incorrect data and just draw the latest frame as-is. + // Possible causes for failure of this condition include seeks, pausing, + // end of playback or start of playback. + bool valid = true; + for (int i = surface_bse, ii; valid && i != surface_end; i = ii) { + ii = surface_wrap(i + 1); + if (p->surfaces[i].id == 0 || p->surfaces[ii].id == 0) { + valid = false; + } else if (p->surfaces[ii].id < p->surfaces[i].id) { + valid = false; + MP_DBG(p, "interpolation queue underrun\n"); + } + } + + // Update OSD PTS to synchronize subtitles with the displayed frame + p->osd_pts = p->surfaces[surface_now].pts; + + // Finally, draw the right mix of frames to the screen. + if (!is_new) + pass_info_reset(p, true); + pass_describe(p, "interpolation"); + if (!valid || t->still) { + // surface_now is guaranteed to be valid, so we can safely use it. + pass_read_tex(p, p->surfaces[surface_now].tex); + p->is_interpolated = false; + } else { + double mix = t->vsync_offset / t->ideal_frame_duration; + // The scaler code always wants the fcoord to be between 0 and 1, + // so we try to adjust by using the previous set of N frames instead + // (which requires some extra checking to make sure it's valid) + if (mix < 0.0) { + int prev = surface_wrap(surface_bse - 1); + if (p->surfaces[prev].id != 0 && + p->surfaces[prev].id < p->surfaces[surface_bse].id) + { + mix += 1.0; + surface_bse = prev; + } else { + mix = 0.0; // at least don't blow up, this should only + // ever happen at the start of playback + } + } + + if (oversample) { + // Oversample uses the frame area as mix ratio, not the vsync + // position itself + double vsync_dist = t->vsync_interval / t->ideal_frame_duration, + threshold = tscale->conf.kernel.params[0]; + threshold = isnan(threshold) ? 0.0 : threshold; + mix = (1 - mix) / vsync_dist; + mix = mix <= 0 + threshold ? 0 : mix; + mix = mix >= 1 - threshold ? 1 : mix; + mix = 1 - mix; + } + + // Blend the frames together + if (oversample || linear) { + gl_sc_uniform_dynamic(p->sc); + gl_sc_uniform_f(p->sc, "inter_coeff", mix); + GLSL(color = mix(texture(texture0, texcoord0), + texture(texture1, texcoord1), + inter_coeff);) + } else { + gl_sc_uniform_dynamic(p->sc); + gl_sc_uniform_f(p->sc, "fcoord", mix); + pass_sample_separated_gen(p->sc, tscale, 0, 0); + } + + // Load all the required frames + for (int i = 0; i < size; i++) { + struct image img = + image_wrap(p->surfaces[surface_wrap(surface_bse+i)].tex, + PLANE_RGB, p->components); + // Since the code in pass_sample_separated currently assumes + // the textures are bound in-order and starting at 0, we just + // assert to make sure this is the case (which it should always be) + int id = pass_bind(p, img); + assert(id == i); + } + + MP_TRACE(p, "inter frame dur: %f vsync: %f, mix: %f\n", + t->ideal_frame_duration, t->vsync_interval, mix); + p->is_interpolated = true; + } + pass_draw_to_screen(p, fbo, flags); + + p->frames_drawn += 1; +} + +void gl_video_render_frame(struct gl_video *p, struct vo_frame *frame, + struct ra_fbo fbo, int flags) +{ + gl_video_update_options(p); + + struct mp_rect target_rc = {0, 0, fbo.tex->params.w, fbo.tex->params.h}; + + p->broken_frame = false; + + bool has_frame = !!frame->current; + + struct m_color c = p->clear_color; + float clear_color[4] = {c.r / 255.0, c.g / 255.0, c.b / 255.0, c.a / 255.0}; + p->ra->fns->clear(p->ra, fbo.tex, clear_color, &target_rc); + + if (p->hwdec_overlay) { + if (has_frame) { + float *color = p->hwdec_overlay->overlay_colorkey; + p->ra->fns->clear(p->ra, fbo.tex, color, &p->dst_rect); + } + + p->hwdec_overlay->driver->overlay_frame(p->hwdec_overlay, frame->current, + &p->src_rect, &p->dst_rect, + frame->frame_id != p->image.id); + + if (frame->current) + p->osd_pts = frame->current->pts; + + // Disable GL rendering + has_frame = false; + } + + if (has_frame) { + bool interpolate = p->opts.interpolation && frame->display_synced && + (p->frames_drawn || !frame->still); + if (interpolate) { + double ratio = frame->ideal_frame_duration / frame->vsync_interval; + if (fabs(ratio - 1.0) < p->opts.interpolation_threshold) + interpolate = false; + } + + if (interpolate) { + gl_video_interpolate_frame(p, frame, fbo, flags); + } else { + bool is_new = frame->frame_id != p->image.id; + + // Redrawing a frame might update subtitles. + if (frame->still && p->opts.blend_subs) + is_new = true; + + if (is_new || !p->output_tex_valid) { + p->output_tex_valid = false; + + pass_info_reset(p, !is_new); + if (!pass_render_frame(p, frame->current, frame->frame_id, flags)) + goto done; + + // For the non-interpolation case, we draw to a single "cache" + // texture to speed up subsequent re-draws (if any exist) + struct ra_fbo dest_fbo = fbo; + bool repeats = frame->num_vsyncs > 1 && frame->display_synced; + if ((repeats || frame->still) && !p->dumb_mode && + (p->ra->caps & RA_CAP_BLIT) && fbo.tex->params.blit_dst) + { + // Attempt to use the same format as the destination FBO + // if possible. Some RAs use a wrapped dummy format here, + // so fall back to the fbo_format in that case. + const struct ra_format *fmt = fbo.tex->params.format; + if (fmt->dummy_format) + fmt = p->fbo_format; + + bool r = ra_tex_resize(p->ra, p->log, &p->output_tex, + fbo.tex->params.w, fbo.tex->params.h, + fmt); + if (r) { + dest_fbo = (struct ra_fbo) { p->output_tex }; + p->output_tex_valid = true; + } + } + pass_draw_to_screen(p, dest_fbo, flags); + } + + // "output tex valid" and "output tex needed" are equivalent + if (p->output_tex_valid && fbo.tex->params.blit_dst) { + pass_info_reset(p, true); + pass_describe(p, "redraw cached frame"); + struct mp_rect src = p->dst_rect; + struct mp_rect dst = src; + if (fbo.flip) { + dst.y0 = fbo.tex->params.h - src.y0; + dst.y1 = fbo.tex->params.h - src.y1; + } + timer_pool_start(p->blit_timer); + p->ra->fns->blit(p->ra, fbo.tex, p->output_tex, &dst, &src); + timer_pool_stop(p->blit_timer); + pass_record(p, timer_pool_measure(p->blit_timer)); + } + } + } + +done: + + debug_check_gl(p, "after video rendering"); + + if (p->osd && (flags & (RENDER_FRAME_SUBS | RENDER_FRAME_OSD))) { + // If we haven't actually drawn anything so far, then we technically + // need to consider this the start of a new pass. Let's call it a + // redraw just because, since it's basically a blank frame anyway + if (!has_frame) + pass_info_reset(p, true); + + int osd_flags = p->opts.blend_subs ? OSD_DRAW_OSD_ONLY : 0; + if (!(flags & RENDER_FRAME_SUBS)) + osd_flags |= OSD_DRAW_OSD_ONLY; + if (!(flags & RENDER_FRAME_OSD)) + osd_flags |= OSD_DRAW_SUB_ONLY; + + pass_draw_osd(p, osd_flags, flags, p->osd_pts, p->osd_rect, fbo, true); + debug_check_gl(p, "after OSD rendering"); + } + + p->broken_frame |= gl_sc_error_state(p->sc); + if (p->broken_frame) { + // Make the screen solid blue to make it visually clear that an + // error has occurred + float color[4] = {0.0, 0.05, 0.5, 1.0}; + p->ra->fns->clear(p->ra, fbo.tex, color, &target_rc); + } + + p->frames_rendered++; + pass_report_performance(p); +} + +void gl_video_screenshot(struct gl_video *p, struct vo_frame *frame, + struct voctrl_screenshot *args) +{ + if (!p->ra->fns->tex_download) + return; + + bool ok = false; + struct mp_image *res = NULL; + struct ra_tex *target = NULL; + struct mp_rect old_src = p->src_rect; + struct mp_rect old_dst = p->dst_rect; + struct mp_osd_res old_osd = p->osd_rect; + struct vo_frame *nframe = vo_frame_ref(frame); + + // Disable interpolation and such. + nframe->redraw = true; + nframe->repeat = false; + nframe->still = true; + nframe->pts = 0; + nframe->duration = -1; + + if (!args->scaled) { + int w, h; + mp_image_params_get_dsize(&p->image_params, &w, &h); + if (w < 1 || h < 1) + return; + + int src_w = p->image_params.w; + int src_h = p->image_params.h; + struct mp_rect src = {0, 0, src_w, src_h}; + struct mp_rect dst = {0, 0, w, h}; + + if (mp_image_crop_valid(&p->image_params)) + src = p->image_params.crop; + + if (p->image_params.rotate % 180 == 90) { + MPSWAP(int, w, h); + MPSWAP(int, src_w, src_h); + } + mp_rect_rotate(&src, src_w, src_h, p->image_params.rotate); + mp_rect_rotate(&dst, w, h, p->image_params.rotate); + + struct mp_osd_res osd = { + .display_par = 1.0, + .w = mp_rect_w(dst), + .h = mp_rect_h(dst), + }; + gl_video_resize(p, &src, &dst, &osd); + } + + gl_video_reset_surfaces(p); + + struct ra_tex_params params = { + .dimensions = 2, + .downloadable = true, + .w = p->osd_rect.w, + .h = p->osd_rect.h, + .d = 1, + .render_dst = true, + }; + + params.format = ra_find_unorm_format(p->ra, 1, 4); + int mpfmt = IMGFMT_RGB0; + if (args->high_bit_depth && p->ra_format.component_bits > 8) { + const struct ra_format *fmt = ra_find_unorm_format(p->ra, 2, 4); + if (fmt && fmt->renderable) { + params.format = fmt; + mpfmt = IMGFMT_RGBA64; + } + } + + if (!params.format || !params.format->renderable) + goto done; + target = ra_tex_create(p->ra, ¶ms); + if (!target) + goto done; + + int flags = 0; + if (args->subs) + flags |= RENDER_FRAME_SUBS; + if (args->osd) + flags |= RENDER_FRAME_OSD; + if (args->scaled) + flags |= RENDER_SCREEN_COLOR; + gl_video_render_frame(p, nframe, (struct ra_fbo){target}, flags); + + res = mp_image_alloc(mpfmt, params.w, params.h); + if (!res) + goto done; + + struct ra_tex_download_params download_params = { + .tex = target, + .dst = res->planes[0], + .stride = res->stride[0], + }; + if (!p->ra->fns->tex_download(p->ra, &download_params)) + goto done; + + if (p->broken_frame) + goto done; + + ok = true; +done: + talloc_free(nframe); + ra_tex_free(p->ra, &target); + gl_video_resize(p, &old_src, &old_dst, &old_osd); + gl_video_reset_surfaces(p); + if (!ok) + TA_FREEP(&res); + args->res = res; +} + +// Use