react-native-vision-camera-spoof-detector 1.0.21 → 1.0.23
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.gitattributes +2 -0
- package/.gitignore +27 -0
- package/CHANGELOG.md +2 -2
- package/CONTRIBUTING.md +2 -2
- package/README.md +442 -443
- package/android/build.gradle +12 -4
- package/android/src/main/cpp/CMakeLists.txt +101 -94
- package/android/src/main/cpp/libyuv/.clang-format +6 -6
- package/android/src/main/cpp/libyuv/.gn +40 -40
- package/android/src/main/cpp/libyuv/.vpython3 +410 -410
- package/android/src/main/cpp/libyuv/AUTHORS +7 -7
- package/android/src/main/cpp/libyuv/Android.bp +202 -202
- package/android/src/main/cpp/libyuv/Android.mk +106 -106
- package/android/src/main/cpp/libyuv/BUILD.gn +490 -490
- package/android/src/main/cpp/libyuv/CM_linux_packages.cmake +70 -70
- package/android/src/main/cpp/libyuv/CMakeLists.txt +259 -252
- package/android/src/main/cpp/libyuv/DEPS +935 -935
- package/android/src/main/cpp/libyuv/DIR_METADATA +3 -3
- package/android/src/main/cpp/libyuv/LICENSE +29 -29
- package/android/src/main/cpp/libyuv/OWNERS +11 -11
- package/android/src/main/cpp/libyuv/PATENTS +23 -23
- package/android/src/main/cpp/libyuv/PRESUBMIT.py +65 -65
- package/android/src/main/cpp/libyuv/README.chromium +11 -11
- package/android/src/main/cpp/libyuv/README.md +19 -19
- package/android/src/main/cpp/libyuv/build_overrides/build.gni +60 -60
- package/android/src/main/cpp/libyuv/build_overrides/gtest.gni +19 -19
- package/android/src/main/cpp/libyuv/build_overrides/partition_alloc.gni +17 -17
- package/android/src/main/cpp/libyuv/codereview.settings +5 -5
- package/android/src/main/cpp/libyuv/docs/deprecated_builds.md +409 -409
- package/android/src/main/cpp/libyuv/docs/environment_variables.md +64 -64
- package/android/src/main/cpp/libyuv/docs/feature_detection.md +108 -108
- package/android/src/main/cpp/libyuv/docs/filtering.md +196 -196
- package/android/src/main/cpp/libyuv/docs/formats.md +208 -208
- package/android/src/main/cpp/libyuv/docs/getting_started.md +296 -296
- package/android/src/main/cpp/libyuv/docs/rotation.md +107 -107
- package/android/src/main/cpp/libyuv/download_vs_toolchain.py +29 -29
- package/android/src/main/cpp/libyuv/include/libyuv/basic_types.h +68 -68
- package/android/src/main/cpp/libyuv/include/libyuv/compare.h +111 -111
- package/android/src/main/cpp/libyuv/include/libyuv/compare_row.h +112 -112
- package/android/src/main/cpp/libyuv/include/libyuv/convert.h +1117 -1117
- package/android/src/main/cpp/libyuv/include/libyuv/convert_argb.h +2335 -2335
- package/android/src/main/cpp/libyuv/include/libyuv/convert_from.h +203 -203
- package/android/src/main/cpp/libyuv/include/libyuv/convert_from_argb.h +394 -394
- package/android/src/main/cpp/libyuv/include/libyuv/cpu_id.h +147 -147
- package/android/src/main/cpp/libyuv/include/libyuv/cpu_support.h +99 -99
- package/android/src/main/cpp/libyuv/include/libyuv/loongson_intrinsics.h +1949 -1949
- package/android/src/main/cpp/libyuv/include/libyuv/mjpeg_decoder.h +195 -195
- package/android/src/main/cpp/libyuv/include/libyuv/planar_functions.h +1131 -1131
- package/android/src/main/cpp/libyuv/include/libyuv/rotate.h +296 -296
- package/android/src/main/cpp/libyuv/include/libyuv/rotate_argb.h +37 -37
- package/android/src/main/cpp/libyuv/include/libyuv/rotate_row.h +265 -265
- package/android/src/main/cpp/libyuv/include/libyuv/row.h +6738 -6738
- package/android/src/main/cpp/libyuv/include/libyuv/row_sve.h +2154 -2154
- package/android/src/main/cpp/libyuv/include/libyuv/scale.h +336 -336
- package/android/src/main/cpp/libyuv/include/libyuv/scale_argb.h +76 -76
- package/android/src/main/cpp/libyuv/include/libyuv/scale_rgb.h +42 -42
- package/android/src/main/cpp/libyuv/include/libyuv/scale_row.h +1726 -1726
- package/android/src/main/cpp/libyuv/include/libyuv/scale_uv.h +51 -51
- package/android/src/main/cpp/libyuv/include/libyuv/version.h +16 -16
- package/android/src/main/cpp/libyuv/include/libyuv/video_common.h +222 -222
- package/android/src/main/cpp/libyuv/include/libyuv.h +33 -33
- package/android/src/main/cpp/libyuv/infra/config/OWNERS +3 -3
- package/android/src/main/cpp/libyuv/infra/config/PRESUBMIT.py +17 -17
- package/android/src/main/cpp/libyuv/infra/config/README.md +2 -2
- package/android/src/main/cpp/libyuv/infra/config/codereview.settings +6 -6
- package/android/src/main/cpp/libyuv/infra/config/commit-queue.cfg +144 -144
- package/android/src/main/cpp/libyuv/infra/config/cr-buildbucket.cfg +1185 -1185
- package/android/src/main/cpp/libyuv/infra/config/luci-logdog.cfg +9 -9
- package/android/src/main/cpp/libyuv/infra/config/luci-milo.cfg +246 -246
- package/android/src/main/cpp/libyuv/infra/config/luci-scheduler.cfg +385 -385
- package/android/src/main/cpp/libyuv/infra/config/main.star +402 -402
- package/android/src/main/cpp/libyuv/infra/config/project.cfg +16 -16
- package/android/src/main/cpp/libyuv/infra/config/realms.cfg +111 -111
- package/android/src/main/cpp/libyuv/libyuv.gni +34 -34
- package/android/src/main/cpp/libyuv/libyuv.gyp +149 -149
