gecko/media/libyuv/util/psnr.cc
Randell Jesup c5fd234270 Bug 813645 - Move libyuv to media/libyuv r=glandium
--HG--
rename : media/webrtc/trunk/third_party/libyuv/AUTHORS => media/libyuv/AUTHORS
rename : media/webrtc/trunk/third_party/libyuv/Android.mk => media/libyuv/Android.mk
rename : media/webrtc/trunk/third_party/libyuv/DEPS => media/libyuv/DEPS
rename : media/webrtc/trunk/third_party/libyuv/LICENSE => media/libyuv/LICENSE
rename : media/webrtc/trunk/third_party/libyuv/LICENSE_THIRD_PARTY => media/libyuv/LICENSE_THIRD_PARTY
rename : media/webrtc/trunk/third_party/libyuv/OWNERS => media/libyuv/OWNERS
rename : media/webrtc/trunk/third_party/libyuv/PATENTS => media/libyuv/PATENTS
rename : media/webrtc/trunk/third_party/libyuv/README.chromium => media/libyuv/README.chromium
rename : media/webrtc/trunk/third_party/libyuv/all.gyp => media/libyuv/all.gyp
rename : media/webrtc/trunk/third_party/libyuv/codereview.settings => media/libyuv/codereview.settings
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv.h => media/libyuv/include/libyuv.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/basic_types.h => media/libyuv/include/libyuv/basic_types.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/compare.h => media/libyuv/include/libyuv/compare.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/convert.h => media/libyuv/include/libyuv/convert.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/convert_argb.h => media/libyuv/include/libyuv/convert_argb.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/convert_from.h => media/libyuv/include/libyuv/convert_from.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/convert_from_argb.h => media/libyuv/include/libyuv/convert_from_argb.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/cpu_id.h => media/libyuv/include/libyuv/cpu_id.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/format_conversion.h => media/libyuv/include/libyuv/format_conversion.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/mjpeg_decoder.h => media/libyuv/include/libyuv/mjpeg_decoder.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/planar_functions.h => media/libyuv/include/libyuv/planar_functions.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/rotate.h => media/libyuv/include/libyuv/rotate.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/rotate_argb.h => media/libyuv/include/libyuv/rotate_argb.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/row.h => media/libyuv/include/libyuv/row.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/scale.h => media/libyuv/include/libyuv/scale.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/scale_argb.h => media/libyuv/include/libyuv/scale_argb.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/scale_row.h => media/libyuv/include/libyuv/scale_row.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/version.h => media/libyuv/include/libyuv/version.h
rename : media/webrtc/trunk/third_party/libyuv/include/libyuv/video_common.h => media/libyuv/include/libyuv/video_common.h
rename : media/webrtc/trunk/third_party/libyuv/libyuv.gyp => media/libyuv/libyuv.gyp
rename : media/webrtc/trunk/third_party/libyuv/libyuv.gypi => media/libyuv/libyuv.gypi
rename : media/webrtc/trunk/third_party/libyuv/libyuv_nacl.gyp => media/libyuv/libyuv_nacl.gyp
rename : media/webrtc/trunk/third_party/libyuv/libyuv_test.gyp => media/libyuv/libyuv_test.gyp
rename : media/webrtc/trunk/third_party/libyuv/linux.mk => media/libyuv/linux.mk
