path: root/src/gpu.c

/*
 * This file is part of libplacebo.
 *
 * libplacebo 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.
 *
 * libplacebo 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 libplacebo. If not, see <http://www.gnu.org/licenses/>.
 */

#include "common.h"
#include "gpu.h"

#define require(expr) pl_require(gpu, expr)

void pl_gpu_destroy(pl_gpu gpu)
{
    if (!gpu)
        return;

    struct pl_gpu_fns *impl = PL_PRIV(gpu);
    pl_dispatch_destroy(&impl->dp);
    impl->destroy(gpu);
}

pl_dispatch pl_gpu_dispatch(pl_gpu gpu)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->dp;
}

pl_cache pl_gpu_cache(pl_gpu gpu)
{
    if (!gpu)
        return NULL;
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return atomic_load(&impl->cache);
}

void pl_gpu_set_cache(pl_gpu gpu, pl_cache cache)
{
    struct pl_gpu_fns *impl = PL_PRIV(gpu);
    atomic_store(&impl->cache, cache);
}

bool pl_fmt_is_ordered(pl_fmt fmt)
{
    bool ret = !fmt->opaque;
    for (int i = 0; i < fmt->num_components; i++)
        ret &= fmt->sample_order[i] == i;
    return ret;
}

bool pl_fmt_is_float(pl_fmt fmt)
{
    switch (fmt->type) {
    case PL_FMT_UNKNOWN: // more likely than not
    case PL_FMT_FLOAT:
    case PL_FMT_UNORM:
    case PL_FMT_SNORM:
        return true;

    case PL_FMT_UINT:
    case PL_FMT_SINT:
        return false;

    case PL_FMT_TYPE_COUNT:
        break;
    }

    pl_unreachable();
}

bool pl_fmt_has_modifier(pl_fmt fmt, uint64_t modifier)
{
    if (!fmt)
        return false;

    for (int i = 0; i < fmt->num_modifiers; i++) {
        if (fmt->modifiers[i] == modifier)
            return true;
    }

    return false;
}

pl_fmt pl_find_fmt(pl_gpu gpu, enum pl_fmt_type type, int num_components,
                    int min_depth, int host_bits, enum pl_fmt_caps caps)
{
    for (int n = 0; n < gpu->num_formats; n++) {
        pl_fmt fmt = gpu->formats[n];
        if (fmt->type != type || fmt->num_components != num_components)
            continue;
        if ((fmt->caps & caps) != caps)
            continue;

        // When specifying some particular host representation, ensure the
        // format is non-opaque, ordered and unpadded
        if (host_bits && fmt->opaque)
            continue;
        if (host_bits && fmt->texel_size * 8 != host_bits * num_components)
            continue;
        if (host_bits && !pl_fmt_is_ordered(fmt))
            continue;

        for (int i = 0; i < fmt->num_components; i++) {
            if (fmt->component_depth[i] < min_depth)
                goto next_fmt;
            if (host_bits && fmt->host_bits[i] != host_bits)
                goto next_fmt;
        }

        return fmt;

next_fmt: ; // equivalent to `continue`
    }

    // ran out of formats
    PL_TRACE(gpu, "No matching format found");
    return NULL;
}

pl_fmt pl_find_vertex_fmt(pl_gpu gpu, enum pl_fmt_type type, int comps)
{
    static const size_t sizes[] = {
        [PL_FMT_FLOAT] = sizeof(float),
        [PL_FMT_UNORM] = sizeof(unsigned),
        [PL_FMT_UINT]  = sizeof(unsigned),
        [PL_FMT_SNORM] = sizeof(int),
        [PL_FMT_SINT]  = sizeof(int),
    };

    return pl_find_fmt(gpu, type, comps, 0, 8 * sizes[type], PL_FMT_CAP_VERTEX);
}

pl_fmt pl_find_named_fmt(pl_gpu gpu, const char *name)
{
    if (!name)
        return NULL;

    for (int i = 0; i < gpu->num_formats; i++) {
        pl_fmt fmt = gpu->formats[i];
        if (strcmp(name, fmt->name) == 0)
            return fmt;
    }

    // ran out of formats
    return NULL;
}

pl_fmt pl_find_fourcc(pl_gpu gpu, uint32_t fourcc)
{
    if (!fourcc)
        return NULL;

    for (int i = 0; i < gpu->num_formats; i++) {
        pl_fmt fmt = gpu->formats[i];
        if (fourcc == fmt->fourcc)
            return fmt;
    }

    // ran out of formats
    return NULL;
}

static inline bool check_mod(pl_gpu gpu, pl_fmt fmt, uint64_t mod)
{
    for (int i = 0; i < fmt->num_modifiers; i++) {
        if (fmt->modifiers[i] == mod)
            return true;
    }


    PL_ERR(gpu, "DRM modifier %s not available for format %s. Available modifiers:",
           PRINT_DRM_MOD(mod), fmt->name);
    for (int i = 0; i < fmt->num_modifiers; i++)
        PL_ERR(gpu, "    %s", PRINT_DRM_MOD(fmt->modifiers[i]));

    return false;
}

pl_tex pl_tex_create(pl_gpu gpu, const struct pl_tex_params *params)
{
    require(params->format);
    require(!params->import_handle || !params->export_handle);
    require(!params->import_handle || !params->initial_data);
    if (params->export_handle) {
        require(params->export_handle & gpu->export_caps.tex);
        require(PL_ISPOT(params->export_handle));
    }
    if (params->import_handle) {
        require(params->import_handle & gpu->import_caps.tex);
        require(PL_ISPOT(params->import_handle));
        if (params->import_handle == PL_HANDLE_DMA_BUF) {
            if (!check_mod(gpu, params->format, params->shared_mem.drm_format_mod))
                goto error;
            if (params->shared_mem.stride_w)
                require(params->w && params->shared_mem.stride_w >= params->w);
            if (params->shared_mem.stride_h)
                require(params->h && params->shared_mem.stride_h >= params->h);
        } else if (params->import_handle == PL_HANDLE_MTL_TEX) {
            require(params->shared_mem.plane <= 2);
        }
    }