- package/android/src/main/cpp/libyuv/libyuv.gypi +87 -87
- package/android/src/main/cpp/libyuv/linux.mk +96 -96
- package/android/src/main/cpp/libyuv/public.mk +13 -13
- package/android/src/main/cpp/libyuv/pylintrc +49 -49
- package/android/src/main/cpp/libyuv/riscv_script/prepare_toolchain_qemu.sh +74 -74
- package/android/src/main/cpp/libyuv/riscv_script/riscv-clang.cmake +56 -56
- package/android/src/main/cpp/libyuv/riscv_script/run_qemu.sh +15 -15
- package/android/src/main/cpp/libyuv/source/compare.cc +435 -435
- package/android/src/main/cpp/libyuv/source/compare_common.cc +74 -74
- package/android/src/main/cpp/libyuv/source/compare_gcc.cc +362 -362
- package/android/src/main/cpp/libyuv/source/compare_neon.cc +96 -96
- package/android/src/main/cpp/libyuv/source/compare_neon64.cc +223 -223
- package/android/src/main/cpp/libyuv/source/compare_win.cc +241 -241
- package/android/src/main/cpp/libyuv/source/convert.cc +4746 -4746
- package/android/src/main/cpp/libyuv/source/convert_argb.cc +9179 -9179
- package/android/src/main/cpp/libyuv/source/convert_from.cc +866 -866
- package/android/src/main/cpp/libyuv/source/convert_from_argb.cc +3671 -3671
- package/android/src/main/cpp/libyuv/source/convert_jpeg.cc +602 -602
- package/android/src/main/cpp/libyuv/source/convert_to_argb.cc +391 -391
- package/android/src/main/cpp/libyuv/source/convert_to_i420.cc +288 -288
- package/android/src/main/cpp/libyuv/source/cpu_id.cc +496 -496
- package/android/src/main/cpp/libyuv/source/mjpeg_decoder.cc +580 -580
- package/android/src/main/cpp/libyuv/source/mjpeg_validate.cc +71 -71
- package/android/src/main/cpp/libyuv/source/planar_functions.cc +5663 -5663
- package/android/src/main/cpp/libyuv/source/rotate.cc +1241 -1241
- package/android/src/main/cpp/libyuv/source/rotate_any.cc +76 -76
- package/android/src/main/cpp/libyuv/source/rotate_argb.cc +259 -259
- package/android/src/main/cpp/libyuv/source/rotate_common.cc +208 -208
- package/android/src/main/cpp/libyuv/source/rotate_gcc.cc +505 -505
- package/android/src/main/cpp/libyuv/source/rotate_lsx.cc +233 -233
- package/android/src/main/cpp/libyuv/source/rotate_neon.cc +219 -219
- package/android/src/main/cpp/libyuv/source/rotate_neon64.cc +273 -273
- package/android/src/main/cpp/libyuv/source/rotate_sme.cc +174 -174
- package/android/src/main/cpp/libyuv/source/rotate_win.cc +253 -253
- package/android/src/main/cpp/libyuv/source/row_any.cc +2519 -2519
- package/android/src/main/cpp/libyuv/source/row_common.cc +4461 -4461
- package/android/src/main/cpp/libyuv/source/row_gcc.cc +9570 -9570
- package/android/src/main/cpp/libyuv/source/row_lasx.cc +2343 -2343
- package/android/src/main/cpp/libyuv/source/row_lsx.cc +3030 -3030
- package/android/src/main/cpp/libyuv/source/row_neon.cc +4026 -4026
- package/android/src/main/cpp/libyuv/source/row_neon64.cc +5617 -5617
- package/android/src/main/cpp/libyuv/source/row_rvv.cc +2599 -2599
- package/android/src/main/cpp/libyuv/source/row_sme.cc +1183 -1183
- package/android/src/main/cpp/libyuv/source/row_sve.cc +1088 -1088
- package/android/src/main/cpp/libyuv/source/row_win.cc +6453 -6453
- package/android/src/main/cpp/libyuv/source/scale.cc +2710 -2710
- package/android/src/main/cpp/libyuv/source/scale_any.cc +991 -991
- package/android/src/main/cpp/libyuv/source/scale_argb.cc +1158 -1158
- package/android/src/main/cpp/libyuv/source/scale_common.cc +1977 -1977
- package/android/src/main/cpp/libyuv/source/scale_gcc.cc +2947 -2947
- package/android/src/main/cpp/libyuv/source/scale_lsx.cc +739 -739
- package/android/src/main/cpp/libyuv/source/scale_neon.cc +1449 -1449
- package/android/src/main/cpp/libyuv/source/scale_neon64.cc +1552 -1552
- package/android/src/main/cpp/libyuv/source/scale_rgb.cc +82 -82
- package/android/src/main/cpp/libyuv/source/scale_rvv.cc +1971 -1971
- package/android/src/main/cpp/libyuv/source/scale_sme.cc +555 -555
- package/android/src/main/cpp/libyuv/source/scale_uv.cc +1159 -1159
- package/android/src/main/cpp/libyuv/source/scale_win.cc +1392 -1392
- package/android/src/main/cpp/libyuv/source/test.sh +35 -35
- package/android/src/main/cpp/libyuv/source/video_common.cc +62 -62
- package/android/src/main/cpp/libyuv/tools_libyuv/OWNERS +4 -4
- package/android/src/main/cpp/libyuv/tools_libyuv/autoroller/roll_deps.py +931 -931
- package/android/src/main/cpp/libyuv/tools_libyuv/autoroller/unittests/roll_deps_test.py +164 -164
- package/android/src/main/cpp/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS +21 -21
- package/android/src/main/cpp/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.new +13 -13
- package/android/src/main/cpp/libyuv/tools_libyuv/autoroller/unittests/testdata/DEPS.chromium.old +13 -13
- package/android/src/main/cpp/libyuv/tools_libyuv/get_landmines.py +38 -38
- package/android/src/main/cpp/libyuv/tools_libyuv/msan/OWNERS +3 -3
- package/android/src/main/cpp/libyuv/tools_libyuv/msan/blacklist.txt +9 -9
- package/android/src/main/cpp/libyuv/tools_libyuv/ubsan/OWNERS +3 -3