this color instead of the global option. +void gl_video_set_clear_color(struct gl_video *p, struct m_color c) +{ + p->force_clear_color = true; + p->clear_color = c; +} + +void gl_video_set_osd_pts(struct gl_video *p, double pts) +{ + p->osd_pts = pts; +} + +bool gl_video_check_osd_change(struct gl_video *p, struct mp_osd_res *res, + double pts) +{ + return p->osd ? mpgl_osd_check_change(p->osd, res, pts) : false; +} + +void gl_video_resize(struct gl_video *p, + struct mp_rect *src, struct mp_rect *dst, + struct mp_osd_res *osd) +{ + if (mp_rect_equals(&p->src_rect, src) && + mp_rect_equals(&p->dst_rect, dst) && + osd_res_equals(p->osd_rect, *osd)) + return; + + p->src_rect = *src; + p->dst_rect = *dst; + p->osd_rect = *osd; + + gl_video_reset_surfaces(p); + + if (p->osd) + mpgl_osd_resize(p->osd, p->osd_rect, p->image_params.stereo3d); +} + +static void frame_perf_data(struct pass_info pass[], struct mp_frame_perf *out) +{ + for (int i = 0; i < VO_PASS_PERF_MAX; i++) { + if (!pass[i].desc.len) + break; + out->perf[out->count] = pass[i].perf; + strncpy(out->desc[out->count], pass[i].desc.start, + sizeof(out->desc[out->count]) - 1); + out->desc[out->count][sizeof(out->desc[out->count]) - 1] = '\0'; + out->count++; + } +} + +void gl_video_perfdata(struct gl_video *p, struct voctrl_performance_data *out) +{ + *out = (struct voctrl_performance_data){0}; + frame_perf_data(p->pass_fresh, &out->fresh); + frame_perf_data(p->pass_redraw, &out->redraw); +} + +// Returns false on failure. +static bool pass_upload_image(struct gl_video *p, struct mp_image *mpi, uint64_t id) +{ + struct video_image *vimg = &p->image; + + if (vimg->id == id) + return true; + + unref_current_image(p); + + mpi = mp_image_new_ref(mpi); + if (!mpi) + goto error; + + vimg->mpi = mpi; + vimg->id = id; + p->osd_pts = mpi->pts; + p->frames_uploaded++; + + if (p->hwdec_active) { + // Hardware decoding + + if (!p->hwdec_mapper) + goto error; + + pass_describe(p, "map frame (hwdec)"); + timer_pool_start(p->upload_timer); + bool ok = ra_hwdec_mapper_map(p->hwdec_mapper, vimg->mpi) >= 0; + timer_pool_stop(p->upload_timer); + pass_record(p, timer_pool_measure(p->upload_timer)); + + vimg->hwdec_mapped = true; + if (ok) { + struct mp_image layout = {0}; + mp_image_set_params(&layout, &p->image_params); + struct ra_tex **tex = p->hwdec_mapper->tex; + for (int n = 0; n < p->plane_count; n++) { + vimg->planes[n] = (struct texplane){ + .w = mp_image_plane_w(&layout, n), + .h = mp_image_plane_h(&layout, n), + .tex = tex[n], + }; + } + } else { + MP_FATAL(p, "Mapping hardware decoded surface failed.\n"); + goto error; + } + return true; + } + + // Software decoding + assert(mpi->num_planes == p->plane_count); + + timer_pool_start(p->upload_timer); + for (int n = 0; n < p->plane_count; n++) { + struct texplane *plane = &vimg->planes[n]; + if (!plane->tex) { + timer_pool_stop(p->upload_timer); + goto error; + } + + struct ra_tex_upload_params params = { + .tex = plane->tex, + .src = mpi->planes[n], + .invalidate = true, + .stride = mpi->stride[n], + }; + + plane->flipped = params.stride < 0; + if (plane->flipped) { + int h = mp_image_plane_h(mpi, n); + params.src = (char *)params.src + (h - 1) * params.stride; + params.stride = -params.stride; + } + + struct dr_buffer *mapped = gl_find_dr_buffer(p, mpi->planes[n]); + if (mapped) { + params.buf = mapped->buf; + params.buf_offset = (uintptr_t)params.src - + (uintptr_t)mapped->buf->data; + params.src = NULL; + } + + if (p->using_dr_path != !!mapped) { + p->using_dr_path = !!mapped; + MP_VERBOSE(p, "DR enabled: %s\n", p->using_dr_path ? "yes" : "no"); + } + + if (!p->ra->fns->tex_upload(p->ra, ¶ms)) { + timer_pool_stop(p->upload_timer); + goto error; + } + + if (mapped && !mapped->mpi) + mapped->mpi = mp_image_new_ref(mpi); + } + timer_pool_stop(p->upload_timer); + + bool using_pbo = p->ra->use_pbo || !(p->ra->caps & RA_CAP_DIRECT_UPLOAD); + const char *mode = p->using_dr_path ? "DR" : using_pbo ? "PBO" : "naive"; + pass_describe(p, "upload frame (%s)", mode); + pass_record(p, timer_pool_measure(p->upload_timer)); + + return true; + +error: + unref_current_image(p); + p->broken_frame = true; + return false; +} + +static bool test_fbo(struct gl_video *p, const struct ra_format *fmt) +{ + MP_VERBOSE(p, "Testing FBO format %s\n", fmt->name); + struct ra_tex *tex = NULL; + bool success = ra_tex_resize(p->ra, p->log, &tex, 16, 16, fmt); + ra_tex_free(p->ra, &tex); + return success; +} + +// Return whether dumb-mode can be used without disabling any features. +// Essentially, vo_gpu with mostly default settings will return true. +static bool check_dumb_mode(struct gl_video *p) +{ + struct gl_video_opts *o = &p->opts; + if (p->use_integer_conversion) + return false; + if (o->dumb_mode > 0) // requested by user + return true; + if (o->dumb_mode < 0) // disabled by user + return false; + + // otherwise, use auto-detection + if (o->correct_downscaling || o->linear_downscaling || + o->linear_upscaling || o->sigmoid_upscaling || o->interpolation || + o->blend_subs || o->deband || o->unsharp) + return false; + // check remaining scalers (tscale is already implicitly excluded above) + for (int i = 0; i < SCALER_COUNT; i++) { + if (i != SCALER_TSCALE) { + const char *name = o->scaler[i].kernel.name; + if (name && strcmp(name, "bilinear") != 0) + return false; + } + } + if (o->user_shaders && o->user_shaders[0]) + return false; + return true; +} + +// Disable features that are not supported with the current OpenGL version. +static void check_gl_features(struct gl_video *p) +{ + struct ra *ra = p->ra; + bool have_float_tex = !!ra_find_float16_format(ra, 1); + bool have_mglsl = ra->glsl_version >= 130; // modern GLSL + const struct ra_format *rg_tex = ra_find_unorm_format(p->ra, 1, 2); + bool have_texrg = rg_tex && !rg_tex->luminance_alpha; + bool have_compute = ra->caps & RA_CAP_COMPUTE; + bool have_ssbo = ra->caps & RA_CAP_BUF_RW; + bool have_fragcoord = ra->caps & RA_CAP_FRAGCOORD; + + const char *auto_fbo_fmts[] = {"rgba16f", "rgba16hf", "rgba16", + "rgb10_a2", "rgba8", 0}; + const char *user_fbo_fmts[] = {p->opts.fbo_format, 0}; + const char **fbo_fmts = user_fbo_fmts[0] && strcmp(user_fbo_fmts[0], "auto") + ? user_fbo_fmts : auto_fbo_fmts; + bool user_specified_fbo_fmt = fbo_fmts == user_fbo_fmts; + bool fbo_test_result = false; + bool have_fbo = false; + p->fbo_format = NULL; + for (int n = 0; fbo_fmts[n]; n++) { + const char *fmt = fbo_fmts[n]; + const struct ra_format *f = ra_find_named_format(p->ra, fmt); + if (!f && user_specified_fbo_fmt) + MP_WARN(p, "FBO format '%s' not found!\n", fmt); + if (f && f->renderable && f->linear_filter && + (fbo_test_result = test_fbo(p, f))) { + MP_VERBOSE(p, "Using FBO format %s.\n", f->name); + have_fbo = true; + p->fbo_format = f; + break; + } + + if (user_specified_fbo_fmt) { + MP_WARN(p, "User-specified FBO format '%s' failed to initialize! " + "(exists=%d, renderable=%d, linear_filter=%d, " + "fbo_test_result=%d)\n", + fmt, !!f, f ? f->renderable : 0, f ? f->linear_filter : 0, + fbo_test_result); + } + } + + if (!have_fragcoord && p->opts.dither_depth >= 0 && + p->opts.dither_algo != DITHER_NONE) + { + p->opts.dither_algo = DITHER_NONE; + MP_WARN(p, "Disabling dithering (no gl_FragCoord).\n"); + } + if (!have_fragcoord && p->opts.alpha_mode == ALPHA_BLEND_TILES) { + p->opts.alpha_mode = ALPHA_BLEND; + // Verbose, since this is the default setting + MP_VERBOSE(p, "Disabling alpha checkerboard (no gl_FragCoord).\n"); + } + if (!have_fbo && have_compute) { + have_compute = false; + MP_WARN(p, "Force-disabling compute shaders as an FBO format was not " + "available! See your FBO format configuration!\n"); + } + + if (have_compute && have_fbo && !p->fbo_format->storable) { + have_compute = false; + MP_WARN(p, "Force-disabling compute shaders as the chosen FBO format " + "is not storable! See your FBO format configuration!\n"); + } + + if (!have_compute && p->opts.dither_algo == DITHER_ERROR_DIFFUSION) { + MP_WARN(p, "Disabling error diffusion dithering because compute shader " + "was not supported. Fallback to dither=fruit instead.\n"); + p->opts.dither_algo = DITHER_FRUIT; + } + + bool have_compute_peak = have_compute && have_ssbo; + if (!have_compute_peak && p->opts.tone_map.compute_peak >= 0) { + int msgl = p->opts.tone_map.compute_peak == 1 ? MSGL_WARN : MSGL_V; + MP_MSG(p, msgl, "Disabling HDR peak computation (one or more of the " + "following is not supported: compute shaders=%d, " + "SSBO=%d).\n", have_compute, have_ssbo); + p->opts.tone_map.compute_peak = -1; + } + + p->forced_dumb_mode = p->opts.dumb_mode > 0 || !have_fbo || !have_texrg; + bool voluntarily_dumb = check_dumb_mode(p); + if (p->forced_dumb_mode || voluntarily_dumb) { + if (voluntarily_dumb) { + MP_VERBOSE(p, "No advanced processing required. Enabling dumb mode.\n"); + } else if (p->opts.dumb_mode <= 0) { + MP_WARN(p, "High bit depth FBOs unsupported. Enabling dumb mode.\n" + "Most extended features will be disabled.