    switch (pl_tex_params_dimension(*params)) {
    case 1:
        require(params->w > 0);
        require(params->w <= gpu->limits.max_tex_1d_dim);
        require(!params->renderable);
        require(!params->blit_src || gpu->limits.blittable_1d_3d);
        require(!params->blit_dst || gpu->limits.blittable_1d_3d);
        require(!params->format->num_planes);
        break;
    case 2:
        require(params->w > 0 && params->h > 0);
        require(params->w <= gpu->limits.max_tex_2d_dim);
        require(params->h <= gpu->limits.max_tex_2d_dim);
        break;
    case 3:
        require(params->w > 0 && params->h > 0 && params->d > 0);
        require(params->w <= gpu->limits.max_tex_3d_dim);
        require(params->h <= gpu->limits.max_tex_3d_dim);
        require(params->d <= gpu->limits.max_tex_3d_dim);
        require(!params->renderable);
        require(!params->blit_src || gpu->limits.blittable_1d_3d);
        require(!params->blit_dst || gpu->limits.blittable_1d_3d);
        require(!params->format->num_planes);
        break;
    }

    enum pl_fmt_caps fmt_caps = params->format->caps;
    bool fmt_opaque = params->format->opaque;
    for (int i = 0; i < params->format->num_planes; i++) {
        pl_fmt pfmt = params->format->planes[i].format;
        fmt_caps |= pfmt->caps;
        fmt_opaque &= pfmt->opaque;
    }

    require(!params->host_readable || fmt_caps & PL_FMT_CAP_HOST_READABLE);
    require(!params->host_writable || !fmt_opaque);
    require(!params->sampleable || fmt_caps & PL_FMT_CAP_SAMPLEABLE);
    require(!params->renderable || fmt_caps & PL_FMT_CAP_RENDERABLE);
    require(!params->storable   || fmt_caps & PL_FMT_CAP_STORABLE);
    require(!params->blit_src   || fmt_caps & PL_FMT_CAP_BLITTABLE);
    require(!params->blit_dst   || fmt_caps & PL_FMT_CAP_BLITTABLE);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->tex_create(gpu, params);

error:
    if (params->debug_tag)
        PL_ERR(gpu, "  for texture: %s", params->debug_tag);
    return NULL;
}

void pl_tex_destroy(pl_gpu gpu, pl_tex *tex)
{
    if (!*tex)
        return;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->tex_destroy(gpu, *tex);
    *tex = NULL;
}

static bool pl_tex_params_superset(struct pl_tex_params a, struct pl_tex_params b)
{
    return a.w == b.w && a.h == b.h && a.d == b.d &&
           a.format          == b.format &&
           (a.sampleable     || !b.sampleable) &&
           (a.renderable     || !b.renderable) &&
           (a.storable       || !b.storable) &&
           (a.blit_src       || !b.blit_src) &&
           (a.blit_dst       || !b.blit_dst) &&
           (a.host_writable  || !b.host_writable) &&
           (a.host_readable  || !b.host_readable);
}

bool pl_tex_recreate(pl_gpu gpu, pl_tex *tex, const struct pl_tex_params *params)
{
    if (params->initial_data) {
        PL_ERR(gpu, "pl_tex_recreate may not be used with `initial_data`!");
        return false;
    }

    if (params->import_handle) {
        PL_ERR(gpu, "pl_tex_recreate may not be used with `import_handle`!");
        return false;
    }

    if (*tex && pl_tex_params_superset((*tex)->params, *params)) {
        pl_tex_invalidate(gpu, *tex);
        return true;
    }

    PL_DEBUG(gpu, "(Re)creating %dx%dx%d texture with format %s: %s",
             params->w, params->h, params->d, params->format->name,
             PL_DEF(params->debug_tag, "unknown"));

    pl_tex_destroy(gpu, tex);
    *tex = pl_tex_create(gpu, params);

    return !!*tex;
}

void pl_tex_clear_ex(pl_gpu gpu, pl_tex dst, const union pl_clear_color color)
{
    require(dst->params.blit_dst);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    if (impl->tex_invalidate)
        impl->tex_invalidate(gpu, dst);
    impl->tex_clear_ex(gpu, dst, color);
    return;

error:
    if (dst->params.debug_tag)
        PL_ERR(gpu, "  for texture: %s", dst->params.debug_tag);
}

void pl_tex_clear(pl_gpu gpu, pl_tex dst, const float color[4])
{
    if (!pl_fmt_is_float(dst->params.format)) {
        PL_ERR(gpu, "Cannot call `pl_tex_clear` on integer textures, please "
               "use `pl_tex_clear_ex` instead.");
        return;
    }

    const union pl_clear_color col = {
        .f = { color[0], color[1], color[2], color[3] },
    };

    pl_tex_clear_ex(gpu, dst, col);
}

void pl_tex_invalidate(pl_gpu gpu, pl_tex tex)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    if (impl->tex_invalidate)
        impl->tex_invalidate(gpu, tex);
}

static void strip_coords(pl_tex tex, pl_rect3d *rc)
{
    if (!tex->params.d) {
        rc->z0 = 0;
        rc->z1 = 1;
    }

    if (!tex->params.h) {
        rc->y0 = 0;
        rc->y1 = 1;
    }
}

static void infer_rc(pl_tex tex, pl_rect3d *rc)
{
    if (!rc->x0 && !rc->x1)
        rc->x1 = tex->params.w;
    if (!rc->y0 && !rc->y1)
        rc->y1 = tex->params.h;
    if (!rc->z0 && !rc->z1)
        rc->z1 = tex->params.d;
}

void pl_tex_blit(pl_gpu gpu, const struct pl_tex_blit_params *params)
{
    pl_tex src = params->src, dst = params->dst;
    require(src && dst);
    pl_fmt src_fmt = src->params.format;
    pl_fmt dst_fmt = dst->params.format;
    require(src_fmt->internal_size == dst_fmt->internal_size);
    require((src_fmt->type == PL_FMT_UINT) == (dst_fmt->type == PL_FMT_UINT));
    require((src_fmt->type == PL_FMT_SINT) == (dst_fmt->type == PL_FMT_SINT));
    require(src->params.blit_src);
    require(dst->params.blit_dst);
    require(params->sample_mode != PL_TEX_SAMPLE_LINEAR || (src_fmt->caps & PL_FMT_CAP_LINEAR));