- package/android/src/main/cpp/libyuv/tools_libyuv/ubsan/blacklist.txt +15 -15
- package/android/src/main/cpp/libyuv/tools_libyuv/ubsan/vptr_blacklist.txt +25 -25
- package/android/src/main/cpp/libyuv/unit_test/basictypes_test.cc +43 -43
- package/android/src/main/cpp/libyuv/unit_test/color_test.cc +848 -848
- package/android/src/main/cpp/libyuv/unit_test/compare_test.cc +739 -739
- package/android/src/main/cpp/libyuv/unit_test/convert_argb_test.cc +2867 -2867
- package/android/src/main/cpp/libyuv/unit_test/convert_test.cc +2133 -2133
- package/android/src/main/cpp/libyuv/unit_test/cpu_test.cc +427 -427
- package/android/src/main/cpp/libyuv/unit_test/cpu_thread_test.cc +63 -63
- package/android/src/main/cpp/libyuv/unit_test/math_test.cc +160 -160
- package/android/src/main/cpp/libyuv/unit_test/planar_test.cc +4731 -4731
- package/android/src/main/cpp/libyuv/unit_test/rotate_argb_test.cc +334 -334
- package/android/src/main/cpp/libyuv/unit_test/rotate_test.cc +962 -962
- package/android/src/main/cpp/libyuv/unit_test/scale_argb_test.cc +590 -590
- package/android/src/main/cpp/libyuv/unit_test/scale_plane_test.cc +465 -465
- package/android/src/main/cpp/libyuv/unit_test/scale_rgb_test.cc +280 -280
- package/android/src/main/cpp/libyuv/unit_test/scale_test.cc +1135 -1135
- package/android/src/main/cpp/libyuv/unit_test/scale_uv_test.cc +249 -249
- package/android/src/main/cpp/libyuv/unit_test/testdata/arm_v7.txt +12 -12
- package/android/src/main/cpp/libyuv/unit_test/testdata/mips.txt +7 -7
- package/android/src/main/cpp/libyuv/unit_test/testdata/mips_loongson2k.txt +5 -5
- package/android/src/main/cpp/libyuv/unit_test/testdata/mips_loongson3.txt +10 -10
- package/android/src/main/cpp/libyuv/unit_test/testdata/mips_loongson_mmi.txt +7 -7
- package/android/src/main/cpp/libyuv/unit_test/testdata/riscv64.txt +3 -3
- package/android/src/main/cpp/libyuv/unit_test/testdata/riscv64_rvv.txt +3 -3
- package/android/src/main/cpp/libyuv/unit_test/testdata/riscv64_rvv_zvfh.txt +3 -3
- package/android/src/main/cpp/libyuv/unit_test/testdata/tegra3.txt +23 -23
- package/android/src/main/cpp/libyuv/unit_test/unit_test.cc +581 -581
- package/android/src/main/cpp/libyuv/unit_test/unit_test.h +232 -232
- package/android/src/main/cpp/libyuv/unit_test/video_common_test.cc +112 -112
- package/android/src/main/cpp/libyuv/util/Makefile +9 -9
- package/android/src/main/cpp/libyuv/util/color.cc +120 -120
- package/android/src/main/cpp/libyuv/util/compare.cc +67 -67
- package/android/src/main/cpp/libyuv/util/cpuid.c +211 -211
- package/android/src/main/cpp/libyuv/util/i444tonv12_eg.cc +27 -27
- package/android/src/main/cpp/libyuv/util/psnr.cc +291 -291
- package/android/src/main/cpp/libyuv/util/psnr.h +47 -47
- package/android/src/main/cpp/libyuv/util/psnr_main.cc +620 -620
- package/android/src/main/cpp/libyuv/util/ssim.cc +364 -364
- package/android/src/main/cpp/libyuv/util/ssim.h +38 -38
- package/android/src/main/cpp/libyuv/util/yuvconstants.c +114 -114
- package/android/src/main/cpp/libyuv/util/yuvconvert.cc +367 -367
- package/android/src/main/cpp/libyuv/winarm.mk +47 -47
- package/package.json +86 -84
- package/android/.gradle/8.9/checksums/checksums.lock +0 -0
- package/android/.gradle/8.9/dependencies-accessors/gc.properties +0 -0
- package/android/.gradle/8.9/fileChanges/last-build.bin +0 -0
- package/android/.gradle/8.9/fileHashes/fileHashes.lock +0 -0
- package/android/.gradle/8.9/gc.properties +0 -0
- package/android/.gradle/buildOutputCleanup/buildOutputCleanup.lock +0 -0
- package/android/.gradle/buildOutputCleanup/cache.properties +0 -2
- package/android/.gradle/vcs-1/gc.properties +0 -0
|
@@ -1,291 +1,291 @@
|
|
|
1
|
-
/*
|
|
2
|
-
* Copyright 2013 The LibYuv Project Authors. All rights reserved.
|
|
3
|
-
*
|
|
4
|
-
* Use of this source code is governed by a BSD-style license
|
|
5
|
-
* that can be found in the LICENSE file in the root of the source
|
|
6
|
-
* tree. An additional intellectual property rights grant can be found
|
|
7
|
-
* in the file PATENTS. All contributing project authors may
|
|
8
|
-
* be found in the AUTHORS file in the root of the source tree.
|
|
9
|
-
*/
|
|
10
|
-
|
|
11
|
-
#include "./psnr.h" // NOLINT
|
|
12
|
-
|
|
13
|
-
#ifdef _OPENMP
|
|
14
|
-
#include <omp.h>
|
|
15
|
-
#endif
|
|
16
|
-
#ifdef _MSC_VER
|
|
17
|
-
#include <intrin.h> // For __cpuid()
|
|
18
|
-
#endif
|
|
19
|
-
|
|
20
|
-
#ifdef __cplusplus
|
|
21
|
-
extern "C" {
|
|
22
|
-
#endif
|
|
23
|
-
|
|
24
|
-
typedef unsigned int uint32_t; // NOLINT
|
|
25
|
-
#ifdef _MSC_VER
|
|
26
|
-
typedef unsigned __int64 uint64_t;
|
|
27
|
-
#else // COMPILER_MSVC
|
|
28
|
-
#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
|
|
29
|
-
typedef unsigned long uint64_t; // NOLINT
|
|
30
|
-
#else // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
|
|
31
|
-
typedef unsigned long long uint64_t; // NOLINT
|
|
32
|
-
#endif // __LP64__
|
|
33
|
-
#endif // _MSC_VER
|
|
34
|
-
|
|
35
|
-
// libyuv provides this function when linking library for jpeg support.