\n"); + } + p->dumb_mode = true; + static const struct scaler_config dumb_scaler_config = { + {"bilinear", .params = {NAN, NAN}}, + {.params = {NAN, NAN}}, + }; + // Most things don't work, so whitelist all options that still work. + p->opts = (struct gl_video_opts){ + .scaler = { + [SCALER_SCALE] = dumb_scaler_config, + [SCALER_DSCALE] = dumb_scaler_config, + [SCALER_CSCALE] = dumb_scaler_config, + [SCALER_TSCALE] = dumb_scaler_config, + }, + .gamma = p->opts.gamma, + .gamma_auto = p->opts.gamma_auto, + .pbo = p->opts.pbo, + .fbo_format = p->opts.fbo_format, + .alpha_mode = p->opts.alpha_mode, + .use_rectangle = p->opts.use_rectangle, + .background = p->opts.background, + .dither_algo = p->opts.dither_algo, + .dither_depth = p->opts.dither_depth, + .dither_size = p->opts.dither_size, + .error_diffusion = p->opts.error_diffusion, + .temporal_dither = p->opts.temporal_dither, + .temporal_dither_period = p->opts.temporal_dither_period, + .tex_pad_x = p->opts.tex_pad_x, + .tex_pad_y = p->opts.tex_pad_y, + .tone_map = p->opts.tone_map, + .early_flush = p->opts.early_flush, + .icc_opts = p->opts.icc_opts, + .hwdec_interop = p->opts.hwdec_interop, + .target_trc = p->opts.target_trc, + .target_prim = p->opts.target_prim, + .target_peak = p->opts.target_peak, + }; + if (!have_fbo) + p->use_lut_3d = false; + return; + } + p->dumb_mode = false; + + // Normally, we want to disable them by default if FBOs are unavailable, + // because they will be slow (not critically slow, but still slower). + // Without FP textures, we must always disable them. + // I don't know if luminance alpha float textures exist, so disregard them. + for (int n = 0; n < SCALER_COUNT; n++) { + const struct filter_kernel *kernel = + mp_find_filter_kernel(p->opts.scaler[n].kernel.name); + if (kernel) { + char *reason = NULL; + if (!have_float_tex) + reason = "(float tex. missing)"; + if (!have_mglsl) + reason = "(GLSL version too old)"; + if (reason) { + MP_WARN(p, "Disabling scaler #%d %s %s.\n", n, + p->opts.scaler[n].kernel.name, reason); + // p->opts is a copy => we can just mess with it. + p->opts.scaler[n].kernel.name = "bilinear"; + if (n == SCALER_TSCALE) + p->opts.interpolation = false; + } + } + } + + int use_cms = p->opts.target_prim != MP_CSP_PRIM_AUTO || + p->opts.target_trc != MP_CSP_TRC_AUTO || p->use_lut_3d; + + // mix() is needed for some gamma functions + if (!have_mglsl && (p->opts.linear_downscaling || + p->opts.linear_upscaling || p->opts.sigmoid_upscaling)) + { + p->opts.linear_downscaling = false; + p->opts.linear_upscaling = false; + p->opts.sigmoid_upscaling = false; + MP_WARN(p, "Disabling linear/sigmoid scaling (GLSL version too old).\n"); + } + if (!have_mglsl && use_cms) { + p->opts.target_prim = MP_CSP_PRIM_AUTO; + p->opts.target_trc = MP_CSP_TRC_AUTO; + p->use_lut_3d = false; + MP_WARN(p, "Disabling color management (GLSL version too old).\n"); + } + if (!have_mglsl && p->opts.deband) { + p->opts.deband = false; + MP_WARN(p, "Disabling debanding (GLSL version too old).\n"); + } +} + +static void init_gl(struct gl_video *p) +{ + debug_check_gl(p, "before init_gl"); + + p->upload_timer = timer_pool_create(p->ra); + p->blit_timer = timer_pool_create(p->ra); + p->osd_timer = timer_pool_create(p->ra); + + debug_check_gl(p, "after init_gl"); + + ra_dump_tex_formats(p->ra, MSGL_DEBUG); + ra_dump_img_formats(p->ra, MSGL_DEBUG); +} + +void gl_video_uninit(struct gl_video *p) +{ + if (!p) + return; + + uninit_video(p); + ra_hwdec_ctx_uninit(&p->hwdec_ctx); + gl_sc_destroy(p->sc); + + ra_tex_free(p->ra, &p->lut_3d_texture); + ra_buf_free(p->ra, &p->hdr_peak_ssbo); + + timer_pool_destroy(p->upload_timer); + timer_pool_destroy(p->blit_timer); + timer_pool_destroy(p->osd_timer); + + for (int i = 0; i < VO_PASS_PERF_MAX; i++) { + talloc_free(p->pass_fresh[i].desc.start); + talloc_free(p->pass_redraw[i].desc.start); + } + + mpgl_osd_destroy(p->osd); + + // Forcibly destroy possibly remaining image references. This should also + // cause gl_video_dr_free_buffer() to be called for the remaining buffers. + gc_pending_dr_fences(p, true); + + // Should all have been unreffed already. + assert(!p->num_dr_buffers); + + talloc_free(p); +} + +void gl_video_reset(struct gl_video *p) +{ + gl_video_reset_surfaces(p); +} + +bool gl_video_showing_interpolated_frame(struct gl_video *p) +{ + return p->is_interpolated; +} + +static bool is_imgfmt_desc_supported(struct gl_video *p, + const struct ra_imgfmt_desc *desc) +{ + if (!desc->num_planes) + return false; + + if (desc->planes[0]->ctype == RA_CTYPE_UINT && p->forced_dumb_mode) + return false; + + return true; +} + +bool gl_video_check_format(struct gl_video *p, int mp_format) +{ + struct ra_imgfmt_desc desc; + if (ra_get_imgfmt_desc(p->ra, mp_format, &desc) && + is_imgfmt_desc_supported(p, &desc)) + return true; + if (ra_hwdec_get(&p->hwdec_ctx, mp_format)) + return true; + return false; +} + +void gl_video_config(struct gl_video *p, struct mp_image_params *params) +{ + unmap_overlay(p); + unref_current_image(p); + + if (!mp_image_params_equal(&p->real_image_params, params)) { + uninit_video(p); + p->real_image_params = *params; + p->image_params = *params; + if (params->imgfmt) + init_video(p); + } + + gl_video_reset_surfaces(p); +} + +void gl_video_set_osd_source(struct gl_video *p, struct osd_state *osd) +{ + mpgl_osd_destroy(p->osd); + p->osd = NULL; + p->osd_state = osd; + reinit_osd(p); +} + +struct gl_video *gl_video_init(struct ra *ra, struct mp_log *log, + struct mpv_global *g) +{ + struct gl_video *p = talloc_ptrtype(NULL, p); + *p = (struct gl_video) { + .ra = ra, + .global = g, + .log = log, + .sc = gl_sc_create(ra, g, log), + .video_eq = mp_csp_equalizer_create(p, g), + .opts_cache = m_config_cache_alloc(p, g, &gl_video_conf), + }; + // make sure this variable is initialized to *something* + p->pass = p->pass_fresh; + struct gl_video_opts *opts = p->opts_cache->opts; + p->cms = gl_lcms_init(p, log, g, opts->icc_opts), + p->opts = *opts; + for (int n = 0; n < SCALER_COUNT; n++) + p->scaler[n] = (struct scaler){.index = n}; + // our VAO always has the vec2 position as the first element + MP_TARRAY_APPEND(p, p->vao, p->vao_len, (struct ra_renderpass_input) { + .name = "position", + .type = RA_VARTYPE_FLOAT, + .dim_v = 2, + .dim_m = 1, + .offset = 0, + }); + init_gl(p); + reinit_from_options(p); + return p; +} + +// Get static string for scaler shader. If "tscale" is set to true, the +// scaler must be a separable convolution filter. +static const char *handle_scaler_opt(const char *name, bool tscale) +{ + if (name && name[0]) { + const struct filter_kernel *kernel = mp_find_filter_kernel(name); + if (kernel && (!tscale || !kernel->polar)) + return kernel->f.name; + + const struct filter_window *window = mp_find_filter_window(name); + if (window) + return window->name; + + for (const char *const *filter = tscale ? fixed_tscale_filters + : fixed_scale_filters; + *filter; filter++) { + if (strcmp(*filter, name) == 0) + return *filter; + } + } + return NULL; +} + +static void gl_video_update_options(struct gl_video *p) +{ + if (m_config_cache_update(p->opts_cache)) { + gl_lcms_update_options(p->cms); + reinit_from_options(p); + } + + if (mp_csp_equalizer_state_changed(p->video_eq)) + p->output_tex_valid = false; +} + +static void reinit_from_options(struct gl_video *p) +{ + p->use_lut_3d = gl_lcms_has_profile(p->cms); + + // Copy the option fields, so that check_gl_features() can mutate them. + // This works only for the fields themselves of course, not for any memory + // referenced by them. + p->opts = *(struct gl_video_opts *)p->opts_cache->opts; + + if (!p->force_clear_color) + p->clear_color = p->opts.background; + + check_gl_features(p); + uninit_rendering(p); + if (p->opts.shader_cache) + gl_sc_set_cache_dir(p->sc, p->opts.shader_cache_dir); + p->ra->use_pbo = p->opts.pbo; + gl_video_setup_hooks(p); + reinit_osd(p); + + struct mp_vo_opts *vo_opts = mp_get_config_group(p, p->global, &vo_sub_opts); + if (p->opts.interpolation && !vo_opts->video_sync && !p->dsi_warned) { + MP_WARN(p, "Interpolation now requires enabling display-sync mode.\n" + "E.g.: --video-sync=display-resample\n"); + p->dsi_warned = true; + } + talloc_free(vo_opts); + + if (p->opts.correct_downscaling && !p->correct_downscaling_warned) { + const char *name = p->opts.scaler[SCALER_DSCALE].kernel.name; + if (!name) + name = p->opts.scaler[SCALER_SCALE].kernel.name; + if (!name || !strcmp(name, "bilinear")) { + MP_WARN(p, "correct-downscaling requires non-bilinear scaler.\n"); + p->correct_downscaling_warned = true; + } + } +} + +void gl_video_configure_queue(struct gl_video *p, struct vo *vo) +{ + gl_video_update_options(p); + + int queue_size = 1; + + // Figure out an adequate size for the interpolation queue. The larger + // the radius, the earlier we need to queue frames. + if (p->opts.interpolation) { + const struct filter_kernel *kernel = + mp_find_filter_kernel(p->opts.scaler[SCALER_TSCALE].kernel.name); + if (kernel) { + // filter_scale wouldn't be correctly initialized were we to use it here. + // This is fine since we're always upsampling, but beware if downsampling + // is added! + double radius = kernel->f.radius; + radius = radius > 0 ? radius : p->opts.scaler[SCALER_TSCALE].radius; + queue_size += 1 + ceil(radius); + } else { + // Oversample/linear case + queue_size += 2; + } + } + + vo_set_queue_params(vo, 0, queue_size); +} + +static int validate_scaler_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value) +{ + struct bstr param = bstr0(*value); + char s[32] = {0}; + int r = 1; + bool tscale = bstr_equals0(name, "tscale"); + if (bstr_equals0(param, "help")) { + r = M_OPT_EXIT; + } else if (bstr_equals0(name, "dscale") && !param.len) { + return r; // empty dscale means "use same as upscaler" + } else if (bstr_equals0(name, "cscale") && !param.len) { + return r; // empty cscale means "use same as upscaler" + } else { + snprintf(s, sizeof(s), "%.*s", BSTR_P(param)); + if (!handle_scaler_opt(s, tscale)) + r = M_OPT_INVALID; + } + if (r < 1) { + mp_info(log, "Available scalers:\n"); + for (const char *const *filter = tscale ? fixed_tscale_filters + : fixed_scale_filters; + *filter; filter++) { + mp_info(log, " %s\n", *filter); + } + for (int n = 0; mp_filter_kernels[n].f.name; n++) { + if (!tscale || !mp_filter_kernels[n].polar) + mp_info(log, " %s\n", mp_filter_kernels[n].f.name); + } + for (int n = 0; mp_filter_windows[n].name; n++) { + for (int m = 0; mp_filter_kernels[m].f.name; m++) { + if (!strcmp(mp_filter_windows[n].name, mp_filter_kernels[m].f.name)) + goto next_window; // don't log duplicates + } + mp_info(log, " %s\n", mp_filter_windows[n].name); +next_window: ; + } + if (s[0]) + mp_fatal(log, "No scaler named '%s' found!