    struct pl_tex_blit_params fixed = *params;
    infer_rc(src, &fixed.src_rc);
    infer_rc(dst, &fixed.dst_rc);
    strip_coords(src, &fixed.src_rc);
    strip_coords(dst, &fixed.dst_rc);

    require(fixed.src_rc.x0 >= 0 && fixed.src_rc.x0 < src->params.w);
    require(fixed.src_rc.x1 > 0 && fixed.src_rc.x1 <= src->params.w);
    require(fixed.dst_rc.x0 >= 0 && fixed.dst_rc.x0 < dst->params.w);
    require(fixed.dst_rc.x1 > 0 && fixed.dst_rc.x1 <= dst->params.w);

    if (src->params.h) {
        require(fixed.src_rc.y0 >= 0 && fixed.src_rc.y0 < src->params.h);
        require(fixed.src_rc.y1 > 0 && fixed.src_rc.y1 <= src->params.h);
    }

    if (dst->params.h) {
        require(fixed.dst_rc.y0 >= 0 && fixed.dst_rc.y0 < dst->params.h);
        require(fixed.dst_rc.y1 > 0 && fixed.dst_rc.y1 <= dst->params.h);
    }

    if (src->params.d) {
        require(fixed.src_rc.z0 >= 0 && fixed.src_rc.z0 < src->params.d);
        require(fixed.src_rc.z1 > 0 && fixed.src_rc.z1 <= src->params.d);
    }

    if (dst->params.d) {
        require(fixed.dst_rc.z0 >= 0 && fixed.dst_rc.z0 < dst->params.d);
        require(fixed.dst_rc.z1 > 0 && fixed.dst_rc.z1 <= dst->params.d);
    }

    pl_rect3d full = {0, 0, 0, dst->params.w, dst->params.h, dst->params.d};
    strip_coords(dst, &full);

    pl_rect3d rcnorm = fixed.dst_rc;
    pl_rect3d_normalize(&rcnorm);
    if (pl_rect3d_eq(rcnorm, full))
        pl_tex_invalidate(gpu, dst);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->tex_blit(gpu, &fixed);
    return;

error:
    if (src->params.debug_tag || dst->params.debug_tag) {
        PL_ERR(gpu, "  for textures: src %s, dst %s",
               PL_DEF(src->params.debug_tag, "(unknown)"),
               PL_DEF(dst->params.debug_tag, "(unknown)"));
    }
}

static bool fix_tex_transfer(pl_gpu gpu, struct pl_tex_transfer_params *params)
{
    pl_tex tex = params->tex;
    pl_fmt fmt = tex->params.format;
    pl_rect3d rc = params->rc;

    // Infer the default values
    infer_rc(tex, &rc);
    strip_coords(tex, &rc);

    if (!params->row_pitch || !tex->params.w)
        params->row_pitch = pl_rect_w(rc) * fmt->texel_size;
    if (!params->depth_pitch || !tex->params.d)
        params->depth_pitch = pl_rect_h(rc) * params->row_pitch;

    require(params->row_pitch);
    require(params->depth_pitch);
    params->rc = rc;

    // Check the parameters for sanity
    switch (pl_tex_params_dimension(tex->params))
    {
    case 3:
        require(rc.z1 > rc.z0);
        require(rc.z0 >= 0 && rc.z0 <  tex->params.d);
        require(rc.z1 >  0 && rc.z1 <= tex->params.d);
        require(params->depth_pitch >= pl_rect_h(rc) * params->row_pitch);
        require(params->depth_pitch % params->row_pitch == 0);
        // fall through
    case 2:
        require(rc.y1 > rc.y0);
        require(rc.y0 >= 0 && rc.y0 <  tex->params.h);
        require(rc.y1 >  0 && rc.y1 <= tex->params.h);
        require(params->row_pitch >= pl_rect_w(rc) * fmt->texel_size);
        require(params->row_pitch % fmt->texel_align == 0);
        // fall through
    case 1:
        require(rc.x1 > rc.x0);
        require(rc.x0 >= 0 && rc.x0 <  tex->params.w);
        require(rc.x1 >  0 && rc.x1 <= tex->params.w);
        break;
    }

    require(!params->buf ^ !params->ptr); // exactly one
    if (params->buf) {
        pl_buf buf = params->buf;
        size_t size = pl_tex_transfer_size(params);
        require(params->buf_offset + size >= params->buf_offset); // overflow check
        require(params->buf_offset + size <= buf->params.size);
        require(gpu->limits.buf_transfer);
    }

    require(!params->callback || gpu->limits.callbacks);
    return true;

error:
    if (tex->params.debug_tag)
        PL_ERR(gpu, "  for texture: %s", tex->params.debug_tag);
    return false;
}

bool pl_tex_upload(pl_gpu gpu, const struct pl_tex_transfer_params *params)
{
    pl_tex tex = params->tex;
    require(tex->params.host_writable);

    struct pl_tex_transfer_params fixed = *params;
    if (!fix_tex_transfer(gpu, &fixed))
        goto error;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->tex_upload(gpu, &fixed);

error:
    if (tex->params.debug_tag)
        PL_ERR(gpu, "  for texture: %s", tex->params.debug_tag);
    return false;
}

bool pl_tex_download(pl_gpu gpu, const struct pl_tex_transfer_params *params)
{
    pl_tex tex = params->tex;
    require(tex->params.host_readable);

    struct pl_tex_transfer_params fixed = *params;
    if (!fix_tex_transfer(gpu, &fixed))
        goto error;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->tex_download(gpu, &fixed);

error:
    if (tex->params.debug_tag)
        PL_ERR(gpu, "  for texture: %s", tex->params.debug_tag);
    return false;
}

bool pl_tex_poll(pl_gpu gpu, pl_tex tex, uint64_t t)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->tex_poll ? impl->tex_poll(gpu, tex, t) : false;
}

pl_buf pl_buf_create(pl_gpu gpu, const struct pl_buf_params *params)
{
    struct pl_buf_params params_rounded;

    require(!params->import_handle || !params->export_handle);
    if (params->export_handle) {
        require(PL_ISPOT(params->export_handle));
        require(params->export_handle & gpu->export_caps.buf);
    }
    if (params->import_handle) {
        require(PL_ISPOT(params->import_handle));
        require(params->import_handle & gpu->import_caps.buf);
        const struct pl_shared_mem *shmem = &params->shared_mem;
        require(shmem->offset + params->size <= shmem->size);
        require(params->import_handle != PL_HANDLE_DMA_BUF || !shmem->drm_format_mod);