|
|
36
|
-
#if !defined(HAVE_JPEG)
|
|
37
|
-
|
|
38
|
-
#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
|
|
39
|
-
!defined(__aarch64__)
|
|
40
|
-
#define HAS_SUMSQUAREERROR_NEON
|
|
41
|
-
static uint32_t SumSquareError_NEON(const uint8_t* src_a,
|
|
42
|
-
const uint8_t* src_b,
|
|
43
|
-
int count) {
|
|
44
|
-
volatile uint32_t sse;
|
|
45
|
-
asm volatile(
|
|
46
|
-
"vmov.u8 q7, #0 \n"
|
|
47
|
-
"vmov.u8 q9, #0 \n"
|
|
48
|
-
"vmov.u8 q8, #0 \n"
|
|
49
|
-
"vmov.u8 q10, #0 \n"
|
|
50
|
-
|
|
51
|
-
"1: \n"
|
|
52
|
-
"vld1.u8 {q0}, [%0]! \n"
|
|
53
|
-
"vld1.u8 {q1}, [%1]! \n"
|
|
54
|
-
"vsubl.u8 q2, d0, d2 \n"
|
|
55
|
-
"vsubl.u8 q3, d1, d3 \n"
|
|
56
|
-
"vmlal.s16 q7, d4, d4 \n"
|
|
57
|
-
"vmlal.s16 q8, d6, d6 \n"
|
|
58
|
-
"vmlal.s16 q8, d5, d5 \n"
|
|
59
|
-
"vmlal.s16 q10, d7, d7 \n"
|
|
60
|
-
"subs %2, %2, #16 \n"
|
|
61
|
-
"bhi 1b \n"
|
|
62
|
-
|
|
63
|
-
"vadd.u32 q7, q7, q8 \n"
|
|
64
|
-
"vadd.u32 q9, q9, q10 \n"
|
|
65
|
-
"vadd.u32 q10, q7, q9 \n"
|
|
66
|
-
"vpaddl.u32 q1, q10 \n"
|
|
67
|
-
"vadd.u64 d0, d2, d3 \n"
|
|
68
|
-
"vmov.32 %3, d0[0] \n"
|
|
69
|
-
: "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
|
|
70
|
-
:
|
|
71
|
-
: "memory", "cc", "q0", "q1", "q2", "q3", "q7", "q8", "q9", "q10");
|
|
72
|
-
return sse;
|
|
73
|
-
}
|
|
74
|
-
#elif !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
|
|
75
|
-
#define HAS_SUMSQUAREERROR_NEON
|
|
76
|
-
static uint32_t SumSquareError_NEON(const uint8_t* src_a,
|
|
77
|
-
const uint8_t* src_b,
|
|
78
|
-
int count) {
|
|
79
|
-
volatile uint32_t sse;
|
|
80
|
-
asm volatile(
|
|
81
|
-
"eor v16.16b, v16.16b, v16.16b \n"
|
|
82
|
-
"eor v18.16b, v18.16b, v18.16b \n"
|
|
83
|
-
"eor v17.16b, v17.16b, v17.16b \n"
|
|
84
|
-
"eor v19.16b, v19.16b, v19.16b \n"
|
|
85
|
-
|
|
86
|
-
"1: \n"
|
|
87
|
-
"ld1 {v0.16b}, [%0], #16 \n"
|
|
88
|
-
"ld1 {v1.16b}, [%1], #16 \n"
|
|
89
|
-
"subs %w2, %w2, #16 \n"
|
|
90
|
-
"usubl v2.8h, v0.8b, v1.8b \n"
|
|
91
|
-
"usubl2 v3.8h, v0.16b, v1.16b \n"
|
|
92
|
-
"smlal v16.4s, v2.4h, v2.4h \n"
|
|
93
|
-
"smlal v17.4s, v3.4h, v3.4h \n"
|
|
94
|
-
"smlal2 v18.4s, v2.8h, v2.8h \n"
|
|
95
|
-
"smlal2 v19.4s, v3.8h, v3.8h \n"
|
|
96
|
-
"b.gt 1b \n"
|
|
97
|
-
|
|
98
|
-
"add v16.4s, v16.4s, v17.4s \n"
|
|
99
|
-
"add v18.4s, v18.4s, v19.4s \n"
|
|
100
|
-
"add v19.4s, v16.4s, v18.4s \n"
|
|
101
|
-
"addv s0, v19.4s \n"
|
|
102
|
-
"fmov %w3, s0 \n"
|
|
103
|
-
: "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
|
|
104
|
-
:
|
|
105
|
-
: "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
|
|
106
|
-
return sse;
|
|
107
|
-
}
|
|
108
|
-
#elif !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
|
|
109
|
-
#define HAS_SUMSQUAREERROR_SSE2
|
|
110
|
-
__declspec(naked) static uint32_t SumSquareError_SSE2(const uint8_t* /*src_a*/,
|
|
111
|
-
const uint8_t* /*src_b*/,
|
|
112
|
-
int /*count*/) {
|
|
113
|
-
__asm {
|
|
114
|
-
mov eax, [esp + 4] // src_a
|
|
115
|
-
mov edx, [esp + 8] // src_b
|
|
116
|
-
mov ecx, [esp + 12] // count
|
|
117
|
-
pxor xmm0, xmm0
|
|
118
|
-
pxor xmm5, xmm5
|
|
119
|
-
sub edx, eax
|
|
120
|
-
|
|
121
|
-
wloop:
|
|
122
|
-
movdqu xmm1, [eax]
|
|
123
|
-
movdqu xmm2, [eax + edx]
|
|
124
|
-
lea eax, [eax + 16]
|
|
125
|
-
movdqu xmm3, xmm1
|
|
126
|
-
psubusb xmm1, xmm2
|
|
127
|
-
psubusb xmm2, xmm3
|
|
128
|
-
por xmm1, xmm2
|
|
129
|
-
movdqu xmm2, xmm1
|
|
130
|
-
punpcklbw xmm1, xmm5
|
|
131
|
-
punpckhbw xmm2, xmm5
|
|
132
|
-
pmaddwd xmm1, xmm1
|
|
133
|
-
pmaddwd xmm2, xmm2
|
|
134
|
-
paddd xmm0, xmm1
|
|
135
|
-
paddd xmm0, xmm2
|
|
136
|
-
sub ecx, 16
|
|
137
|
-
ja wloop
|
|
138
|
-
|
|
139
|
-
pshufd xmm1, xmm0, 0EEh
|
|
140
|
-
paddd xmm0, xmm1
|
|
141
|
-
pshufd xmm1, xmm0, 01h
|
|
142
|
-
paddd xmm0, xmm1
|
|
143
|
-
movd eax, xmm0
|
|
144
|
-
ret
|
|
145
|
-
}
|
|
146
|
-
}
|
|
147
|
-
#elif !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
|
|
148
|
-
#define HAS_SUMSQUAREERROR_SSE2
|
|
149
|
-
static uint32_t SumSquareError_SSE2(const uint8_t* src_a,
|
|
150
|
-
const uint8_t* src_b,
|
|
151
|
-
int count) {
|
|
152
|
-
uint32_t sse;
|
|
153
|
-
asm volatile( // NOLINT
|
|
154
|
-
"pxor %%xmm0,%%xmm0 \n"
|
|
155
|
-
"pxor %%xmm5,%%xmm5 \n"
|
|
156
|
-
"sub %0,%1 \n"
|
|
157
|
-
|
|
158
|
-
"1: \n"
|
|
159
|
-
"movdqu (%0),%%xmm1 \n"
|
|
160
|
-
"movdqu (%0,%1,1),%%xmm2 \n"
|