\n", s); + } + return r; +} + +static int validate_window_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value) +{ + struct bstr param = bstr0(*value); + char s[32] = {0}; + int r = 1; + if (bstr_equals0(param, "help")) { + r = M_OPT_EXIT; + } else if (!param.len) { + return r; // empty string means "use preferred window" + } else { + snprintf(s, sizeof(s), "%.*s", BSTR_P(param)); + const struct filter_window *window = mp_find_filter_window(s); + if (!window) + r = M_OPT_INVALID; + } + if (r < 1) { + mp_info(log, "Available windows:\n"); + for (int n = 0; mp_filter_windows[n].name; n++) + mp_info(log, " %s\n", mp_filter_windows[n].name); + if (s[0]) + mp_fatal(log, "No window named '%s' found!\n", s); + } + return r; +} + +static int validate_error_diffusion_opt(struct mp_log *log, const m_option_t *opt, + struct bstr name, const char **value) +{ + struct bstr param = bstr0(*value); + char s[32] = {0}; + int r = 1; + if (bstr_equals0(param, "help")) { + r = M_OPT_EXIT; + } else { + snprintf(s, sizeof(s), "%.*s", BSTR_P(param)); + const struct error_diffusion_kernel *k = mp_find_error_diffusion_kernel(s); + if (!k) + r = M_OPT_INVALID; + } + if (r < 1) { + mp_info(log, "Available error diffusion kernels:\n"); + for (int n = 0; mp_error_diffusion_kernels[n].name; n++) + mp_info(log, " %s\n", mp_error_diffusion_kernels[n].name); + if (s[0]) + mp_fatal(log, "No error diffusion kernel named '%s' found!\n", s); + } + return r; +} + +void gl_video_set_ambient_lux(struct gl_video *p, int lux) +{ + if (p->opts.gamma_auto) { + p->opts.gamma = gl_video_scale_ambient_lux(16.0, 256.0, 1.0, 1.2, lux); + MP_TRACE(p, "ambient light changed: %d lux (gamma: %f)\n", lux, + p->opts.gamma); + } +} + +static void *gl_video_dr_alloc_buffer(struct gl_video *p, size_t size) +{ + struct ra_buf_params params = { + .type = RA_BUF_TYPE_TEX_UPLOAD, + .host_mapped = true, + .size = size, + }; + + struct ra_buf *buf = ra_buf_create(p->ra, ¶ms); + if (!buf) + return NULL; + + MP_TARRAY_GROW(p, p->dr_buffers, p->num_dr_buffers); + p->dr_buffers[p->num_dr_buffers++] = (struct dr_buffer){ .buf = buf }; + + return buf->data; +} + +static void gl_video_dr_free_buffer(void *opaque, uint8_t *data) +{ + struct gl_video *p = opaque; + + for (int n = 0; n < p->num_dr_buffers; n++) { + struct dr_buffer *buffer = &p->dr_buffers[n]; + if (buffer->buf->data == data) { + assert(!buffer->mpi); // can't be freed while it has a ref + ra_buf_free(p->ra, &buffer->buf); + MP_TARRAY_REMOVE_AT(p->dr_buffers, p->num_dr_buffers, n); + return; + } + } + // not found - must not happen + MP_ASSERT_UNREACHABLE(); +} + +struct mp_image *gl_video_get_image(struct gl_video *p, int imgfmt, int w, int h, + int stride_align, int flags) +{ + if (flags & VO_DR_FLAG_HOST_CACHED) { + if (p->ra->caps & RA_CAP_SLOW_DR) { + MP_VERBOSE(p, "DR path suspected slow/uncached, disabling.\n"); + return NULL; + } + } + + if (!gl_video_check_format(p, imgfmt)) + return NULL; + + int size = mp_image_get_alloc_size(imgfmt, w, h, stride_align); + if (size < 0) + return NULL; + + int alloc_size = size + stride_align; + void *ptr = gl_video_dr_alloc_buffer(p, alloc_size); + if (!ptr) + return NULL; + + // (we expect vo.c to proxy the free callback, so it happens in the same + // thread it was allocated in, removing the need for synchronization) + struct mp_image *res = mp_image_from_buffer(imgfmt, w, h, stride_align, + ptr, alloc_size, p, + gl_video_dr_free_buffer); + if (!res) + gl_video_dr_free_buffer(p, ptr); + return res; +} + +void gl_video_init_hwdecs(struct gl_video *p, struct ra_ctx *ra_ctx, + struct mp_hwdec_devices *devs, + bool load_all_by_default) +{ + assert(!p->hwdec_ctx.ra_ctx); + p->hwdec_ctx = (struct ra_hwdec_ctx) { + .log = p->log, + .global = p->global, + .ra_ctx = ra_ctx, + }; + + ra_hwdec_ctx_init(&p->hwdec_ctx, devs, p->opts.hwdec_interop, load_all_by_default); +} + +void gl_video_load_hwdecs_for_img_fmt(struct gl_video *p, struct mp_hwdec_devices *devs, + struct hwdec_imgfmt_request *params) +{ + assert(p->hwdec_ctx.ra_ctx); + ra_hwdec_ctx_load_fmt(&p->hwdec_ctx, devs, params); +} diff --git a/video/out/gpu/video.h b/video/out/gpu/video.h new file mode 100644 index 0000000..411d336 --- /dev/null +++ b/video/out/gpu/video.h @@ -0,0 +1,238 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#ifndef MP_GL_VIDEO_H +#define MP_GL_VIDEO_H + +#include <stdbool.h> + +#include "options/m_option.h" +#include "sub/osd.h" +#include "utils.h" +#include "lcms.h" +#include "shader_cache.h" +#include "video/csputils.h" +#include "video/out/filter_kernels.h" + +struct scaler_fun { + char *name; + float params[2]; + float blur; + float taper; +}; + +struct scaler_config { + struct scaler_fun kernel; + struct scaler_fun window; + float radius; + float antiring; + float clamp; +}; + +struct scaler { + int index; + struct scaler_config conf; + double scale_factor; + bool initialized; + struct filter_kernel *kernel; + struct ra_tex *lut; + struct ra_tex *sep_fbo; + bool insufficient; + + // kernel points here + struct filter_kernel kernel_storage; +}; + +enum scaler_unit { + SCALER_SCALE, // luma/video + SCALER_DSCALE, // luma-video downscaling + SCALER_CSCALE, // chroma upscaling + SCALER_TSCALE, // temporal scaling (interpolation) + SCALER_COUNT +}; + +enum dither_algo { + DITHER_NONE = 0, + DITHER_FRUIT, + DITHER_ORDERED, + DITHER_ERROR_DIFFUSION, +}; + +enum alpha_mode { + ALPHA_NO = 0, + ALPHA_YES, + ALPHA_BLEND, + ALPHA_BLEND_TILES, +}; + +enum blend_subs_mode { + BLEND_SUBS_NO = 0, + BLEND_SUBS_YES, + BLEND_SUBS_VIDEO, +}; + +enum tone_mapping { + TONE_MAPPING_AUTO, + TONE_MAPPING_CLIP, + TONE_MAPPING_MOBIUS, + TONE_MAPPING_REINHARD, + TONE_MAPPING_HABLE, + TONE_MAPPING_GAMMA, + TONE_MAPPING_LINEAR, + TONE_MAPPING_SPLINE, + TONE_MAPPING_BT_2390, + TONE_MAPPING_BT_2446A, + TONE_MAPPING_ST2094_40, + TONE_MAPPING_ST2094_10, +}; + +enum gamut_mode { + GAMUT_AUTO, + GAMUT_CLIP, + GAMUT_PERCEPTUAL, + GAMUT_RELATIVE, + GAMUT_SATURATION, + GAMUT_ABSOLUTE, + GAMUT_DESATURATE, + GAMUT_DARKEN, + GAMUT_WARN, + GAMUT_LINEAR, +}; + +struct gl_tone_map_opts { + int curve; + float curve_param; + float max_boost; + bool inverse; + int compute_peak; + float decay_rate; + float scene_threshold_low; + float scene_threshold_high; + float peak_percentile; + float contrast_recovery; + float contrast_smoothness; + int gamut_mode; + bool visualize; +}; + +struct gl_video_opts { + int dumb_mode; + struct scaler_config scaler[4]; + float gamma; + bool gamma_auto; + int target_prim; + int target_trc; + int target_peak; + int target_contrast; + int target_gamut; + struct gl_tone_map_opts tone_map; + bool correct_downscaling; + bool linear_downscaling; + bool linear_upscaling; + bool sigmoid_upscaling; + float sigmoid_center; + float sigmoid_slope; + bool scaler_resizes_only; + bool pbo; + int dither_depth; + int dither_algo; + int dither_size; + bool temporal_dither; + int temporal_dither_period; + char *error_diffusion; + char *fbo_format; + int alpha_mode; + bool use_rectangle; + struct m_color background; + bool interpolation; + float interpolation_threshold; + int blend_subs; + char **user_shaders; + char **user_shader_opts; + bool deband; + struct deband_opts *deband_opts; + float unsharp; + int tex_pad_x, tex_pad_y; + struct mp_icc_opts *icc_opts; + bool shader_cache; + int early_flush; + char *shader_cache_dir; + char *hwdec_interop; +}; + +extern const struct m_sub_options gl_video_conf; + +struct gl_video; +struct vo_frame; +struct voctrl_screenshot; + +enum { + RENDER_FRAME_SUBS = 1 << 0, + RENDER_FRAME_OSD = 1 << 1, + RENDER_FRAME_VF_SUBS = 1 << 2, + RENDER_SCREEN_COLOR = 1 << 3, // 3D LUT and dithering + RENDER_FRAME_DEF = RENDER_FRAME_SUBS | RENDER_FRAME_OSD | RENDER_SCREEN_COLOR, +}; + +struct gl_video *gl_video_init(struct ra *ra, struct mp_log *log, + struct mpv_global *g); +void gl_video_uninit(struct gl_video *p); +void gl_video_set_osd_source(struct gl_video *p, struct osd_state *osd); +bool gl_video_check_format(struct gl_video *p, int mp_format); +void gl_video_config(struct gl_video *p, struct mp_image_params *params); +void gl_video_render_frame(struct gl_video *p, struct vo_frame *frame, + struct ra_fbo fbo, int flags); +void gl_video_resize(struct gl_video *p, + struct mp_rect *src, struct mp_rect *dst, + struct mp_osd_res *osd); +void gl_video_set_fb_depth(struct gl_video *p, int fb_depth); +void gl_video_perfdata(struct gl_video *p, struct voctrl_performance_data *out); +void gl_video_set_clear_color(struct gl_video *p, struct m_color color); +void gl_video_set_osd_pts(struct gl_video *p, double pts); +bool gl_video_check_osd_change(struct gl_video *p, struct mp_osd_res *osd, + double pts); + +void gl_video_screenshot(struct gl_video *p, struct vo_frame *frame, + struct voctrl_screenshot *args); + +float gl_video_scale_ambient_lux(float lmin, float lmax, + float rmin, float rmax, float lux); +void gl_video_set_ambient_lux(struct gl_video *p, int lux); +void gl_video_set_icc_profile(struct gl_video *p, bstr icc_data); +bool gl_video_icc_auto_enabled(struct gl_video *p); +bool gl_video_gamma_auto_enabled(struct gl_video *p); +struct mp_colorspace gl_video_get_output_colorspace(struct gl_video *p); + +void gl_video_reset(struct gl_video *p); +bool gl_video_showing_interpolated_frame(struct