        // Fix misalignment on host pointer imports
        if (params->import_handle == PL_HANDLE_HOST_PTR) {
            uintptr_t page_mask = ~(gpu->limits.align_host_ptr - 1);
            uintptr_t ptr_base = (uintptr_t) shmem->handle.ptr & page_mask;
            size_t ptr_offset = (uintptr_t) shmem->handle.ptr - ptr_base;
            size_t buf_offset = ptr_offset + shmem->offset;
            size_t ptr_size = PL_ALIGN2(ptr_offset + shmem->size,
                                        gpu->limits.align_host_ptr);

            if (ptr_base != (uintptr_t) shmem->handle.ptr || ptr_size > shmem->size) {
                static bool warned_rounding = false;
                if (!warned_rounding) {
                    warned_rounding = true;
                    PL_WARN(gpu, "Imported host pointer is not page-aligned. "
                            "This should normally be fine on most platforms, "
                            "but may cause issues in some rare circumstances.");
                }

                PL_TRACE(gpu, "Rounding imported host pointer %p + %zu -> %zu to "
                         "nearest page boundaries: %p + %zu -> %zu",
                          shmem->handle.ptr, shmem->offset, shmem->size,
                          (void *) ptr_base, buf_offset, ptr_size);
            }

            params_rounded = *params;
            params_rounded.shared_mem.handle.ptr = (void *) ptr_base;
            params_rounded.shared_mem.offset = buf_offset;
            params_rounded.shared_mem.size = ptr_size;
            params = &params_rounded;
        }
    }

    require(params->size > 0 && params->size <= gpu->limits.max_buf_size);
    require(!params->uniform || params->size <= gpu->limits.max_ubo_size);
    require(!params->storable || params->size <= gpu->limits.max_ssbo_size);
    require(!params->drawable || params->size <= gpu->limits.max_vbo_size);
    require(!params->host_mapped || params->size <= gpu->limits.max_mapped_size);

    if (params->format) {
        pl_fmt fmt = params->format;
        require(params->size <= gpu->limits.max_buffer_texels * fmt->texel_size);
        require(!params->uniform || (fmt->caps & PL_FMT_CAP_TEXEL_UNIFORM));
        require(!params->storable || (fmt->caps & PL_FMT_CAP_TEXEL_STORAGE));
    }

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    pl_buf buf = impl->buf_create(gpu, params);
    if (buf)
        require(!params->host_mapped || buf->data);

    return buf;

error:
    if (params->debug_tag)
        PL_ERR(gpu, "  for buffer: %s", params->debug_tag);
    return NULL;
}

void pl_buf_destroy(pl_gpu gpu, pl_buf *buf)
{
    if (!*buf)
        return;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->buf_destroy(gpu, *buf);
    *buf = NULL;
}

static bool pl_buf_params_superset(struct pl_buf_params a, struct pl_buf_params b)
{
    return a.size            >= b.size &&
           a.memory_type     == b.memory_type &&
           a.format          == b.format &&
           (a.host_writable  || !b.host_writable) &&
           (a.host_readable  || !b.host_readable) &&
           (a.host_mapped    || !b.host_mapped) &&
           (a.uniform        || !b.uniform) &&
           (a.storable       || !b.storable) &&
           (a.drawable       || !b.drawable);
}

bool pl_buf_recreate(pl_gpu gpu, pl_buf *buf, const struct pl_buf_params *params)
{

    if (params->initial_data) {
        PL_ERR(gpu, "pl_buf_recreate may not be used with `initial_data`!");
        return false;
    }

    if (*buf && pl_buf_params_superset((*buf)->params, *params))
        return true;

    PL_INFO(gpu, "(Re)creating %zu buffer", params->size);
    pl_buf_destroy(gpu, buf);
    *buf = pl_buf_create(gpu, params);

    return !!*buf;
}

void pl_buf_write(pl_gpu gpu, pl_buf buf, size_t buf_offset,
                  const void *data, size_t size)
{
    require(buf->params.host_writable);
    require(buf_offset + size <= buf->params.size);
    require(buf_offset == PL_ALIGN2(buf_offset, 4));

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->buf_write(gpu, buf, buf_offset, data, size);
    return;

error:
    if (buf->params.debug_tag)
        PL_ERR(gpu, "  for buffer: %s", buf->params.debug_tag);
}

bool pl_buf_read(pl_gpu gpu, pl_buf buf, size_t buf_offset,
                 void *dest, size_t size)
{
    require(buf->params.host_readable);
    require(buf_offset + size <= buf->params.size);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->buf_read(gpu, buf, buf_offset, dest, size);

error:
    if (buf->params.debug_tag)
        PL_ERR(gpu, "  for buffer: %s", buf->params.debug_tag);
    return false;
}

void pl_buf_copy(pl_gpu gpu, pl_buf dst, size_t dst_offset,
                 pl_buf src, size_t src_offset, size_t size)
{
    require(src_offset + size <= src->params.size);
    require(dst_offset + size <= dst->params.size);
    require(src != dst);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->buf_copy(gpu, dst, dst_offset, src, src_offset, size);
    return;