rename : media/webrtc/trunk/third_party/libyuv/public.mk => media/libyuv/public.mk
rename : media/webrtc/trunk/third_party/libyuv/source/compare.cc => media/libyuv/source/compare.cc
rename : media/webrtc/trunk/third_party/libyuv/source/compare_common.cc => media/libyuv/source/compare_common.cc
rename : media/webrtc/trunk/third_party/libyuv/source/compare_neon.cc => media/libyuv/source/compare_neon.cc
rename : media/webrtc/trunk/third_party/libyuv/source/compare_posix.cc => media/libyuv/source/compare_posix.cc
rename : media/webrtc/trunk/third_party/libyuv/source/compare_win.cc => media/libyuv/source/compare_win.cc
rename : media/webrtc/trunk/third_party/libyuv/source/convert.cc => media/libyuv/source/convert.cc
rename : media/webrtc/trunk/third_party/libyuv/source/convert_argb.cc => media/libyuv/source/convert_argb.cc
rename : media/webrtc/trunk/third_party/libyuv/source/convert_from.cc => media/libyuv/source/convert_from.cc
rename : media/webrtc/trunk/third_party/libyuv/source/convert_from_argb.cc => media/libyuv/source/convert_from_argb.cc
rename : media/webrtc/trunk/third_party/libyuv/source/convert_jpeg.cc => media/libyuv/source/convert_jpeg.cc
rename : media/webrtc/trunk/third_party/libyuv/source/convert_to_argb.cc => media/libyuv/source/convert_to_argb.cc
rename : media/webrtc/trunk/third_party/libyuv/source/convert_to_i420.cc => media/libyuv/source/convert_to_i420.cc
rename : media/webrtc/trunk/third_party/libyuv/source/cpu_id.cc => media/libyuv/source/cpu_id.cc
rename : media/webrtc/trunk/third_party/libyuv/source/format_conversion.cc => media/libyuv/source/format_conversion.cc
rename : media/webrtc/trunk/third_party/libyuv/source/mjpeg_decoder.cc => media/libyuv/source/mjpeg_decoder.cc
rename : media/webrtc/trunk/third_party/libyuv/source/mjpeg_validate.cc => media/libyuv/source/mjpeg_validate.cc
rename : media/webrtc/trunk/third_party/libyuv/source/planar_functions.cc => media/libyuv/source/planar_functions.cc
rename : media/webrtc/trunk/third_party/libyuv/source/rotate.cc => media/libyuv/source/rotate.cc
rename : media/webrtc/trunk/third_party/libyuv/source/rotate_argb.cc => media/libyuv/source/rotate_argb.cc
rename : media/webrtc/trunk/third_party/libyuv/source/rotate_mips.cc => media/libyuv/source/rotate_mips.cc
rename : media/webrtc/trunk/third_party/libyuv/source/rotate_neon.cc => media/libyuv/source/rotate_neon.cc
rename : media/webrtc/trunk/third_party/libyuv/source/row_any.cc => media/libyuv/source/row_any.cc
rename : media/webrtc/trunk/third_party/libyuv/source/row_common.cc => media/libyuv/source/row_common.cc
rename : media/webrtc/trunk/third_party/libyuv/source/row_mips.cc => media/libyuv/source/row_mips.cc
rename : media/webrtc/trunk/third_party/libyuv/source/row_neon.cc => media/libyuv/source/row_neon.cc
rename : media/webrtc/trunk/third_party/libyuv/source/row_posix.cc => media/libyuv/source/row_posix.cc
rename : media/webrtc/trunk/third_party/libyuv/source/row_win.cc => media/libyuv/source/row_win.cc
rename : media/webrtc/trunk/third_party/libyuv/source/row_x86.asm => media/libyuv/source/row_x86.asm
rename : media/webrtc/trunk/third_party/libyuv/source/scale.cc => media/libyuv/source/scale.cc
rename : media/webrtc/trunk/third_party/libyuv/source/scale_argb.cc => media/libyuv/source/scale_argb.cc
rename : media/webrtc/trunk/third_party/libyuv/source/scale_common.cc => media/libyuv/source/scale_common.cc
rename : media/webrtc/trunk/third_party/libyuv/source/scale_mips.cc => media/libyuv/source/scale_mips.cc