|
161
|
-
"lea 0x10(%0),%0 \n"
|
|
162
|
-
"movdqu %%xmm1,%%xmm3 \n"
|
|
163
|
-
"psubusb %%xmm2,%%xmm1 \n"
|
|
164
|
-
"psubusb %%xmm3,%%xmm2 \n"
|
|
165
|
-
"por %%xmm2,%%xmm1 \n"
|
|
166
|
-
"movdqu %%xmm1,%%xmm2 \n"
|
|
167
|
-
"punpcklbw %%xmm5,%%xmm1 \n"
|
|
168
|
-
"punpckhbw %%xmm5,%%xmm2 \n"
|
|
169
|
-
"pmaddwd %%xmm1,%%xmm1 \n"
|
|
170
|
-
"pmaddwd %%xmm2,%%xmm2 \n"
|
|
171
|
-
"paddd %%xmm1,%%xmm0 \n"
|
|
172
|
-
"paddd %%xmm2,%%xmm0 \n"
|
|
173
|
-
"sub $0x10,%2 \n"
|
|
174
|
-
"ja 1b \n"
|
|
175
|
-
|
|
176
|
-
"pshufd $0xee,%%xmm0,%%xmm1 \n"
|
|
177
|
-
"paddd %%xmm1,%%xmm0 \n"
|
|
178
|
-
"pshufd $0x1,%%xmm0,%%xmm1 \n"
|
|
179
|
-
"paddd %%xmm1,%%xmm0 \n"
|
|
180
|
-
"movd %%xmm0,%3 \n"
|
|
181
|
-
|
|
182
|
-
: "+r"(src_a), // %0
|
|
183
|
-
"+r"(src_b), // %1
|
|
184
|
-
"+r"(count), // %2
|
|
185
|
-
"=g"(sse) // %3
|
|
186
|
-
:
|
|
187
|
-
: "memory", "cc"
|
|
188
|
-
#if defined(__SSE2__)
|
|
189
|
-
,
|
|
190
|
-
"xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
|
|
191
|
-
#endif
|
|
192
|
-
); // NOLINT
|
|
193
|
-
return sse;
|
|
194
|
-
}
|
|
195
|
-
#endif // LIBYUV_DISABLE_X86 etc
|
|
196
|
-
|
|
197
|
-
#if defined(HAS_SUMSQUAREERROR_SSE2)
|
|
198
|
-
#if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
|
|
199
|
-
static __inline void __cpuid(int cpu_info[4], int info_type) {
|
|
200
|
-
asm volatile( // NOLINT
|
|
201
|
-
"mov %%ebx, %%edi \n"
|
|
202
|
-
"cpuid \n"
|
|
203
|
-
"xchg %%edi, %%ebx \n"
|
|
204
|
-
: "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]),
|
|
205
|
-
"=d"(cpu_info[3])
|
|
206
|
-
: "a"(info_type));
|
|
207
|
-
}
|
|
208
|
-
// For gcc/clang but not clangcl.
|
|
209
|
-
#elif !defined(_MSC_VER) && (defined(__i386__) || defined(__x86_64__))
|
|
210
|
-
static __inline void __cpuid(int cpu_info[4], int info_type) {
|
|
211
|
-
asm volatile( // NOLINT
|
|
212
|
-
"cpuid \n"
|
|
213
|
-
: "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]),
|
|
214
|
-
"=d"(cpu_info[3])
|
|
215
|
-
: "a"(info_type));
|
|
216
|
-
}
|
|
217
|
-
#endif
|
|
218
|
-
|
|
219
|
-
static int CpuHasSSE2() {
|
|
220
|
-
#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86)
|
|
221
|
-
int cpu_info[4];
|
|
222
|
-
__cpuid(cpu_info, 1);
|
|
223
|
-
if (cpu_info[3] & 0x04000000) {
|
|
224
|
-
return 1;
|
|
225
|
-
}
|
|
226
|
-
#endif
|
|
227
|
-
return 0;
|
|
228
|
-
}
|
|
229
|
-
#endif // HAS_SUMSQUAREERROR_SSE2
|
|
230
|
-
|
|
231
|
-
static uint32_t SumSquareError_C(const uint8_t* src_a,
|
|
232
|
-
const uint8_t* src_b,
|
|
233
|
-
int count) {
|
|
234
|
-
uint32_t sse = 0u;
|
|
235
|
-
for (int x = 0; x < count; ++x) {
|
|
236
|
-
int diff = src_a[x] - src_b[x];
|
|
237
|
-
sse += static_cast<uint32_t>(diff * diff);
|
|
238
|
-
}
|
|
239
|
-
return sse;
|
|
240
|
-
}
|
|
241
|
-
|
|
242
|
-
double ComputeSumSquareError(const uint8_t* src_a,
|
|
243
|
-
const uint8_t* src_b,
|
|
244
|
-
int count) {
|
|
245
|
-
uint32_t (*SumSquareError)(const uint8_t* src_a, const uint8_t* src_b,
|
|
246
|
-
int count) = SumSquareError_C;
|
|
247
|
-
#if defined(HAS_SUMSQUAREERROR_NEON)
|
|
248
|
-
SumSquareError = SumSquareError_NEON;
|
|
249
|
-
#endif
|
|
250
|
-
#if defined(HAS_SUMSQUAREERROR_SSE2)
|
|
251
|
-
if (CpuHasSSE2()) {
|
|
252
|
-
SumSquareError = SumSquareError_SSE2;
|
|
253
|
-
}
|
|
254
|
-
#endif
|
|
255
|
-
const int kBlockSize = 1 << 15;
|
|
256
|
-
uint64_t sse = 0;
|
|
257
|
-
#ifdef _OPENMP
|
|
258
|
-
#pragma omp parallel for reduction(+ : sse)
|
|
259
|
-
#endif
|
|
260
|
-
for (int i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
|
|
261
|
-
sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
|
|
262
|
-
}
|
|
263
|
-
src_a += count & ~(kBlockSize - 1);
|
|
264
|
-
src_b += count & ~(kBlockSize - 1);
|
|
265
|
-
int remainder = count & (kBlockSize - 1) & ~15;
|
|
266
|
-
if (remainder) {
|
|
267
|
-
sse += SumSquareError(src_a, src_b, remainder);
|
|
268
|
-
src_a += remainder;
|
|
269
|
-
src_b += remainder;
|
|
270
|
-
}
|
|
271
|
-
remainder = count & 15;
|
|
272
|
-
if (remainder) {
|
|
273
|
-
sse += SumSquareError_C(src_a, src_b, remainder);
|
|
274
|
-
}
|
|
275
|
-
return static_cast<double>(sse);
|
|
276
|
-
}
|
|
277
|
-
#endif
|
|
278
|
-
|
|
279
|
-
// PSNR formula: psnr = 10 * log10 (Peak Signal^2 * size / sse)
|
|
280
|
-
// Returns 128.0 (kMaxPSNR) if sse is 0 (perfect match).