error:
    if (src->params.debug_tag || dst->params.debug_tag) {
        PL_ERR(gpu, "  for buffers: src %s, dst %s",
               src->params.debug_tag, dst->params.debug_tag);
    }
}

bool pl_buf_export(pl_gpu gpu, pl_buf buf)
{
    require(buf->params.export_handle || buf->params.import_handle);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->buf_export(gpu, buf);

error:
    if (buf->params.debug_tag)
        PL_ERR(gpu, "  for buffer: %s", buf->params.debug_tag);
    return false;
}

bool pl_buf_poll(pl_gpu gpu, pl_buf buf, uint64_t t)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->buf_poll ? impl->buf_poll(gpu, buf, t) : false;
}

size_t pl_var_type_size(enum pl_var_type type)
{
    switch (type) {
    case PL_VAR_SINT:  return sizeof(int);
    case PL_VAR_UINT:  return sizeof(unsigned int);
    case PL_VAR_FLOAT: return sizeof(float);
    case PL_VAR_INVALID: // fall through
    case PL_VAR_TYPE_COUNT: break;
    }

    pl_unreachable();
}

#define PL_VAR(TYPE, NAME, M, V)                        \
    struct pl_var pl_var_##NAME(const char *name) {     \
        return (struct pl_var) {                        \
            .name  = name,                              \
            .type  = PL_VAR_##TYPE,                     \
            .dim_m = M,                                 \
            .dim_v = V,                                 \
            .dim_a = 1,                                 \
        };                                              \
    }

PL_VAR(FLOAT, float,    1, 1)
PL_VAR(FLOAT, vec2,     1, 2)
PL_VAR(FLOAT, vec3,     1, 3)
PL_VAR(FLOAT, vec4,     1, 4)
PL_VAR(FLOAT, mat2,     2, 2)
PL_VAR(FLOAT, mat2x3,   2, 3)
PL_VAR(FLOAT, mat2x4,   2, 4)
PL_VAR(FLOAT, mat3,     3, 3)
PL_VAR(FLOAT, mat3x4,   3, 4)
PL_VAR(FLOAT, mat4x2,   4, 2)
PL_VAR(FLOAT, mat4x3,   4, 3)
PL_VAR(FLOAT, mat4,     4, 4)
PL_VAR(SINT,  int,      1, 1)
PL_VAR(SINT,  ivec2,    1, 2)
PL_VAR(SINT,  ivec3,    1, 3)
PL_VAR(SINT,  ivec4,    1, 4)
PL_VAR(UINT,  uint,     1, 1)
PL_VAR(UINT,  uvec2,    1, 2)
PL_VAR(UINT,  uvec3,    1, 3)
PL_VAR(UINT,  uvec4,    1, 4)

#undef PL_VAR

const struct pl_named_var pl_var_glsl_types[] = {
    // float vectors
    { "float",  { .type = PL_VAR_FLOAT, .dim_m = 1, .dim_v = 1, .dim_a = 1, }},
    { "vec2",   { .type = PL_VAR_FLOAT, .dim_m = 1, .dim_v = 2, .dim_a = 1, }},
    { "vec3",   { .type = PL_VAR_FLOAT, .dim_m = 1, .dim_v = 3, .dim_a = 1, }},
    { "vec4",   { .type = PL_VAR_FLOAT, .dim_m = 1, .dim_v = 4, .dim_a = 1, }},
    // float matrices
    { "mat2",   { .type = PL_VAR_FLOAT, .dim_m = 2, .dim_v = 2, .dim_a = 1, }},
    { "mat2x3", { .type = PL_VAR_FLOAT, .dim_m = 2, .dim_v = 3, .dim_a = 1, }},
    { "mat2x4", { .type = PL_VAR_FLOAT, .dim_m = 2, .dim_v = 4, .dim_a = 1, }},
    { "mat3",   { .type = PL_VAR_FLOAT, .dim_m = 3, .dim_v = 3, .dim_a = 1, }},
    { "mat3x4", { .type = PL_VAR_FLOAT, .dim_m = 3, .dim_v = 4, .dim_a = 1, }},
    { "mat4x2", { .type = PL_VAR_FLOAT, .dim_m = 4, .dim_v = 2, .dim_a = 1, }},
    { "mat4x3", { .type = PL_VAR_FLOAT, .dim_m = 4, .dim_v = 3, .dim_a = 1, }},
    { "mat4",   { .type = PL_VAR_FLOAT, .dim_m = 4, .dim_v = 4, .dim_a = 1, }},
    // integer vectors
    { "int",    { .type = PL_VAR_SINT,  .dim_m = 1, .dim_v = 1, .dim_a = 1, }},
    { "ivec2",  { .type = PL_VAR_SINT,  .dim_m = 1, .dim_v = 2, .dim_a = 1, }},
    { "ivec3",  { .type = PL_VAR_SINT,  .dim_m = 1, .dim_v = 3, .dim_a = 1, }},
    { "ivec4",  { .type = PL_VAR_SINT,  .dim_m = 1, .dim_v = 4, .dim_a = 1, }},
    // unsigned integer vectors
    { "uint",   { .type = PL_VAR_UINT,  .dim_m = 1, .dim_v = 1, .dim_a = 1, }},
    { "uvec2",  { .type = PL_VAR_UINT,  .dim_m = 1, .dim_v = 2, .dim_a = 1, }},
    { "uvec3",  { .type = PL_VAR_UINT,  .dim_m = 1, .dim_v = 3, .dim_a = 1, }},
    { "uvec4",  { .type = PL_VAR_UINT,  .dim_m = 1, .dim_v = 4, .dim_a = 1, }},

    {0},
};