rename : media/webrtc/trunk/third_party/libyuv/source/scale_neon.cc => media/libyuv/source/scale_neon.cc
rename : media/webrtc/trunk/third_party/libyuv/source/scale_posix.cc => media/libyuv/source/scale_posix.cc
rename : media/webrtc/trunk/third_party/libyuv/source/scale_win.cc => media/libyuv/source/scale_win.cc
rename : media/webrtc/trunk/third_party/libyuv/source/video_common.cc => media/libyuv/source/video_common.cc
rename : media/webrtc/trunk/third_party/libyuv/source/x86inc.asm => media/libyuv/source/x86inc.asm
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/libyuv_tests.bat => media/libyuv/tools/valgrind-libyuv/libyuv_tests.bat
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/libyuv_tests.py => media/libyuv/tools/valgrind-libyuv/libyuv_tests.py
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/libyuv_tests.sh => media/libyuv/tools/valgrind-libyuv/libyuv_tests.sh
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/memcheck/OWNERS => media/libyuv/tools/valgrind-libyuv/memcheck/OWNERS
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/memcheck/PRESUBMIT.py => media/libyuv/tools/valgrind-libyuv/memcheck/PRESUBMIT.py
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/memcheck/suppressions.txt => media/libyuv/tools/valgrind-libyuv/memcheck/suppressions.txt
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/memcheck/suppressions_mac.txt => media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_mac.txt
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/memcheck/suppressions_win32.txt => media/libyuv/tools/valgrind-libyuv/memcheck/suppressions_win32.txt
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/tsan/OWNERS => media/libyuv/tools/valgrind-libyuv/tsan/OWNERS
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/tsan/PRESUBMIT.py => media/libyuv/tools/valgrind-libyuv/tsan/PRESUBMIT.py
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/tsan/suppressions.txt => media/libyuv/tools/valgrind-libyuv/tsan/suppressions.txt
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/tsan/suppressions_mac.txt => media/libyuv/tools/valgrind-libyuv/tsan/suppressions_mac.txt
rename : media/webrtc/trunk/third_party/libyuv/tools/valgrind-libyuv/tsan/suppressions_win32.txt => media/libyuv/tools/valgrind-libyuv/tsan/suppressions_win32.txt
rename : media/webrtc/trunk/third_party/libyuv/unit_test/basictypes_test.cc => media/libyuv/unit_test/basictypes_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/compare_test.cc => media/libyuv/unit_test/compare_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/convert_test.cc => media/libyuv/unit_test/convert_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/cpu_test.cc => media/libyuv/unit_test/cpu_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/math_test.cc => media/libyuv/unit_test/math_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/planar_test.cc => media/libyuv/unit_test/planar_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/rotate_argb_test.cc => media/libyuv/unit_test/rotate_argb_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/rotate_test.cc => media/libyuv/unit_test/rotate_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/scale_argb_test.cc => media/libyuv/unit_test/scale_argb_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/scale_test.cc => media/libyuv/unit_test/scale_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/testdata/arm_v7.txt => media/libyuv/unit_test/testdata/arm_v7.txt
