From 26a029d407be480d791972afb5975cf62c9360a6 Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Fri, 19 Apr 2024 02:47:55 +0200 Subject: Adding upstream version 124.0.1. Signed-off-by: Daniel Baumann --- third_party/jpeg-xl/lib/jxl/base/arch_macros.h | 33 ++ third_party/jpeg-xl/lib/jxl/base/bits.h | 147 +++++++ third_party/jpeg-xl/lib/jxl/base/byte_order.h | 274 +++++++++++++ .../jpeg-xl/lib/jxl/base/c_callback_support.h | 32 ++ third_party/jpeg-xl/lib/jxl/base/common.h | 95 +++++ .../jpeg-xl/lib/jxl/base/compiler_specific.h | 157 +++++++ third_party/jpeg-xl/lib/jxl/base/data_parallel.h | 124 ++++++ third_party/jpeg-xl/lib/jxl/base/exif.h | 91 ++++ third_party/jpeg-xl/lib/jxl/base/fast_math-inl.h | 236 +++++++++++ third_party/jpeg-xl/lib/jxl/base/float.h | 102 +++++ third_party/jpeg-xl/lib/jxl/base/iaca.h | 65 +++ third_party/jpeg-xl/lib/jxl/base/matrix_ops.h | 84 ++++ third_party/jpeg-xl/lib/jxl/base/os_macros.h | 50 +++ third_party/jpeg-xl/lib/jxl/base/override.h | 29 ++ third_party/jpeg-xl/lib/jxl/base/printf_macros.h | 34 ++ third_party/jpeg-xl/lib/jxl/base/random.h | 99 +++++ .../jpeg-xl/lib/jxl/base/rational_polynomial-inl.h | 102 +++++ .../jpeg-xl/lib/jxl/base/sanitizer_definitions.h | 44 ++ third_party/jpeg-xl/lib/jxl/base/scope_guard.h | 48 +++ third_party/jpeg-xl/lib/jxl/base/span.h | 80 ++++ third_party/jpeg-xl/lib/jxl/base/status.h | 456 +++++++++++++++++++++ 21 files changed, 2382 insertions(+) create mode 100644 third_party/jpeg-xl/lib/jxl/base/arch_macros.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/bits.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/byte_order.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/c_callback_support.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/common.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/compiler_specific.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/data_parallel.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/exif.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/fast_math-inl.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/float.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/iaca.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/matrix_ops.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/os_macros.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/override.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/printf_macros.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/random.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/rational_polynomial-inl.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/sanitizer_definitions.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/scope_guard.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/span.h create mode 100644 third_party/jpeg-xl/lib/jxl/base/status.h (limited to 'third_party/jpeg-xl/lib/jxl/base') diff --git a/third_party/jpeg-xl/lib/jxl/base/arch_macros.h b/third_party/jpeg-xl/lib/jxl/base/arch_macros.h new file mode 100644 index 0000000000..a98301915e --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/arch_macros.h @@ -0,0 +1,33 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_ARCH_MACROS_H_ +#define LIB_JXL_BASE_ARCH_MACROS_H_ + +// Defines the JXL_ARCH_* macros. + +namespace jxl { + +#if defined(__x86_64__) || defined(_M_X64) +#define JXL_ARCH_X64 1 +#else +#define JXL_ARCH_X64 0 +#endif + +#if defined(__powerpc64__) || defined(_M_PPC) +#define JXL_ARCH_PPC 1 +#else +#define JXL_ARCH_PPC 0 +#endif + +#if defined(__aarch64__) || defined(__arm__) +#define JXL_ARCH_ARM 1 +#else +#define JXL_ARCH_ARM 0 +#endif + +} // namespace jxl + +#endif // LIB_JXL_BASE_ARCH_MACROS_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/bits.h b/third_party/jpeg-xl/lib/jxl/base/bits.h new file mode 100644 index 0000000000..9f86118e72 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/bits.h @@ -0,0 +1,147 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_BITS_H_ +#define LIB_JXL_BASE_BITS_H_ + +// Specialized instructions for processing register-sized bit arrays. + +#include "lib/jxl/base/compiler_specific.h" +#include "lib/jxl/base/status.h" + +#if JXL_COMPILER_MSVC +#include +#endif + +#include +#include + +namespace jxl { + +// Empty struct used as a size tag type. +template +struct SizeTag {}; + +template +constexpr bool IsSigned() { + return T(0) > T(-1); +} + +// Undefined results for x == 0. +static JXL_INLINE JXL_MAYBE_UNUSED size_t +Num0BitsAboveMS1Bit_Nonzero(SizeTag<4> /* tag */, const uint32_t x) { + JXL_DASSERT(x != 0); +#if JXL_COMPILER_MSVC + unsigned long index; + _BitScanReverse(&index, x); + return 31 - index; +#else + return static_cast(__builtin_clz(x)); +#endif +} +static JXL_INLINE JXL_MAYBE_UNUSED size_t +Num0BitsAboveMS1Bit_Nonzero(SizeTag<8> /* tag */, const uint64_t x) { + JXL_DASSERT(x != 0); +#if JXL_COMPILER_MSVC +#if JXL_ARCH_X64 + unsigned long index; + _BitScanReverse64(&index, x); + return 63 - index; +#else // JXL_ARCH_X64 + // _BitScanReverse64 not available + uint32_t msb = static_cast(x >> 32u); + unsigned long index; + if (msb == 0) { + uint32_t lsb = static_cast(x & 0xFFFFFFFF); + _BitScanReverse(&index, lsb); + return 63 - index; + } else { + _BitScanReverse(&index, msb); + return 31 - index; + } +#endif // JXL_ARCH_X64 +#else + return static_cast(__builtin_clzll(x)); +#endif +} +template +static JXL_INLINE JXL_MAYBE_UNUSED size_t +Num0BitsAboveMS1Bit_Nonzero(const T x) { + static_assert(!IsSigned(), "Num0BitsAboveMS1Bit_Nonzero: use unsigned"); + return Num0BitsAboveMS1Bit_Nonzero(SizeTag(), x); +} + +// Undefined results for x == 0. +static JXL_INLINE JXL_MAYBE_UNUSED size_t +Num0BitsBelowLS1Bit_Nonzero(SizeTag<4> /* tag */, const uint32_t x) { + JXL_DASSERT(x != 0); +#if JXL_COMPILER_MSVC + unsigned long index; + _BitScanForward(&index, x); + return index; +#else + return static_cast(__builtin_ctz(x)); +#endif +} +static JXL_INLINE JXL_MAYBE_UNUSED size_t +Num0BitsBelowLS1Bit_Nonzero(SizeTag<8> /* tag */, const uint64_t x) { + JXL_DASSERT(x != 0); +#if JXL_COMPILER_MSVC +#if JXL_ARCH_X64 + unsigned long index; + _BitScanForward64(&index, x); + return index; +#else // JXL_ARCH_64 + // _BitScanForward64 not available + uint32_t lsb = static_cast(x & 0xFFFFFFFF); + unsigned long index; + if (lsb == 0) { + uint32_t msb = static_cast(x >> 32u); + _BitScanForward(&index, msb); + return 32 + index; + } else { + _BitScanForward(&index, lsb); + return index; + } +#endif // JXL_ARCH_X64 +#else + return static_cast(__builtin_ctzll(x)); +#endif +} +template +static JXL_INLINE JXL_MAYBE_UNUSED size_t Num0BitsBelowLS1Bit_Nonzero(T x) { + static_assert(!IsSigned(), "Num0BitsBelowLS1Bit_Nonzero: use unsigned"); + return Num0BitsBelowLS1Bit_Nonzero(SizeTag(), x); +} + +// Returns bit width for x == 0. +template +static JXL_INLINE JXL_MAYBE_UNUSED size_t Num0BitsAboveMS1Bit(const T x) { + return (x == 0) ? sizeof(T) * 8 : Num0BitsAboveMS1Bit_Nonzero(x); +} + +// Returns bit width for x == 0. +template +static JXL_INLINE JXL_MAYBE_UNUSED size_t Num0BitsBelowLS1Bit(const T x) { + return (x == 0) ? sizeof(T) * 8 : Num0BitsBelowLS1Bit_Nonzero(x); +} + +// Returns base-2 logarithm, rounded down. +template +static JXL_INLINE JXL_MAYBE_UNUSED size_t FloorLog2Nonzero(const T x) { + return (sizeof(T) * 8 - 1) ^ Num0BitsAboveMS1Bit_Nonzero(x); +} + +// Returns base-2 logarithm, rounded up. +template +static JXL_INLINE JXL_MAYBE_UNUSED size_t CeilLog2Nonzero(const T x) { + const size_t floor_log2 = FloorLog2Nonzero(x); + if ((x & (x - 1)) == 0) return floor_log2; // power of two + return floor_log2 + 1; +} + +} // namespace jxl + +#endif // LIB_JXL_BASE_BITS_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/byte_order.h b/third_party/jpeg-xl/lib/jxl/base/byte_order.h new file mode 100644 index 0000000000..8966834e08 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/byte_order.h @@ -0,0 +1,274 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_BYTE_ORDER_H_ +#define LIB_JXL_BASE_BYTE_ORDER_H_ + +#include +#include +#include // memcpy + +#include "lib/jxl/base/compiler_specific.h" + +#if JXL_COMPILER_MSVC +#include // _byteswap_* +#endif + +#if (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) +#define JXL_BYTE_ORDER_LITTLE 1 +#else +// This means that we don't know that the byte order is little endian, in +// this case we use endian-neutral code that works for both little- and +// big-endian. +#define JXL_BYTE_ORDER_LITTLE 0 +#endif + +// Returns whether the system is little-endian (least-significant byte first). +#if JXL_BYTE_ORDER_LITTLE +static constexpr bool IsLittleEndian() { return true; } +#else +static inline bool IsLittleEndian() { + const uint32_t multibyte = 1; + uint8_t byte; + memcpy(&byte, &multibyte, 1); + return byte == 1; +} +#endif + +static inline bool SwapEndianness(JxlEndianness endianness) { + return ((endianness == JXL_BIG_ENDIAN && IsLittleEndian()) || + (endianness == JXL_LITTLE_ENDIAN && !IsLittleEndian())); +} + +#if JXL_COMPILER_MSVC +#define JXL_BSWAP16(x) _byteswap_ushort(x) +#define JXL_BSWAP32(x) _byteswap_ulong(x) +#define JXL_BSWAP64(x) _byteswap_uint64(x) +#else +#define JXL_BSWAP16(x) __builtin_bswap16(x) +#define JXL_BSWAP32(x) __builtin_bswap32(x) +#define JXL_BSWAP64(x) __builtin_bswap64(x) +#endif + +static JXL_INLINE uint32_t LoadBE16(const uint8_t* p) { + const uint32_t