diff options
| -rw-r--r-- | source/planar_functions.cc | 4 | ||||
| -rw-r--r-- | unit_test/color_test.cc | 38 | ||||
| -rw-r--r-- | unit_test/compare_test.cc | 81 | ||||
| -rw-r--r-- | unit_test/convert_test.cc | 504 | ||||
| -rw-r--r-- | unit_test/cpu_test.cc | 8 | ||||
| -rw-r--r-- | unit_test/math_test.cc | 1 | ||||
| -rw-r--r-- | unit_test/planar_test.cc | 509 | ||||
| -rw-r--r-- | unit_test/rotate_argb_test.cc | 13 | ||||
| -rw-r--r-- | unit_test/rotate_test.cc | 25 | ||||
| -rw-r--r-- | unit_test/scale_argb_test.cc | 16 | ||||
| -rw-r--r-- | unit_test/unit_test.h | 10 |
11 files changed, 504 insertions, 705 deletions
diff --git a/source/planar_functions.cc b/source/planar_functions.cc index b1b1f2c8..2e2351aa 100644 --- a/source/planar_functions.cc +++ b/source/planar_functions.cc @@ -1744,7 +1744,9 @@ int ARGBBlur(const uint8* src_argb, int src_stride_argb, #if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2) if (TestCpuFlag(kCpuHasSSE2)) { ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2; - CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2; + if (IS_ALIGNED(dst_cumsum, 16)) { + CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2; + } } #endif // Compute enough CumulativeSum for first row to be blurred. After this diff --git a/unit_test/color_test.cc b/unit_test/color_test.cc index 555413f7..0852a91d 100644 --- a/unit_test/color_test.cc +++ b/unit_test/color_test.cc @@ -10,12 +10,12 @@ #include <stdlib.h> +#include "libyuv/basic_types.h" #include "libyuv/convert.h" #include "libyuv/convert_argb.h" #include "libyuv/convert_from.h" #include "libyuv/convert_from_argb.h" #include "libyuv/cpu_id.h" -#include "libyuv/row.h" // For Sobel #include "../unit_test/unit_test.h" namespace libyuv { @@ -41,15 +41,15 @@ namespace libyuv { const int kPixels = benchmark_width_ * benchmark_height_; \ const int kHalfPixels = ((benchmark_width_ + 1) / 2) * \ ((benchmark_height_ + HS1) / HS); \ - align_buffer_64(orig_y, kPixels); \ - align_buffer_64(orig_u, kHalfPixels); \ - align_buffer_64(orig_v, kHalfPixels); \ - align_buffer_64(orig_pixels, kPixels * 4); \ - align_buffer_64(temp_y, kPixels); \ - align_buffer_64(temp_u, kHalfPixels); \ - align_buffer_64(temp_v, kHalfPixels); \ - align_buffer_64(dst_pixels_opt, kPixels * 4); \ - align_buffer_64(dst_pixels_c, kPixels * 4); \ + align_buffer_page_end(orig_y, kPixels); \ + align_buffer_page_end(orig_u, kHalfPixels); \ + align_buffer_page_end(orig_v, kHalfPixels); \ + align_buffer_page_end(orig_pixels, kPixels * 4); \ + align_buffer_page_end(temp_y, kPixels); \ + align_buffer_page_end(temp_u, kHalfPixels); \ + align_buffer_page_end(temp_v, kHalfPixels); \ + align_buffer_page_end(dst_pixels_opt, kPixels * 4); \ + align_buffer_page_end(dst_pixels_c, kPixels * 4); \ \ MemRandomize(orig_pixels, kPixels * 4); \ MemRandomize(orig_y, kPixels); \ @@ -132,15 +132,15 @@ namespace libyuv { static_cast<int>(dst_pixels_opt[i]), DIFF); \ } \ \ - free_aligned_buffer_64(orig_pixels); \ - free_aligned_buffer_64(orig_y); \ - free_aligned_buffer_64(orig_u); \ - free_aligned_buffer_64(orig_v); \ - free_aligned_buffer_64(temp_y); \ - free_aligned_buffer_64(temp_u); \ - free_aligned_buffer_64(temp_v); \ - free_aligned_buffer_64(dst_pixels_opt); \ - free_aligned_buffer_64(dst_pixels_c); \ + free_aligned_buffer_page_end(orig_pixels); \ + free_aligned_buffer_page_end(orig_y); \ + free_aligned_buffer_page_end(orig_u); \ + free_aligned_buffer_page_end(orig_v); \ + free_aligned_buffer_page_end(temp_y); \ + free_aligned_buffer_page_end(temp_u); \ + free_aligned_buffer_page_end(temp_v); \ + free_aligned_buffer_page_end(dst_pixels_opt); \ + free_aligned_buffer_page_end(dst_pixels_c); \ } \ TESTCS(TestI420, I420ToARGB, ARGBToI420, 1, 2, benchmark_width_, ERROR_FULL) diff --git a/unit_test/compare_test.cc b/unit_test/compare_test.cc index 832f7312..a8ce671d 100644 --- a/unit_test/compare_test.cc +++ b/unit_test/compare_test.cc @@ -16,7 +16,6 @@ #include "libyuv/basic_types.h" #include "libyuv/compare.h" #include "libyuv/cpu_id.h" -#include "libyuv/row.h" #include "libyuv/video_common.h" namespace libyuv { @@ -34,8 +33,8 @@ static uint32 ReferenceHashDjb2(const uint8* src, uint64 count, uint32 seed) { TEST_F(LibYUVBaseTest, Djb2_Test) { const int kMaxTest = benchmark_width_ * benchmark_height_; - align_buffer_64(src_a, kMaxTest); - align_buffer_64(src_b, kMaxTest); + align_buffer_page_end(src_a, kMaxTest); + align_buffer_page_end(src_b, kMaxTest); const char* fox = "The quick brown fox jumps over the lazy dog" " and feels as if he were in the seventh heaven of typography" @@ -112,13 +111,13 @@ TEST_F(LibYUVBaseTest, Djb2_Test) { h2 = HashDjb2(src_a, kMaxTest / 2, 0); EXPECT_EQ(h1, h2); - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } TEST_F(LibYUVBaseTest, BenchmarkDjb2_Opt) { const int kMaxTest = benchmark_width_ * benchmark_height_; - align_buffer_64(src_a, kMaxTest); + align_buffer_page_end(src_a, kMaxTest); for (int i = 0; i < kMaxTest; ++i) { src_a[i] = i; @@ -129,12 +128,12 @@ TEST_F(LibYUVBaseTest, BenchmarkDjb2_Opt) { h1 = HashDjb2(src_a, kMaxTest, 5381); } EXPECT_EQ(h1, h2); - free_aligned_buffer_64(src_a); + free_aligned_buffer_page_end(src_a); } TEST_F(LibYUVBaseTest, BenchmarkDjb2_Unaligned) { const int kMaxTest = benchmark_width_ * benchmark_height_; - align_buffer_64(src_a, kMaxTest + 1); + align_buffer_page_end(src_a, kMaxTest + 1); for (int i = 0; i < kMaxTest; ++i) { src_a[i + 1] = i; } @@ -144,13 +143,13 @@ TEST_F(LibYUVBaseTest, BenchmarkDjb2_Unaligned) { h1 = HashDjb2(src_a + 1, kMaxTest, 5381); } EXPECT_EQ(h1, h2); - free_aligned_buffer_64(src_a); + free_aligned_buffer_page_end(src_a); } TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Opt) { uint32 fourcc; const int kMaxTest = benchmark_width_ * benchmark_height_ * 4; - align_buffer_64(src_a, kMaxTest); + align_buffer_page_end(src_a, kMaxTest); for (int i = 0; i < kMaxTest; ++i) { src_a[i] = 255; } @@ -172,13 +171,13 @@ TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Opt) { } EXPECT_EQ(0, fourcc); - free_aligned_buffer_64(src_a); + free_aligned_buffer_page_end(src_a); } TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Unaligned) { uint32 fourcc; const int kMaxTest = benchmark_width_ * benchmark_height_ * 4 + 1; - align_buffer_64(src_a, kMaxTest); + align_buffer_page_end(src_a, kMaxTest); for (int i = 1; i < kMaxTest; ++i) { src_a[i] = 255; } @@ -200,12 +199,12 @@ TEST_F(LibYUVBaseTest, BenchmarkARGBDetect_Unaligned) { } EXPECT_EQ(0, fourcc); - free_aligned_buffer_64(src_a); + free_aligned_buffer_page_end(src_a); } TEST_F(LibYUVBaseTest, BenchmarkSumSquareError_Opt) { const int kMaxWidth = 4096 * 3; - align_buffer_64(src_a, kMaxWidth); - align_buffer_64(src_b, kMaxWidth); + align_buffer_page_end(src_a, kMaxWidth); + align_buffer_page_end(src_b, kMaxWidth); memset(src_a, 0, kMaxWidth); memset(src_b, 0, kMaxWidth); @@ -229,14 +228,14 @@ TEST_F(LibYUVBaseTest, BenchmarkSumSquareError_Opt) { EXPECT_EQ(0, h1); - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } TEST_F(LibYUVBaseTest, SumSquareError) { const int kMaxWidth = 4096 * 3; - align_buffer_64(src_a, kMaxWidth); - align_buffer_64(src_b, kMaxWidth); + align_buffer_page_end(src_a, kMaxWidth); + align_buffer_page_end(src_b, kMaxWidth); memset(src_a, 0, kMaxWidth); memset(src_b, 0, kMaxWidth); @@ -269,13 +268,13 @@ TEST_F(LibYUVBaseTest, SumSquareError) { EXPECT_EQ(c_err, opt_err); - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } TEST_F(LibYUVBaseTest, BenchmarkPsnr_Opt) { - align_buffer_64(src_a, benchmark_width_ * benchmark_height_); - align_buffer_64(src_b, benchmark_width_ * benchmark_height_); + align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_); + align_buffer_page_end(src_b, benchmark_width_ * benchmark_height_); for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) { src_a[i] = i; src_b[i] = i; @@ -294,13 +293,13 @@ TEST_F(LibYUVBaseTest, BenchmarkPsnr_Opt) { EXPECT_EQ(0, 0); - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } TEST_F(LibYUVBaseTest, BenchmarkPsnr_Unaligned) { - align_buffer_64(src_a, benchmark_width_ * benchmark_height_ + 1); - align_buffer_64(src_b, benchmark_width_ * benchmark_height_); + align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_ + 1); + align_buffer_page_end(src_b, benchmark_width_ * benchmark_height_); for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) { src_a[i + 1] = i; src_b[i] = i; @@ -319,8 +318,8 @@ TEST_F(LibYUVBaseTest, BenchmarkPsnr_Unaligned) { EXPECT_EQ(0, 0); - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } TEST_F(LibYUVBaseTest, Psnr) { @@ -329,8 +328,8 @@ TEST_F(LibYUVBaseTest, Psnr) { const int b = 128; const int kSrcPlaneSize = (kSrcWidth + b * 2) * (kSrcHeight + b * 2); const int kSrcStride = 2 * b + kSrcWidth; - align_buffer_64(src_a, kSrcPlaneSize); - align_buffer_64(src_b, kSrcPlaneSize); + align_buffer_page_end(src_a, kSrcPlaneSize); + align_buffer_page_end(src_b, kSrcPlaneSize); memset(src_a, 0, kSrcPlaneSize); memset(src_b, 0, kSrcPlaneSize); @@ -396,13 +395,13 @@ TEST_F(LibYUVBaseTest, Psnr) { EXPECT_EQ(opt_err, c_err); - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } TEST_F(LibYUVBaseTest, DISABLED_BenchmarkSsim_Opt) { - align_buffer_64(src_a, benchmark_width_ * benchmark_height_); - align_buffer_64(src_b, benchmark_width_ * benchmark_height_); + align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_); + align_buffer_page_end(src_b, benchmark_width_ * benchmark_height_); for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) { src_a[i] = i; src_b[i] = i; @@ -421,8 +420,8 @@ TEST_F(LibYUVBaseTest, DISABLED_BenchmarkSsim_Opt) { EXPECT_EQ(0, 0); // Pass if we get this far. - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } TEST_F(LibYUVBaseTest, Ssim) { @@ -431,8 +430,8 @@ TEST_F(LibYUVBaseTest, Ssim) { const int b = 128; const int kSrcPlaneSize = (kSrcWidth + b * 2) * (kSrcHeight + b * 2); const int kSrcStride = 2 * b + kSrcWidth; - align_buffer_64(src_a, kSrcPlaneSize); - align_buffer_64(src_b, kSrcPlaneSize); + align_buffer_page_end(src_a, kSrcPlaneSize); + align_buffer_page_end(src_b, kSrcPlaneSize); memset(src_a, 0, kSrcPlaneSize); memset(src_b, 0, kSrcPlaneSize); @@ -507,8 +506,8 @@ TEST_F(LibYUVBaseTest, Ssim) { EXPECT_EQ(opt_err, c_err); } - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(src_b); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(src_b); } } // namespace libyuv diff --git a/unit_test/convert_test.cc b/unit_test/convert_test.cc index c4d264a4..56a2bfd8 100644 --- a/unit_test/convert_test.cc +++ b/unit_test/convert_test.cc @@ -11,6 +11,7 @@ #include <stdlib.h> #include <time.h> +#include "libyuv/basic_types.h" #include "libyuv/compare.h" #include "libyuv/convert.h" #include "libyuv/convert_argb.h" @@ -22,7 +23,6 @@ #endif #include "libyuv/planar_functions.h" #include "libyuv/rotate.h" -#include "libyuv/row.h" #include "libyuv/video_common.h" #include "../unit_test/unit_test.h" @@ -35,27 +35,27 @@ namespace libyuv { TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) { \ const int kWidth = ((W1280) > 0) ? (W1280) : 1; \ const int kHeight = benchmark_height_; \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_u, \ - SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ - align_buffer_64(src_v, \ - SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ - align_buffer_64(dst_y_c, kWidth * kHeight); \ - align_buffer_64(dst_u_c, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_v_c, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_y_opt, kWidth * kHeight); \ - align_buffer_64(dst_u_opt, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_v_opt, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_u, \ + SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ + align_buffer_page_end(src_v, \ + SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ + align_buffer_page_end(dst_y_c, kWidth * kHeight); \ + align_buffer_page_end(dst_u_c, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_v_c, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_y_opt, kWidth * kHeight); \ + align_buffer_page_end(dst_u_opt, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_v_opt, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ for (int i = 0; i < kHeight; ++i) \ for (int j = 0; j < kWidth; ++j) \ src_y[i * kWidth + j + OFF] = (fastrand() & 0xff); \ @@ -137,15 +137,15 @@ TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) { \ } \ } \ EXPECT_LE(max_diff, 3); \ - free_aligned_buffer_64(dst_y_c); \ - free_aligned_buffer_64(dst_u_c); \ - free_aligned_buffer_64(dst_v_c); \ - free_aligned_buffer_64(dst_y_opt); \ - free_aligned_buffer_64(dst_u_opt); \ - free_aligned_buffer_64(dst_v_opt); \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_u); \ - free_aligned_buffer_64(src_v); \ + free_aligned_buffer_page_end(dst_y_c); \ + free_aligned_buffer_page_end(dst_u_c); \ + free_aligned_buffer_page_end(dst_v_c); \ + free_aligned_buffer_page_end(dst_y_opt); \ + free_aligned_buffer_page_end(dst_u_opt); \ + free_aligned_buffer_page_end(dst_v_opt); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_u); \ + free_aligned_buffer_page_end(src_v); \ } #define TESTPLANARTOP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \ @@ -179,19 +179,19 @@ TESTPLANARTOP(I444, 1, 1, I444, 1, 1) TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) { \ const int kWidth = ((W1280) > 0) ? (W1280) : 1; \ const int kHeight = benchmark_height_; \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_u, \ - SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ - align_buffer_64(src_v, \ - SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ - align_buffer_64(dst_y_c, kWidth * kHeight); \ - align_buffer_64(dst_uv_c, SUBSAMPLE(kWidth * 2, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_y_opt, kWidth * kHeight); \ - align_buffer_64(dst_uv_opt, SUBSAMPLE(kWidth * 2, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_u, \ + SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ + align_buffer_page_end(src_v, \ + SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ + align_buffer_page_end(dst_y_c, kWidth * kHeight); \ + align_buffer_page_end(dst_uv_c, SUBSAMPLE(kWidth * 2, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_y_opt, kWidth * kHeight); \ + align_buffer_page_end(dst_uv_opt, SUBSAMPLE(kWidth * 2, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ for (int i = 0; i < kHeight; ++i) \ for (int j = 0; j < kWidth; ++j) \ src_y[i * kWidth + j + OFF] = (fastrand() & 0xff); \ @@ -255,13 +255,13 @@ TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) { \ } \ } \ EXPECT_LE(max_diff, 1); \ - free_aligned_buffer_64(dst_y_c); \ - free_aligned_buffer_64(dst_uv_c); \ - free_aligned_buffer_64(dst_y_opt); \ - free_aligned_buffer_64(dst_uv_opt); \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_u); \ - free_aligned_buffer_64(src_v); \ + free_aligned_buffer_page_end(dst_y_c); \ + free_aligned_buffer_page_end(dst_uv_c); \ + free_aligned_buffer_page_end(dst_y_opt); \ + free_aligned_buffer_page_end(dst_uv_opt); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_u); \ + free_aligned_buffer_page_end(src_v); \ } #define TESTPLANARTOBP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \ @@ -287,23 +287,23 @@ TESTPLANARTOBP(I420, 2, 2, NV21, 2, 2) TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) { \ const int kWidth = ((W1280) > 0) ? (W1280) : 1; \ const int kHeight = benchmark_height_; \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_uv, 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ - align_buffer_64(dst_y_c, kWidth * kHeight); \ - align_buffer_64(dst_u_c, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_v_c, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_y_opt, kWidth * kHeight); \ - align_buffer_64(dst_u_opt, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_v_opt, \ - SUBSAMPLE(kWidth, SUBSAMP_X) * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_uv, 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) + OFF); \ + align_buffer_page_end(dst_y_c, kWidth * kHeight); \ + align_buffer_page_end(dst_u_c, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_v_c, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_y_opt, kWidth * kHeight); \ + align_buffer_page_end(dst_u_opt, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_v_opt, \ + SUBSAMPLE(kWidth, SUBSAMP_X) * \ + SUBSAMPLE(kHeight, SUBSAMP_Y)); \ for (int i = 0; i < kHeight; ++i) \ for (int j = 0; j < kWidth; ++j) \ src_y[i * kWidth + j + OFF] = (fastrand() & 0xff); \ @@ -379,14 +379,14 @@ TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) { \ } \ } \ EXPECT_LE(max_diff, 1); \ - free_aligned_buffer_64(dst_y_c); \ - free_aligned_buffer_64(dst_u_c); \ - free_aligned_buffer_64(dst_v_c); \ - free_aligned_buffer_64(dst_y_opt); \ - free_aligned_buffer_64(dst_u_opt); \ - free_aligned_buffer_64(dst_v_opt); \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_uv); \ + free_aligned_buffer_page_end(dst_y_c); \ + free_aligned_buffer_page_end(dst_u_c); \ + free_aligned_buffer_page_end(dst_v_c); \ + free_aligned_buffer_page_end(dst_y_opt); \ + free_aligned_buffer_page_end(dst_u_opt); \ + free_aligned_buffer_page_end(dst_v_opt); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_uv); \ } #define TESTBIPLANARTOP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \ @@ -417,11 +417,11 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN); \ const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X); \ const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y); \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_u, kSizeUV + OFF); \ - align_buffer_64(src_v, kSizeUV + OFF); \ - align_buffer_64(dst_argb_c, kStrideB * kHeight + OFF); \ - align_buffer_64(dst_argb_opt, kStrideB * kHeight + OFF); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_u, kSizeUV + OFF); \ + align_buffer_page_end(src_v, kSizeUV + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideB * kHeight + OFF); \ + align_buffer_page_end(dst_argb_opt, kStrideB * kHeight + OFF); \ for (int i = 0; i < kWidth * kHeight; ++i) { \ src_y[i + OFF] = (fastrand() & 0xff); \ } \ @@ -447,8 +447,8 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ } \ int max_diff = 0; \ /* Convert to ARGB so 565 is expanded to bytes that can be compared. */ \ - align_buffer_64(dst_argb32_c, kWidth * BPP_C * kHeight); \ - align_buffer_64(dst_argb32_opt, kWidth * BPP_C * kHeight); \ + align_buffer_page_end(dst_argb32_c, kWidth * BPP_C * kHeight); \ + align_buffer_page_end(dst_argb32_opt, kWidth * BPP_C * kHeight); \ memset(dst_argb32_c, 2, kWidth * BPP_C * kHeight); \ memset(dst_argb32_opt, 102, kWidth * BPP_C * kHeight); \ FMT_B##To##FMT_C(dst_argb_c + OFF, kStrideB, \ @@ -466,13 +466,13 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ } \ } \ EXPECT_LE(max_diff, DIFF); \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_u); \ - free_aligned_buffer_64(src_v); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ - free_aligned_buffer_64(dst_argb32_c); \ - free_aligned_buffer_64(dst_argb32_opt); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_u); \ + free_aligned_buffer_page_end(src_v); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ + free_aligned_buffer_page_end(dst_argb32_c); \ + free_aligned_buffer_page_end(dst_argb32_opt); \ } #define TESTPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \ @@ -526,12 +526,12 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN); \ const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X); \ const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y); \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_u, kSizeUV + OFF); \ - align_buffer_64(src_v, kSizeUV + OFF); \ - align_buffer_64(src_a, kWidth * kHeight + OFF); \ - align_buffer_64(dst_argb_c, kStrideB * kHeight + OFF); \ - align_buffer_64(dst_argb_opt, kStrideB * kHeight + OFF); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_u, kSizeUV + OFF); \ + align_buffer_page_end(src_v, kSizeUV + OFF); \ + align_buffer_page_end(src_a, kWidth * kHeight + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideB * kHeight + OFF); \ + align_buffer_page_end(dst_argb_opt, kStrideB * kHeight + OFF); \ for (int i = 0; i < kWidth * kHeight; ++i) { \ src_y[i + OFF] = (fastrand() & 0xff); \ src_a[i + OFF] = (fastrand() & 0xff); \ @@ -568,12 +568,12 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ } \ } \ EXPECT_LE(max_diff, DIFF); \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_u); \ - free_aligned_buffer_64(src_v); \ - free_aligned_buffer_64(src_a); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_u); \ + free_aligned_buffer_page_end(src_v); \ + free_aligned_buffer_page_end(src_a); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ } #define TESTQPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \ @@ -599,11 +599,11 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ const int kHeight = benchmark_height_; \ const int kStrideB = kWidth * BPP_B; \ const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X); \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_uv, \ - kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y) * 2 + OFF); \ - align_buffer_64(dst_argb_c, kStrideB * kHeight); \ - align_buffer_64(dst_argb_opt, kStrideB * kHeight); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_uv, \ + kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y) * 2 + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideB * kHeight); \ + align_buffer_page_end(dst_argb_opt, kStrideB * kHeight); \ for (int i = 0; i < kHeight; ++i) \ for (int j = 0; j < kWidth; ++j) \ src_y[i * kWidth + j + OFF] = (fastrand() & 0xff); \ @@ -627,8 +627,8 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ kWidth, NEG kHeight); \ } \ /* Convert to ARGB so 565 is expanded to bytes that can be compared. */ \ - align_buffer_64(dst_argb32_c, kWidth * 4 * kHeight); \ - align_buffer_64(dst_argb32_opt, kWidth * 4 * kHeight); \ + align_buffer_page_end(dst_argb32_c, kWidth * 4 * kHeight); \ + align_buffer_page_end(dst_argb32_opt, kWidth * 4 * kHeight); \ memset(dst_argb32_c, 2, kWidth * 4 * kHeight); \ memset(dst_argb32_opt, 102, kWidth * 4 * kHeight); \ FMT_B##ToARGB(dst_argb_c, kStrideB, \ @@ -649,12 +649,12 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) { \ } \ } \ EXPECT_LE(max_diff, DIFF); \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_uv); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ - free_aligned_buffer_64(dst_argb32_c); \ - free_aligned_buffer_64(dst_argb32_opt); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_uv); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ + free_aligned_buffer_page_end(dst_argb32_c); \ + free_aligned_buffer_page_end(dst_argb32_opt); \ } #define TESTBIPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, DIFF) \ @@ -679,19 +679,16 @@ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) { \ const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X); \ const int kStride = \ (kStrideUV * SUBSAMP_X * 8 * BPP_A + 7) / 8; \ - align_buffer_64(src_argb, kStride * kHeight + OFF); \ - align_buffer_64(dst_y_c, kWidth * kHeight); \ - align_buffer_64(dst_u_c, \ - kStrideUV * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_v_c, \ - kStrideUV * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_y_opt, kWidth * kHeight); \ - align_buffer_64(dst_u_opt, \ - kStrideUV * \ - SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_v_opt, \ + align_buffer_page_end(src_argb, kStride * kHeight + OFF); \ + align_buffer_page_end(dst_y_c, kWidth * kHeight); \ + align_buffer_page_end(dst_u_c, \ + kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_v_c, \ + kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_y_opt, kWidth * kHeight); \ + align_buffer_page_end(dst_u_opt, \ + kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_v_opt, \ kStrideUV * \ SUBSAMPLE(kHeight, SUBSAMP_Y)); \ memset(dst_y_c, 1, kWidth * kHeight); \ @@ -741,13 +738,13 @@ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) { \ kStrideUV + j]), DIFF); \ } \ } \ - free_aligned_buffer_64(dst_y_c); \ - free_aligned_buffer_64(dst_u_c); \ - free_aligned_buffer_64(dst_v_c); \ - free_aligned_buffer_64(dst_y_opt); \ - free_aligned_buffer_64(dst_u_opt); \ - free_aligned_buffer_64(dst_v_opt); \ - free_aligned_buffer_64(src_argb); \ + free_aligned_buffer_page_end(dst_y_c); \ + free_aligned_buffer_page_end(dst_u_c); \ + free_aligned_buffer_page_end(dst_v_c); \ + free_aligned_buffer_page_end(dst_y_opt); \ + free_aligned_buffer_page_end(dst_u_opt); \ + free_aligned_buffer_page_end(dst_v_opt); \ + free_aligned_buffer_page_end(src_argb); \ } #define TESTATOPLANAR(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, \ @@ -790,18 +787,20 @@ TESTATOPLANAR(UYVY, 2, 1, I422, 2, 1, 2) TESTATOPLANAR(I400, 1, 1, I420, 2, 2, 2) TESTATOPLANAR(J400, 1, 1, J420, 2, 2, 2) -#define TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,\ - W1280, N, NEG, OFF) \ +#define TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, \ + SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG, OFF) \ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) { \ const int kWidth = ((W1280) > 0) ? (W1280) : 1; \ const int kHeight = benchmark_height_; \ const int kStride = SUBSAMPLE(kWidth, SUB_A) * BPP_A; \ const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X); \ - align_buffer_64(src_argb, kStride * kHeight + OFF); \ - align_buffer_64(dst_y_c, kWidth * kHeight); \ - align_buffer_64(dst_uv_c, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y)); \ - align_buffer_64(dst_y_opt, kWidth * kHeight); \ - align_buffer_64(dst_uv_opt, kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(src_argb, kStride * kHeight + OFF); \ + align_buffer_page_end(dst_y_c, kWidth * kHeight); \ + align_buffer_page_end(dst_uv_c, \ + kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y)); \ + align_buffer_page_end(dst_y_opt, kWidth * kHeight); \ + align_buffer_page_end(dst_uv_opt, \ + kStrideUV * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y)); \ for (int i = 0; i < kHeight; ++i) \ for (int j = 0; j < kStride; ++j) \ src_argb[(i * kStride) + j + OFF] = (fastrand() & 0xff); \ @@ -842,11 +841,11 @@ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) { \ } \ } \ EXPECT_LE(max_diff, 4); \ - free_aligned_buffer_64(dst_y_c); \ - free_aligned_buffer_64(dst_uv_c); \ - free_aligned_buffer_64(dst_y_opt); \ - free_aligned_buffer_64(dst_uv_opt); \ - free_aligned_buffer_64(src_argb); \ + free_aligned_buffer_page_end(dst_y_c); \ + free_aligned_buffer_page_end(dst_uv_c); \ + free_aligned_buffer_page_end(dst_y_opt); \ + free_aligned_buffer_page_end(dst_uv_opt); \ + free_aligned_buffer_page_end(src_argb); \ } #define TESTATOBIPLANAR(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \ @@ -874,9 +873,9 @@ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##N) { \ const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B; \ const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A; \ const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B; \ - align_buffer_64(src_argb, kStrideA * kHeightA + OFF); \ - align_buffer_64(dst_argb_c, kStrideB * kHeightB); \ - align_buffer_64(dst_argb_opt, kStrideB * kHeightB); \ + align_buffer_page_end(src_argb, kStrideA * kHeightA + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideB * kHeightB); \ + align_buffer_page_end(dst_argb_opt, kStrideB * kHeightB); \ for (int i = 0; i < kStrideA * kHeightA; ++i) { \ src_argb[i + OFF] = (fastrand() & 0xff); \ } \ @@ -902,9 +901,9 @@ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##N) { \ } \ } \ EXPECT_LE(max_diff, DIFF); \ - free_aligned_buffer_64(src_argb); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ + free_aligned_buffer_page_end(src_argb); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ } #define TESTATOBRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, \ @@ -1007,9 +1006,9 @@ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither##N) { \ const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B; \ const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A; \ const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B; \ - align_buffer_64(src_argb, kStrideA * kHeightA + OFF); \ - align_buffer_64(dst_argb_c, kStrideB * kHeightB); \ - align_buffer_64(dst_argb_opt, kStrideB * kHeightB); \ + align_buffer_page_end(src_argb, kStrideA * kHeightA + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideB * kHeightB); \ + align_buffer_page_end(dst_argb_opt, kStrideB * kHeightB); \ for (int i = 0; i < kStrideA * kHeightA; ++i) { \ src_argb[i + OFF] = (fastrand() & 0xff); \ } \ @@ -1035,9 +1034,9 @@ TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither##N) { \ } \ } \ EXPECT_LE(max_diff, DIFF); \ - free_aligned_buffer_64(src_argb); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ + free_aligned_buffer_page_end(src_argb); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ } #define TESTATOBDRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, \ @@ -1108,9 +1107,9 @@ TEST_F(LibYUVConvertTest, FMT_ATOB##_Symetric##N) { \ const int kHeight = benchmark_height_; \ const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A; \ const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A; \ - align_buffer_64(src_argb, kStrideA * kHeightA + OFF); \ - align_buffer_64(dst_argb_c, kStrideA * kHeightA); \ - align_buffer_64(dst_argb_opt, kStrideA * kHeightA); \ + align_buffer_page_end(src_argb, kStrideA * kHeightA + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideA * kHeightA); \ + align_buffer_page_end(dst_argb_opt, kStrideA * kHeightA); \ for (int i = 0; i < kStrideA * kHeightA; ++i) { \ src_argb[i + OFF] = (fastrand() & 0xff); \ } \ @@ -1138,9 +1137,9 @@ TEST_F(LibYUVConvertTest, FMT_ATOB##_Symetric##N) { \ EXPECT_EQ(src_argb[i + OFF], dst_argb_opt[i]); \ EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]); \ } \ - free_aligned_buffer_64(src_argb); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ + free_aligned_buffer_page_end(src_argb); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ } #define TESTSYM(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A) \ @@ -1361,22 +1360,22 @@ TEST_F(LibYUVConvertTest, NV12Crop) { const int sample_size = kWidth * kHeight + kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y) * 2; - align_buffer_64(src_y, sample_size); + align_buffer_page_end(src_y, sample_size); uint8* src_uv = src_y + kWidth * kHeight; - align_buffer_64(dst_y, kDestWidth * kDestHeight); - align_buffer_64(dst_u, + align_buffer_page_end(dst_y, kDestWidth * kDestHeight); + align_buffer_page_end(dst_u, SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y)); - align_buffer_64(dst_v, + align_buffer_page_end(dst_v, SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y)); - align_buffer_64(dst_y_2, kDestWidth * kDestHeight); - align_buffer_64(dst_u_2, + align_buffer_page_end(dst_y_2, kDestWidth * kDestHeight); + align_buffer_page_end(dst_u_2, SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y)); - align_buffer_64(dst_v_2, + align_buffer_page_end(dst_v_2, SUBSAMPLE(kDestWidth, SUBSAMP_X) * SUBSAMPLE(kDestHeight, SUBSAMP_Y)); @@ -1432,13 +1431,13 @@ TEST_F(LibYUVConvertTest, NV12Crop) { dst_v_2[i * SUBSAMPLE(kDestWidth, SUBSAMP_X) + j]); } } - free_aligned_buffer_64(dst_y); - free_aligned_buffer_64(dst_u); - free_aligned_buffer_64(dst_v); - free_aligned_buffer_64(dst_y_2); - free_aligned_buffer_64(dst_u_2); - free_aligned_buffer_64(dst_v_2); - free_aligned_buffer_64(src_y); + free_aligned_buffer_page_end(dst_y); + free_aligned_buffer_page_end(dst_u); + free_aligned_buffer_page_end(dst_v); + free_aligned_buffer_page_end(dst_y_2); + free_aligned_buffer_page_end(dst_u_2); + free_aligned_buffer_page_end(dst_v_2); + free_aligned_buffer_page_end(src_y); } TEST_F(LibYUVConvertTest, TestYToARGB) { @@ -1471,9 +1470,10 @@ static const uint8 kNoDither4x4[16] = { }; TEST_F(LibYUVConvertTest, TestNoDither) { - align_buffer_64(src_argb, benchmark_width_ * benchmark_height_ * 4); - align_buffer_64(dst_rgb565, benchmark_width_ * benchmark_height_ * 2); - align_buffer_64(dst_rgb565dither, benchmark_width_ * benchmark_height_ * 2); + align_buffer_page_end(src_argb, benchmark_width_ * benchmark_height_ * 4); + align_buffer_page_end(dst_rgb565, benchmark_width_ * benchmark_height_ * 2); + align_buffer_page_end(dst_rgb565dither, + benchmark_width_ * benchmark_height_ * 2); MemRandomize(src_argb, benchmark_width_ * benchmark_height_ * 4); MemRandomize(dst_rgb565, benchmark_width_ * benchmark_height_ * 2); MemRandomize(dst_rgb565dither, benchmark_width_ * benchmark_height_ * 2); @@ -1487,9 +1487,9 @@ TEST_F(LibYUVConvertTest, TestNoDither) { EXPECT_EQ(dst_rgb565[i], dst_rgb565dither[i]); } - free_aligned_buffer_64(src_argb); - free_aligned_buffer_64(dst_rgb565); - free_aligned_buffer_64(dst_rgb565dither); + free_aligned_buffer_page_end(src_argb); + free_aligned_buffer_page_end(dst_rgb565); + free_aligned_buffer_page_end(dst_rgb565dither); } // Ordered 4x4 dither for 888 to 565. Values from 0 to 7. @@ -1501,11 +1501,13 @@ static const uint8 kDither565_4x4[16] = { }; TEST_F(LibYUVConvertTest, TestDither) { - align_buffer_64(src_argb, benchmark_width_ * benchmark_height_ * 4); - align_buffer_64(dst_rgb565, benchmark_width_ * benchmark_height_ * 2); - align_buffer_64(dst_rgb565dither, benchmark_width_ * benchmark_height_ * 2); - align_buffer_64(dst_argb, benchmark_width_ * benchmark_height_ * 4); - align_buffer_64(dst_argbdither, benchmark_width_ * benchmark_height_ * 4); + align_buffer_page_end(src_argb, benchmark_width_ * benchmark_height_ * 4); + align_buffer_page_end(dst_rgb565, benchmark_width_ * benchmark_height_ * 2); + align_buffer_page_end(dst_rgb565dither, + benchmark_width_ * benchmark_height_ * 2); + align_buffer_page_end(dst_argb, benchmark_width_ * benchmark_height_ * 4); + align_buffer_page_end(dst_argbdither, + benchmark_width_ * benchmark_height_ * 4); MemRandomize(src_argb, benchmark_width_ * benchmark_height_ * 4); MemRandomize(dst_rgb565, benchmark_width_ * benchmark_height_ * 2); MemRandomize(dst_rgb565dither, benchmark_width_ * benchmark_height_ * 2); @@ -1527,11 +1529,11 @@ TEST_F(LibYUVConvertTest, TestDither) { for (int i = 0; i < benchmark_width_ * benchmark_height_ * 4; ++i) { EXPECT_NEAR(dst_argb[i], dst_argbdither[i], 9); } - free_aligned_buffer_64(src_argb); - free_aligned_buffer_64(dst_rgb565); - free_aligned_buffer_64(dst_rgb565dither); - free_aligned_buffer_64(dst_argb); - free_aligned_buffer_64(dst_argbdither); + free_aligned_buffer_page_end(src_argb); + free_aligned_buffer_page_end(dst_rgb565); + free_aligned_buffer_page_end(dst_rgb565dither); + free_aligned_buffer_page_end(dst_argb); + free_aligned_buffer_page_end(dst_argbdither); } #define TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \ @@ -1542,11 +1544,11 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##Dither##N) { \ const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN); \ const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X); \ const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y); \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_u, kSizeUV + OFF); \ - align_buffer_64(src_v, kSizeUV + OFF); \ - align_buffer_64(dst_argb_c, kStrideB * kHeight + OFF); \ - align_buffer_64(dst_argb_opt, kStrideB * kHeight + OFF); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_u, kSizeUV + OFF); \ + align_buffer_page_end(src_v, kSizeUV + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideB * kHeight + OFF); \ + align_buffer_page_end(dst_argb_opt, kStrideB * kHeight + OFF); \ for (int i = 0; i < kWidth * kHeight; ++i) { \ src_y[i + OFF] = (fastrand() & 0xff); \ } \ @@ -1572,8 +1574,8 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##Dither##N) { \ } \ int max_diff = 0; \ /* Convert to ARGB so 565 is expanded to bytes that can be compared. */ \ - align_buffer_64(dst_argb32_c, kWidth * BPP_C * kHeight); \ - align_buffer_64(dst_argb32_opt, kWidth * BPP_C * kHeight); \ + align_buffer_page_end(dst_argb32_c, kWidth * BPP_C * kHeight); \ + align_buffer_page_end(dst_argb32_opt, kWidth * BPP_C * kHeight); \ memset(dst_argb32_c, 2, kWidth * BPP_C * kHeight); \ memset(dst_argb32_opt, 102, kWidth * BPP_C * kHeight); \ FMT_B##To##FMT_C(dst_argb_c + OFF, kStrideB, \ @@ -1591,13 +1593,13 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##Dither##N) { \ } \ } \ EXPECT_LE(max_diff, DIFF); \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_u); \ - free_aligned_buffer_64(src_v); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ - free_aligned_buffer_64(dst_argb32_c); \ - free_aligned_buffer_64(dst_argb32_opt); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_u); \ + free_aligned_buffer_page_end(src_v); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ + free_aligned_buffer_page_end(dst_argb32_c); \ + free_aligned_buffer_page_end(dst_argb32_opt); \ } #define TESTPLANARTOBD(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \ @@ -1618,23 +1620,23 @@ TEST_F(LibYUVConvertTest, NAME) { \ const int kWidth = benchmark_width_; \ const int kHeight = benchmark_height_; \ \ - align_buffer_64(orig_uyvy, \ + align_buffer_page_end(orig_uyvy, \ 4 * SUBSAMPLE(kWidth, 2) * kHeight); \ - align_buffer_64(orig_y, kWidth * kHeight); \ - align_buffer_64(orig_u, \ + align_buffer_page_end(orig_y, kWidth * kHeight); \ + align_buffer_page_end(orig_u, \ SUBSAMPLE(kWidth, 2) * \ SUBSAMPLE(kHeight, 2)); \ - align_buffer_64(orig_v, \ + align_buffer_page_end(orig_v, \ SUBSAMPLE(kWidth, 2) * \ SUBSAMPLE(kHeight, 2)); \ \ - align_buffer_64(dst_y_orig, kWidth * kHeight); \ - align_buffer_64(dst_uv_orig, 2 * \ + align_buffer_page_end(dst_y_orig, kWidth * kHeight); \ + align_buffer_page_end(dst_uv_orig, 2 * \ SUBSAMPLE(kWidth, 2) * \ SUBSAMPLE(kHeight, 2)); \ \ - align_buffer_64(dst_y, kWidth * kHeight); \ - align_buffer_64(dst_uv, 2 * \ + align_buffer_page_end(dst_y, kWidth * kHeight); \ + align_buffer_page_end(dst_uv, 2 * \ SUBSAMPLE(kWidth, 2) * \ SUBSAMPLE(kHeight, 2)); \ \ @@ -1671,14 +1673,14 @@ TEST_F(LibYUVConvertTest, NAME) { \ EXPECT_EQ(dst_uv_orig[i], dst_uv[i]); \ } \ \ - free_aligned_buffer_64(orig_uyvy); \ - free_aligned_buffer_64(orig_y); \ - free_aligned_buffer_64(orig_u); \ - free_aligned_buffer_64(orig_v); \ - free_aligned_buffer_64(dst_y_orig); \ - free_aligned_buffer_64(dst_uv_orig); \ - free_aligned_buffer_64(dst_y); \ - free_aligned_buffer_64(dst_uv); \ + free_aligned_buffer_page_end(orig_uyvy); \ + free_aligned_buffer_page_end(orig_y); \ + free_aligned_buffer_page_end(orig_u); \ + free_aligned_buffer_page_end(orig_v); \ + free_aligned_buffer_page_end(dst_y_orig); \ + free_aligned_buffer_page_end(dst_uv_orig); \ + free_aligned_buffer_page_end(dst_y); \ + free_aligned_buffer_page_end(dst_uv); \ } TESTPTOB(TestYUY2ToNV12, YUY2ToI420, YUY2ToNV12) @@ -1692,10 +1694,10 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) { \ const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B; \ const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X); \ const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y); \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_u, kSizeUV + OFF); \ - align_buffer_64(src_v, kSizeUV + OFF); \ - align_buffer_64(dst_argb_b, kStrideB * kHeight + OFF); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_u, kSizeUV + OFF); \ + align_buffer_page_end(src_v, kSizeUV + OFF); \ + align_buffer_page_end(dst_argb_b, kStrideB * kHeight + OFF); \ for (int i = 0; i < kWidth * kHeight; ++i) { \ src_y[i + OFF] = (fastrand() & 0xff); \ } \ @@ -1713,8 +1715,8 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) { \ } \ /* Convert to a 3rd format in 1 step and 2 steps and compare */ \ const int kStrideC = kWidth * BPP_C; \ - align_buffer_64(dst_argb_c, kStrideC * kHeight + OFF); \ - align_buffer_64(dst_argb_bc, kStrideC * kHeight + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideC * kHeight + OFF); \ + align_buffer_page_end(dst_argb_bc, kStrideC * kHeight + OFF); \ memset(dst_argb_c + OFF, 2, kStrideC * kHeight); \ memset(dst_argb_bc + OFF, 3, kStrideC * kHeight); \ FMT_PLANAR##To##FMT_C(src_y + OFF, kWidth, \ @@ -1729,12 +1731,12 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) { \ for (int i = 0; i < kStrideC * kHeight; ++i) { \ EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]); \ } \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_u); \ - free_aligned_buffer_64(src_v); \ - free_aligned_buffer_64(dst_argb_b); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_bc); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_u); \ + free_aligned_buffer_page_end(src_v); \ + free_aligned_buffer_page_end(dst_argb_b); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_bc); \ } #define TESTPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \ @@ -1789,11 +1791,11 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) { \ const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B; \ const int kSizeUV = \ SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y); \ - align_buffer_64(src_y, kWidth * kHeight + OFF); \ - align_buffer_64(src_u, kSizeUV + OFF); \ - align_buffer_64(src_v, kSizeUV + OFF); \ - align_buffer_64(src_a, kWidth * kHeight + OFF); \ - align_buffer_64(dst_argb_b, kStrideB * kHeight + OFF); \ + align_buffer_page_end(src_y, kWidth * kHeight + OFF); \ + align_buffer_page_end(src_u, kSizeUV + OFF); \ + align_buffer_page_end(src_v, kSizeUV + OFF); \ + align_buffer_page_end(src_a, kWidth * kHeight + OFF); \ + align_buffer_page_end(dst_argb_b, kStrideB * kHeight + OFF); \ for (int i = 0; i < kWidth * kHeight; ++i) { \ src_y[i + OFF] = (fastrand() & 0xff); \ src_a[i + OFF] = (fastrand() & 0xff); \ @@ -1814,8 +1816,8 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) { \ int max_diff = 0; \ /* Convert to a 3rd format in 1 step and 2 steps and compare */ \ const int kStrideC = kWidth * BPP_C; \ - align_buffer_64(dst_argb_c, kStrideC * kHeight + OFF); \ - align_buffer_64(dst_argb_bc, kStrideC * kHeight + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideC * kHeight + OFF); \ + align_buffer_page_end(dst_argb_bc, kStrideC * kHeight + OFF); \ memset(dst_argb_c + OFF, 2, kStrideC * kHeight); \ memset(dst_argb_bc + OFF, 3, kStrideC * kHeight); \ FMT_PLANAR##To##FMT_C(src_y + OFF, kWidth, \ @@ -1831,13 +1833,13 @@ TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) { \ for (int i = 0; i < kStrideC * kHeight; ++i) { \ EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]); \ } \ - free_aligned_buffer_64(src_y); \ - free_aligned_buffer_64(src_u); \ - free_aligned_buffer_64(src_v); \ - free_aligned_buffer_64(src_a); \ - free_aligned_buffer_64(dst_argb_b); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_bc); \ + free_aligned_buffer_page_end(src_y); \ + free_aligned_buffer_page_end(src_u); \ + free_aligned_buffer_page_end(src_v); \ + free_aligned_buffer_page_end(src_a); \ + free_aligned_buffer_page_end(dst_argb_b); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_bc); \ } #define TESTQPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \ diff --git a/unit_test/cpu_test.cc b/unit_test/cpu_test.cc index 5933ee44..0cd06f9b 100644 --- a/unit_test/cpu_test.cc +++ b/unit_test/cpu_test.cc @@ -13,7 +13,6 @@ #include "libyuv/basic_types.h" #include "libyuv/cpu_id.h" -#include "libyuv/row.h" // For HAS_ARGBSHUFFLEROW_AVX2. #include "libyuv/version.h" #include "../unit_test/unit_test.h" @@ -68,16 +67,9 @@ printf("_MSC_VER %d\n", _MSC_VER); #if !defined(LIBYUV_DISABLE_X86) && (defined(GCC_HAS_AVX2) || \ defined(CLANG_HAS_AVX2) || defined(VISUALC_HAS_AVX2)) printf("Has AVX2 1\n"); - // If compiler supports AVX2, the following function is expected to exist: -#if !defined(HAS_ARGBSHUFFLEROW_AVX2) - EXPECT_TRUE(0); // HAS_ARGBSHUFFLEROW_AVX2 was expected. -#endif #else printf("Has AVX2 0\n"); // If compiler does not support AVX2, the following function not expected: -#if defined(HAS_ARGBSHUFFLEROW_AVX2) - EXPECT_TRUE(0); // HAS_ARGBSHUFFLEROW_AVX2 was not expected. -#endif #endif } diff --git a/unit_test/math_test.cc b/unit_test/math_test.cc index 62979542..19af9f6b 100644 --- a/unit_test/math_test.cc +++ b/unit_test/math_test.cc @@ -14,7 +14,6 @@ #include "libyuv/basic_types.h" #include "libyuv/cpu_id.h" -#include "libyuv/row.h" #include "libyuv/scale.h" #include "libyuv/scale_row.h" #include "../unit_test/unit_test.h" diff --git a/unit_test/planar_test.cc b/unit_test/planar_test.cc index 1974a033..d2ae35d0 100644 --- a/unit_test/planar_test.cc +++ b/unit_test/planar_test.cc @@ -19,17 +19,16 @@ #include "libyuv/cpu_id.h" #include "libyuv/planar_functions.h" #include "libyuv/rotate.h" -#include "libyuv/row.h" // For Sobel #include "../unit_test/unit_test.h" namespace libyuv { TEST_F(LibYUVPlanarTest, TestAttenuate) { const int kSize = 1280 * 4; - align_buffer_64(orig_pixels, kSize); - align_buffer_64(atten_pixels, kSize); - align_buffer_64(unatten_pixels, kSize); - align_buffer_64(atten2_pixels, kSize); + align_buffer_page_end(orig_pixels, kSize); + align_buffer_page_end(atten_pixels, kSize); + align_buffer_page_end(unatten_pixels, kSize); + align_buffer_page_end(atten2_pixels, kSize); // Test unattenuation clamps orig_pixels[0 * 4 + 0] = 200u; @@ -98,10 +97,10 @@ TEST_F(LibYUVPlanarTest, TestAttenuate) { EXPECT_NEAR(85, atten_pixels[255 * 4 + 2], 1); EXPECT_EQ(255, atten_pixels[255 * 4 + 3]); - free_aligned_buffer_64(atten2_pixels); - free_aligned_buffer_64(unatten_pixels); - free_aligned_buffer_64(atten_pixels); - free_aligned_buffer_64(orig_pixels); + free_aligned_buffer_page_end(atten2_pixels); + free_aligned_buffer_page_end(unatten_pixels); + free_aligned_buffer_page_end(atten_pixels); + free_aligned_buffer_page_end(orig_pixels); } static int TestAttenuateI(int width, int height, int benchmark_iterations, @@ -112,9 +111,9 @@ static int TestAttenuateI(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); for (int i = 0; i < kStride * height; ++i) { src_argb[i + off] = (fastrand() & 0xff); } @@ -140,9 +139,9 @@ static int TestAttenuateI(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -186,9 +185,9 @@ static int TestUnattenuateI(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); for (int i = 0; i < kStride * height; ++i) { src_argb[i + off] = (fastrand() & 0xff); } @@ -217,9 +216,9 @@ static int TestUnattenuateI(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -1023,16 +1022,16 @@ TEST_F(LibYUVPlanarTest, TestInterpolatePlane) { #define TESTTERP(FMT_A, BPP_A, STRIDE_A, \ FMT_B, BPP_B, STRIDE_B, \ - W1280, TERP, N, NEG, OFF) \ + W1280, TERP, N, NEG, OFF) \ TEST_F(LibYUVPlanarTest, ARGBInterpolate##TERP##N) { \ const int kWidth = ((W1280) > 0) ? (W1280) : 1; \ const int kHeight = benchmark_height_; \ const int kStrideA = (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A; \ const int kStrideB = (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B; \ - align_buffer_64(src_argb_a, kStrideA * kHeight + OFF); \ - align_buffer_64(src_argb_b, kStrideA * kHeight + OFF); \ - align_buffer_64(dst_argb_c, kStrideB * kHeight); \ - align_buffer_64(dst_argb_opt, kStrideB * kHeight); \ + align_buffer_page_end(src_argb_a, kStrideA * kHeight + OFF); \ + align_buffer_page_end(src_argb_b, kStrideA * kHeight + OFF); \ + align_buffer_page_end(dst_argb_c, kStrideB * kHeight); \ + align_buffer_page_end(dst_argb_opt, kStrideB * kHeight); \ for (int i = 0; i < kStrideA * kHeight; ++i) { \ src_argb_a[i + OFF] = (fastrand() & 0xff); \ src_argb_b[i + OFF] = (fastrand() & 0xff); \ @@ -1052,10 +1051,10 @@ TEST_F(LibYUVPlanarTest, ARGBInterpolate##TERP##N) { \ for (int i = 0; i < kStrideB * kHeight; ++i) { \ EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]); \ } \ - free_aligned_buffer_64(src_argb_a); \ - free_aligned_buffer_64(src_argb_b); \ - free_aligned_buffer_64(dst_argb_c); \ - free_aligned_buffer_64(dst_argb_opt); \ + free_aligned_buffer_page_end(src_argb_a); \ + free_aligned_buffer_page_end(src_argb_b); \ + free_aligned_buffer_page_end(dst_argb_c); \ + free_aligned_buffer_page_end(dst_argb_opt); \ } #define TESTINTERPOLATE(TERP) \ @@ -1078,10 +1077,10 @@ static int TestBlend(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(src_argb_b, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(src_argb_b, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); for (int i = 0; i < kStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); src_argb_b[i + off] = (fastrand() & 0xff); @@ -1114,10 +1113,10 @@ static int TestBlend(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(src_argb_b); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(src_argb_b); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -1157,11 +1156,11 @@ static void TestBlendPlane(int width, int height, int benchmark_iterations, } const int kBpp = 1; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(src_argb_b, kStride * height + off); - align_buffer_64(src_argb_alpha, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height + off); - align_buffer_64(dst_argb_opt, kStride * height + off); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(src_argb_b, kStride * height + off); + align_buffer_page_end(src_argb_alpha, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height + off); + align_buffer_page_end(dst_argb_opt, kStride * height + off); memset(dst_argb_c, 255, kStride * height + off); memset(dst_argb_opt, 255, kStride * height + off); @@ -1212,11 +1211,11 @@ static void TestBlendPlane(int width, int height, int benchmark_iterations, for (int i = 0; i < kStride * height; ++i) { EXPECT_EQ(dst_argb_c[i + off], dst_argb_opt[i + off]); } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(src_argb_b); - free_aligned_buffer_64(src_argb_alpha); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(src_argb_b); + free_aligned_buffer_page_end(src_argb_alpha); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return; } @@ -1245,19 +1244,19 @@ static void TestI420Blend(int width, int height, int benchmark_iterations, width = ((width) > 0) ? (width) : 1; const int kStrideUV = SUBSAMPLE(width, 2); const int kSizeUV = kStrideUV * SUBSAMPLE(height, 2); - align_buffer_64(src_y0, width * height + off); - align_buffer_64(src_u0, kSizeUV + off); - align_buffer_64(src_v0, kSizeUV + off); - align_buffer_64(src_y1, width * height + off); - align_buffer_64(src_u1, kSizeUV + off); - align_buffer_64(src_v1, kSizeUV + off); - align_buffer_64(src_a, width * height + off); - align_buffer_64(dst_y_c, width * height + off); - align_buffer_64(dst_u_c, kSizeUV + off); - align_buffer_64(dst_v_c, kSizeUV + off); - align_buffer_64(dst_y_opt, width * height + off); - align_buffer_64(dst_u_opt, kSizeUV + off); - align_buffer_64(dst_v_opt, kSizeUV + off); + align_buffer_page_end(src_y0, width * height + off); + align_buffer_page_end(src_u0, kSizeUV + off); + align_buffer_page_end(src_v0, kSizeUV + off); + align_buffer_page_end(src_y1, width * height + off); + align_buffer_page_end(src_u1, kSizeUV + off); + align_buffer_page_end(src_v1, kSizeUV + off); + align_buffer_page_end(src_a, width * height + off); + align_buffer_page_end(dst_y_c, width * height + off); + align_buffer_page_end(dst_u_c, kSizeUV + off); + align_buffer_page_end(dst_v_c, kSizeUV + off); + align_buffer_page_end(dst_y_opt, width * height + off); + align_buffer_page_end(dst_u_opt, kSizeUV + off); + align_buffer_page_end(dst_v_opt, kSizeUV + off); MemRandomize(src_y0, width * height + off); MemRandomize(src_u0, kSizeUV + off); @@ -1306,19 +1305,19 @@ static void TestI420Blend(int width, int height, int benchmark_iterations, EXPECT_EQ(dst_u_c[i + off], dst_u_opt[i + off]); EXPECT_EQ(dst_v_c[i + off], dst_v_opt[i + off]); } - free_aligned_buffer_64(src_y0); - free_aligned_buffer_64(src_u0); - free_aligned_buffer_64(src_v0); - free_aligned_buffer_64(src_y1); - free_aligned_buffer_64(src_u1); - free_aligned_buffer_64(src_v1); - free_aligned_buffer_64(src_a); - free_aligned_buffer_64(dst_y_c); - free_aligned_buffer_64(dst_u_c); - free_aligned_buffer_64(dst_v_c); - free_aligned_buffer_64(dst_y_opt); - free_aligned_buffer_64(dst_u_opt); - free_aligned_buffer_64(dst_v_opt); + free_aligned_buffer_page_end(src_y0); + free_aligned_buffer_page_end(src_u0); + free_aligned_buffer_page_end(src_v0); + free_aligned_buffer_page_end(src_y1); + free_aligned_buffer_page_end(src_u1); + free_aligned_buffer_page_end(src_v1); + free_aligned_buffer_page_end(src_a); + free_aligned_buffer_page_end(dst_y_c); + free_aligned_buffer_page_end(dst_u_c); + free_aligned_buffer_page_end(dst_v_c); + free_aligned_buffer_page_end(dst_y_opt); + free_aligned_buffer_page_end(dst_u_opt); + free_aligned_buffer_page_end(dst_v_opt); return; } @@ -1375,206 +1374,6 @@ TEST_F(LibYUVPlanarTest, TestAffine) { #endif } -TEST_F(LibYUVPlanarTest, TestSobelX) { - SIMD_ALIGNED(uint8 orig_pixels_0[1280 + 2]); - SIMD_ALIGNED(uint8 orig_pixels_1[1280 + 2]); - SIMD_ALIGNED(uint8 orig_pixels_2[1280 + 2]); - SIMD_ALIGNED(uint8 sobel_pixels_c[1280]); - SIMD_ALIGNED(uint8 sobel_pixels_opt[1280]); - - for (int i = 0; i < 1280 + 2; ++i) { - orig_pixels_0[i] = i; - orig_pixels_1[i] = i * 2; - orig_pixels_2[i] = i * 3; - } - - SobelXRow_C(orig_pixels_0, orig_pixels_1, orig_pixels_2, - sobel_pixels_c, 1280); - - EXPECT_EQ(16u, sobel_pixels_c[0]); - EXPECT_EQ(16u, sobel_pixels_c[100]); - EXPECT_EQ(255u, sobel_pixels_c[255]); - - void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1, - const uint8* src_y2, uint8* dst_sobely, int width) = - SobelXRow_C; -#if defined(HAS_SOBELXROW_SSE2) - if (TestCpuFlag(kCpuHasSSE2)) { - SobelXRow = SobelXRow_SSE2; - } -#endif -#if defined(HAS_SOBELXROW_NEON) - if (TestCpuFlag(kCpuHasNEON)) { - SobelXRow = SobelXRow_NEON; - } -#endif - for (int i = 0; i < benchmark_pixels_div1280_; ++i) { - SobelXRow(orig_pixels_0, orig_pixels_1, orig_pixels_2, - sobel_pixels_opt, 1280); - } - for (int i = 0; i < 1280; ++i) { - EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]); - } -} - -TEST_F(LibYUVPlanarTest, TestSobelY) { - SIMD_ALIGNED(uint8 orig_pixels_0[1280 + 2]); - SIMD_ALIGNED(uint8 orig_pixels_1[1280 + 2]); - SIMD_ALIGNED(uint8 sobel_pixels_c[1280]); - SIMD_ALIGNED(uint8 sobel_pixels_opt[1280]); - - for (int i = 0; i < 1280 + 2; ++i) { - orig_pixels_0[i] = i; - orig_pixels_1[i] = i * 2; - } - - SobelYRow_C(orig_pixels_0, orig_pixels_1, sobel_pixels_c, 1280); - - EXPECT_EQ(4u, sobel_pixels_c[0]); - EXPECT_EQ(255u, sobel_pixels_c[100]); - EXPECT_EQ(0u, sobel_pixels_c[255]); - void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1, - uint8* dst_sobely, int width) = SobelYRow_C; -#if defined(HAS_SOBELYROW_SSE2) - if (TestCpuFlag(kCpuHasSSE2)) { - SobelYRow = SobelYRow_SSE2; - } -#endif -#if defined(HAS_SOBELYROW_NEON) - if (TestCpuFlag(kCpuHasNEON)) { - SobelYRow = SobelYRow_NEON; - } -#endif - for (int i = 0; i < benchmark_pixels_div1280_; ++i) { - SobelYRow(orig_pixels_0, orig_pixels_1, sobel_pixels_opt, 1280); - } - for (int i = 0; i < 1280; ++i) { - EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]); - } -} - -TEST_F(LibYUVPlanarTest, TestSobel) { - SIMD_ALIGNED(uint8 orig_sobelx[1280]); - SIMD_ALIGNED(uint8 orig_sobely[1280]); - SIMD_ALIGNED(uint8 sobel_pixels_c[1280 * 4]); - SIMD_ALIGNED(uint8 sobel_pixels_opt[1280 * 4]); - - for (int i = 0; i < 1280; ++i) { - orig_sobelx[i] = i; - orig_sobely[i] = i * 2; - } - - SobelRow_C(orig_sobelx, orig_sobely, sobel_pixels_c, 1280); - - EXPECT_EQ(0u, sobel_pixels_c[0]); - EXPECT_EQ(3u, sobel_pixels_c[4]); - EXPECT_EQ(3u, sobel_pixels_c[5]); - EXPECT_EQ(3u, sobel_pixels_c[6]); - EXPECT_EQ(255u, sobel_pixels_c[7]); - EXPECT_EQ(6u, sobel_pixels_c[8]); - EXPECT_EQ(6u, sobel_pixels_c[9]); - EXPECT_EQ(6u, sobel_pixels_c[10]); - EXPECT_EQ(255u, sobel_pixels_c[7]); - EXPECT_EQ(255u, sobel_pixels_c[100 * 4 + 1]); - EXPECT_EQ(255u, sobel_pixels_c[255 * 4 + 1]); - void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely, - uint8* dst_argb, int width) = SobelRow_C; -#if defined(HAS_SOBELROW_SSE2) - if (TestCpuFlag(kCpuHasSSE2)) { - SobelRow = SobelRow_SSE2; - } -#endif -#if defined(HAS_SOBELROW_NEON) - if (TestCpuFlag(kCpuHasNEON)) { - SobelRow = SobelRow_NEON; - } -#endif - for (int i = 0; i < benchmark_pixels_div1280_; ++i) { - SobelRow(orig_sobelx, orig_sobely, sobel_pixels_opt, 1280); - } - for (int i = 0; i < 1280 * 4; ++i) { - EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]); - } -} - -TEST_F(LibYUVPlanarTest, TestSobelToPlane) { - SIMD_ALIGNED(uint8 orig_sobelx[1280]); - SIMD_ALIGNED(uint8 orig_sobely[1280]); - SIMD_ALIGNED(uint8 sobel_pixels_c[1280]); - SIMD_ALIGNED(uint8 sobel_pixels_opt[1280]); - - for (int i = 0; i < 1280; ++i) { - orig_sobelx[i] = i; - orig_sobely[i] = i * 2; - } - - SobelToPlaneRow_C(orig_sobelx, orig_sobely, sobel_pixels_c, 1280); - - EXPECT_EQ(0u, sobel_pixels_c[0]); - EXPECT_EQ(3u, sobel_pixels_c[1]); - EXPECT_EQ(6u, sobel_pixels_c[2]); - EXPECT_EQ(99u, sobel_pixels_c[33]); - EXPECT_EQ(255u, sobel_pixels_c[100]); - void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely, - uint8* dst_y, int width) = SobelToPlaneRow_C; -#if defined(HAS_SOBELTOPLANEROW_SSE2) - if (TestCpuFlag(kCpuHasSSE2)) { - SobelToPlaneRow = SobelToPlaneRow_SSE2; - } -#endif -#if defined(HAS_SOBELTOPLANEROW_NEON) - if (TestCpuFlag(kCpuHasNEON)) { - SobelToPlaneRow = SobelToPlaneRow_NEON; - } -#endif - for (int i = 0; i < benchmark_pixels_div1280_; ++i) { - SobelToPlaneRow(orig_sobelx, orig_sobely, sobel_pixels_opt, 