#define MAX_DIM 4

const char *pl_var_glsl_type_name(struct pl_var var)
{
    static const char *types[PL_VAR_TYPE_COUNT][MAX_DIM+1][MAX_DIM+1] = {
    // float vectors
    [PL_VAR_FLOAT][1][1] = "float",
    [PL_VAR_FLOAT][1][2] = "vec2",
    [PL_VAR_FLOAT][1][3] = "vec3",
    [PL_VAR_FLOAT][1][4] = "vec4",
    // float matrices
    [PL_VAR_FLOAT][2][2] = "mat2",
    [PL_VAR_FLOAT][2][3] = "mat2x3",
    [PL_VAR_FLOAT][2][4] = "mat2x4",
    [PL_VAR_FLOAT][3][2] = "mat3x2",
    [PL_VAR_FLOAT][3][3] = "mat3",
    [PL_VAR_FLOAT][3][4] = "mat3x4",
    [PL_VAR_FLOAT][4][2] = "mat4x2",
    [PL_VAR_FLOAT][4][3] = "mat4x3",
    [PL_VAR_FLOAT][4][4] = "mat4",
    // integer vectors
    [PL_VAR_SINT][1][1] = "int",
    [PL_VAR_SINT][1][2] = "ivec2",
    [PL_VAR_SINT][1][3] = "ivec3",
    [PL_VAR_SINT][1][4] = "ivec4",
    // unsigned integer vectors
    [PL_VAR_UINT][1][1] = "uint",
    [PL_VAR_UINT][1][2] = "uvec2",
    [PL_VAR_UINT][1][3] = "uvec3",
    [PL_VAR_UINT][1][4] = "uvec4",
    };

    if (var.dim_v > MAX_DIM || var.dim_m > MAX_DIM)
        return NULL;

    return types[var.type][var.dim_m][var.dim_v];
}

struct pl_var pl_var_from_fmt(pl_fmt fmt, const char *name)
{
    static const enum pl_var_type vartypes[] = {
        [PL_FMT_FLOAT] = PL_VAR_FLOAT,
        [PL_FMT_UNORM] = PL_VAR_FLOAT,
        [PL_FMT_SNORM] = PL_VAR_FLOAT,
        [PL_FMT_UINT]  = PL_VAR_UINT,
        [PL_FMT_SINT]  = PL_VAR_SINT,
    };

    pl_assert(fmt->type < PL_ARRAY_SIZE(vartypes));
    return (struct pl_var) {
        .type  = vartypes[fmt->type],
        .name  = name,
        .dim_v = fmt->num_components,
        .dim_m = 1,
        .dim_a = 1,
    };
}

struct pl_var_layout pl_var_host_layout(size_t offset, const struct pl_var *var)
{
    size_t col_size = pl_var_type_size(var->type) * var->dim_v;
    return (struct pl_var_layout) {
        .offset = offset,
        .stride = col_size,
        .size   = col_size * var->dim_m * var->dim_a,
    };
}

struct pl_var_layout pl_std140_layout(size_t offset, const struct pl_var *var)
{
    size_t el_size = pl_var_type_size(var->type);

    // std140 packing rules:
    // 1. The size of generic values is their size in bytes
    // 2. The size of vectors is the vector length * the base count
    // 3. Matrices are treated like arrays of column vectors
    // 4. The size of array rows is that of the element size rounded up to
    // the nearest multiple of vec4
    // 5. All values are aligned to a multiple of their size (stride for arrays),
    // with the exception of vec3 which is aligned like vec4
    size_t stride = el_size * var->dim_v;
    size_t align = stride;
    if (var->dim_v == 3)
        align += el_size;
    if (var->dim_m * var->dim_a > 1)
        stride = align = PL_ALIGN2(align, sizeof(float[4]));

    return (struct pl_var_layout) {
        .offset = PL_ALIGN2(offset, align),
        .stride = stride,
        .size   = stride * var->dim_m * var->dim_a,
    };
}

struct pl_var_layout pl_std430_layout(size_t offset, const struct pl_var *var)
{
    size_t el_size = pl_var_type_size(var->type);

    // std430 packing rules: like std140, except arrays/matrices are always
    // "tightly" packed, even arrays/matrices of vec3s
    size_t stride = el_size * var->dim_v;
    size_t align = stride;
    if (var->dim_v == 3)
        align += el_size;
    if (var->dim_m * var->dim_a > 1)
        stride = align;

    return (struct pl_var_layout) {
        .offset = PL_ALIGN2(offset, align),
        .stride = stride,
        .size   = stride * var->dim_m * var->dim_a,
    };
}

void memcpy_layout(void *dst_p, struct pl_var_layout dst_layout,
                   const void *src_p, struct pl_var_layout src_layout)
{
    uintptr_t src = (uintptr_t) src_p + src_layout.offset;
    uintptr_t dst = (uintptr_t) dst_p + dst_layout.offset;

    if (src_layout.stride == dst_layout.stride) {
        pl_assert(dst_layout.size == src_layout.size);
        memcpy((void *) dst, (const void *) src, src_layout.size);
        return;
    }

    size_t stride = PL_MIN(src_layout.stride, dst_layout.stride);
    uintptr_t end = src + src_layout.size;
    while (src < end) {
        pl_assert(dst < dst + dst_layout.size);
        memcpy((void *) dst, (const void *) src, stride);
        src += src_layout.stride;
        dst += dst_layout.stride;
    }
}

int pl_desc_namespace(pl_gpu gpu, enum pl_desc_type type)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    int ret = impl->desc_namespace(gpu, type);
    pl_assert(ret >= 0 && ret < PL_DESC_TYPE_COUNT);
    return ret;
}

const char *pl_desc_access_glsl_name(enum pl_desc_access mode)
{
    switch (mode) {
    case PL_DESC_ACCESS_READWRITE: return "";
    case PL_DESC_ACCESS_READONLY:  return "readonly";
    case PL_DESC_ACCESS_WRITEONLY: return "writeonly";
    case PL_DESC_ACCESS_COUNT: break;
    }

    pl_unreachable();
}

const struct pl_blend_params pl_alpha_overlay = {
    .src_rgb    = PL_BLEND_SRC_ALPHA,
    .dst_rgb    = PL_BLEND_ONE_MINUS_SRC_ALPHA,
    .src_alpha  = PL_BLEND_ONE,
    .dst_alpha  = PL_BLEND_ONE_MINUS_SRC_ALPHA,
};