byte1 = p[0]; + const uint32_t byte0 = p[1]; + return (byte1 << 8) | byte0; +} + +static JXL_INLINE uint32_t LoadLE16(const uint8_t* p) { + const uint32_t byte0 = p[0]; + const uint32_t byte1 = p[1]; + return (byte1 << 8) | byte0; +} + +static JXL_INLINE uint32_t LoadBE32(const uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + uint32_t big; + memcpy(&big, p, 4); + return JXL_BSWAP32(big); +#else + // Byte-order-independent - can't assume this machine is big endian. + const uint32_t byte3 = p[0]; + const uint32_t byte2 = p[1]; + const uint32_t byte1 = p[2]; + const uint32_t byte0 = p[3]; + return (byte3 << 24) | (byte2 << 16) | (byte1 << 8) | byte0; +#endif +} + +static JXL_INLINE uint64_t LoadBE64(const uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + uint64_t big; + memcpy(&big, p, 8); + return JXL_BSWAP64(big); +#else + // Byte-order-independent - can't assume this machine is big endian. + const uint64_t byte7 = p[0]; + const uint64_t byte6 = p[1]; + const uint64_t byte5 = p[2]; + const uint64_t byte4 = p[3]; + const uint64_t byte3 = p[4]; + const uint64_t byte2 = p[5]; + const uint64_t byte1 = p[6]; + const uint64_t byte0 = p[7]; + return (byte7 << 56ull) | (byte6 << 48ull) | (byte5 << 40ull) | + (byte4 << 32ull) | (byte3 << 24ull) | (byte2 << 16ull) | + (byte1 << 8ull) | byte0; +#endif +} + +static JXL_INLINE uint32_t LoadLE32(const uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + uint32_t little; + memcpy(&little, p, 4); + return little; +#else + // Byte-order-independent - can't assume this machine is big endian. + const uint32_t byte0 = p[0]; + const uint32_t byte1 = p[1]; + const uint32_t byte2 = p[2]; + const uint32_t byte3 = p[3]; + return (byte3 << 24) | (byte2 << 16) | (byte1 << 8) | byte0; +#endif +} + +static JXL_INLINE uint64_t LoadLE64(const uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + uint64_t little; + memcpy(&little, p, 8); + return little; +#else + // Byte-order-independent - can't assume this machine is big endian. + const uint64_t byte0 = p[0]; + const uint64_t byte1 = p[1]; + const uint64_t byte2 = p[2]; + const uint64_t byte3 = p[3]; + const uint64_t byte4 = p[4]; + const uint64_t byte5 = p[5]; + const uint64_t byte6 = p[6]; + const uint64_t byte7 = p[7]; + return (byte7 << 56) | (byte6 << 48) | (byte5 << 40) | (byte4 << 32) | + (byte3 << 24) | (byte2 << 16) | (byte1 << 8) | byte0; +#endif +} + +// Loads a Big-Endian float +static JXL_INLINE float LoadBEFloat(const uint8_t* p) { + uint32_t u = LoadBE32(p); + float result; + memcpy(&result, &u, 4); + return result; +} + +// Loads a Little-Endian float +static JXL_INLINE float LoadLEFloat(const uint8_t* p) { + uint32_t u = LoadLE32(p); + float result; + memcpy(&result, &u, 4); + return result; +} + +static JXL_INLINE void StoreBE16(const uint32_t native, uint8_t* p) { + p[0] = (native >> 8) & 0xFF; + p[1] = native & 0xFF; +} + +static JXL_INLINE void StoreLE16(const uint32_t native, uint8_t* p) { + p[1] = (native >> 8) & 0xFF; + p[0] = native & 0xFF; +} + +static JXL_INLINE void StoreBE32(const uint32_t native, uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + const uint32_t big = JXL_BSWAP32(native); + memcpy(p, &big, 4); +#else + // Byte-order-independent - can't assume this machine is big endian. + p[0] = native >> 24; + p[1] = (native >> 16) & 0xFF; + p[2] = (native >> 8) & 0xFF; + p[3] = native & 0xFF; +#endif +} + +static JXL_INLINE void StoreBE64(const uint64_t native, uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + const uint64_t big = JXL_BSWAP64(native); + memcpy(p, &big, 8); +#else + // Byte-order-independent - can't assume this machine is big endian. + p[0] = native >> 56ull; + p[1] = (native >> 48ull) & 0xFF; + p[2] = (native >> 40ull) & 0xFF; + p[3] = (native >> 32ull) & 0xFF; + p[4] = (native >> 24ull) & 0xFF; + p[5] = (native >> 16ull) & 0xFF; + p[6] = (native >> 8ull) & 0xFF; + p[7] = native & 0xFF; +#endif +} + +static JXL_INLINE void StoreLE32(const uint32_t native, uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + const uint32_t little = native; + memcpy(p, &little, 4); +#else + // Byte-order-independent - can't assume this machine is big endian. + p[3] = native >> 24; + p[2] = (native >> 16) & 0xFF; + p[1] = (native >> 8) & 0xFF; + p[0] = native & 0xFF; +#endif +} + +static JXL_INLINE void StoreLE64(const uint64_t native, uint8_t* p) { +#if JXL_BYTE_ORDER_LITTLE + const uint64_t little = native; + memcpy(p, &little, 8); +#else + // Byte-order-independent - can't assume this machine is big endian. + p[7] = native >> 56; + p[6] = (native >> 48) & 0xFF; + p[5] = (native >> 40) & 0xFF; + p[4] = (native >> 32) & 0xFF; + p[3] = (native >> 24) & 0xFF; + p[2] = (native >> 16) & 0xFF; + p[1] = (native >> 8) & 0xFF; + p[0] = native & 0xFF; +#endif +} + +static JXL_INLINE float BSwapFloat(float x) { + uint32_t u; + memcpy(&u, &x, 4); + uint32_t uswap = JXL_BSWAP32(u); + float xswap; + memcpy(&xswap, &uswap, 4); + return xswap; +} + +// Big/Little Endian order. +struct OrderBE {}; +struct OrderLE {}; + +// Wrappers for calling from generic code. +static JXL_INLINE void Store16(OrderBE /*tag*/, const uint32_t native, + uint8_t* p) { + return StoreBE16(native, p); +} + +static JXL_INLINE void Store16(OrderLE /*tag*/, const uint32_t native, + uint8_t* p) { + return StoreLE16(native, p); +} + +static JXL_INLINE void Store32(OrderBE /*tag*/, const uint32_t native, + uint8_t* p) { + return StoreBE32(native, p); +} + +static JXL_INLINE void Store32(OrderLE /*tag*/, const uint32_t native, + uint8_t* p) { + return StoreLE32(native, p); +} + +static JXL_INLINE uint32_t Load16(OrderBE /*tag*/, const uint8_t* p) { + return LoadBE16(p); +} + +static JXL_INLINE uint32_t Load16(OrderLE /*tag*/, const uint8_t* p) { + return LoadLE16(p); +} + +static JXL_INLINE uint32_t Load32(OrderBE /*tag*/, const uint8_t* p) { + return LoadBE32(p); +} + +static JXL_INLINE uint32_t Load32(OrderLE /*tag*/, const uint8_t* p) { + return LoadLE32(p); +} + +#endif // LIB_JXL_BASE_BYTE_ORDER_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/c_callback_support.h b/third_party/jpeg-xl/lib/jxl/base/c_callback_support.h new file mode 100644 index 0000000000..aee0ce5346 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/c_callback_support.h @@ -0,0 +1,32 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_C_CALLBACK_SUPPORT_H_ +#define LIB_JXL_BASE_C_CALLBACK_SUPPORT_H_ + +#include + +namespace jxl { +namespace detail { + +template +struct MethodToCCallbackHelper {}; + +template +struct MethodToCCallbackHelper { + template + static R Call(void *opaque, Args... args) { + return (reinterpret_cast(opaque)->*method)( + std::forward(args)...); + } +}; + +} // namespace detail +} // namespace jxl + +#define METHOD_TO_C_CALLBACK(method) \ + ::jxl::detail::MethodToCCallbackHelper::Call + +#endif // LIB_JXL_BASE_C_CALLBACK_SUPPORT_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/common.h b/third_party/jpeg-xl/lib/jxl/base/common.h new file mode 100644 index 0000000000..b7fe6ab0bc --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/common.h @@ -0,0 +1,95 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_COMMON_H_ +#define LIB_JXL_BASE_COMMON_H_ + +// Shared constants and helper functions. + +#include +#include +#include +#include +#include + +#include "lib/jxl/base/compiler_specific.h" + +namespace jxl { +// Some enums and typedefs used by more than one header file. + +constexpr size_t kBitsPerByte = 8; // more clear than CHAR_BIT + +constexpr inline size_t RoundUpBitsToByteMultiple(size_t bits) { + return (bits + 7) & ~size_t(7); +} + +constexpr inline size_t RoundUpToBlockDim(size_t dim) { + return (dim + 7) & ~size_t(7); +} + +static inline bool JXL_MAYBE_UNUSED SafeAdd(const uint64_t a, const uint64_t b, + uint64_t& sum) { + sum = a + b; + return sum >= a; // no need to check b - either sum >= both or < both. +} + +template +constexpr inline T1 DivCeil(T1 a, T2 b) { + return (a + b - 1) / b; +} + +// Works for any `align`; if a power of two, compiler emits ADD+AND. +constexpr inline size_t RoundUpTo(size_t what, size_t align) { + return DivCeil(what, align) * align; +} + +constexpr double kPi = 3.14159265358979323846264338327950288; + +// Reasonable default for sRGB, matches common monitors. We map white to this +// many nits (cd/m^2) by default. Butteraugli was tuned for 250 nits, which is +// very close. +// NB: This constant is not very "base", but it is shared between modules. +static constexpr float kDefaultIntensityTarget = 255; + +template +constexpr T Pi(T multiplier) { + return static_cast(multiplier * kPi); +} + +// Prior to C++14 (i.e. C++11): provide our own make_unique +#if __cplusplus < 201402L +template +std::unique_ptr make_unique(Args&&... args) { + return std::unique_ptr(new T(std::forward(args)...)); +} +#else +using std::make_unique; +#endif + +template +JXL_INLINE T Clamp1(T val, T low, T hi) { + return val < low ? low : val > hi ? hi : val; +} + +// conversion from integer to string. +template +std::string ToString(T n) { + char data[32] = {}; + if (T(0.1) != T(0)) { + // float + snprintf(data, sizeof(data), "%g", static_cast(n)); + } else if (T(-1) > T(0)) { + // unsigned + snprintf(data, sizeof(data), "%llu", static_cast(n)); + } else { + // signed + snprintf(data, sizeof(data), "%lld", static_cast(n)); + } + return