gl_video *p); + +struct mp_hwdec_devices; +void gl_video_init_hwdecs(struct gl_video *p, struct ra_ctx *ra_ctx, + struct mp_hwdec_devices *devs, + bool load_all_by_default); +struct hwdec_imgfmt_request; +void gl_video_load_hwdecs_for_img_fmt(struct gl_video *p, struct mp_hwdec_devices *devs, + struct hwdec_imgfmt_request *params); + +struct vo; +void gl_video_configure_queue(struct gl_video *p, struct vo *vo); + +struct mp_image *gl_video_get_image(struct gl_video *p, int imgfmt, int w, int h, + int stride_align, int flags); + + +#endif diff --git a/video/out/gpu/video_shaders.c b/video/out/gpu/video_shaders.c new file mode 100644 index 0000000..6c0e8a8 --- /dev/null +++ b/video/out/gpu/video_shaders.c @@ -0,0 +1,1033 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#include <math.h> + +#include "video_shaders.h" +#include "video.h" + +#define GLSL(x) gl_sc_add(sc, #x "\n"); +#define GLSLF(...) gl_sc_addf(sc, __VA_ARGS__) +#define GLSLH(x) gl_sc_hadd(sc, #x "\n"); +#define GLSLHF(...) gl_sc_haddf(sc, __VA_ARGS__) + +// Set up shared/commonly used variables and macros +void sampler_prelude(struct gl_shader_cache *sc, int tex_num) +{ + GLSLF("#undef tex\n"); + GLSLF("#undef texmap\n"); + GLSLF("#define tex texture%d\n", tex_num); + GLSLF("#define texmap texmap%d\n", tex_num); + GLSLF("vec2 pos = texcoord%d;\n", tex_num); + GLSLF("vec2 size = texture_size%d;\n", tex_num); + GLSLF("vec2 pt = pixel_size%d;\n", tex_num); +} + +static void pass_sample_separated_get_weights(struct gl_shader_cache *sc, + struct scaler *scaler) +{ + gl_sc_uniform_texture(sc, "lut", scaler->lut); + GLSLF("float ypos = LUT_POS(fcoord, %d.0);\n", scaler->lut->params.h); + + int N = scaler->kernel->size; + int width = (N + 3) / 4; // round up + + GLSLF("float weights[%d];\n", N); + for (int i = 0; i < N; i++) { + if (i % 4 == 0) + GLSLF("c = texture(lut, vec2(%f, ypos));\n", (i / 4 + 0.5) / width); + GLSLF("weights[%d] = c[%d];\n", i, i % 4); + } +} + +// Handle a single pass (either vertical or horizontal). The direction is given +// by the vector (d_x, d_y). If the vector is 0, then planar interpolation is +// used instead (samples from texture0 through textureN) +void pass_sample_separated_gen(struct gl_shader_cache *sc, struct scaler *scaler, + int d_x, int d_y) +{ + int N = scaler->kernel->size; + bool use_ar = scaler->conf.antiring > 0; + bool planar = d_x == 0 && d_y == 0; + GLSL(color = vec4(0.0);) + GLSLF("{\n"); + if (!planar) { + GLSLF("vec2 dir = vec2(%d.0, %d.0);\n", d_x, d_y); + GLSL(pt *= dir;) + GLSL(float fcoord = dot(fract(pos * size - vec2(0.5)), dir);) + GLSLF("vec2 base = pos - fcoord * pt - pt * vec2(%d.0);\n", N / 2 - 1); + } + GLSL(vec4 c;) + if (use_ar) { + GLSL(vec4 hi = vec4(0.0);) + GLSL(vec4 lo = vec4(1.0);) + } + pass_sample_separated_get_weights(sc, scaler); + GLSLF("// scaler samples\n"); + for (int n = 0; n < N; n++) { + if (planar) { + GLSLF("c = texture(texture%d, texcoord%d);\n", n, n); + } else { + GLSLF("c = texture(tex, base + pt * vec2(%d.0));\n", n); + } + GLSLF("color += vec4(weights[%d]) * c;\n", n); + if (use_ar && (n == N/2-1 || n == N/2)) { + GLSL(lo = min(lo, c);) + GLSL(hi = max(hi, c);) + } + } + if (use_ar) + GLSLF("color = mix(color, clamp(color, lo, hi), %f);\n", + scaler->conf.antiring); + GLSLF("}\n"); +} + +// Subroutine for computing and adding an individual texel contribution +// If planar is false, samples directly +// If planar is true, takes the pixel from inX[idx] where X is the component and +// `idx` must be defined by the caller +static void polar_sample(struct gl_shader_cache *sc, struct scaler *scaler, + int x, int y, int components, bool planar) +{ + double radius = scaler->kernel->radius * scaler->kernel->filter_scale; + double radius_cutoff = scaler->kernel->radius_cutoff; + + // Since we can't know the subpixel position in advance, assume a + // worst case scenario + int yy = y > 0 ? y-1 : y; + int xx = x > 0 ? x-1 : x; + double dmax = sqrt(xx*xx + yy*yy); + // Skip samples definitely outside the radius + if (dmax >= radius_cutoff) + return; + GLSLF("d = length(vec2(%d.0, %d.0) - fcoord);\n", x, y); + // Check for samples that might be skippable + bool maybe_skippable = dmax >= radius_cutoff - M_SQRT2; + if (maybe_skippable) + GLSLF("if (d < %f) {\n", radius_cutoff); + + // get the weight for this pixel + if (scaler->lut->params.dimensions == 1) { + GLSLF("w = tex1D(lut, LUT_POS(d * 1.0/%f, %d.0)).r;\n", + radius, scaler->lut->params.w); + } else { + GLSLF("w = texture(lut, vec2(0.5, LUT_POS(d * 1.0/%f, %d.0))).r;\n", + radius, scaler->lut->params.h); + } + GLSL(wsum += w;) + + if (planar) { + for (int n = 0; n < components; n++) + GLSLF("color[%d] += w * in%d[idx];\n", n, n); + } else { + GLSLF("in0 = texture(tex, base + pt * vec2(%d.0, %d.0));\n", x, y); + GLSL(color += vec4(w) * in0;) + } + + if (maybe_skippable) + GLSLF("}\n"); +} + +void pass_sample_polar(struct gl_shader_cache *sc, struct scaler *scaler, + int components, bool sup_gather) +{ + GLSL(color = vec4(0.0);) + GLSLF("{\n"); + GLSL(vec2 fcoord = fract(pos * size - vec2(0.5));) + GLSL(vec2 base = pos - fcoord * pt;) + GLSLF("float w, d, wsum = 0.0;\n"); + for (int n = 0; n < components; n++) + GLSLF("vec4 in%d;\n", n); + GLSL(int idx;) + + gl_sc_uniform_texture(sc, "lut", scaler->lut); + + GLSLF("// scaler samples\n"); + int bound = ceil(scaler->kernel->radius_cutoff); + for (int y = 1-bound; y <= bound; y += 2) { + for (int x = 1-bound; x <= bound; x += 2) { + // First we figure out whether it's more efficient to use direct + // sampling or gathering. The problem is that gathering 4 texels + // only to discard some of them is very wasteful, so only do it if + // we suspect it will be a win rather than a loss. This is the case + // exactly when all four texels are within bounds + bool use_gather = sqrt(x*x + y*y) < scaler->kernel->radius_cutoff; + + if (!sup_gather) + use_gather = false; + + if (use_gather) { + // Gather the four surrounding texels simultaneously + for (int n = 0; n < components; n++) { + GLSLF("in%d = textureGatherOffset(tex, base, " + "ivec2(%d, %d), %d);\n", n, x, y, n); + } + + // Mix in all of the points with their weights + for (int p = 0; p < 4; p++) { + // The four texels are gathered counterclockwise starting + // from the bottom left + static const int xo[4] = {0, 1, 1, 0}; + static const int yo[4] = {1, 1, 0, 0}; + if (x+xo[p] > bound || y+yo[p] > bound) + continue; + GLSLF("idx = %d;\n", p); + polar_sample(sc, scaler, x+xo[p], y+yo[p], components, true); + } + } else { + // switch to direct sampling instead, for efficiency/compatibility + for (int yy = y; yy <= bound && yy <= y+1; yy++) { + for (int xx = x; xx <= bound && xx <= x+1; xx++) + polar_sample(sc, scaler, xx, yy, components, false); + } + } + } + } + + GLSL(color = color / vec4(wsum);) + GLSLF("}\n"); +} + +// bw/bh: block size +// iw/ih: input size (pre-calculated to fit all required texels) +void pass_compute_polar(struct gl_shader_cache *sc, struct scaler *scaler, + int components, int bw, int bh, int iw, int ih) +{ + int bound = ceil(scaler->kernel->radius_cutoff); + int offset = bound - 1; // padding top/left + + GLSL(color = vec4(0.0);) + GLSLF("{\n"); + GLSL(vec2 wpos = texmap(gl_WorkGroupID * gl_WorkGroupSize);) + GLSL(vec2 wbase = wpos - pt * fract(wpos * size - vec2(0.5));) + GLSL(vec2 fcoord = fract(pos * size - vec2(0.5));) + GLSL(vec2 base = pos - pt * fcoord;) + GLSL(ivec2 rel = ivec2(round((base - wbase) * size));) + GLSL(int idx;) + GLSLF("float w, d, wsum = 0.0;\n"); + gl_sc_uniform_texture(sc, "lut", scaler->lut); + + // Load all relevant texels into shmem + for (int c = 0; c < components; c++) + GLSLHF("shared float in%d[%d];\n", c, ih * iw); + + GLSL(vec4 c;) + GLSLF("for (int y = int(gl_LocalInvocationID.y); y < %d; y += %d) {\n", ih, bh); + GLSLF("for (int x = int(gl_LocalInvocationID.x); x < %d; x += %d) {\n", iw, bw); + GLSLF("c = texture(tex, wbase + pt * vec2(x - %d, y - %d));\n", offset, offset); + for (int c = 0; c < components; c++) + GLSLF("in%d[%d * y + x] = c[%d];\n", c, iw, c); + GLSLF("}}\n"); + GLSL(groupMemoryBarrier();) + GLSL(barrier();) + + // Dispatch the actual samples + GLSLF("// scaler samples\n"); + for (int y = 1-bound; y <= bound; y++) { + for (int x = 1-bound; x <= bound; x++) { + GLSLF("idx = %d * rel.y + rel.x + %d;\n", iw, + iw * (y + offset) + x + offset); + polar_sample(sc, scaler, x, y, components, true); + } + } + + GLSL(color = color / vec4(wsum);) + GLSLF("}\n"); +} + +static void bicubic_calcweights(struct gl_shader_cache *sc, const char *t, const char *s) +{ + // Explanation of how bicubic scaling with only 4 texel fetches is done: + // http://www.mate.tue.nl/mate/pdfs/10318.pdf + // 'Efficient GPU-Based Texture Interpolation using Uniform B-Splines' + // Explanation why this algorithm normally always blurs, even with unit + // scaling: + // http://bigwww.epfl.ch/preprints/ruijters1001p.pdf + // 'GPU Prefilter for Accurate Cubic B-spline Interpolation' + GLSLF("vec4 %s = vec4(-0.5, 0.1666, 0.3333, -0.3333) * %s" + " + vec4(1, 0, -0.5, 0.5);\n", t, s); + GLSLF("%s = %s * %s + vec4(0, 0, -0.5, 0.5);\n", t, t, s); + GLSLF("%s = %s * %s + vec4(-0.6666, 0, 0.8333, 0.1666);\n", t, t, s); + GLSLF("%s.xy *= vec2(1, 1) / vec2(%s.z, %s.w);\n", t, t, t); + GLSLF("%s.xy += vec2(1.0 + %s, 1.0 - %s);\n", t, s, s); +} + +void pass_sample_bicubic_fast(struct gl_shader_cache *sc) +{ + GLSLF("{\n"); + GLSL(vec2 fcoord = fract(pos * size + vec2(0.5, 0.5));) + bicubic_calcweights(sc, "parmx", "fcoord.x"); + bicubic_calcweights(sc, "parmy", "fcoord.y"); + GLSL(vec4 cdelta;) + GLSL(cdelta.xz = parmx.rg * vec2(-pt.x, pt.x);) + GLSL(cdelta.yw = parmy.rg * vec2(-pt.y, pt.y);) + // first y-interpolation + GLSL(vec4 ar = texture(tex, pos + cdelta.xy);) + GLSL(vec4 ag = texture(tex, pos + cdelta.xw);) + GLSL(vec4 ab = mix(ag, ar, parmy.b);) + // second y-interpolation + GLSL(vec4 br = texture(tex, pos + cdelta.zy);) + GLSL(vec4 bg = texture(tex, pos + cdelta.zw);) + GLSL(vec4 aa = mix(bg, br, parmy.b);) + // x-interpolation + GLSL(color = mix(aa, ab, parmx.b);) + GLSLF("}\n"); +} + +void pass_sample_oversample(struct gl_shader_cache *sc, struct scaler *scaler, + int w, int h) +{ + GLSLF("{\n"); + GLSL(vec2 pos = pos - vec2(0.5) * pt;) // round to nearest + GLSL(vec2 fcoord = fract(pos * size - vec2(0.5));) + // Determine the mixing coefficient vector + gl_sc_uniform_vec2(sc, "output_size", (float[2]){w, h}); + GLSL(vec2 coeff = fcoord * output_size/size;) + float threshold = scaler->conf.kernel.params[0]; + threshold = isnan(threshold) ? 