rename : media/webrtc/trunk/third_party/libyuv/unit_test/testdata/tegra3.txt => media/libyuv/unit_test/testdata/tegra3.txt
rename : media/webrtc/trunk/third_party/libyuv/unit_test/unit_test.cc => media/libyuv/unit_test/unit_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/unit_test.h => media/libyuv/unit_test/unit_test.h
rename : media/webrtc/trunk/third_party/libyuv/unit_test/version_test.cc => media/libyuv/unit_test/version_test.cc
rename : media/webrtc/trunk/third_party/libyuv/unit_test/video_common_test.cc => media/libyuv/unit_test/video_common_test.cc
rename : media/webrtc/trunk/third_party/libyuv/util/Makefile => media/libyuv/util/Makefile
rename : media/webrtc/trunk/third_party/libyuv/util/compare.cc => media/libyuv/util/compare.cc
rename : media/webrtc/trunk/third_party/libyuv/util/convert.cc => media/libyuv/util/convert.cc
rename : media/webrtc/trunk/third_party/libyuv/util/cpuid.c => media/libyuv/util/cpuid.c
rename : media/webrtc/trunk/third_party/libyuv/util/psnr.cc => media/libyuv/util/psnr.cc
rename : media/webrtc/trunk/third_party/libyuv/util/psnr.h => media/libyuv/util/psnr.h
rename : media/webrtc/trunk/third_party/libyuv/util/psnr_main.cc => media/libyuv/util/psnr_main.cc
rename : media/webrtc/trunk/third_party/libyuv/util/ssim.cc => media/libyuv/util/ssim.cc
rename : media/webrtc/trunk/third_party/libyuv/util/ssim.h => media/libyuv/util/ssim.h
rename : media/webrtc/trunk/third_party/libyuv/winarm.mk => media/libyuv/winarm.mk
2014-01-30 19:55:04 -05:00

248 lines
7.9 KiB
C++

/*
* Copyright 2013 The LibYuv 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 in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "./psnr.h" // NOLINT
#include <math.h>
#ifdef _OPENMP
#include <omp.h>
#endif
#ifdef _MSC_VER
#include <intrin.h> // For __cpuid()
#endif
#ifdef __cplusplus
extern "C" {
#endif
typedef unsigned int uint32; // NOLINT
#ifdef _MSC_VER
typedef unsigned __int64 uint64;
#else // COMPILER_MSVC
#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
typedef unsigned long uint64; // NOLINT
#else // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
typedef unsigned long long uint64; // NOLINT
#endif // __LP64__
#endif // _MSC_VER
// PSNR formula: psnr = 10 * log10 (Peak Signal^2 * size / sse)
double ComputePSNR(double sse, double size) {
const double kMINSSE = 255.0 * 255.0 * size / pow(10., kMaxPSNR / 10.);
if (sse <= kMINSSE)
sse = kMINSSE; // Produces max PSNR of 128
return 10.0 * log10(65025.0 * size / sse);
}
#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
#define HAS_SUMSQUAREERROR_NEON
static uint32 SumSquareError_NEON(const uint8* src_a,
const uint8* src_b, int count) {
volatile uint32 sse;
asm volatile ( // NOLINT
"vmov.u8 q7, #0 \n"
"vmov.u8 q9, #0 \n"
"vmov.u8 q8, #0 \n"
"vmov.u8 q10, #0 \n"
"1: \n"
"vld1.u8 {q0}, [%0]! \n"
"vld1.u8 {q1}, [%1]! \n"
"vsubl.u8 q2, d0, d2 \n"
"vsubl.u8 q3, d1, d3 \n"
"vmlal.s16 q7, d4, d4 \n"
"vmlal.s16 q8, d6, d6 \n"
"vmlal.s16 q8, d5, d5 \n"
"vmlal.s16 q10, d7, d7 \n"
"subs %2, %2, #16 \n"
"bhi 1b \n"
"vadd.u32 q7, q7, q8 \n"
"vadd.u32 q9, q9, q10 \n"
"vadd.u32 q10, q7, q9 \n"
"vpaddl.u32 q1, q10 \n"
"vadd.u64 d0, d2, d3 \n"
"vmov.32 %3, d0[0] \n"
: "+r"(src_a),
"+r"(src_b),
"+r"(count),
"=r"(sse)
:
: "memory", "cc", "q0", "q1", "q2", "q3", "q7", "q8", "q9", "q10");
return sse;
}
#elif !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
#define HAS_SUMSQUAREERROR_SSE2
__declspec(naked)
static uint32 SumSquareError_SSE2(const uint8* /*src_a*/,
const uint8* /*src_b*/, int /*count*/) {