data; +} + +} // namespace jxl + +#endif // LIB_JXL_BASE_COMMON_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/compiler_specific.h b/third_party/jpeg-xl/lib/jxl/base/compiler_specific.h new file mode 100644 index 0000000000..702ff8e058 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/compiler_specific.h @@ -0,0 +1,157 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_COMPILER_SPECIFIC_H_ +#define LIB_JXL_BASE_COMPILER_SPECIFIC_H_ + +// Macros for compiler version + nonstandard keywords, e.g. __builtin_expect. + +#include +#include + +#include "lib/jxl/base/sanitizer_definitions.h" + +// #if is shorter and safer than #ifdef. *_VERSION are zero if not detected, +// otherwise 100 * major + minor version. Note that other packages check for +// #ifdef COMPILER_MSVC, so we cannot use that same name. + +#ifdef _MSC_VER +#define JXL_COMPILER_MSVC _MSC_VER +#else +#define JXL_COMPILER_MSVC 0 +#endif + +#ifdef __GNUC__ +#define JXL_COMPILER_GCC (__GNUC__ * 100 + __GNUC_MINOR__) +#else +#define JXL_COMPILER_GCC 0 +#endif + +#ifdef __clang__ +#define JXL_COMPILER_CLANG (__clang_major__ * 100 + __clang_minor__) +// Clang pretends to be GCC for compatibility. +#undef JXL_COMPILER_GCC +#define JXL_COMPILER_GCC 0 +#else +#define JXL_COMPILER_CLANG 0 +#endif + +#if JXL_COMPILER_MSVC +#define JXL_RESTRICT __restrict +#elif JXL_COMPILER_GCC || JXL_COMPILER_CLANG +#define JXL_RESTRICT __restrict__ +#else +#define JXL_RESTRICT +#endif + +#if JXL_COMPILER_MSVC +#define JXL_INLINE __forceinline +#define JXL_NOINLINE __declspec(noinline) +#else +#define JXL_INLINE inline __attribute__((always_inline)) +#define JXL_NOINLINE __attribute__((noinline)) +#endif + +#if JXL_COMPILER_MSVC +#define JXL_NORETURN __declspec(noreturn) +#elif JXL_COMPILER_GCC || JXL_COMPILER_CLANG +#define JXL_NORETURN __attribute__((noreturn)) +#else +#define JXL_NORETURN +#endif + +#if JXL_COMPILER_MSVC +#define JXL_UNREACHABLE_BUILTIN __assume(false) +#elif JXL_COMPILER_CLANG || JXL_COMPILER_GCC >= 405 +#define JXL_UNREACHABLE_BUILTIN __builtin_unreachable() +#else +#define JXL_UNREACHABLE_BUILTIN +#endif + +#if JXL_COMPILER_MSVC +#define JXL_MAYBE_UNUSED +#else +// Encountered "attribute list cannot appear here" when using the C++17 +// [[maybe_unused]], so only use the old style attribute for now. +#define JXL_MAYBE_UNUSED __attribute__((unused)) +#endif + +// MSAN execution won't hurt if some code it not inlined, but this can greatly +// improve compilation time. Unfortunately this macro can not be used just +// everywhere - inside header files it leads to "multiple definition" error; +// though it would be better not to have JXL_INLINE in header overall. +#if JXL_MEMORY_SANITIZER || JXL_ADDRESS_SANITIZER || JXL_THREAD_SANITIZER +#define JXL_MAYBE_INLINE JXL_MAYBE_UNUSED +#else +#define JXL_MAYBE_INLINE JXL_INLINE +#endif + +#if JXL_COMPILER_MSVC +// Unsupported, __assume is not the same. +#define JXL_LIKELY(expr) expr +#define JXL_UNLIKELY(expr) expr +#else +#define JXL_LIKELY(expr) __builtin_expect(!!(expr), 1) +#define JXL_UNLIKELY(expr) __builtin_expect(!!(expr), 0) +#endif + +// Returns a void* pointer which the compiler then assumes is N-byte aligned. +// Example: float* JXL_RESTRICT aligned = (float*)JXL_ASSUME_ALIGNED(in, 32); +// +// The assignment semantics are required by GCC/Clang. ICC provides an in-place +// __assume_aligned, whereas MSVC's __assume appears unsuitable. +#if JXL_COMPILER_CLANG +// Early versions of Clang did not support __builtin_assume_aligned. +#define JXL_HAS_ASSUME_ALIGNED __has_builtin(__builtin_assume_aligned) +#elif JXL_COMPILER_GCC +#define JXL_HAS_ASSUME_ALIGNED 1 +#else +#define JXL_HAS_ASSUME_ALIGNED 0 +#endif + +#if JXL_HAS_ASSUME_ALIGNED +#define JXL_ASSUME_ALIGNED(ptr, align) __builtin_assume_aligned((ptr), (align)) +#else +#define JXL_ASSUME_ALIGNED(ptr, align) (ptr) /* not supported */ +#endif + +#ifdef __has_attribute +#define JXL_HAVE_ATTRIBUTE(x) __has_attribute(x) +#else +#define JXL_HAVE_ATTRIBUTE(x) 0 +#endif + +// Raises warnings if the function return value is unused. Should appear as the +// first part of a function definition/declaration. +#if JXL_HAVE_ATTRIBUTE(nodiscard) +#define JXL_MUST_USE_RESULT [[nodiscard]] +#elif JXL_COMPILER_CLANG && JXL_HAVE_ATTRIBUTE(warn_unused_result) +#define JXL_MUST_USE_RESULT __attribute__((warn_unused_result)) +#else +#define JXL_MUST_USE_RESULT +#endif + +// Disable certain -fsanitize flags for functions that are expected to include +// things like unsigned integer overflow. For example use in the function +// declaration JXL_NO_SANITIZE("unsigned-integer-overflow") to silence unsigned +// integer overflow ubsan messages. +#if JXL_COMPILER_CLANG && JXL_HAVE_ATTRIBUTE(no_sanitize) +#define JXL_NO_SANITIZE(X) __attribute__((no_sanitize(X))) +#else +#define JXL_NO_SANITIZE(X) +#endif + +#if JXL_HAVE_ATTRIBUTE(__format__) +#define JXL_FORMAT(idx_fmt, idx_arg) \ + __attribute__((__format__(__printf__, idx_fmt, idx_arg))) +#else +#define JXL_FORMAT(idx_fmt, idx_arg) +#endif + +#if JXL_COMPILER_MSVC +using ssize_t = intptr_t; +#endif + +#endif // LIB_JXL_BASE_COMPILER_SPECIFIC_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/data_parallel.h b/third_party/jpeg-xl/lib/jxl/base/data_parallel.h new file mode 100644 index 0000000000..a7f977b7ce --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/data_parallel.h @@ -0,0 +1,124 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_DATA_PARALLEL_H_ +#define LIB_JXL_BASE_DATA_PARALLEL_H_ + +// Portable, low-overhead C++11 ThreadPool alternative to OpenMP for +// data-parallel computations. + +#include +#include +#include + +#include "lib/jxl/base/compiler_specific.h" +#include "lib/jxl/base/status.h" +#if JXL_COMPILER_MSVC +// suppress warnings about the const & applied to function types +#pragma warning(disable : 4180) +#endif + +namespace jxl { + +class ThreadPool { + public: + ThreadPool(JxlParallelRunner runner, void* runner_opaque) + : runner_(runner), + runner_opaque_(runner ? runner_opaque : static_cast(this)) {} + + ThreadPool(const ThreadPool&) = delete; + ThreadPool& operator&(const ThreadPool&) = delete; + + JxlParallelRunner runner() const { return runner_; } + void* runner_opaque() const { return runner_opaque_; } + + // Runs init_func(num_threads) followed by data_func(task, thread) on worker + // thread(s) for every task in [begin, end). init_func() must return a Status + // indicating whether the initialization succeeded. + // "thread" is an integer smaller than num_threads. + // Not thread-safe - no two calls to Run may overlap. + // Subsequent calls will reuse the same threads. + // + // Precondition: begin <= end. + template + Status Run(uint32_t begin, uint32_t end, const InitFunc& init_func, + const DataFunc& data_func, const char* caller = "") { + JXL_ASSERT(begin <= end); + if (begin == end) return true; + RunCallState call_state(init_func, data_func); + // The runner_ uses the C convention and returns 0 in case of error, so we + // convert it to a Status. + if (!runner_) { + void* jpegxl_opaque = static_cast(&call_state); + if (call_state.CallInitFunc(jpegxl_opaque, 1) != 0) { + return JXL_FAILURE("Failed to initialize thread"); + } + for (uint32_t i = begin; i < end; i++) { + call_state.CallDataFunc(jpegxl_opaque, i, 0); + } + return true; + } + return (*runner_)(runner_opaque_, static_cast(&call_state), + &call_state.CallInitFunc, &call_state.CallDataFunc, begin, + end) == 0; + } + + // Use this as init_func when no initialization is needed. + static Status NoInit(size_t num_threads) { return true; } + + private: + // class holding the state of a Run() call to pass to the runner_ as an + // opaque_jpegxl pointer. + template + class RunCallState final { + public: + RunCallState(const InitFunc& init_func, const DataFunc& data_func) + : init_func_(init_func), data_func_(data_func) {} + + // JxlParallelRunInit interface. + static int CallInitFunc(void* jpegxl_opaque, size_t num_threads) { + const auto* self = + static_cast*>(jpegxl_opaque); + // Returns -1 when the internal init function returns false Status to + // indicate an error. + return self->init_func_(num_threads) ? 