|
|
281
|
-
double ComputePSNR(double sse, double size) {
|
|
282
|
-
const double kMINSSE = 255.0 * 255.0 * size / pow(10.0, kMaxPSNR / 10.0);
|
|
283
|
-
if (sse <= kMINSSE) {
|
|
284
|
-
sse = kMINSSE; // Produces max PSNR of 128
|
|
285
|
-
}
|
|
286
|
-
return 10.0 * log10(255.0 * 255.0 * size / sse);
|
|
287
|
-
}
|
|
288
|
-
|
|
289
|
-
#ifdef __cplusplus
|
|
290
|
-
} // extern "C"
|
|
291
|
-
#endif
|
|
1
|
+
/*
|
|
2
|
+
* Copyright 2013 The LibYuv Project Authors. All rights reserved.
|
|
3
|
+
*
|
|
4
|
+
* Use of this source code is governed by a BSD-style license
|
|
5
|
+
* that can be found in the LICENSE file in the root of the source
|
|
6
|
+
* tree. An additional intellectual property rights grant can be found
|
|
7
|
+
* in the file PATENTS. All contributing project authors may
|
|
8
|
+
* be found in the AUTHORS file in the root of the source tree.
|
|
9
|
+
*/
|
|
10
|
+
|
|
11
|
+
#include "./psnr.h" // NOLINT
|
|
12
|
+
|
|
13
|
+
#ifdef _OPENMP
|
|
14
|
+
#include <omp.h>
|
|
15
|
+
#endif
|
|
16
|
+
#ifdef _MSC_VER
|
|
17
|
+
#include <intrin.h> // For __cpuid()
|
|
18
|
+
#endif
|
|
19
|
+
|
|
20
|
+
#ifdef __cplusplus
|
|
21
|
+
extern "C" {
|
|
22
|
+
#endif
|
|
23
|
+
|
|
24
|
+
typedef unsigned int uint32_t; // NOLINT
|
|
25
|
+
#ifdef _MSC_VER
|
|
26
|
+
typedef unsigned __int64 uint64_t;
|
|
27
|
+
#else // COMPILER_MSVC
|
|
28
|
+
#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
|
|
29
|
+
typedef unsigned long uint64_t; // NOLINT
|
|
30
|
+
#else // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
|
|
31
|
+
typedef unsigned long long uint64_t; // NOLINT
|
|
32
|
+
#endif // __LP64__
|
|
33
|
+
#endif // _MSC_VER
|
|
34
|
+
|
|
35
|
+
// libyuv provides this function when linking library for jpeg support.
|
|
36
|
+
#if !defined(HAVE_JPEG)
|
|
37
|
+
|
|
38
|
+
#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
|
|
39
|
+
!defined(__aarch64__)
|
|
40
|
+
#define HAS_SUMSQUAREERROR_NEON
|
|
41
|
+
static uint32_t SumSquareError_NEON(const uint8_t* src_a,
|
|
42
|
+
const uint8_t* src_b,
|
|
43
|
+
int count) {
|
|
44
|
+
volatile uint32_t sse;
|
|
45
|
+
asm volatile(
|
|
46
|
+
"vmov.u8 q7, #0 \n"
|
|
47
|
+
"vmov.u8 q9, #0 \n"
|
|
48
|
+
"vmov.u8 q8, #0 \n"
|
|
49
|
+
"vmov.u8 q10, #0 \n"
|
|
50
|
+
|
|
51
|
+
"1: \n"
|
|
52
|
+
"vld1.u8 {q0}, [%0]! \n"
|
|
53
|
+
"vld1.u8 {q1}, [%1]! \n"
|
|
54
|
+
"vsubl.u8 q2, d0, d2 \n"
|
|
55
|
+
"vsubl.u8 q3, d1, d3 \n"
|
|
56
|
+
"vmlal.s16 q7, d4, d4 \n"
|
|
57
|
+
"vmlal.s16 q8, d6, d6 \n"
|
|
58
|
+
"vmlal.s16 q8, d5, d5 \n"
|
|
59
|
+
"vmlal.s16 q10, d7, d7 \n"
|
|
60
|
+
"subs %2, %2, #16 \n"
|
|
61
|
+
"bhi 1b \n"
|
|
62
|
+
|
|
63
|
+
"vadd.u32 q7, q7, q8 \n"
|
|
64
|
+
"vadd.u32 q9, q9, q10 \n"
|
|
65
|
+
"vadd.u32 q10, q7, q9 \n"
|
|
66
|
+
"vpaddl.u32 q1, q10 \n"
|
|
67
|
+
"vadd.u64 d0, d2, d3 \n"
|
|
68
|
+
"vmov.32 %3, d0[0] \n"
|
|
69
|
+
: "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
|
|
70
|
+
:
|
|
71
|
+
: "memory", "cc", "q0", "q1", "q2", "q3", "q7", "q8", "q9", "q10");
|
|
72
|
+
return sse;
|
|
73
|
+
}
|
|
74
|
+
#elif !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
|
|
75
|
+
#define HAS_SUMSQUAREERROR_NEON
|
|
76
|
+
static uint32_t SumSquareError_NEON(const uint8_t* src_a,
|
|
77
|
+
const uint8_t* src_b,
|
|
78
|
+
int count) {
|
|
79
|
+
volatile uint32_t sse;
|
|
80
|
+
asm volatile(
|
|
81
|
+
"eor v16.16b, v16.16b, v16.16b \n"
|
|
82
|
+
"eor v18.16b, v18.16b, v18.16b \n"
|
|
83
|
+
"eor v17.16b, v17.16b, v17.16b \n"
|
|
84
|
+
"eor v19.16b, v19.16b, v19.16b \n"
|
|
85
|
+
|
|
86
|
+
"1: \n"
|
|
87
|
+
"ld1 {v0.16b}, [%0], #16 \n"
|
|
88
|
+
"ld1 {v1.16b}, [%1], #16 \n"
|
|
89
|
+
"subs %w2, %w2, #16 \n"
|
|
90
|
+
"usubl v2.8h, v0.8b, v1.8b \n"
|
|
91
|
+
"usubl2 v3.8h, v0.16b, v1.16b \n"
|
|
92
|
+
"smlal v16.4s, v2.4h, v2.4h \n"
|
|
93
|
+
"smlal v17.4s, v3.4h, v3.4h \n"
|
|
94
|
+