static inline void log_shader_sources(pl_log log, enum pl_log_level level,
                                      const struct pl_pass_params *params)
{
    if (!pl_msg_test(log, level) || !params->glsl_shader)
        return;

    switch (params->type) {
    case PL_PASS_RASTER:
        if (!params->vertex_shader)
            return;
        pl_msg(log, level, "vertex shader source:");
        pl_msg_source(log, level, params->vertex_shader);
        pl_msg(log, level, "fragment shader source:");
        pl_msg_source(log, level, params->glsl_shader);
        return;

    case PL_PASS_COMPUTE:
        pl_msg(log, level, "compute shader source:");
        pl_msg_source(log, level, params->glsl_shader);
        return;

    case PL_PASS_INVALID:
    case PL_PASS_TYPE_COUNT:
        break;
    }

    pl_unreachable();
}

static void log_spec_constants(pl_log log, enum pl_log_level lev,
                               const struct pl_pass_params *params,
                               const void *constant_data)
{
    if (!constant_data || !params->num_constants || !pl_msg_test(log, lev))
        return;

    pl_msg(log, lev, "Specialization constant values:");

    uintptr_t data_base = (uintptr_t) constant_data;
    for (int i = 0; i < params->num_constants; i++) {
        union {
            int i;
            unsigned u;
            float f;
        } *data = (void *) (data_base + params->constants[i].offset);
        int id = params->constants[i].id;

        switch (params->constants[i].type) {
        case PL_VAR_SINT:  pl_msg(log, lev, "  constant_id=%d: %d", id, data->i); break;
        case PL_VAR_UINT:  pl_msg(log, lev, "  constant_id=%d: %u", id, data->u); break;
        case PL_VAR_FLOAT: pl_msg(log, lev, "  constant_id=%d: %f", id, data->f); break;
        default: pl_unreachable();
        }
    }
}

pl_pass pl_pass_create(pl_gpu gpu, const struct pl_pass_params *params)
{
    require(params->glsl_shader);
    switch(params->type) {
    case PL_PASS_RASTER:
        require(params->vertex_shader);
        require(params->vertex_stride % gpu->limits.align_vertex_stride == 0);
        for (int i = 0; i < params->num_vertex_attribs; i++) {
            struct pl_vertex_attrib va = params->vertex_attribs[i];
            require(va.name);
            require(va.fmt);
            require(va.fmt->caps & PL_FMT_CAP_VERTEX);
            require(va.offset + va.fmt->texel_size <= params->vertex_stride);
        }

        require(params->target_format);
        require(params->target_format->caps & PL_FMT_CAP_RENDERABLE);
        require(!params->blend_params || params->target_format->caps & PL_FMT_CAP_BLENDABLE);
        require(!params->blend_params || params->load_target);
        break;
    case PL_PASS_COMPUTE:
        require(gpu->glsl.compute);
        break;
    case PL_PASS_INVALID:
    case PL_PASS_TYPE_COUNT:
        pl_unreachable();
    }

    size_t num_var_comps = 0;
    for (int i = 0; i < params->num_variables; i++) {
        struct pl_var var = params->variables[i];
        num_var_comps += var.dim_v * var.dim_m * var.dim_a;
        require(var.name);
        require(pl_var_glsl_type_name(var));
    }
    require(num_var_comps <= gpu->limits.max_variable_comps);

    require(params->num_constants <= gpu->limits.max_constants);
    for (int i = 0; i < params->num_constants; i++)
        require(params->constants[i].type);

    for (int i = 0; i < params->num_descriptors; i++) {
        struct pl_desc desc = params->descriptors[i];
        require(desc.name);

        // enforce disjoint descriptor bindings for each namespace
        int namespace = pl_desc_namespace(gpu, desc.type);
        for (int j = i+1; j < params->num_descriptors; j++) {
            struct pl_desc other = params->descriptors[j];
            require(desc.binding != other.binding ||
                    namespace != pl_desc_namespace(gpu, other.type));
        }
    }

    require(params->push_constants_size <= gpu->limits.max_pushc_size);
    require(params->push_constants_size == PL_ALIGN2(params->push_constants_size, 4));

    log_shader_sources(gpu->log, PL_LOG_DEBUG, params);
    log_spec_constants(gpu->log, PL_LOG_DEBUG, params, params->constant_data);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    pl_pass pass = impl->pass_create(gpu, params);
    if (!pass)
        goto error;

    return pass;

error:
    log_shader_sources(gpu->log, PL_LOG_ERR, params);
    pl_log_stack_trace(gpu->log, PL_LOG_ERR);
    pl_debug_abort();
    return NULL;
}

void pl_pass_destroy(pl_gpu gpu, pl_pass *pass)
{
    if (!*pass)
        return;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->pass_destroy(gpu, *pass);
    *pass = NULL;
}

void pl_pass_run(pl_gpu gpu, const struct pl_pass_run_params *params)
{
    pl_pass pass = params->pass;
    struct pl_pass_run_params new = *params;

    for (int i = 0; i < pass->params.num_descriptors; i++) {
        struct pl_desc desc = pass->params.descriptors[i];
        struct pl_desc_binding db = params->desc_bindings[i];
        require(db.object);
        switch (desc.type) {
        case PL_DESC_SAMPLED_TEX: {
            pl_tex tex = db.object;
            pl_fmt fmt = tex->params.format;
            require(tex->params.sampleable);
            require(db.sample_mode != PL_TEX_SAMPLE_LINEAR || (fmt->caps & PL_FMT_CAP_LINEAR));
            break;
        }
        case PL_DESC_STORAGE_IMG: {
            pl_tex tex = db.object;
            pl_fmt fmt = tex->params.format;
            require(tex->params.storable);
            require(desc.access != PL_DESC_ACCESS_READWRITE || (fmt->caps & PL_FMT_CAP_READWRITE));
            break;
        }
        case PL_DESC_BUF_UNIFORM: {
            pl_buf buf = db.object;
            require(buf->params.uniform);
            break;
        }
        case PL_DESC_BUF_STORAGE: {
            pl_buf buf = db.object;
            require(buf->params.storable);
            break;
        }
        case PL_DESC_BUF_TEXEL_UNIFORM: {
            pl_buf buf = db.object;
            require(buf->params.uniform && buf->params.format);
            break;
        }
        case PL_DESC_BUF_TEXEL_STORAGE: {
            pl_buf buf = db.object;
            pl_fmt fmt = buf->params.format;
            require(buf->params.storable && buf->params.format);
            require(desc.access != PL_DESC_ACCESS_READWRITE || (fmt->caps & PL_FMT_CAP_READWRITE));
            break;
        }
        case PL_DESC_INVALID:
        case PL_DESC_TYPE_COUNT:
            pl_unreachable();
        }
    }