0.0 : threshold; + GLSLF("coeff = (coeff - %f) * 1.0/%f;\n", threshold, 1.0 - 2 * threshold); + GLSL(coeff = clamp(coeff, 0.0, 1.0);) + // Compute the right blend of colors + GLSL(color = texture(tex, pos + pt * (coeff - fcoord));) + GLSLF("}\n"); +} + +// Common constants for SMPTE ST.2084 (HDR) +static const float PQ_M1 = 2610./4096 * 1./4, + PQ_M2 = 2523./4096 * 128, + PQ_C1 = 3424./4096, + PQ_C2 = 2413./4096 * 32, + PQ_C3 = 2392./4096 * 32; + +// Common constants for ARIB STD-B67 (HLG) +static const float HLG_A = 0.17883277, + HLG_B = 0.28466892, + HLG_C = 0.55991073; + +// Common constants for Panasonic V-Log +static const float VLOG_B = 0.00873, + VLOG_C = 0.241514, + VLOG_D = 0.598206; + +// Common constants for Sony S-Log +static const float SLOG_A = 0.432699, + SLOG_B = 0.037584, + SLOG_C = 0.616596 + 0.03, + SLOG_P = 3.538813, + SLOG_Q = 0.030001, + SLOG_K2 = 155.0 / 219.0; + +// Linearize (expand), given a TRC as input. In essence, this is the ITU-R +// EOTF, calculated on an idealized (reference) monitor with a white point of +// MP_REF_WHITE and infinite contrast. +// +// These functions always output to a normalized scale of [0,1], for +// convenience of the video.c code that calls it. To get the values in an +// absolute scale, multiply the result by `mp_trc_nom_peak(trc)` +void pass_linearize(struct gl_shader_cache *sc, enum mp_csp_trc trc) +{ + if (trc == MP_CSP_TRC_LINEAR) + return; + + GLSLF("// linearize\n"); + + // Note that this clamp may technically violate the definition of + // ITU-R BT.2100, which allows for sub-blacks and super-whites to be + // displayed on the display where such would be possible. That said, the + // problem is that not all gamma curves are well-defined on the values + // outside this range, so we ignore it and just clip anyway for sanity. + GLSL(color.rgb = clamp(color.rgb, 0.0, 1.0);) + + switch (trc) { + case MP_CSP_TRC_SRGB: + GLSLF("color.rgb = mix(color.rgb * vec3(1.0/12.92), \n" + " pow((color.rgb + vec3(0.055))/vec3(1.055), vec3(2.4)), \n" + " %s(lessThan(vec3(0.04045), color.rgb))); \n", + gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_BT_1886: + GLSL(color.rgb = pow(color.rgb, vec3(2.4));) + break; + case MP_CSP_TRC_GAMMA18: + GLSL(color.rgb = pow(color.rgb, vec3(1.8));) + break; + case MP_CSP_TRC_GAMMA20: + GLSL(color.rgb = pow(color.rgb, vec3(2.0));) + break; + case MP_CSP_TRC_GAMMA22: + GLSL(color.rgb = pow(color.rgb, vec3(2.2));) + break; + case MP_CSP_TRC_GAMMA24: + GLSL(color.rgb = pow(color.rgb, vec3(2.4));) + break; + case MP_CSP_TRC_GAMMA26: + GLSL(color.rgb = pow(color.rgb, vec3(2.6));) + break; + case MP_CSP_TRC_GAMMA28: + GLSL(color.rgb = pow(color.rgb, vec3(2.8));) + break; + case MP_CSP_TRC_PRO_PHOTO: + GLSLF("color.rgb = mix(color.rgb * vec3(1.0/16.0), \n" + " pow(color.rgb, vec3(1.8)), \n" + " %s(lessThan(vec3(0.03125), color.rgb))); \n", + gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_PQ: + GLSLF("color.rgb = pow(color.rgb, vec3(1.0/%f));\n", PQ_M2); + GLSLF("color.rgb = max(color.rgb - vec3(%f), vec3(0.0)) \n" + " / (vec3(%f) - vec3(%f) * color.rgb);\n", + PQ_C1, PQ_C2, PQ_C3); + GLSLF("color.rgb = pow(color.rgb, vec3(%f));\n", 1.0 / PQ_M1); + // PQ's output range is 0-10000, but we need it to be relative to + // MP_REF_WHITE instead, so rescale + GLSLF("color.rgb *= vec3(%f);\n", 10000 / MP_REF_WHITE); + break; + case MP_CSP_TRC_HLG: + GLSLF("color.rgb = mix(vec3(4.0) * color.rgb * color.rgb,\n" + " exp((color.rgb - vec3(%f)) * vec3(1.0/%f)) + vec3(%f),\n" + " %s(lessThan(vec3(0.5), color.rgb)));\n", + HLG_C, HLG_A, HLG_B, gl_sc_bvec(sc, 3)); + GLSLF("color.rgb *= vec3(1.0/%f);\n", MP_REF_WHITE_HLG); + break; + case MP_CSP_TRC_V_LOG: + GLSLF("color.rgb = mix((color.rgb - vec3(0.125)) * vec3(1.0/5.6), \n" + " pow(vec3(10.0), (color.rgb - vec3(%f)) * vec3(1.0/%f)) \n" + " - vec3(%f), \n" + " %s(lessThanEqual(vec3(0.181), color.rgb))); \n", + VLOG_D, VLOG_C, VLOG_B, gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_S_LOG1: + GLSLF("color.rgb = pow(vec3(10.0), (color.rgb - vec3(%f)) * vec3(1.0/%f))\n" + " - vec3(%f);\n", + SLOG_C, SLOG_A, SLOG_B); + break; + case MP_CSP_TRC_S_LOG2: + GLSLF("color.rgb = mix((color.rgb - vec3(%f)) * vec3(1.0/%f), \n" + " (pow(vec3(10.0), (color.rgb - vec3(%f)) * vec3(1.0/%f)) \n" + " - vec3(%f)) * vec3(1.0/%f), \n" + " %s(lessThanEqual(vec3(%f), color.rgb))); \n", + SLOG_Q, SLOG_P, SLOG_C, SLOG_A, SLOG_B, SLOG_K2, gl_sc_bvec(sc, 3), SLOG_Q); + break; + case MP_CSP_TRC_ST428: + GLSL(color.rgb = vec3(52.37/48.0) * pow(color.rgb, vec3(2.6));); + break; + default: + abort(); + } + + // Rescale to prevent clipping on non-float textures + GLSLF("color.rgb *= vec3(1.0/%f);\n", mp_trc_nom_peak(trc)); +} + +// Delinearize (compress), given a TRC as output. This corresponds to the +// inverse EOTF (not the OETF) in ITU-R terminology, again assuming a +// reference monitor. +// +// Like pass_linearize, this functions ingests values on an normalized scale +void pass_delinearize(struct gl_shader_cache *sc, enum mp_csp_trc trc) +{ + if (trc == MP_CSP_TRC_LINEAR) + return; + + GLSLF("// delinearize\n"); + GLSL(color.rgb = clamp(color.rgb, 0.0, 1.0);) + GLSLF("color.rgb *= vec3(%f);\n", mp_trc_nom_peak(trc)); + + switch (trc) { + case MP_CSP_TRC_SRGB: + GLSLF("color.rgb = mix(color.rgb * vec3(12.92), \n" + " vec3(1.055) * pow(color.rgb, vec3(1.0/2.4)) \n" + " - vec3(0.055), \n" + " %s(lessThanEqual(vec3(0.0031308), color.rgb))); \n", + gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_BT_1886: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/2.4));) + break; + case MP_CSP_TRC_GAMMA18: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/1.8));) + break; + case MP_CSP_TRC_GAMMA20: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/2.0));) + break; + case MP_CSP_TRC_GAMMA22: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/2.2));) + break; + case MP_CSP_TRC_GAMMA24: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/2.4));) + break; + case MP_CSP_TRC_GAMMA26: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/2.6));) + break; + case MP_CSP_TRC_GAMMA28: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/2.8));) + break; + case MP_CSP_TRC_PRO_PHOTO: + GLSLF("color.rgb = mix(color.rgb * vec3(16.0), \n" + " pow(color.rgb, vec3(1.0/1.8)), \n" + " %s(lessThanEqual(vec3(0.001953), color.rgb))); \n", + gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_PQ: + GLSLF("color.rgb *= vec3(1.0/%f);\n", 10000 / MP_REF_WHITE); + GLSLF("color.rgb = pow(color.rgb, vec3(%f));\n", PQ_M1); + GLSLF("color.rgb = (vec3(%f) + vec3(%f) * color.rgb) \n" + " / (vec3(1.0) + vec3(%f) * color.rgb);\n", + PQ_C1, PQ_C2, PQ_C3); + GLSLF("color.rgb = pow(color.rgb, vec3(%f));\n", PQ_M2); + break; + case MP_CSP_TRC_HLG: + GLSLF("color.rgb *= vec3(%f);\n", MP_REF_WHITE_HLG); + GLSLF("color.rgb = mix(vec3(0.5) * sqrt(color.rgb),\n" + " vec3(%f) * log(color.rgb - vec3(%f)) + vec3(%f),\n" + " %s(lessThan(vec3(1.0), color.rgb)));\n", + HLG_A, HLG_B, HLG_C, gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_V_LOG: + GLSLF("color.rgb = mix(vec3(5.6) * color.rgb + vec3(0.125), \n" + " vec3(%f) * log(color.rgb + vec3(%f)) \n" + " + vec3(%f), \n" + " %s(lessThanEqual(vec3(0.01), color.rgb))); \n", + VLOG_C / M_LN10, VLOG_B, VLOG_D, gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_S_LOG1: + GLSLF("color.rgb = vec3(%f) * log(color.rgb + vec3(%f)) + vec3(%f);\n", + SLOG_A / M_LN10, SLOG_B, SLOG_C); + break; + case MP_CSP_TRC_S_LOG2: + GLSLF("color.rgb = mix(vec3(%f) * color.rgb + vec3(%f), \n" + " vec3(%f) * log(vec3(%f) * color.rgb + vec3(%f)) \n" + " + vec3(%f), \n" + " %s(lessThanEqual(vec3(0.0), color.rgb))); \n", + SLOG_P, SLOG_Q, SLOG_A / M_LN10, SLOG_K2, SLOG_B, SLOG_C, gl_sc_bvec(sc, 3)); + break; + case MP_CSP_TRC_ST428: + GLSL(color.rgb = pow(color.rgb * vec3(48.0/52.37), vec3(1.0/2.6));); + break; + default: + abort(); + } +} + +// Apply the OOTF mapping from a given light type to display-referred light. +// Assumes absolute scale values. `peak` is used to tune the OOTF where +// applicable (currently only HLG). +static void pass_ootf(struct gl_shader_cache *sc, enum mp_csp_light light, + float peak) +{ + if (light == MP_CSP_LIGHT_DISPLAY) + return; + + GLSLF("// apply ootf\n"); + + switch (light) + { + case MP_CSP_LIGHT_SCENE_HLG: { + // HLG OOTF from BT.2100, scaled to the chosen display peak + float gamma = MPMAX(1.0, 1.2 + 0.42 * log10(peak * MP_REF_WHITE / 1000.0)); + GLSLF("color.rgb *= vec3(%f * pow(dot(src_luma, color.rgb), %f));\n", + peak / pow(12.0 / MP_REF_WHITE_HLG, gamma), gamma - 1.0); + break; + } + case MP_CSP_LIGHT_SCENE_709_1886: + // This OOTF is defined by encoding the result as 709 and then decoding + // it as 1886; although this is called 709_1886 we actually use the + // more