__asm {
mov eax, [esp + 4] // src_a
mov edx, [esp + 8] // src_b
mov ecx, [esp + 12] // count
pxor xmm0, xmm0
pxor xmm5, xmm5
sub edx, eax
wloop:
movdqu xmm1, [eax]
movdqu xmm2, [eax + edx]
lea eax, [eax + 16]
movdqu xmm3, xmm1
psubusb xmm1, xmm2
psubusb xmm2, xmm3
por xmm1, xmm2
movdqu xmm2, xmm1
punpcklbw xmm1, xmm5
punpckhbw xmm2, xmm5
pmaddwd xmm1, xmm1
pmaddwd xmm2, xmm2
paddd xmm0, xmm1
paddd xmm0, xmm2
sub ecx, 16
ja wloop
pshufd xmm1, xmm0, 0EEh
paddd xmm0, xmm1
pshufd xmm1, xmm0, 01h
paddd xmm0, xmm1
movd eax, xmm0
ret
}
}
#elif !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
#define HAS_SUMSQUAREERROR_SSE2
static uint32 SumSquareError_SSE2(const uint8* src_a,
const uint8* src_b, int count) {
uint32 sse;
asm volatile ( // NOLINT
"pxor %%xmm0,%%xmm0 \n"
"pxor %%xmm5,%%xmm5 \n"
"sub %0,%1 \n"
"1: \n"
"movdqu (%0),%%xmm1 \n"
"movdqu (%0,%1,1),%%xmm2 \n"
"lea 0x10(%0),%0 \n"
"movdqu %%xmm1,%%xmm3 \n"
"psubusb %%xmm2,%%xmm1 \n"
"psubusb %%xmm3,%%xmm2 \n"
"por %%xmm2,%%xmm1 \n"
"movdqu %%xmm1,%%xmm2 \n"
"punpcklbw %%xmm5,%%xmm1 \n"
"punpckhbw %%xmm5,%%xmm2 \n"
"pmaddwd %%xmm1,%%xmm1 \n"
"pmaddwd %%xmm2,%%xmm2 \n"
"paddd %%xmm1,%%xmm0 \n"
"paddd %%xmm2,%%xmm0 \n"
"sub $0x10,%2 \n"
"ja 1b \n"
"pshufd $0xee,%%xmm0,%%xmm1 \n"
"paddd %%xmm1,%%xmm0 \n"
"pshufd $0x1,%%xmm0,%%xmm1 \n"
"paddd %%xmm1,%%xmm0 \n"
"movd %%xmm0,%3 \n"
: "+r"(src_a), // %0
"+r"(src_b), // %1
"+r"(count), // %2
"=g"(sse) // %3
:
: "memory", "cc"
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
#endif
); // NOLINT
return sse;
}
#endif // LIBYUV_DISABLE_X86 etc
#if defined(HAS_SUMSQUAREERROR_SSE2)
#if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
static __inline void __cpuid(int cpu_info[4], int info_type) {
asm volatile ( // NOLINT
"mov %%ebx, %%edi \n"
"cpuid \n"
"xchg %%edi, %%ebx \n"
: "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
: "a"(info_type));
}
#elif defined(__i386__) || defined(__x86_64__)
static __inline void __cpuid(int cpu_info[4], int info_type) {
asm volatile ( // NOLINT
"cpuid \n"
: "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
: "a"(info_type));
}
#endif
static int CpuHasSSE2() {
#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86)
int cpu_info[4];
__cpuid(cpu_info, 1);
if (cpu_info[3] & 0x04000000) {
return 1;
}
#endif
return 0;
}
#endif // HAS_SUMSQUAREERROR_SSE2
static uint32 SumSquareError_C(const uint8* src_a,
const uint8* src_b, int count) {
uint32 sse = 0u;
for (int x = 0; x < count; ++x) {
int diff = src_a[x] - src_b[x];
sse += static_cast<uint32>(diff * diff);
}
return sse;
}
double ComputeSumSquareError(const uint8* src_a,
const uint8* src_b, int count) {
uint32 (*SumSquareError)(const uint8* src_a,
const uint8* src_b, int count) = SumSquareError_C;
#if defined(HAS_SUMSQUAREERROR_NEON)
SumSquareError = SumSquareError_NEON;
#endif
#if defined(HAS_SUMSQUAREERROR_SSE2)
if (CpuHasSSE2()) {
SumSquareError = SumSquareError_SSE2;
}
#endif
const int kBlockSize = 1 << 15;
uint64 sse = 0;
#ifdef _OPENMP
#pragma omp parallel for reduction(+: sse)
#endif
for (int i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
}
src_a += count & ~(kBlockSize - 1);
src_b += count & ~(kBlockSize - 1);
int remainder = count & (kBlockSize - 1) & ~15;
if (remainder) {
sse += SumSquareError(src_a, src_b, remainder);
src_a += remainder;
src_b += remainder;
}
remainder = count & 15;
if (remainder) {
sse += SumSquareError_C(src_a, src_b, remainder);
}
return static_cast<double>(sse);
}
#ifdef __cplusplus
} // extern "C"
#endif