"smlal2 v18.4s, v2.8h, v2.8h \n"
|
|
95
|
+
"smlal2 v19.4s, v3.8h, v3.8h \n"
|
|
96
|
+
"b.gt 1b \n"
|
|
97
|
+
|
|
98
|
+
"add v16.4s, v16.4s, v17.4s \n"
|
|
99
|
+
"add v18.4s, v18.4s, v19.4s \n"
|
|
100
|
+
"add v19.4s, v16.4s, v18.4s \n"
|
|
101
|
+
"addv s0, v19.4s \n"
|
|
102
|
+
"fmov %w3, s0 \n"
|
|
103
|
+
: "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
|
|
104
|
+
:
|
|
105
|
+
: "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
|
|
106
|
+
return sse;
|
|
107
|
+
}
|
|
108
|
+
#elif !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
|
|
109
|
+
#define HAS_SUMSQUAREERROR_SSE2
|
|
110
|
+
__declspec(naked) static uint32_t SumSquareError_SSE2(const uint8_t* /*src_a*/,
|
|
111
|
+
const uint8_t* /*src_b*/,
|
|
112
|
+
int /*count*/) {
|
|
113
|
+
__asm {
|
|
114
|
+
mov eax, [esp + 4] // src_a
|
|
115
|
+
mov edx, [esp + 8] // src_b
|
|
116
|
+
mov ecx, [esp + 12] // count
|
|
117
|
+
pxor xmm0, xmm0
|
|
118
|
+
pxor xmm5, xmm5
|
|
119
|
+
sub edx, eax
|
|
120
|
+
|
|
121
|
+
wloop:
|
|
122
|
+
movdqu xmm1, [eax]
|
|
123
|
+
movdqu xmm2, [eax + edx]
|
|
124
|
+
lea eax, [eax + 16]
|
|
125
|
+
movdqu xmm3, xmm1
|
|
126
|
+
psubusb xmm1, xmm2
|
|
127
|
+
psubusb xmm2, xmm3
|
|
128
|
+
por xmm1, xmm2
|
|
129
|
+
movdqu xmm2, xmm1
|
|
130
|
+
punpcklbw xmm1, xmm5
|
|
131
|
+
punpckhbw xmm2, xmm5
|
|
132
|
+
pmaddwd xmm1, xmm1
|
|
133
|
+
pmaddwd xmm2, xmm2
|
|
134
|
+
paddd xmm0, xmm1
|
|
135
|
+
paddd xmm0, xmm2
|
|
136
|
+
sub ecx, 16
|
|
137
|
+
ja wloop
|
|
138
|
+
|
|
139
|
+
pshufd xmm1, xmm0, 0EEh
|
|
140
|
+
paddd xmm0, xmm1
|
|
141
|
+
pshufd xmm1, xmm0, 01h
|
|
142
|
+
paddd xmm0, xmm1
|
|
143
|
+
movd eax, xmm0
|
|
144
|
+
ret
|
|
145
|
+
}
|
|
146
|
+
}
|
|
147
|
+
#elif !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
|
|
148
|
+
#define HAS_SUMSQUAREERROR_SSE2
|
|
149
|
+
static uint32_t SumSquareError_SSE2(const uint8_t* src_a,
|
|
150
|
+
const uint8_t* src_b,
|
|
151
|
+
int count) {
|
|
152
|
+
uint32_t sse;
|
|
153
|
+
asm volatile( // NOLINT
|
|
154
|
+
"pxor %%xmm0,%%xmm0 \n"
|
|
155
|
+
"pxor %%xmm5,%%xmm5 \n"
|
|
156
|
+
"sub %0,%1 \n"
|
|
157
|
+
|
|
158
|
+
"1: \n"
|
|
159
|
+
"movdqu (%0),%%xmm1 \n"
|
|
160
|
+
"movdqu (%0,%1,1),%%xmm2 \n"
|
|
161
|
+
"lea 0x10(%0),%0 \n"
|
|
162
|
+
"movdqu %%xmm1,%%xmm3 \n"
|
|
163
|
+
"psubusb %%xmm2,%%xmm1 \n"
|
|
164
|
+
"psubusb %%xmm3,%%xmm2 \n"
|
|
165
|
+
"por %%xmm2,%%xmm1 \n"
|
|
166
|
+
"movdqu %%xmm1,%%xmm2 \n"
|
|
167
|
+
"punpcklbw %%xmm5,%%xmm1 \n"
|
|
168
|
+
"punpckhbw %%xmm5,%%xmm2 \n"
|
|
169
|
+
"pmaddwd %%xmm1,%%xmm1 \n"
|
|
170
|
+
"pmaddwd %%xmm2,%%xmm2 \n"
|
|
171
|
+
"paddd %%xmm1,%%xmm0 \n"
|
|
172
|
+
"paddd %%xmm2,%%xmm0 \n"
|
|
173
|
+
"sub $0x10,%2 \n"
|
|
174
|
+
"ja 1b \n"
|
|
175
|
+
|
|
176
|
+
"pshufd $0xee,%%xmm0,%%xmm1 \n"
|
|
177
|
+
"paddd %%xmm1,%%xmm0 \n"
|
|
178
|
+
"pshufd $0x1,%%xmm0,%%xmm1 \n"
|
|
179
|
+
"paddd %%xmm1,%%xmm0 \n"
|
|
180
|
+
"movd %%xmm0,%3 \n"
|
|
181
|
+
|
|
182
|
+
: "+r"(src_a), // %0
|
|
183
|
+
"+r"(src_b), // %1
|
|
184
|
+
"+r"(count), // %2
|
|
185
|
+
"=g"(sse) // %3
|
|
186
|
+
:
|
|
187
|
+
: "memory", "cc"
|
|
188
|
+
#if defined(__SSE2__)
|
|
189
|
+
,
|
|
190
|
+
"xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
|
|
191
|
+
#endif
|
|
192
|
+
); // NOLINT
|
|
193
|
+
return sse;
|
|
194
|
+
}
|
|
195
|
+
#endif // LIBYUV_DISABLE_X86 etc
|
|
196
|
+
|
|
197
|
+
#if defined(HAS_SUMSQUAREERROR_SSE2)
|
|
198
|
+
#if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
|
|
199
|
+
static __inline void __cpuid(int cpu_info[4], int info_type) {
|
|
200
|
+
asm volatile( // NOLINT
|
|
201
|
+
"mov %%ebx, %%edi \n"
|
|
202
|
+
"cpuid \n"
|
|
203
|
+
"xchg %%edi, %%ebx \n"
|
|
204
|
+
: "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]),
|
|
205
|
+
"=d"(cpu_info[3])
|
|
206
|
+
: "a"(info_type));
|
|
207
|
+
}
|
|
208
|
+
// For gcc/clang but not clangcl.