0 : -1; + } + + // JxlParallelRunFunction interface. + static void CallDataFunc(void* jpegxl_opaque, uint32_t value, + size_t thread_id) { + const auto* self = + static_cast*>(jpegxl_opaque); + return self->data_func_(value, thread_id); + } + + private: + const InitFunc& init_func_; + const DataFunc& data_func_; + }; + + // The caller supplied runner function and its opaque void*. + const JxlParallelRunner runner_; + void* const runner_opaque_; +}; + +template +Status RunOnPool(ThreadPool* pool, const uint32_t begin, const uint32_t end, + const InitFunc& init_func, const DataFunc& data_func, + const char* caller) { + if (pool == nullptr) { + ThreadPool default_pool(nullptr, nullptr); + return default_pool.Run(begin, end, init_func, data_func, caller); + } else { + return pool->Run(begin, end, init_func, data_func, caller); + } +} + +} // namespace jxl +#if JXL_COMPILER_MSVC +#pragma warning(default : 4180) +#endif + +#endif // LIB_JXL_BASE_DATA_PARALLEL_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/exif.h b/third_party/jpeg-xl/lib/jxl/base/exif.h new file mode 100644 index 0000000000..2caafddc04 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/exif.h @@ -0,0 +1,91 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_EXIF_H_ +#define LIB_JXL_EXIF_H_ + +// Basic parsing of Exif (just enough for the render-impacting things +// like orientation) + +#include + +#include +#include +#include + +#include "lib/jxl/base/byte_order.h" +#include "lib/jxl/base/compiler_specific.h" + +namespace jxl { + +constexpr uint16_t kExifOrientationTag = 274; + +// Checks if a blob looks like Exif, and if so, sets bigendian +// according to the tiff endianness +JXL_INLINE bool IsExif(const std::vector& exif, bool* bigendian) { + if (exif.size() < 12) return false; // not enough bytes for a valid exif blob + const uint8_t* t = exif.data(); + if (LoadLE32(t) == 0x2A004D4D) { + *bigendian = true; + return true; + } else if (LoadLE32(t) == 0x002A4949) { + *bigendian = false; + return true; + } + return false; // not a valid tiff header +} + +// Finds the position of an Exif tag, or 0 if it is not found +JXL_INLINE size_t FindExifTagPosition(const std::vector& exif, + uint16_t tagname) { + bool bigendian; + if (!IsExif(exif, &bigendian)) return 0; + const uint8_t* t = exif.data() + 4; + uint64_t offset = (bigendian ? LoadBE32(t) : LoadLE32(t)); + if (exif.size() < 12 + offset + 2 || offset < 8) return 0; + t += offset - 4; + if (offset + 2 >= exif.size()) return 0; + uint16_t nb_tags = (bigendian ? LoadBE16(t) : LoadLE16(t)); + t += 2; + while (nb_tags > 0) { + if (t + 12 >= exif.data() + exif.size()) return 0; + uint16_t tag = (bigendian ? LoadBE16(t) : LoadLE16(t)); + t += 2; + if (tag == tagname) return static_cast(t - exif.data()); + t += 10; + nb_tags--; + } + return 0; +} + +// TODO(jon): tag 1 can be used to represent Adobe RGB 1998 if it has value +// "R03" +// TODO(jon): set intrinsic dimensions according to +// https://discourse.wicg.io/t/proposal-exif-image-resolution-auto-and-from-image/4326/24 +// Parses the Exif data just enough to extract any render-impacting info. +// If the Exif data is invalid or could not be parsed, then it is treated +// as a no-op. +JXL_INLINE void InterpretExif(const std::vector& exif, + JxlOrientation* orientation) { + bool bigendian; + if (!IsExif(exif, &bigendian)) return; + size_t o_pos = FindExifTagPosition(exif, kExifOrientationTag); + if (o_pos) { + const uint8_t* t = exif.data() + o_pos; + uint16_t type = (bigendian ? LoadBE16(t) : LoadLE16(t)); + t += 2; + uint32_t count = (bigendian ? LoadBE32(t) : LoadLE32(t)); + t += 4; + uint16_t value = (bigendian ? LoadBE16(t) : LoadLE16(t)); + t += 4; + if (type == 3 && count == 1 && value >= 1 && value <= 8) { + *orientation = static_cast(value); + } + } +} + +} // namespace jxl + +#endif // LIB_JXL_EXIF_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/fast_math-inl.h b/third_party/jpeg-xl/lib/jxl/base/fast_math-inl.h new file mode 100644 index 0000000000..fa749cc257 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/fast_math-inl.h @@ -0,0 +1,236 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +// Fast SIMD math ops (log2, encoder only, cos, erf for splines) + +#if defined(LIB_JXL_BASE_FAST_MATH_INL_H_) == defined(HWY_TARGET_TOGGLE) +#ifdef LIB_JXL_BASE_FAST_MATH_INL_H_ +#undef LIB_JXL_BASE_FAST_MATH_INL_H_ +#else +#define LIB_JXL_BASE_FAST_MATH_INL_H_ +#endif + +#include + +#include "lib/jxl/base/common.h" +#include "lib/jxl/base/rational_polynomial-inl.h" +HWY_BEFORE_NAMESPACE(); +namespace jxl { +namespace HWY_NAMESPACE { + +// These templates are not found via ADL. +using hwy::HWY_NAMESPACE::Abs; +using hwy::HWY_NAMESPACE::Add; +using hwy::HWY_NAMESPACE::Eq; +using hwy::HWY_NAMESPACE::Floor; +using hwy::HWY_NAMESPACE::Ge; +using hwy::HWY_NAMESPACE::GetLane; +using hwy::HWY_NAMESPACE::IfThenElse; +using hwy::HWY_NAMESPACE::IfThenZeroElse; +using hwy::HWY_NAMESPACE::Le; +using hwy::HWY_NAMESPACE::Min; +using hwy::HWY_NAMESPACE::Mul; +using hwy::HWY_NAMESPACE::MulAdd; +using hwy::HWY_NAMESPACE::NegMulAdd; +using hwy::HWY_NAMESPACE::Rebind; +using hwy::HWY_NAMESPACE::ShiftLeft; +using hwy::HWY_NAMESPACE::ShiftRight; +using hwy::HWY_NAMESPACE::Sub; +using hwy::HWY_NAMESPACE::Xor; + +// Computes base-2 logarithm like std::log2. Undefined if negative / NaN. +// L1 error ~3.9E-6 +template +V FastLog2f(const DF df, V x) { + // 2,2 rational polynomial approximation of std::log1p(x) / std::log(2). + HWY_ALIGN const float p[4 * (2 + 1)] = {HWY_REP4(-1.8503833400518310E-06f), + HWY_REP4(1.4287160470083755E+00f), + HWY_REP4(7.4245873327820566E-01f)}; + HWY_ALIGN const float q[4 * (2 + 1)] = {HWY_REP4(9.9032814277590719E-01f), + HWY_REP4(1.0096718572241148E+00f), + HWY_REP4(1.7409343003366853E-01f)}; + + const Rebind di; + const auto x_bits = BitCast(di, x); + + // Range reduction to [-1/3, 1/3] - 3 integer, 2 float ops + const auto exp_bits = Sub(x_bits, Set(di, 0x3f2aaaab)); // = 2/3 + // Shifted exponent = log2; also used to clear mantissa. + const auto exp_shifted = ShiftRight<23>(exp_bits); + const auto mantissa = BitCast(df, Sub(x_bits, ShiftLeft<23>(exp_shifted))); + const auto exp_val = ConvertTo(df, exp_shifted); + return Add(EvalRationalPolynomial(df, Sub(mantissa, Set(df, 1.0f)), p, q), + exp_val); +} + +// max relative error ~3e-7 +template +V FastPow2f(const DF df, V x) { + const Rebind di; + auto floorx = Floor(x); + auto exp = + BitCast(df, ShiftLeft<23>(Add(ConvertTo(di, floorx), Set(di, 127)))); + auto frac = Sub(x, floorx); + auto num = Add(frac, Set(df, 1.01749063e+01)); + num = MulAdd(num, frac, Set(df, 4.88687798e+01)); + num = MulAdd(num, frac, Set(df, 9.85506591e+01)); + num = Mul(num, exp); + auto den = MulAdd(frac, Set(df, 2.10242958e-01), Set(df, -2.22328856e-02)); + den = MulAdd(den, frac, Set(df, -1.94414990e+01)); + den = MulAdd(den, frac, Set(df, 9.85506633e+01)); + return Div(num, den); +} + +// max relative error ~3e-5 +template +V FastPowf(const DF df, V base, V exponent) { + return FastPow2f(df, Mul(FastLog2f(df, base), exponent)); +} + +// Computes cosine like std::cos. +// L1 error 7e-5. +template +V FastCosf(const DF df, V x) { + // Step 1: range reduction to [0, 2pi) + const auto pi2 = Set(df, kPi * 2.0f); + const auto pi2_inv = Set(df, 0.5f / kPi); + const auto npi2 = Mul(Floor(Mul(x, pi2_inv)), pi2); + const auto xmodpi2 = Sub(x, npi2); + // Step 2: range reduction to [0, pi] + const auto x_pi = Min(xmodpi2, Sub(pi2, xmodpi2)); + // Step 3: range reduction to [0, pi/2] + const auto above_pihalf = Ge(x_pi, Set(df, kPi / 2.0f)); + const auto x_pihalf = IfThenElse(above_pihalf, Sub(Set(df, kPi), x_pi), x_pi); + // Step 4: Taylor-like approximation, scaled by 2**0.75 to make angle + // duplication steps faster, on x/4. + const auto xs = Mul(x_pihalf, Set(df, 0.25f)); + const auto x2 = Mul(xs, xs); + const auto x4 = Mul(x2, x2); + const auto cosx_prescaling = + MulAdd(x4, Set(df, 0.06960438), + MulAdd(x2, Set(df, -0.84087373), Set(df, 1.68179268))); + // Step 5: angle duplication. + const auto cosx_scale1 = + MulAdd(cosx_prescaling, cosx_prescaling, Set(df, -1.414213562)); + const auto cosx_scale2 = MulAdd(cosx_scale1, cosx_scale1, Set(df, -1)); + // Step 6: change sign if needed. + const Rebind du; + auto signbit = ShiftLeft<31>(BitCast(du, VecFromMask(df, above_pihalf))); + return BitCast(df, Xor(signbit, BitCast(du, cosx_scale2))); +} + +// Computes the error function like std::erf. +// L1 error 7e-4. +template +V FastErff(const DF df, V x) { + // Formula from + // https://en.wikipedia.org/wiki/Error_function#Numerical_approximations + // but constants have been recomputed. + const auto xle0 = Le(x, Zero(df)); + const auto absx = Abs(x); + // Compute 1 - 1 / ((((x * a + b) * x + c) * x + d) * x + 1)**4 + const auto denom1 = + MulAdd(absx, Set(df, 7.77394369e-02), Set(df, 2.05260015e-04)); + const auto denom2 = MulAdd(denom1, absx, Set(df, 2.32120216e-01)); + const auto denom3 = MulAdd(denom2, absx, Set(df, 2.77820801e-01)); + const auto denom4 = MulAdd(denom3, absx, Set(df, 1.0f)); + const auto denom5 = Mul(denom4, denom4); + const auto inv_denom5 = Div(Set(df, 1.0f), denom5); + const auto result = NegMulAdd(inv_denom5, inv_denom5, Set(df, 1.0f)); + // Change sign if needed. + const Rebind du; + auto signbit = ShiftLeft<31>(BitCast(du, VecFromMask(df, xle0))); + return BitCast(df, Xor(signbit, BitCast(du, result))); +} + +inline float FastLog2f(float f) { + HWY_CAPPED(float, 1) D; + return GetLane(FastLog2f(D, Set(D, f))); +} + +inline float FastPow2f(float f) { + HWY_CAPPED(float, 1) D; + return GetLane(FastPow2f(D, Set(D, f))); +} + +inline float FastPowf(float b, float e) { + HWY_CAPPED(float, 1) D; + return GetLane(FastPowf(D, Set(D, b), Set(D, e))); +} + +inline float FastCosf(float f) { + HWY_CAPPED(float, 1) D; + return GetLane(FastCosf(D, Set(D, f))); +} + +inline float FastErff(float f) { + HWY_CAPPED(float, 1) D; + return GetLane(FastErff(D, Set(D, f))); +} + +// Returns cbrt(x) + add with 6 ulp max error. +// Modified from vectormath_exp.h, Apache 2 license. +// https://www.agner.org/optimize/vectorclass.zip +template +V CubeRootAndAdd(const V x, const V add) { + const HWY_FULL(float) df; + const HWY_FULL(int32_t) di; + + const auto kExpBias = Set(di, 0x54800000); // cast(1.) + cast(1.) / 3 + const auto kExpMul = Set(di, 0x002AAAAA); // shifted 1/3 + const auto k1_3 = Set(df, 1.0f / 3); + const auto k4_3 = Set(df, 4.0f / 3); + + const auto xa = x; // assume inputs never negative + const auto xa_3 = Mul(k1_3, xa); + + // Multiply exponent by -1/3 + const auto m1 = BitCast(di, xa); + // Special case for 0. 