precise (by one decimal) values from BT.2020 instead + GLSLF("color.rgb = mix(color.rgb * vec3(4.5), \n" + " vec3(1.0993) * pow(color.rgb, vec3(0.45)) - vec3(0.0993), \n" + " %s(lessThan(vec3(0.0181), color.rgb))); \n", + gl_sc_bvec(sc, 3)); + GLSL(color.rgb = pow(color.rgb, vec3(2.4));) + break; + case MP_CSP_LIGHT_SCENE_1_2: + GLSL(color.rgb = pow(color.rgb, vec3(1.2));) + break; + default: + abort(); + } +} + +// Inverse of the function pass_ootf, for completeness' sake. +static void pass_inverse_ootf(struct gl_shader_cache *sc, enum mp_csp_light light, + float peak) +{ + if (light == MP_CSP_LIGHT_DISPLAY) + return; + + GLSLF("// apply inverse ootf\n"); + + switch (light) + { + case MP_CSP_LIGHT_SCENE_HLG: { + float gamma = MPMAX(1.0, 1.2 + 0.42 * log10(peak * MP_REF_WHITE / 1000.0)); + GLSLF("color.rgb *= vec3(1.0/%f);\n", peak / pow(12.0 / MP_REF_WHITE_HLG, gamma)); + GLSLF("color.rgb /= vec3(max(1e-6, pow(dot(src_luma, color.rgb), %f)));\n", + (gamma - 1.0) / gamma); + break; + } + case MP_CSP_LIGHT_SCENE_709_1886: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/2.4));) + GLSLF("color.rgb = mix(color.rgb * vec3(1.0/4.5), \n" + " pow((color.rgb + vec3(0.0993)) * vec3(1.0/1.0993), \n" + " vec3(1/0.45)), \n" + " %s(lessThan(vec3(0.08145), color.rgb))); \n", + gl_sc_bvec(sc, 3)); + break; + case MP_CSP_LIGHT_SCENE_1_2: + GLSL(color.rgb = pow(color.rgb, vec3(1.0/1.2));) + break; + default: + abort(); + } +} + +// Average light level for SDR signals. This is equal to a signal level of 0.5 +// under a typical presentation gamma of about 2.0. +static const float sdr_avg = 0.25; + +static void hdr_update_peak(struct gl_shader_cache *sc, + const struct gl_tone_map_opts *opts) +{ + // Update the sig_peak/sig_avg from the old SSBO state + GLSL(if (average.y > 0.0) {) + GLSL( sig_avg = max(1e-3, average.x);) + GLSL( sig_peak = max(1.00, average.y);) + GLSL(}) + + // Chosen to avoid overflowing on an 8K buffer + const float log_min = 1e-3, log_scale = 400.0, sig_scale = 10000.0; + + // For performance, and to avoid overflows, we tally up the sub-results per + // pixel using shared memory first + GLSLH(shared int wg_sum;) + GLSLH(shared uint wg_max;) + GLSL(wg_sum = 0; wg_max = 0u;) + GLSL(barrier();) + GLSLF("float sig_log = log(max(sig_max, %f));\n", log_min); + GLSLF("atomicAdd(wg_sum, int(sig_log * %f));\n", log_scale); + GLSLF("atomicMax(wg_max, uint(sig_max * %f));\n", sig_scale); + + // Have one thread per work group update the global atomics + GLSL(memoryBarrierShared();) + GLSL(barrier();) + GLSL(if (gl_LocalInvocationIndex == 0u) {) + GLSL( int wg_avg = wg_sum / int(gl_WorkGroupSize.x * gl_WorkGroupSize.y);) + GLSL( atomicAdd(frame_sum, wg_avg);) + GLSL( atomicMax(frame_max, wg_max);) + GLSL( memoryBarrierBuffer();) + GLSL(}) + GLSL(barrier();) + + // Finally, to update the global state, we increment a counter per dispatch + GLSL(uint num_wg = gl_NumWorkGroups.x * gl_NumWorkGroups.y;) + GLSL(if (gl_LocalInvocationIndex == 0u && atomicAdd(counter, 1u) == num_wg - 1u) {) + GLSL( counter = 0u;) + GLSL( vec2 cur = vec2(float(frame_sum) / float(num_wg), frame_max);) + GLSLF(" cur *= vec2(1.0/%f, 1.0/%f);\n", log_scale, sig_scale); + GLSL( cur.x = exp(cur.x);) + GLSL( if (average.y == 0.0)) + GLSL( average = cur;) + + // Use an IIR low-pass filter to smooth out the detected values, with a + // configurable decay rate based on the desired time constant (tau) + if (opts->decay_rate) { + float decay = 1.0f - expf(-1.0f / opts->decay_rate); + GLSLF(" average += %f * (cur - average);\n", decay); + } else { + GLSLF(" average = cur;\n"); + } + + // Scene change hysteresis + float log_db = 10.0 / log(10.0); + GLSLF(" float weight = smoothstep(%f, %f, abs(log(cur.x / average.x)));\n", + opts->scene_threshold_low / log_db, + opts->scene_threshold_high / log_db); + GLSL( average = mix(average, cur, weight);) + + // Reset SSBO state for the next frame + GLSL( frame_sum = 0; frame_max = 0u;) + GLSL( memoryBarrierBuffer();) + GLSL(}) +} + +static inline float pq_delinearize(float x) +{ + x *= MP_REF_WHITE / 10000.0; + x = powf(x, PQ_M1); + x = (PQ_C1 + PQ_C2 * x) / (1.0 + PQ_C3 * x); + x = pow(x, PQ_M2); + return x; +} + +// Tone map from a known peak brightness to the range [0,1]. If ref_peak +// is 0, we will use peak detection instead +static void pass_tone_map(struct gl_shader_cache *sc, + float src_peak, float dst_peak, + const struct gl_tone_map_opts *opts) +{ + GLSLF("// HDR tone mapping\n"); + + // To prevent discoloration due to out-of-bounds clipping, we need to make + // sure to reduce the value range as far as necessary to keep the entire + // signal in range, so tone map based on the brightest component. + GLSL(int sig_idx = 0;) + GLSL(if (color[1] > color[sig_idx]) sig_idx = 1;) + GLSL(if (color[2] > color[sig_idx]) sig_idx = 2;) + GLSL(float sig_max = color[sig_idx];) + GLSLF("float sig_peak = %f;\n", src_peak); + GLSLF("float sig_avg = %f;\n", sdr_avg); + + if (opts->compute_peak >= 0) + hdr_update_peak(sc, opts); + + // Always hard-clip the upper bound of the signal range to avoid functions + // exploding on inputs greater than 1.0 + GLSLF("vec3 sig = min(color.rgb, sig_peak);\n"); + + // This function always operates on an absolute scale, so ignore the + // dst_peak normalization for it + float dst_scale = dst_peak; + enum tone_mapping curve = opts->curve ? opts->curve : TONE_MAPPING_BT_2390; + if (curve == TONE_MAPPING_BT_2390) + dst_scale = 1.0; + + // Rescale the variables in order to bring it into a representation where + // 1.0 represents the dst_peak. This is because all of the tone mapping + // algorithms are defined in such a way that they map to the range [0.0, 1.0]. + if (dst_scale > 1.0) { + GLSLF("sig *= 1.0/%f;\n", dst_scale); + GLSLF("sig_peak *= 1.0/%f;\n", dst_scale); + } + + GLSL(float sig_orig = sig[sig_idx];) + GLSLF("float slope = min(%f, %f / sig_avg);\n", opts->max_boost, sdr_avg); + GLSL(sig *= slope;) + GLSL(sig_peak *= slope;) + + float param = opts->curve_param; + switch (curve) { + case TONE_MAPPING_CLIP: + GLSLF("sig = min(%f * sig, 1.0);\n", isnan(param) ? 1.0 : param); + break; + + case TONE_MAPPING_MOBIUS: + GLSLF("if (sig_peak > (1.0 + 1e-6)) {\n"); + GLSLF("const float j = %f;\n", isnan(param) ? 0.3 : param); + // solve for M(j) = j; M(sig_peak) = 1.0; M'(j) = 1.0 + // where M(x) = scale * (x+a)/(x+b) + GLSLF("float a = -j*j * (sig_peak - 1.0) / (j*j - 2.0*j + sig_peak);\n"); + GLSLF("float b = (j*j - 2.0*j*sig_peak + sig_peak) / " + "max(1e-6, sig_peak - 1.0);\n"); + GLSLF("float scale = (b*b + 2.0*b*j + j*j) / (b-a);\n"); + GLSLF("sig = mix(sig, scale * (sig + vec3(a)) / (sig + vec3(b))," + " %s(greaterThan(sig, vec3(j))));\n", + gl_sc_bvec(sc, 3)); + GLSLF("}\n"); + break; + + case TONE_MAPPING_REINHARD: { + float contrast = isnan(param) ? 0.5 : param, + offset = (1.0 - contrast) / contrast; + GLSLF("sig = sig / (sig + vec3(%f));\n", offset); + GLSLF("float scale = (sig_peak + %f) / sig_peak;\n", offset); + GLSL(sig *= scale;) + break; + } + + case TONE_MAPPING_HABLE: { + float A = 0.15, B = 0.50, C = 0.10, D = 0.20, E = 0.02, F = 0.30; + GLSLHF("vec3 hable(vec3 x) {\n"); + GLSLHF("return (x * (%f*x + vec3(%f)) + vec3(%f)) / " + " (x * (%f*x + vec3(%f)) + vec3(%f)) " + " - vec3(%f);\n", + A, C*B, D*E, + A, B, D*F, + E/F); + GLSLHF("}\n"); + GLSLF("sig = hable(max(vec3(0.0), sig)) / hable(vec3(sig_peak)).x;\n"); + break; + } + + case TONE_MAPPING_GAMMA: { + float gamma = isnan(param) ? 1.8 : param; + GLSLF("const float cutoff = 0.05, gamma = 1.0/%f;\n", gamma); + GLSL(float scale = pow(cutoff / sig_peak, gamma.x) / cutoff;) + GLSLF("sig = mix(scale * sig," + " pow(sig / sig_peak, vec3(gamma))," + " %s(greaterThan(sig, vec3(cutoff))));\n", + gl_sc_bvec(sc, 3)); + break; + } + + case TONE_MAPPING_LINEAR: { + float coeff = isnan(param) ? 1.0 : param; + GLSLF("sig = min(%f / sig_peak, 1.0) * sig;\n", coeff); + break; + } + + case TONE_MAPPING_BT_2390: + // We first need to encode both sig and sig_peak into PQ space + GLSLF("vec4 sig_pq = vec4(sig.rgb, sig_peak); \n" + "sig_pq *= vec4(1.0/%f); \n" + "sig_pq = pow(sig_pq, vec4(%f)); \n" + "sig_pq = (vec4(%f) + vec4(%f) * sig_pq) \n" + " / (vec4(1.0) + vec4(%f) * sig_pq); \n" + "sig_pq = pow(sig_pq, vec4(%f)); \n", + 10000.0 / MP_REF_WHITE, PQ_M1, PQ_C1, PQ_C2, PQ_C3, PQ_M2); + // Encode both the signal and the target brightness to be relative to + // the source peak brightness, and figure out the target peak in this space + GLSLF("float scale = 1.0 / sig_pq.a; \n" + "sig_pq.rgb *= vec3(scale); \n" + "float maxLum = %f * scale; \n", + pq_delinearize(dst_peak)); + // Apply piece-wise hermite spline + GLSLF("float ks = 1.5 * maxLum - 0.5; \n" + "vec3 tb = (sig_pq.rgb - vec3(ks)) / vec3(1.0 - ks); \n" + "vec3 tb2 = tb * tb; \n" + "vec3 tb3 = tb2 * tb; \n" + "vec3 pb = (2.0 * tb3 - 3.0 * tb2 + vec3(1.0)) * vec3(ks) + \n" + " (tb3 - 2.0 * tb2 + tb) * vec3(1.0 - ks) + \n" + " (-2.0 * tb3 + 3.0 * tb2) * vec3(maxLum); \n" + "sig = mix(pb, sig_pq.rgb, %s(lessThan(sig_pq.rgb, vec3(ks)))); \n", + gl_sc_bvec(sc, 3)); + // Convert back from PQ space to linear light + GLSLF("sig *= vec3(sig_pq.a); \n" + "sig = pow(sig, vec3(1.0/%f)); \n" + "sig = max(sig - vec3(%f), 0.0) / \n" + " (vec3(%f) - vec3(%f) * sig); \n" + "sig = pow(sig, vec3(1.0/%f)); \n" + "sig *= vec3(%f); \n", + PQ_M2, PQ_C1, PQ_C2, PQ_C3, PQ_M1, 10000.0 / MP_REF_WHITE); + break; + + default: + abort(); + } + + GLSLF("float coeff = max(sig[sig_idx] - %f, 1e-6) / \n" + " max(sig[sig_idx], 1.0); \n" + "coeff = %f * pow(coeff / %f, %f); \n" + "color.rgb *= sig[sig_idx] / sig_orig; \n" + "color.rgb = mix(color.rgb, %f * sig, coeff); \n", + 0.18 / dst_scale, 