|
|
209
|
+
#elif !defined(_MSC_VER) && (defined(__i386__) || defined(__x86_64__))
|
|
210
|
+
static __inline void __cpuid(int cpu_info[4], int info_type) {
|
|
211
|
+
asm volatile( // NOLINT
|
|
212
|
+
"cpuid \n"
|
|
213
|
+
: "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]),
|
|
214
|
+
"=d"(cpu_info[3])
|
|
215
|
+
: "a"(info_type));
|
|
216
|
+
}
|
|
217
|
+
#endif
|
|
218
|
+
|
|
219
|
+
static int CpuHasSSE2() {
|
|
220
|
+
#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86)
|
|
221
|
+
int cpu_info[4];
|
|
222
|
+
__cpuid(cpu_info, 1);
|
|
223
|
+
if (cpu_info[3] & 0x04000000) {
|
|
224
|
+
return 1;
|
|
225
|
+
}
|
|
226
|
+
#endif
|
|
227
|
+
return 0;
|
|
228
|
+
}
|
|
229
|
+
#endif // HAS_SUMSQUAREERROR_SSE2
|
|
230
|
+
|
|
231
|
+
static uint32_t SumSquareError_C(const uint8_t* src_a,
|
|
232
|
+
const uint8_t* src_b,
|
|
233
|
+
int count) {
|
|
234
|
+
uint32_t sse = 0u;
|
|
235
|
+
for (int x = 0; x < count; ++x) {
|
|
236
|
+
int diff = src_a[x] - src_b[x];
|
|
237
|
+
sse += static_cast<uint32_t>(diff * diff);
|
|
238
|
+
}
|
|
239
|
+
return sse;
|
|
240
|
+
}
|
|
241
|
+
|
|
242
|
+
double ComputeSumSquareError(const uint8_t* src_a,
|
|
243
|
+
const uint8_t* src_b,
|
|
244
|
+
int count) {
|
|
245
|
+
uint32_t (*SumSquareError)(const uint8_t* src_a, const uint8_t* src_b,
|
|
246
|
+
int count) = SumSquareError_C;
|
|
247
|
+
#if defined(HAS_SUMSQUAREERROR_NEON)
|
|
248
|
+
SumSquareError = SumSquareError_NEON;
|
|
249
|
+
#endif
|
|
250
|
+
#if defined(HAS_SUMSQUAREERROR_SSE2)
|
|
251
|
+
if (CpuHasSSE2()) {
|
|
252
|
+
SumSquareError = SumSquareError_SSE2;
|
|
253
|
+
}
|
|
254
|
+
#endif
|
|
255
|
+
const int kBlockSize = 1 << 15;
|
|
256
|
+
uint64_t sse = 0;
|
|
257
|
+
#ifdef _OPENMP
|
|
258
|
+
#pragma omp parallel for reduction(+ : sse)
|
|
259
|
+
#endif
|
|
260
|
+
for (int i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
|
|
261
|
+
sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
|
|
262
|
+
}
|
|
263
|
+
src_a += count & ~(kBlockSize - 1);
|
|
264
|
+
src_b += count & ~(kBlockSize - 1);
|
|
265
|
+
int remainder = count & (kBlockSize - 1) & ~15;
|
|
266
|
+
if (remainder) {
|
|
267
|
+
sse += SumSquareError(src_a, src_b, remainder);
|
|
268
|
+
src_a += remainder;
|
|
269
|
+
src_b += remainder;
|
|
270
|
+
}
|
|
271
|
+
remainder = count & 15;
|
|
272
|
+
if (remainder) {
|
|
273
|
+
sse += SumSquareError_C(src_a, src_b, remainder);
|
|
274
|
+
}
|
|
275
|
+
return static_cast<double>(sse);
|
|
276
|
+
}
|
|
277
|
+
#endif
|
|
278
|
+
|
|
279
|
+
// PSNR formula: psnr = 10 * log10 (Peak Signal^2 * size / sse)
|
|
280
|
+
// Returns 128.0 (kMaxPSNR) if sse is 0 (perfect match).
|
|
281
|
+
double ComputePSNR(double sse, double size) {
|
|
282
|
+
const double kMINSSE = 255.0 * 255.0 * size / pow(10.0, kMaxPSNR / 10.0);
|
|
283
|
+
if (sse <= kMINSSE) {
|
|
284
|
+
sse = kMINSSE; // Produces max PSNR of 128
|
|
285
|
+
}
|
|
286
|
+
return 10.0 * log10(255.0 * 255.0 * size / sse);
|
|
287
|
+
}
|
|
288
|
+
|
|
289
|
+
#ifdef __cplusplus
|
|
290
|
+
} // extern "C"
|
|
291
|
+
#endif
|