    for (int i = 0; i < params->num_var_updates; i++) {
        struct pl_var_update vu = params->var_updates[i];
        require(vu.index >= 0 && vu.index < pass->params.num_variables);
        require(vu.data);
    }

    require(params->push_constants || !pass->params.push_constants_size);

    switch (pass->params.type) {
    case PL_PASS_RASTER: {
        switch (pass->params.vertex_type) {
        case PL_PRIM_TRIANGLE_LIST:
            require(params->vertex_count % 3 == 0);
            // fall through
        case PL_PRIM_TRIANGLE_STRIP:
            require(params->vertex_count >= 3);
            break;
        case PL_PRIM_TYPE_COUNT:
            pl_unreachable();
        }

        require(!params->vertex_data ^ !params->vertex_buf);
        if (params->vertex_buf) {
            pl_buf vertex_buf = params->vertex_buf;
            require(vertex_buf->params.drawable);
            if (!params->index_data && !params->index_buf) {
                // Cannot bounds check indexed draws
                size_t vert_size = params->vertex_count * pass->params.vertex_stride;
                require(params->buf_offset + vert_size <= vertex_buf->params.size);
            }
        }

        require(!params->index_data || !params->index_buf);
        if (params->index_buf) {
            pl_buf index_buf = params->index_buf;
            require(!params->vertex_data);
            require(index_buf->params.drawable);
            size_t index_size = pl_index_buf_size(params);
            require(params->index_offset + index_size <= index_buf->params.size);
        }

        pl_tex target = params->target;
        require(target);
        require(pl_tex_params_dimension(target->params) == 2);
        require(target->params.format->signature == pass->params.target_format->signature);
        require(target->params.renderable);
        pl_rect2d *vp = &new.viewport;
        pl_rect2d *sc = &new.scissors;

        // Sanitize viewport/scissors
        if (!vp->x0 && !vp->x1)
            vp->x1 = target->params.w;
        if (!vp->y0 && !vp->y1)
            vp->y1 = target->params.h;

        if (!sc->x0 && !sc->x1)
            sc->x1 = target->params.w;
        if (!sc->y0 && !sc->y1)
            sc->y1 = target->params.h;

        // Constrain the scissors to the target dimension (to sanitize the
        // underlying graphics API calls)
        sc->x0 = PL_CLAMP(sc->x0, 0, target->params.w);
        sc->y0 = PL_CLAMP(sc->y0, 0, target->params.h);
        sc->x1 = PL_CLAMP(sc->x1, 0, target->params.w);
        sc->y1 = PL_CLAMP(sc->y1, 0, target->params.h);

        // Scissors wholly outside target -> silently drop pass (also needed
        // to ensure we don't cause UB by specifying invalid scissors)
        if (!pl_rect_w(*sc) || !pl_rect_h(*sc))
            return;

        require(pl_rect_w(*vp) > 0);
        require(pl_rect_h(*vp) > 0);
        require(pl_rect_w(*sc) > 0);
        require(pl_rect_h(*sc) > 0);

        if (!pass->params.load_target)
            pl_tex_invalidate(gpu, target);
        break;
    }
    case PL_PASS_COMPUTE:
        for (int i = 0; i < PL_ARRAY_SIZE(params->compute_groups); i++) {
            require(params->compute_groups[i] >= 0);
            require(params->compute_groups[i] <= gpu->limits.max_dispatch[i]);
        }
        break;
    case PL_PASS_INVALID:
    case PL_PASS_TYPE_COUNT:
        pl_unreachable();
    }

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->pass_run(gpu, &new);

error:
    return;
}

void pl_gpu_flush(pl_gpu gpu)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    if (impl->gpu_flush)
        impl->gpu_flush(gpu);
}

void pl_gpu_finish(pl_gpu gpu)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->gpu_finish(gpu);
}

bool pl_gpu_is_failed(pl_gpu gpu)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    if (!impl->gpu_is_failed)
        return false;

    return impl->gpu_is_failed(gpu);
}

pl_sync pl_sync_create(pl_gpu gpu, enum pl_handle_type handle_type)
{
    require(handle_type);
    require(handle_type & gpu->export_caps.sync);
    require(PL_ISPOT(handle_type));

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->sync_create(gpu, handle_type);

error:
    return NULL;
}

void pl_sync_destroy(pl_gpu gpu, pl_sync *sync)
{
    if (!*sync)
        return;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->sync_destroy(gpu, *sync);
    *sync = NULL;
}

bool pl_tex_export(pl_gpu gpu, pl_tex tex, pl_sync sync)
{
    require(tex->params.import_handle || tex->params.export_handle);

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->tex_export(gpu, tex, sync);

error:
    if (tex->params.debug_tag)
        PL_ERR(gpu, "  for texture: %s", tex->params.debug_tag);
    return false;
}

pl_timer pl_timer_create(pl_gpu gpu)
{
    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    if (!impl->timer_create)
        return NULL;

    return impl->timer_create(gpu);
}

void pl_timer_destroy(pl_gpu gpu, pl_timer *timer)
{
    if (!*timer)
        return;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    impl->timer_destroy(gpu, *timer);
    *timer = NULL;
}

uint64_t pl_timer_query(pl_gpu gpu, pl_timer timer)
{
    if (!timer)
        return 0;

    const struct pl_gpu_fns *impl = PL_PRIV(gpu);
    return impl->timer_query(gpu, timer);
}