0 is represented with an exponent of 0, so the + // "kExpBias - 1/3 * exp" below gives the wrong result. The IfThenZeroElse() + // sets those values as 0, which prevents having NaNs in the computations + // below. + // TODO(eustas): use fused op + const auto m2 = IfThenZeroElse( + Eq(m1, Zero(di)), Sub(kExpBias, Mul((ShiftRight<23>(m1)), kExpMul))); + auto r = BitCast(df, m2); + + // Newton-Raphson iterations + for (int i = 0; i < 3; i++) { + const auto r2 = Mul(r, r); + r = NegMulAdd(xa_3, Mul(r2, r2), Mul(k4_3, r)); + } + // Final iteration + auto r2 = Mul(r, r); + r = MulAdd(k1_3, NegMulAdd(xa, Mul(r2, r2), r), r); + r2 = Mul(r, r); + r = MulAdd(r2, x, add); + + return r; +} + +// NOLINTNEXTLINE(google-readability-namespace-comments) +} // namespace HWY_NAMESPACE +} // namespace jxl +HWY_AFTER_NAMESPACE(); + +#endif // LIB_JXL_BASE_FAST_MATH_INL_H_ + +#if HWY_ONCE +#ifndef LIB_JXL_BASE_FAST_MATH_ONCE +#define LIB_JXL_BASE_FAST_MATH_ONCE + +namespace jxl { +inline float FastLog2f(float f) { return HWY_STATIC_DISPATCH(FastLog2f)(f); } +inline float FastPow2f(float f) { return HWY_STATIC_DISPATCH(FastPow2f)(f); } +inline float FastPowf(float b, float e) { + return HWY_STATIC_DISPATCH(FastPowf)(b, e); +} +inline float FastCosf(float f) { return HWY_STATIC_DISPATCH(FastCosf)(f); } +inline float FastErff(float f) { return HWY_STATIC_DISPATCH(FastErff)(f); } +} // namespace jxl + +#endif // LIB_JXL_BASE_FAST_MATH_ONCE +#endif // HWY_ONCE diff --git a/third_party/jpeg-xl/lib/jxl/base/float.h b/third_party/jpeg-xl/lib/jxl/base/float.h new file mode 100644 index 0000000000..00e112bb34 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/float.h @@ -0,0 +1,102 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_FLOAT_H_ +#define LIB_JXL_BASE_FLOAT_H_ + +#include +#include +#include +#include + +#include "lib/jxl/base/byte_order.h" +#include "lib/jxl/base/compiler_specific.h" +#include "lib/jxl/base/status.h" + +namespace jxl { + +namespace { +// Based on highway scalar implementation, for testing +float LoadFloat16(uint16_t bits16) { + const uint32_t sign = bits16 >> 15; + const uint32_t biased_exp = (bits16 >> 10) & 0x1F; + const uint32_t mantissa = bits16 & 0x3FF; + + // Subnormal or zero + if (biased_exp == 0) { + const float subnormal = + (1.0f / 16384) * (static_cast(mantissa) * (1.0f / 1024)); + return sign ? -subnormal : subnormal; + } + + // Normalized: convert the representation directly (faster than ldexp/tables). + const uint32_t biased_exp32 = biased_exp + (127 - 15); + const uint32_t mantissa32 = mantissa << (23 - 10); + const uint32_t bits32 = (sign << 31) | (biased_exp32 << 23) | mantissa32; + + float result; + memcpy(&result, &bits32, 4); + return result; +} +} // namespace + +template +static Status JXL_INLINE LoadFloatRow(const uint8_t* src, size_t count, + size_t stride, JxlDataType type, + bool little_endian, float scale, + SaveFloatAtFn callback) { + switch (type) { + case JXL_TYPE_FLOAT: + if (little_endian) { + for (size_t i = 0; i < count; ++i) { + callback(i, LoadLEFloat(src + stride * i)); + } + } else { + for (size_t i = 0; i < count; ++i) { + callback(i, LoadBEFloat(src + stride * i)); + } + } + return true; + + case JXL_TYPE_UINT8: + for (size_t i = 0; i < count; ++i) { + // Integer multiply uint8 value before scaling so that the UINT8 value + // and the corresponding UINT16 value convert to the same float + callback(i, (src[stride * i] * 257) * scale); + } + return true; + + case JXL_TYPE_UINT16: + if (little_endian) { + for (size_t i = 0; i < count; ++i) { + callback(i, LoadLE16(src + stride * i) * scale); + } + } else { + for (size_t i = 0; i < count; ++i) { + callback(i, LoadBE16(src + stride * i) * scale); + } + } + return true; + + case JXL_TYPE_FLOAT16: + if (little_endian) { + for (size_t i = 0; i < count; ++i) { + callback(i, LoadFloat16(LoadLE16(src + stride * i))); + } + } else { + for (size_t i = 0; i < count; ++i) { + callback(i, LoadFloat16(LoadBE16(src + stride * i))); + } + } + return true; + + default: + return JXL_FAILURE("Unsupported sample format"); + } +} + +} // namespace jxl + +#endif // LIB_JXL_BASE_FLOAT_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/iaca.h b/third_party/jpeg-xl/lib/jxl/base/iaca.h new file mode 100644 index 0000000000..e5732dae5c --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/iaca.h @@ -0,0 +1,65 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_IACA_H_ +#define LIB_JXL_BASE_IACA_H_ + +#include "lib/jxl/base/compiler_specific.h" + +// IACA (Intel's Code Analyzer) analyzes instruction latencies, but only for +// code between special markers. These functions embed such markers in an +// executable, but only for reading via IACA - they deliberately trigger a +// crash if executed to ensure they are removed in normal builds. + +#ifndef JXL_IACA_ENABLED +#define JXL_IACA_ENABLED 0 +#endif + +namespace jxl { + +// Call before the region of interest. +static JXL_INLINE void BeginIACA() { +#if JXL_IACA_ENABLED && (JXL_COMPILER_GCC || JXL_COMPILER_CLANG) + asm volatile( + // UD2 "instruction" raises an invalid opcode exception. + ".byte 0x0F, 0x0B\n\t" + // Magic sequence recognized by IACA (MOV + addr32 fs:NOP). This actually + // clobbers EBX, but we don't care because the code won't be run, and we + // want IACA to observe the same code the compiler would have generated + // without this marker. + "movl $111, %%ebx\n\t" + ".byte 0x64, 0x67, 0x90\n\t" + : + : + // (Allegedly) clobbering memory may prevent reordering. + : "memory"); +#endif +} + +// Call after the region of interest. +static JXL_INLINE void EndIACA() { +#if JXL_IACA_ENABLED && (JXL_COMPILER_GCC || JXL_COMPILER_CLANG) + asm volatile( + // See above. + "movl $222, %%ebx\n\t" + ".byte 0x64, 0x67, 0x90\n\t" + // UD2 + ".byte 0x0F, 0x0B\n\t" + : + : + // (Allegedly) clobbering memory may prevent reordering. + : "memory"); +#endif +} + +// Add to a scope to mark a region. +struct ScopeIACA { + JXL_INLINE ScopeIACA() { BeginIACA(); } + JXL_INLINE ~ScopeIACA() { EndIACA(); } +}; + +} // namespace jxl + +#endif // LIB_JXL_BASE_IACA_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/matrix_ops.h b/third_party/jpeg-xl/lib/jxl/base/matrix_ops.h new file mode 100644 index 0000000000..1a969bd4f0 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/matrix_ops.h @@ -0,0 +1,84 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_MATRIX_OPS_H_ +#define LIB_JXL_MATRIX_OPS_H_ + +// 3x3 matrix operations. + +#include // abs +#include + +#include "lib/jxl/base/status.h" + +namespace jxl { + +// Computes C = A * B, where A, B, C are 3x3 matrices. +template +void Mul3x3Matrix(const T* a, const T* b, T* c) { + alignas(16) T temp[3]; // For transposed column + for (size_t x = 0; x < 3; x++) { + for (size_t z = 0; z < 3; z++) { + temp[z] = b[z * 3 + x]; + } + for (size_t y = 0; y < 3; y++) { + double e = 0; + for (size_t z = 0; z < 3; z++) { + e += a[y * 3 + z] * temp[z]; + } + c[y * 3 + x] = e; + } + } +} + +// Computes C = A * B, where A is 3x3 matrix and B is vector. +template +void Mul3x3Vector(const T* a, const T* b, T* c) { + for (size_t y = 0; y < 3; y++) { + double e = 0; + for (size_t x = 0; x < 3; x++) { + e += a[y * 3 + x] * b[x]; + } + c[y] = e; + } +} + +// Inverts a 3x3 matrix in place. +template +Status Inv3x3Matrix(T* matrix) { + // Intermediate computation is done in double precision. + double temp[9]; + temp[0] = static_cast(matrix[4]) * matrix[8] - + static_cast(matrix[5]) * matrix[7]; + temp[1] = static_cast(matrix[2]) * matrix[7] - + static_cast(matrix[1]) * matrix[8]; + temp[2] = static_cast(matrix[1]) * matrix[5] - + static_cast(matrix[2]) * matrix[4]; + temp[3] = static_cast(matrix[5]) * matrix[6] - + static_cast(matrix[3]) * matrix[8]; + temp[4] = static_cast(matrix[0]) * matrix[8] - + static_cast(matrix[2]) * matrix[6]; + temp[5] = static_cast(matrix[2]) * matrix[3] - + static_cast(matrix[0]) * matrix[5]; + temp[6] = static_cast(matrix[3]) * matrix[7] - + static_cast(matrix[4]) * matrix[6]; + temp[7] = static_cast(matrix[1]) * matrix[6] - + static_cast(matrix[0]) * matrix[7]; + temp[8] = static_cast(matrix[0]) * matrix[4] - + static_cast(matrix[1]) * matrix[3]; + double det = matrix[0] * temp[0] + matrix[1] * temp[3] + matrix[2] * temp[6]; + if (std::abs(det) < 1e-10) { + return JXL_FAILURE("Matrix determinant is too close to 0"); + } + double idet = 1.0 / det; + for (size_t i = 0; i < 9; i++) { + matrix[i] = temp[i] * idet; + } + return true; +} + +} // namespace jxl + +#endif // LIB_JXL_MATRIX_OPS_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/os_macros.h b/third_party/jpeg-xl/lib/jxl/base/os_macros.h new file mode 100644 index 0000000000..84d0b82bf5 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/os_macros.h @@ -0,0 +1,50 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_OS_MACROS_H_ +#define LIB_JXL_BASE_OS_MACROS_H_ + +// Defines the JXL_OS_* macros. + +#if defined(_WIN32) || defined(_WIN64) +#define JXL_OS_WIN 1 +#else +#define JXL_OS_WIN 0 +#endif + +#ifdef __linux__ +#define JXL_OS_LINUX 1 +#else +#define JXL_OS_LINUX 0 +#endif + +#ifdef __APPLE__ +#define JXL_OS_MAC 1 +#else +#define JXL_OS_MAC 0 +#endif + +#define JXL_OS_IOS 