0.90, dst_scale, 0.20, dst_scale); +} + +// Map colors from one source space to another. These source spaces must be +// known (i.e. not MP_CSP_*_AUTO), as this function won't perform any +// auto-guessing. If is_linear is true, we assume the input has already been +// linearized (e.g. for linear-scaling). If `opts->compute_peak` is true, we +// will detect the peak instead of relying on metadata. Note that this requires +// the caller to have already bound the appropriate SSBO and set up the compute +// shader metadata +void pass_color_map(struct gl_shader_cache *sc, bool is_linear, + struct mp_colorspace src, struct mp_colorspace dst, + const struct gl_tone_map_opts *opts) +{ + GLSLF("// color mapping\n"); + + // Some operations need access to the video's luma coefficients, so make + // them available + float rgb2xyz[3][3]; + mp_get_rgb2xyz_matrix(mp_get_csp_primaries(src.primaries), rgb2xyz); + gl_sc_uniform_vec3(sc, "src_luma", rgb2xyz[1]); + mp_get_rgb2xyz_matrix(mp_get_csp_primaries(dst.primaries), rgb2xyz); + gl_sc_uniform_vec3(sc, "dst_luma", rgb2xyz[1]); + + bool need_ootf = src.light != dst.light; + if (src.light == MP_CSP_LIGHT_SCENE_HLG && src.hdr.max_luma != dst.hdr.max_luma) + need_ootf = true; + + // All operations from here on require linear light as a starting point, + // so we linearize even if src.gamma == dst.gamma when one of the other + // operations needs it + bool need_linear = src.gamma != dst.gamma || + src.primaries != dst.primaries || + src.hdr.max_luma != dst.hdr.max_luma || + need_ootf; + + if (need_linear && !is_linear) { + // We also pull it up so that 1.0 is the reference white + pass_linearize(sc, src.gamma); + is_linear = true; + } + + // Pre-scale the incoming values into an absolute scale + GLSLF("color.rgb *= vec3(%f);\n", mp_trc_nom_peak(src.gamma)); + + if (need_ootf) + pass_ootf(sc, src.light, src.hdr.max_luma / MP_REF_WHITE); + + // Tone map to prevent clipping due to excessive brightness + if (src.hdr.max_luma > dst.hdr.max_luma) { + pass_tone_map(sc, src.hdr.max_luma / MP_REF_WHITE, + dst.hdr.max_luma / MP_REF_WHITE, opts); + } + + // Adapt to the right colorspace if necessary + if (src.primaries != dst.primaries) { + struct mp_csp_primaries csp_src = mp_get_csp_primaries(src.primaries), + csp_dst = mp_get_csp_primaries(dst.primaries); + float m[3][3] = {{0}}; + mp_get_cms_matrix(csp_src, csp_dst, MP_INTENT_RELATIVE_COLORIMETRIC, m); + gl_sc_uniform_mat3(sc, "cms_matrix", true, &m[0][0]); + GLSL(color.rgb = cms_matrix * color.rgb;) + + if (!opts->gamut_mode || opts->gamut_mode == GAMUT_DESATURATE) { + GLSL(float cmin = min(min(color.r, color.g), color.b);) + GLSL(if (cmin < 0.0) { + float luma = dot(dst_luma, color.rgb); + float coeff = cmin / (cmin - luma); + color.rgb = mix(color.rgb, vec3(luma), coeff); + }) + GLSLF("float cmax = 1.0/%f * max(max(color.r, color.g), color.b);\n", + dst.hdr.max_luma / MP_REF_WHITE); + GLSL(if (cmax > 1.0) color.rgb /= cmax;) + } + } + + if (need_ootf) + pass_inverse_ootf(sc, dst.light, dst.hdr.max_luma / MP_REF_WHITE); + + // Post-scale the outgoing values from absolute scale to normalized. + // For SDR, we normalize to the chosen signal peak. For HDR, we normalize + // to the encoding range of the transfer function. + float dst_range = dst.hdr.max_luma / MP_REF_WHITE; + if (mp_trc_is_hdr(dst.gamma)) + dst_range = mp_trc_nom_peak(dst.gamma); + + GLSLF("color.rgb *= vec3(%f);\n", 1.0 / dst_range); + + // Warn for remaining out-of-gamut colors if enabled + if (opts->gamut_mode == GAMUT_WARN) { + GLSL(if (any(greaterThan(color.rgb, vec3(1.005))) || + any(lessThan(color.rgb, vec3(-0.005))))) + GLSL(color.rgb = vec3(1.0) - color.rgb;) // invert + } + + if (is_linear) + pass_delinearize(sc, dst.gamma); +} + +// Wide usage friendly PRNG, shamelessly stolen from a GLSL tricks forum post. +// Obtain random numbers by calling rand(h), followed by h = permute(h) to +// update the state. Assumes the texture was hooked. +// permute() was modified from the original to avoid "large" numbers in +// calculations, since low-end mobile GPUs choke on them (overflow). +static void prng_init(struct gl_shader_cache *sc, AVLFG *lfg) +{ + GLSLH(float mod289(float x) { return x - floor(x * 1.0/289.0) * 289.0; }) + GLSLHF("float permute(float x) {\n"); + GLSLH(return mod289( mod289(34.0*x + 1.0) * (fract(x) + 1.0) );) + GLSLHF("}\n"); + GLSLH(float rand(float x) { return fract(x * 1.0/41.0); }) + + // Initialize the PRNG by hashing the position + a random uniform + GLSL(vec3 _m = vec3(HOOKED_pos, random) + vec3(1.0);) + GLSL(float h = permute(permute(permute(_m.x)+_m.y)+_m.z);) + gl_sc_uniform_dynamic(sc); + gl_sc_uniform_f(sc, "random", (double)av_lfg_get(lfg) / UINT32_MAX); +} + +const struct deband_opts deband_opts_def = { + .iterations = 1, + .threshold = 48.0, + .range = 16.0, + .grain = 32.0, +}; + +#define OPT_BASE_STRUCT struct deband_opts +const struct m_sub_options deband_conf = { + .opts = (const m_option_t[]) { + {"iterations", OPT_INT(iterations), M_RANGE(0, 16)}, + {"threshold", OPT_FLOAT(threshold), M_RANGE(0.0, 4096.0)}, + {"range", OPT_FLOAT(range), M_RANGE(1.0, 64.0)}, + {"grain", OPT_FLOAT(grain), M_RANGE(0.0, 4096.0)}, + {0} + }, + .size = sizeof(struct deband_opts), + .defaults = &deband_opts_def, +}; + +// Stochastically sample a debanded result from a hooked texture. +void pass_sample_deband(struct gl_shader_cache *sc, struct deband_opts *opts, + AVLFG *lfg, enum mp_csp_trc trc) +{ + // Initialize the PRNG + GLSLF("{\n"); + prng_init(sc, lfg); + + // Helper: Compute a stochastic approximation of the avg color around a + // pixel + GLSLHF("vec4 average(float range, inout float h) {\n"); + // Compute a random rangle and distance + GLSLH(float dist = rand(h) * range; h = permute(h);) + GLSLH(float dir = rand(h) * 6.2831853; h = permute(h);) + GLSLH(vec2 o = dist * vec2(cos(dir), sin(dir));) + + // Sample at quarter-turn intervals around the source pixel + GLSLH(vec4 ref[4];) + GLSLH(ref[0] = HOOKED_texOff(vec2( o.x, o.y));) + GLSLH(ref[1] = HOOKED_texOff(vec2(-o.y, o.x));) + GLSLH(ref[2] = HOOKED_texOff(vec2(-o.x, -o.y));) + GLSLH(ref[3] = HOOKED_texOff(vec2( o.y, -o.x));) + + // Return the (normalized) average + GLSLH(return (ref[0] + ref[1] + ref[2] + ref[3])*0.25;) + GLSLHF("}\n"); + + // Sample the source pixel + GLSL(color = HOOKED_tex(HOOKED_pos);) + GLSLF("vec4 avg, diff;\n"); + for (int i = 1; i <= opts->iterations; i++) { + // Sample the average pixel and use it instead of the original if + // the difference is below the given threshold + GLSLF("avg = average(%f, h);\n", i * opts->range); + GLSL(diff = abs(color - avg);) + GLSLF("color = mix(avg, color, %s(greaterThan(diff, vec4(%f))));\n", + gl_sc_bvec(sc, 4), opts->threshold / (i * 16384.0)); + } + + // Add some random noise to smooth out residual differences + GLSL(vec3 noise;) + GLSL(noise.x = rand(h); h = permute(h);) + GLSL(noise.y = rand(h); h = permute(h);) + GLSL(noise.z = rand(h); h = permute(h);) + + // Noise is scaled to the signal level to prevent extreme noise for HDR + float gain = opts->grain/8192.0 / mp_trc_nom_peak(trc); + GLSLF("color.xyz += %f * (noise - vec3(0.5));\n", gain); + GLSLF("}\n"); +} + +// Assumes the texture was hooked +void pass_sample_unsharp(struct gl_shader_cache *sc, float param) { + GLSLF("{\n"); + GLSL(float st1 = 1.2;) + GLSL(vec4 p = HOOKED_tex(HOOKED_pos);) + GLSL(vec4 sum1 = HOOKED_texOff(st1 * vec2(+1, +1)) + + HOOKED_texOff(st1 * vec2(+1, -1)) + + HOOKED_texOff(st1 * vec2(-1, +1)) + + HOOKED_texOff(st1 * vec2(-1, -1));) + GLSL(float st2 = 1.5;) + GLSL(vec4 sum2 = HOOKED_texOff(st2 * vec2(+1, 0)) + + HOOKED_texOff(st2 * vec2( 0, +1)) + + HOOKED_texOff(st2 * vec2(-1, 0)) + + HOOKED_texOff(st2 * vec2( 0, -1));) + GLSL(vec4 t = p * 0.859375 + sum2 * -0.1171875 + sum1 * -0.09765625;) + GLSLF("color = p + t * %f;\n", param); + GLSLF("}\n"); +} diff --git a/video/out/gpu/video_shaders.h b/video/out/gpu/video_shaders.h new file mode 100644 index 0000000..27e7874 --- /dev/null +++ b/video/out/gpu/video_shaders.h @@ -0,0 +1,59 @@ +/* + * This file is part of mpv. + * + * mpv is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * mpv is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with mpv. If not, see <http://www.gnu.org/licenses/>. + */ + +#ifndef MP_GL_VIDEO_SHADERS_H +#define MP_GL_VIDEO_SHADERS_H + +#include <libavutil/lfg.h> + +#include "utils.h" +#include "video.h" + +struct deband_opts { + int iterations; + float threshold; + float range; + float grain; +}; + +extern const struct deband_opts deband_opts_def; +extern const struct m_sub_options deband_conf; + +void sampler_prelude(struct gl_shader_cache *sc, int tex_num); +void pass_sample_separated_gen(struct gl_shader_cache *sc, struct scaler *scaler, + int d_x, int d_y); +void pass_sample_polar(struct gl_shader_cache *sc, struct scaler *scaler, + int components, bool sup_gather); +void pass_compute_polar(struct gl_shader_cache *sc, struct scaler *scaler, + int components, int bw, int bh, int iw, int ih); +void pass_sample_bicubic_fast(struct gl_shader_cache *sc); +void pass_sample_oversample(struct gl_shader_cache *sc, struct scaler *scaler, + int w, int h); + +void pass_linearize(struct gl_shader_cache *sc, enum mp_csp_trc trc); +void pass_delinearize(struct gl_shader_cache *sc, enum mp_csp_trc trc); + +void pass_color_map(struct gl_shader_cache *sc, bool is_linear, + struct mp_colorspace src, struct mp_colorspace dst, + const struct gl_tone_map_opts *opts); + +void pass_sample_deband(struct gl_shader_cache *sc, struct deband_opts *opts, + AVLFG *lfg, enum mp_csp_trc trc); + +void pass_sample_unsharp(struct gl_shader_cache *sc, float param); + +#endif |