0 +#ifdef __APPLE__ +#include +#if TARGET_OS_IPHONE +#undef JXL_OS_IOS +#define JXL_OS_IOS 1 +#endif +#endif + +#ifdef __FreeBSD__ +#define JXL_OS_FREEBSD 1 +#else +#define JXL_OS_FREEBSD 0 +#endif + +#ifdef __HAIKU__ +#define JXL_OS_HAIKU 1 +#else +#define JXL_OS_HAIKU 0 +#endif + +#endif // LIB_JXL_BASE_OS_MACROS_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/override.h b/third_party/jpeg-xl/lib/jxl/base/override.h new file mode 100644 index 0000000000..1f8b657974 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/override.h @@ -0,0 +1,29 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_OVERRIDE_H_ +#define LIB_JXL_BASE_OVERRIDE_H_ + +// 'Trool' for command line arguments: force enable/disable, or use default. + +namespace jxl { + +// No effect if kDefault, otherwise forces a feature (typically a FrameHeader +// flag) on or off. +enum class Override : int { kOn = 1, kOff = 0, kDefault = -1 }; + +static inline Override OverrideFromBool(bool flag) { + return flag ? Override::kOn : Override::kOff; +} + +static inline bool ApplyOverride(Override o, bool default_condition) { + if (o == Override::kOn) return true; + if (o == Override::kOff) return false; + return default_condition; +} + +} // namespace jxl + +#endif // LIB_JXL_BASE_OVERRIDE_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/printf_macros.h b/third_party/jpeg-xl/lib/jxl/base/printf_macros.h new file mode 100644 index 0000000000..3215052afd --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/printf_macros.h @@ -0,0 +1,34 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_PRINTF_MACROS_H_ +#define LIB_JXL_BASE_PRINTF_MACROS_H_ + +// Format string macros. These should be included after any other system +// library since those may unconditionally define these, depending on the +// platform. + +// PRIuS and PRIdS macros to print size_t and ssize_t respectively. +#if !defined(PRIdS) +#if defined(_WIN64) +#define PRIdS "lld" +#elif defined(_WIN32) +#define PRIdS "d" +#else +#define PRIdS "zd" +#endif +#endif // PRIdS + +#if !defined(PRIuS) +#if defined(_WIN64) +#define PRIuS "llu" +#elif defined(_WIN32) +#define PRIuS "u" +#else +#define PRIuS "zu" +#endif +#endif // PRIuS + +#endif // LIB_JXL_BASE_PRINTF_MACROS_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/random.h b/third_party/jpeg-xl/lib/jxl/base/random.h new file mode 100644 index 0000000000..b27815bf00 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/random.h @@ -0,0 +1,99 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_RANDOM_ +#define LIB_JXL_BASE_RANDOM_ + +// Random number generator + distributions. +// We don't use because the implementation (and thus results) differs +// between libstdc++ and libc++. + +#include +#include + +#include +#include + +#include "lib/jxl/base/status.h" + +namespace jxl { +struct Rng { + explicit Rng(size_t seed) + : s{static_cast(0x94D049BB133111EBull), + static_cast(0xBF58476D1CE4E5B9ull) + seed} {} + + // Xorshift128+ adapted from xorshift128+-inl.h + uint64_t operator()() { + uint64_t s1 = s[0]; + const uint64_t s0 = s[1]; + const uint64_t bits = s1 + s0; // b, c + s[0] = s0; + s1 ^= s1 << 23; + s1 ^= s0 ^ (s1 >> 18) ^ (s0 >> 5); + s[1] = s1; + return bits; + } + + // Uniformly distributed int64_t in [begin, end), under the assumption that + // `end-begin` is significantly smaller than 1<<64, otherwise there is some + // bias. + int64_t UniformI(int64_t begin, int64_t end) { + JXL_DASSERT(end > begin); + return static_cast((*this)() % + static_cast(end - begin)) + + begin; + } + + // Same as UniformI, but for uint64_t. + uint64_t UniformU(uint64_t begin, uint64_t end) { + JXL_DASSERT(end > begin); + return (*this)() % (end - begin) + begin; + } + + // Uniformly distributed float in [begin, end) range. Note: only 23 bits of + // randomness. + float UniformF(float begin, float end) { + float f; + // Bits of a random [1, 2) float. + uint32_t u = ((*this)() >> (64 - 23)) | 0x3F800000; + static_assert(sizeof(f) == sizeof(u), + "Float and U32 must have the same size"); + memcpy(&f, &u, sizeof(f)); + // Note: (end-begin) * f + (2*begin-end) may fail to return a number >= + // begin. + return (end - begin) * (f - 1.0f) + begin; + } + + // Bernoulli trial + bool Bernoulli(float p) { return UniformF(0, 1) < p; } + + // State for geometric distributions. + // The stored value is inv_log_1mp + using GeometricDistribution = float; + static GeometricDistribution MakeGeometric(float p) { + return 1.0 / std::log(1 - p); + } + + uint32_t Geometric(const GeometricDistribution& dist) { + float f = UniformF(0, 1); + float inv_log_1mp = dist; + float log = std::log(1 - f) * inv_log_1mp; + return static_cast(log); + } + + template + void Shuffle(T* t, size_t n) { + for (size_t i = 0; i + 1 < n; i++) { + size_t a = UniformU(i, n); + std::swap(t[a], t[i]); + } + } + + private: + uint64_t s[2]; +}; + +} // namespace jxl +#endif // LIB_JXL_BASE_RANDOM_ diff --git a/third_party/jpeg-xl/lib/jxl/base/rational_polynomial-inl.h b/third_party/jpeg-xl/lib/jxl/base/rational_polynomial-inl.h new file mode 100644 index 0000000000..e073937675 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/rational_polynomial-inl.h @@ -0,0 +1,102 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +// Fast SIMD evaluation of rational polynomials for approximating functions. + +#if defined(LIB_JXL_BASE_RATIONAL_POLYNOMIAL_INL_H_) == \ + defined(HWY_TARGET_TOGGLE) +#ifdef LIB_JXL_BASE_RATIONAL_POLYNOMIAL_INL_H_ +#undef LIB_JXL_BASE_RATIONAL_POLYNOMIAL_INL_H_ +#else +#define LIB_JXL_BASE_RATIONAL_POLYNOMIAL_INL_H_ +#endif + +#include + +#include +HWY_BEFORE_NAMESPACE(); +namespace jxl { +namespace HWY_NAMESPACE { +namespace { + +// These templates are not found via ADL. +using hwy::HWY_NAMESPACE::Div; +using hwy::HWY_NAMESPACE::MulAdd; + +// Primary template: default to actual division. +template +struct FastDivision { + HWY_INLINE V operator()(const V n, const V d) const { return n / d; } +}; +// Partial specialization for float vectors. +template +struct FastDivision { + // One Newton-Raphson iteration. + static HWY_INLINE V ReciprocalNR(const V x) { + const auto rcp = ApproximateReciprocal(x); + const auto sum = Add(rcp, rcp); + const auto x_rcp = Mul(x, rcp); + return NegMulAdd(x_rcp, rcp, sum); + } + + V operator()(const V n, const V d) const { +#if 1 // Faster on SKX + return Div(n, d); +#else + return n * ReciprocalNR(d); +#endif + } +}; + +// Approximates smooth functions via rational polynomials (i.e. dividing two +// polynomials). Evaluates polynomials via Horner's scheme, which is faster than +// Clenshaw recurrence for Chebyshev polynomials. LoadDup128 allows us to +// specify constants (replicated 4x) independently of the lane count. +template +HWY_INLINE HWY_MAYBE_UNUSED V EvalRationalPolynomial(const D d, const V x, + const T (&p)[NP], + const T (&q)[NQ]) { + constexpr size_t kDegP = NP / 4 - 1; + constexpr size_t kDegQ = NQ / 4 - 1; + auto yp = LoadDup128(d, &p[kDegP * 4]); + auto yq = LoadDup128(d, &q[kDegQ * 4]); + // We use pointer arithmetic to refer to &p[(kDegP - n) * 4] to avoid a + // compiler warning that the index is out of bounds since we are already + // checking that it is not out of bounds with (kDegP >= n) and the access + // will be optimized away. Similarly with q and kDegQ. + HWY_FENCE; + if (kDegP >= 1) yp = MulAdd(yp, x, LoadDup128(d, p + ((kDegP - 1) * 4))); + if (kDegQ >= 1) yq = MulAdd(yq, x, LoadDup128(d, q + ((kDegQ - 1) * 4))); + HWY_FENCE; + if (kDegP >= 2) yp = MulAdd(yp, x, LoadDup128(d, p + ((kDegP - 2) * 4))); + if (kDegQ >= 2) yq = MulAdd(yq, x, LoadDup128(d, q + ((kDegQ - 2) * 4))); + HWY_FENCE; + if (kDegP >= 3) yp = MulAdd(yp, x, LoadDup128(d, p + ((kDegP - 3) * 4))); + if (kDegQ >= 3) yq = MulAdd(yq, x, LoadDup128(d, q + ((kDegQ - 3) * 4))); + HWY_FENCE; + if (kDegP >= 4) yp = MulAdd(yp, x, LoadDup128(d, p + ((kDegP - 4) * 4))); + if (kDegQ >= 4) yq = MulAdd(yq, x, LoadDup128(d, q + ((kDegQ - 4) * 4))); + HWY_FENCE; + if (kDegP >= 5) yp = MulAdd(yp, x, LoadDup128(d, p + ((kDegP - 5) * 4))); + if (kDegQ >= 5) yq = MulAdd(yq, x, LoadDup128(d, q + ((kDegQ - 5) * 4))); + HWY_FENCE; + if (kDegP >= 6) yp = MulAdd(yp, x, LoadDup128(d, p + ((kDegP - 6) * 4))); + if (kDegQ >= 6) yq = MulAdd(yq, x, LoadDup128(d, q + ((kDegQ - 6) * 4))); + HWY_FENCE; + if (kDegP >= 7) yp = MulAdd(yp, x, LoadDup128(d, p + ((kDegP - 7) * 4))); + if (kDegQ >= 7) yq = MulAdd(yq, x, LoadDup128(d, q + ((kDegQ - 7) * 4))); + + static_assert(kDegP < 8, "Polynomial degree is too high"); + static_assert(kDegQ < 8, "Polynomial degree is too high"); + + return FastDivision()(yp, yq); +} + +} // namespace +// NOLINTNEXTLINE(google-readability-namespace-comments) +} // namespace HWY_NAMESPACE +} // namespace jxl +HWY_AFTER_NAMESPACE(); +#endif // LIB_JXL_BASE_RATIONAL_POLYNOMIAL_INL_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/sanitizer_definitions.h b/third_party/jpeg-xl/lib/jxl/base/sanitizer_definitions.h new file mode 100644 index 0000000000..315f3bd003 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/sanitizer_definitions.h @@ -0,0 +1,44 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_SANITIZER_DEFINITIONS_H_ +#define LIB_JXL_BASE_SANITIZER_DEFINITIONS_H_ + +#ifdef MEMORY_SANITIZER +#define JXL_MEMORY_SANITIZER 1 +#elif defined(__has_feature) +#if __has_feature(memory_sanitizer) +#define JXL_MEMORY_SANITIZER 1 +#else +#define JXL_MEMORY_SANITIZER 0 +#endif +#else +#define JXL_MEMORY_SANITIZER 0 +#endif + +#ifdef ADDRESS_SANITIZER +#define JXL_ADDRESS_SANITIZER 1 +#elif defined(__has_feature) +#if __has_feature(address_sanitizer) +#define JXL_ADDRESS_SANITIZER 1 +#else +#define JXL_ADDRESS_SANITIZER 0 +#endif +#else +#define JXL_ADDRESS_SANITIZER 0 +#endif + +#ifdef THREAD_SANITIZER +#define JXL_THREAD_SANITIZER 1 +#elif defined(__has_feature) +#if __has_feature(thread_sanitizer) +#define JXL_THREAD_SANITIZER 1 +#else +#define JXL_THREAD_SANITIZER 0 +#endif +#else +#define JXL_THREAD_SANITIZER 0 +#endif +#endif // LIB_JXL_BASE_SANITIZER_DEFINITIONS_H diff --git a/third_party/jpeg-xl/lib/jxl/base/scope_guard.h b/third_party/jpeg-xl/lib/jxl/base/scope_guard.h new file mode 100644 index 0000000000..a18a44cb79 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/scope_guard.h @@ -0,0 +1,48 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_SCOPE_GUARD_H_ +#define LIB_JXL_BASE_SCOPE_GUARD_H_ + +#include + +namespace jxl { + +template +class ScopeGuard { + public: + // Discourage unnecessary moves / copies. + ScopeGuard(const ScopeGuard &) = delete; + ScopeGuard &operator=(const ScopeGuard &) = delete; + ScopeGuard &operator=(ScopeGuard &&) = delete; + + // Pre-C++17 does not guarantee RVO -> require move constructor. + ScopeGuard(ScopeGuard &&other) : callback_(std::move(other.callback_)) { + other.armed_ = false; + } + + template + explicit ScopeGuard(CallbackParam &&callback) + : callback_(std::forward(callback)), armed_(true) {} + + ~ScopeGuard() { + if (armed_) callback_(); + } + + void Disarm() { armed_ = false; } + + private: + Callback callback_; + bool armed_; +}; + +template +ScopeGuard MakeScopeGuard(Callback &&callback) { + return ScopeGuard{std::forward(callback)}; +} + +} // namespace jxl + +#endif // LIB_JXL_BASE_SCOPE_GUARD_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/span.h b/third_party/jpeg-xl/lib/jxl/base/span.h new file mode 100644 index 0000000000..dc1c781b9d --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/span.h @@ -0,0 +1,80 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_SPAN_H_ +#define LIB_JXL_BASE_SPAN_H_ + +// Span (array view) is a non-owning container that provides cheap "cut" +// operations and could be used as "ArrayLike" data source for PaddedBytes. + +#include +#include +#include + +#include "lib/jxl/base/status.h" + +namespace jxl { + +template +class Span { + public: + constexpr Span() noexcept : Span(nullptr, 0) {} + + constexpr Span(T* array, size_t length) noexcept + : ptr_(array), len_(length) {} + + template + explicit constexpr Span(T (&a)[N]) noexcept : Span(a, N) {} + + template + constexpr Span(U* array, size_t length) noexcept + : ptr_(reinterpret_cast(array)), len_(length) { + static_assert(sizeof(U) == sizeof(T), "Incompatible type of source."); + } + + template + explicit constexpr Span(const ArrayLike& other) noexcept + : Span(reinterpret_cast(other.data()), other.size()) { + static_assert(sizeof(*other.data()) == sizeof(T), + "Incompatible type of source."); + } + + constexpr T* data() const noexcept { return ptr_; } + + constexpr size_t size() const noexcept { return len_; } + + constexpr bool empty() const noexcept { return len_ == 0; } + + constexpr T* begin() const noexcept { return data(); } + + constexpr T* end() const noexcept { return data() + size(); } + + constexpr T& operator[](size_t i) const noexcept { + // MSVC 2015 accepts this as constexpr, but not ptr_[i] + return *(data() + i); + } + + void remove_prefix(size_t n) noexcept { + JXL_ASSERT(size() >= n); + ptr_ += n; + len_ -= n; + } + + // NCT == non-const-T; compiler will complain if NCT is not compatible with T. + template + void AppendTo(std::vector* dst) const { + dst->insert(dst->end(), begin(), end()); + } + + private: + T* ptr_; + size_t len_; +}; + +typedef Span Bytes; + +} // namespace jxl + +#endif // LIB_JXL_BASE_SPAN_H_ diff --git a/third_party/jpeg-xl/lib/jxl/base/status.h b/third_party/jpeg-xl/lib/jxl/base/status.h new file mode 100644 index 0000000000..b33bd64fc3 --- /dev/null +++ b/third_party/jpeg-xl/lib/jxl/base/status.h @@ -0,0 +1,456 @@ +// Copyright (c) the JPEG XL Project Authors. All rights reserved. +// +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +#ifndef LIB_JXL_BASE_STATUS_H_ +#define LIB_JXL_BASE_STATUS_H_ + +// Error handling: Status return type + helper macros. + +#include +#include +#include +#include + +#include +#include + +#include "lib/jxl/base/compiler_specific.h" +#include "lib/jxl/base/sanitizer_definitions.h" + +#if JXL_ADDRESS_SANITIZER || JXL_MEMORY_SANITIZER || JXL_THREAD_SANITIZER +#include "sanitizer/common_interface_defs.h" // __sanitizer_print_stack_trace +#endif // defined(*_SANITIZER) + +namespace jxl { + +// Uncomment to abort when JXL_FAILURE or JXL_STATUS with a fatal error is +// reached: +// #define JXL_CRASH_ON_ERROR + +#ifndef JXL_ENABLE_ASSERT +#define JXL_ENABLE_ASSERT 1 +#endif + +#ifndef JXL_ENABLE_CHECK +#define JXL_ENABLE_CHECK 1 +#endif + +// Pass -DJXL_DEBUG_ON_ERROR at compile time to print debug messages when a +// function returns JXL_FAILURE or calls JXL_NOTIFY_ERROR. Note that this is +// irrelevant if you also pass -DJXL_CRASH_ON_ERROR. +#if defined(JXL_DEBUG_ON_ERROR) || defined(JXL_CRASH_ON_ERROR) +#undef JXL_DEBUG_ON_ERROR +#define JXL_DEBUG_ON_ERROR 1 +#else // JXL_DEBUG_ON_ERROR || JXL_CRASH_ON_ERROR +#ifdef NDEBUG +#define JXL_DEBUG_ON_ERROR 0 +#else // NDEBUG +#define JXL_DEBUG_ON_ERROR 1 +#endif // NDEBUG +#endif // JXL_DEBUG_ON_ERROR || JXL_CRASH_ON_ERROR + +// Pass -DJXL_DEBUG_ON_ALL_ERROR at compile time to print debug messages on +// all error (fatal and non-fatal) status. This implies JXL_DEBUG_ON_ERROR. +#if defined(JXL_DEBUG_ON_ALL_ERROR) +#undef JXL_DEBUG_ON_ALL_ERROR +#define JXL_DEBUG_ON_ALL_ERROR 1 +// JXL_DEBUG_ON_ALL_ERROR implies JXL_DEBUG_ON_ERROR too. +#undef JXL_DEBUG_ON_ERROR +#define JXL_DEBUG_ON_ERROR 1 +#else // JXL_DEBUG_ON_ALL_ERROR +#define JXL_DEBUG_ON_ALL_ERROR 0 +#endif // JXL_DEBUG_ON_ALL_ERROR + +// The Verbose level for the library +#ifndef JXL_DEBUG_V_LEVEL +#define JXL_DEBUG_V_LEVEL 0 +#endif // JXL_DEBUG_V_LEVEL + +// Pass -DJXL_DEBUG_ON_ABORT={0,1} to force disable/enable the debug messages on +// JXL_ASSERT, JXL_CHECK and JXL_ABORT. +#ifndef JXL_DEBUG_ON_ABORT +#define JXL_DEBUG_ON_ABORT JXL_DEBUG_ON_ERROR +#endif // JXL_DEBUG_ON_ABORT + +#ifdef USE_ANDROID_LOGGER +#include +#define LIBJXL_ANDROID_LOG_TAG ("libjxl") +inline void android_vprintf(const char* format, va_list args) { + char* message = nullptr; + int res = vasprintf(&message, format, args); + if (res != -1) { + __android_log_write(ANDROID_LOG_DEBUG, LIBJXL_ANDROID_LOG_TAG, message); + free(message); + } +} +#endif + +// Print a debug message on standard error or android logs. You should use the +// JXL_DEBUG macro instead of calling Debug directly. This function returns +// false, so it can be used as a return value in JXL_FAILURE. +JXL_FORMAT(1, 2) +inline JXL_NOINLINE bool Debug(const char* format, ...) { + va_list args; + va_start(args, format); +#ifdef USE_ANDROID_LOGGER + android_vprintf(format, args); +#else + vfprintf(stderr, format, args); +#endif + va_end(args); + return false; +} + +// Print a debug message on standard error if "enabled" is true. "enabled" is +// normally a macro that evaluates to 0 or 1 at compile time, so the Debug +// function is never called and optimized out in release builds. Note that the +// arguments are compiled but not evaluated when enabled is false. The format +// string must be a explicit string in the call, for example: +// JXL_DEBUG(JXL_DEBUG_MYMODULE, "my module message: %d", some_var); +// Add a header at the top of your module's .cc or .h file (depending on whether +// you have JXL_DEBUG calls from the .h as well) like this: +// #ifndef JXL_DEBUG_MYMODULE +// #define JXL_DEBUG_MYMODULE 0 +// #endif JXL_DEBUG_MYMODULE +#define JXL_DEBUG_TMP(format, ...) \ + ::jxl::Debug(("%s:%d: " format "\n"), __FILE__, __LINE__, ##__VA_ARGS__) + +#define JXL_DEBUG(enabled, format, ...) \ + do { \ + if (enabled) { \ + JXL_DEBUG_TMP(format, ##__VA_ARGS__); \ + } \ + } while (0) + +// JXL_DEBUG version that prints the debug message if the global verbose level +// defined at compile time by JXL_DEBUG_V_LEVEL is greater or equal than the +// passed level. +#if JXL_DEBUG_V_LEVEL > 0 +#define JXL_DEBUG_V(level, format, ...) \ + JXL_DEBUG(level <= JXL_DEBUG_V_LEVEL, format, ##__VA_ARGS__) +#else +#define JXL_DEBUG_V(level, format, ...) +#endif + +// Warnings (via JXL_WARNING) are enabled by default in debug builds (opt and +// debug). +#ifdef JXL_DEBUG_WARNING +#undef JXL_DEBUG_WARNING +#define JXL_DEBUG_WARNING 1 +#else // JXL_DEBUG_WARNING +#ifdef NDEBUG +#define JXL_DEBUG_WARNING 0 +#else // JXL_DEBUG_WARNING +#define JXL_DEBUG_WARNING 1 +#endif // NDEBUG +#endif // JXL_DEBUG_WARNING +#define JXL_WARNING(format, ...) \ + JXL_DEBUG(JXL_DEBUG_WARNING, format, ##__VA_ARGS__) + +// Exits the program after printing a stack trace when possible. +JXL_NORETURN inline JXL_NOINLINE bool Abort() { +#if JXL_ADDRESS_SANITIZER || JXL_MEMORY_SANITIZER || JXL_THREAD_SANITIZER + // If compiled with any sanitizer print a stack trace. This call doesn't crash + // the program, instead the trap below will crash it also allowing gdb to + // break there. + __sanitizer_print_stack_trace(); +#endif // *_SANITIZER) + +#if JXL_COMPILER_MSVC + __debugbreak(); + abort(); +#else + __builtin_trap(); +#endif +} + +// Exits the program after printing file/line plus a formatted string. +#define JXL_ABORT(format, ...) \ + ((JXL_DEBUG_ON_ABORT) && ::jxl::Debug(("%s:%d: JXL_ABORT: " format "\n"), \ + __FILE__, __LINE__, ##__VA_ARGS__), \ + ::jxl::Abort()) + +// Use this for code paths that are unreachable unless the code would change +// to make it reachable, in which case it will print a warning and abort in +// debug builds. In release builds no code is produced for this, so only use +// this if this path is really unreachable. +#define JXL_UNREACHABLE(format, ...) \ + do { \ + if (JXL_DEBUG_WARNING) { \ + ::jxl::Debug(("%s:%d: JXL_UNREACHABLE: " format "\n"), __FILE__, \ + __LINE__, ##__VA_ARGS__); \ + ::jxl::Abort(); \ + } else { \ + JXL_UNREACHABLE_BUILTIN; \ + } \ + } while (0) + +// Does not guarantee running the code, use only for debug mode checks. +#if JXL_ENABLE_ASSERT +#define JXL_ASSERT(condition) \ + do { \ + if (!(condition)) { \ + JXL_DEBUG(JXL_DEBUG_ON_ABORT, "JXL_ASSERT: %s", #condition); \ + ::jxl::Abort(); \ + } \ + } while (0) +#else +#define JXL_ASSERT(condition) \ + do { \ + } while (0) +#endif + +// Define JXL_IS_DEBUG_BUILD that denotes asan, msan and other debug builds, +// but not opt or release. +#ifndef JXL_IS_DEBUG_BUILD +#if !defined(NDEBUG) || defined(ADDRESS_SANITIZER) || \ + defined(MEMORY_SANITIZER) || defined(THREAD_SANITIZER) || \ + defined(__clang_analyzer__) +#define JXL_IS_DEBUG_BUILD 1 +#else +#define JXL_IS_DEBUG_BUILD 0 +#endif +#endif // JXL_IS_DEBUG_BUILD + +// Same as above, but only runs in debug builds (builds where NDEBUG is not +// defined). This is useful for slower asserts that we want to run more rarely +// than usual. These will run on asan, msan and other debug builds, but not in +// opt or release. +#if JXL_IS_DEBUG_BUILD +#define JXL_DASSERT(condition) \ + do { \ + if (!(condition)) { \ + JXL_DEBUG(JXL_DEBUG_ON_ABORT, "JXL_DASSERT: %s", #condition); \ + ::jxl::Abort(); \ + } \ + } while (0) +#else +#define JXL_DASSERT(condition) \ + do { \ + } while (0) +#endif + +// Always runs the condition, so can be used for non-debug calls. +#if JXL_ENABLE_CHECK +#define JXL_CHECK(condition) \ + do { \ + if (!(condition)) { \ + JXL_DEBUG(JXL_DEBUG_ON_ABORT, "JXL_CHECK: %s", #condition); \ + ::jxl::Abort(); \ + } \ + } while (0) +#else +#define JXL_CHECK(condition) \ + do { \ + (void)(condition); \ + } while (0) +#endif + +// A jxl::Status value from a StatusCode or Status which prints a debug message +// when enabled. +#define JXL_STATUS(status, format, ...) \ + ::jxl::StatusMessage(::jxl::Status(status), "%s:%d: " format "\n", __FILE__, \ + __LINE__, ##__VA_ARGS__) + +// Notify of an error but discard the resulting Status value. This is only +// useful for debug builds or when building with JXL_CRASH_ON_ERROR. +#define JXL_NOTIFY_ERROR(format, ...) \ + (void)JXL_STATUS(::jxl::StatusCode::kGenericError, "JXL_ERROR: " format, \ + ##__VA_ARGS__) + +// An error Status with a message. The JXL_STATUS() macro will return a Status +// object with a kGenericError code, but the comma operator helps with +// clang-tidy inference and potentially with optimizations. +#define JXL_FAILURE(format, ...) \ + ((void)JXL_STATUS(::jxl::StatusCode::kGenericError, "JXL_FAILURE: " format, \ + ##__VA_ARGS__), \ + ::jxl::Status(::jxl::StatusCode::kGenericError)) + +// Always evaluates the status exactly once, so can be used for non-debug calls. +// Returns from the current context if the passed Status expression is an error +// (fatal or non-fatal). The return value is the passed Status. +#define JXL_RETURN_IF_ERROR(status) \ + do { \ + ::jxl::Status jxl_return_if_error_status = (status); \ + if (!jxl_return_if_error_status) { \ + (void)::jxl::StatusMessage( \ + jxl_return_if_error_status, \ + "%s:%d: JXL_RETURN_IF_ERROR code=%d: %s\n", __FILE__, __LINE__, \ + static_cast(jxl_return_if_error_status.code()), #status); \ + return jxl_return_if_error_status; \ + } \ + } while (0) + +// As above, but without calling StatusMessage. Intended for bundles (see +// fields.h), which have numerous call sites (-> relevant for code size) and do +// not want to generate excessive messages when decoding partial headers. +#define JXL_QUIET_RETURN_IF_ERROR(status) \ + do { \ + ::jxl::Status jxl_return_if_error_status = (status); \ + if (!jxl_return_if_error_status) { \ + return jxl_return_if_error_status; \ + } \ + } while (0) + +enum class StatusCode : int32_t { + // Non-fatal errors (negative values). + kNotEnoughBytes = -1, + + // The only non-error status code. + kOk = 0, + + // Fatal-errors (positive values) + kGenericError = 1, +}; + +// Drop-in replacement for bool that raises compiler warnings if not used +// after being returned from a function. Example: +// Status LoadFile(...) { return true; } is more compact than +// bool JXL_MUST_USE_RESULT LoadFile(...) { return true; } +// In case of error, the status can carry an extra error code in its value which +// is split between fatal and non-fatal error codes. +class JXL_MUST_USE_RESULT Status { + public: + // We want implicit constructor from bool to allow returning "true" or "false" + // on a function when using Status. "true" means kOk while "false" means a + // generic fatal error. + // NOLINTNEXTLINE(google-explicit-constructor) + constexpr Status(bool ok) + : code_(ok ? StatusCode::kOk : StatusCode::kGenericError) {} + + // NOLINTNEXTLINE(google-explicit-constructor) + constexpr Status(StatusCode code) : code_(code) {} + + // We also want implicit cast to bool to check for return values of functions. + // NOLINTNEXTLINE(google-explicit-constructor) + constexpr operator bool() const { return code_ == StatusCode::kOk; } + + constexpr StatusCode code() const { return code_; } + + // Returns whether the status code is a fatal error. + constexpr bool IsFatalError() const { + return static_cast(code_) > 0; + } + + private: + StatusCode code_; +}; + +static constexpr Status OkStatus() { return Status(StatusCode::kOk); } + +// Helper function to create a Status and print the debug message or abort when +// needed. +inline JXL_FORMAT(2, 3) Status + StatusMessage(const Status status, const char* format, ...) { + // This block will be optimized out when JXL_DEBUG_ON_ERROR and + // JXL_DEBUG_ON_ALL_ERROR are both disabled. + if ((JXL_DEBUG_ON_ERROR && status.IsFatalError()) || + (JXL_DEBUG_ON_ALL_ERROR && !status)) { + va_list args; + va_start(args, format); +#ifdef USE_ANDROID_LOGGER + android_vprintf(format, args); +#else + vfprintf(stderr, format, args); +#endif + va_end(args); + } +#ifdef JXL_CRASH_ON_ERROR + // JXL_CRASH_ON_ERROR means to Abort() only on non-fatal errors. + if (status.IsFatalError()) { + Abort(); + } +#endif // JXL_CRASH_ON_ERROR + return status; +} + +template +class JXL_MUST_USE_RESULT StatusOr { + static_assert(!std::is_convertible::value && + !std::is_convertible::value, + "You cannot make a StatusOr with a type convertible from or to " + "StatusCode"); + static_assert(std::is_move_constructible::value && + std::is_move_assignable::value, + "T must be move constructible and move assignable"); + + public: + // NOLINTNEXTLINE(google-explicit-constructor) + StatusOr(StatusCode code) : code_(code) { + JXL_ASSERT(code_ != StatusCode::kOk); + } + + // NOLINTNEXTLINE(google-explicit-constructor) + StatusOr(Status status) : StatusOr(status.code()) {} + + // NOLINTNEXTLINE(google-explicit-constructor) + StatusOr(T&& value) : code_(StatusCode::kOk) { + new (&storage_.data_) T(std::move(value)); + } + + StatusOr(StatusOr&& other) noexcept { + if (other.ok()) { + new (&storage_.data_) T(std::move(other.storage_.data_)); + } + code_ = other.code_; + } + + StatusOr& operator=(StatusOr&& other) noexcept { + if (this == &other) return *this; + if (ok() && other.ok()) { + storage_.data_ = std::move(other.storage_.data_); + } else if (other.ok()) { + new (&storage_.data_) T(std::move(other.storage_.data_)); + } else if (ok()) { + storage_.data_.~T(); + } + code_ = other.code_; + return *this; + } + + StatusOr(const StatusOr&) = delete; + StatusOr operator=(const StatusOr&) = delete; + + bool ok() const { return code_ == StatusCode::kOk; } + Status status() const { return code_; } + + // Only call this if you are absolutely sure that `ok()` is true. + // Ideally, never call this manually and rely on JXL_ASSIGN_OR_RETURN. + T value() && { + JXL_ASSERT(ok()); + return std::move(storage_.data_); + } + + ~StatusOr() { + if (code_ == StatusCode::kOk) { + storage_.data_.~T(); + } + } + + private: + union Storage { + char placeholder_; + T data_; + Storage() {} + ~Storage() {} + } storage_; + + StatusCode code_; +}; + +#define JXL_ASSIGN_OR_RETURN(lhs, statusor) \ + PRIVATE_JXL_ASSIGN_OR_RETURN_IMPL( \ + assign_or_return_temporary_variable##__LINE__, lhs, statusor) + +// NOLINTBEGIN(bugprone-macro-parentheses) +#define PRIVATE_JXL_ASSIGN_OR_RETURN_IMPL(name, lhs, statusor) \ + auto name = statusor; \ + JXL_RETURN_IF_ERROR(name.status()); \ + lhs = std::move(name).value(); +// NOLINTEND(bugprone-macro-parentheses) + +} // namespace jxl + +#endif // LIB_JXL_BASE_STATUS_H_ -- cgit v1.2.3