1280); - } - for (int i = 0; i < 1280; ++i) { - EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]); - } -} - -TEST_F(LibYUVPlanarTest, TestSobelXY) { - SIMD_ALIGNED(uint8 orig_sobelx[1280]); - SIMD_ALIGNED(uint8 orig_sobely[1280]); - SIMD_ALIGNED(uint8 sobel_pixels_c[1280 * 4]); - SIMD_ALIGNED(uint8 sobel_pixels_opt[1280 * 4]); - - for (int i = 0; i < 1280; ++i) { - orig_sobelx[i] = i; - orig_sobely[i] = i * 2; - } - - SobelXYRow_C(orig_sobelx, orig_sobely, sobel_pixels_c, 1280); - - EXPECT_EQ(0u, sobel_pixels_c[0]); - EXPECT_EQ(2u, sobel_pixels_c[4]); - EXPECT_EQ(3u, sobel_pixels_c[5]); - EXPECT_EQ(1u, sobel_pixels_c[6]); - EXPECT_EQ(255u, sobel_pixels_c[7]); - EXPECT_EQ(255u, sobel_pixels_c[100 * 4 + 1]); - EXPECT_EQ(255u, sobel_pixels_c[255 * 4 + 1]); - void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely, - uint8* dst_argb, int width) = SobelXYRow_C; -#if defined(HAS_SOBELXYROW_SSE2) - if (TestCpuFlag(kCpuHasSSE2)) { - SobelXYRow = SobelXYRow_SSE2; - } -#endif -#if defined(HAS_SOBELXYROW_NEON) - if (TestCpuFlag(kCpuHasNEON)) { - SobelXYRow = SobelXYRow_NEON; - } -#endif - for (int i = 0; i < benchmark_pixels_div1280_; ++i) { - SobelXYRow(orig_sobelx, orig_sobely, sobel_pixels_opt, 1280); - } - for (int i = 0; i < 1280 * 4; ++i) { - EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]); - } -} - TEST_F(LibYUVPlanarTest, TestCopyPlane) { int err = 0; int yw = benchmark_width_; @@ -1583,9 +1382,9 @@ TEST_F(LibYUVPlanarTest, TestCopyPlane) { int i, j; int y_plane_size = (yw + b * 2) * (yh + b * 2); - align_buffer_64(orig_y, y_plane_size); - align_buffer_64(dst_c, y_plane_size); - align_buffer_64(dst_opt, y_plane_size); + align_buffer_page_end(orig_y, y_plane_size); + align_buffer_page_end(dst_c, y_plane_size); + align_buffer_page_end(dst_opt, y_plane_size); memset(orig_y, 0, y_plane_size); memset(dst_c, 0, y_plane_size); @@ -1631,9 +1430,9 @@ TEST_F(LibYUVPlanarTest, TestCopyPlane) { ++err; } - free_aligned_buffer_64(orig_y); - free_aligned_buffer_64(dst_c); - free_aligned_buffer_64(dst_opt); + free_aligned_buffer_page_end(orig_y); + free_aligned_buffer_page_end(dst_c); + free_aligned_buffer_page_end(dst_opt); EXPECT_EQ(0, err); } @@ -1646,10 +1445,10 @@ static int TestMultiply(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(src_argb_b, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(src_argb_b, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); for (int i = 0; i < kStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); src_argb_b[i + off] = (fastrand() & 0xff); @@ -1678,10 +1477,10 @@ static int TestMultiply(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(src_argb_b); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(src_argb_b); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -1721,10 +1520,10 @@ static int TestAdd(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(src_argb_b, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(src_argb_b, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); for (int i = 0; i < kStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); src_argb_b[i + off] = (fastrand() & 0xff); @@ -1753,10 +1552,10 @@ static int TestAdd(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(src_argb_b); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(src_argb_b); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -1796,10 +1595,10 @@ static int TestSubtract(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(src_argb_b, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(src_argb_b, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); for (int i = 0; i < kStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); src_argb_b[i + off] = (fastrand() & 0xff); @@ -1828,10 +1627,10 @@ static int TestSubtract(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(src_argb_b); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(src_argb_b); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -1871,9 +1670,9 @@ static int TestSobel(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); memset(src_argb_a, 0, kStride * height + off); for (int i = 0; i < kStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); @@ -1900,9 +1699,9 @@ static int TestSobel(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -1944,9 +1743,9 @@ static int TestSobelToPlane(int width, int height, int benchmark_iterations, const int kDstBpp = 1; const int kSrcStride = (width * kSrcBpp + 15) & ~15; const int kDstStride = (width * kDstBpp + 15) & ~15; - align_buffer_64(src_argb_a, kSrcStride * height + off); - align_buffer_64(dst_argb_c, kDstStride * height); - align_buffer_64(dst_argb_opt, kDstStride * height); + align_buffer_page_end(src_argb_a, kSrcStride * height + off); + align_buffer_page_end(dst_argb_c, kDstStride * height); + align_buffer_page_end(dst_argb_opt, kDstStride * height); memset(src_argb_a, 0, kSrcStride * height + off); for (int i = 0; i < kSrcStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); @@ -1973,9 +1772,9 @@ static int TestSobelToPlane(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -2019,9 +1818,9 @@ static int TestSobelXY(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); memset(src_argb_a, 0, kStride * height + off); for (int i = 0; i < kStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); @@ -2048,9 +1847,9 @@ static int TestSobelXY(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } @@ -2090,10 +1889,10 @@ static int TestBlur(int width, int height, int benchmark_iterations, } const int kBpp = 4; const int kStride = width * kBpp; - align_buffer_64(src_argb_a, kStride * height + off); - align_buffer_64(dst_cumsum, width * height * 16); - align_buffer_64(dst_argb_c, kStride * height); - align_buffer_64(dst_argb_opt, kStride * height); + align_buffer_page_end(src_argb_a, kStride * height + off); + align_buffer_page_end(dst_cumsum, width * height * 16); + align_buffer_page_end(dst_argb_c, kStride * height); + align_buffer_page_end(dst_argb_opt, kStride * height); for (int i = 0; i < kStride * height; ++i) { src_argb_a[i + off] = (fastrand() & 0xff); } @@ -2122,15 +1921,15 @@ static int TestBlur(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(src_argb_a); - free_aligned_buffer_64(dst_cumsum); - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(src_argb_a); + free_aligned_buffer_page_end(dst_cumsum); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } static const int kBlurSize = 55; -TEST_F(LibYUVPlanarTest, ARGBBlur_Any) { +TEST_F(LibYUVPlanarTest, DISABLED_ARGBBlur_Any) { int max_diff = TestBlur(benchmark_width_ - 1, benchmark_height_, benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_, @@ -2288,7 +2087,7 @@ TEST_F(LibYUVPlanarTest, TestARGBLumaColorTable) { SIMD_ALIGNED(uint8 dst_pixels_c[1280][4]); memset(orig_pixels, 0, sizeof(orig_pixels)); - align_buffer_64(lumacolortable, 32768); + align_buffer_page_end(lumacolortable, 32768); int v = 0; for (int i = 0; i < 32768; ++i) { lumacolortable[i] = v; @@ -2357,14 +2156,14 @@ TEST_F(LibYUVPlanarTest, TestARGBLumaColorTable) { EXPECT_EQ(dst_pixels_c[i][3], dst_pixels_opt[i][3]); } - free_aligned_buffer_64(lumacolortable); + free_aligned_buffer_page_end(lumacolortable); } TEST_F(LibYUVPlanarTest, TestARGBCopyAlpha) { const int kSize = benchmark_width_ * benchmark_height_ * 4; - align_buffer_64(orig_pixels, kSize); - align_buffer_64(dst_pixels_opt, kSize); - align_buffer_64(dst_pixels_c, kSize); + align_buffer_page_end(orig_pixels, kSize); + align_buffer_page_end(dst_pixels_opt, kSize); + align_buffer_page_end(dst_pixels_c, kSize); MemRandomize(orig_pixels, kSize); MemRandomize(dst_pixels_opt, kSize); @@ -2385,16 +2184,16 @@ TEST_F(LibYUVPlanarTest, TestARGBCopyAlpha) { EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]); } - free_aligned_buffer_64(dst_pixels_c); - free_aligned_buffer_64(dst_pixels_opt); - free_aligned_buffer_64(orig_pixels); + free_aligned_buffer_page_end(dst_pixels_c); + free_aligned_buffer_page_end(dst_pixels_opt); + free_aligned_buffer_page_end(orig_pixels); } TEST_F(LibYUVPlanarTest, TestARGBExtractAlpha) { const int kPixels = benchmark_width_ * benchmark_height_; - align_buffer_64(src_pixels, kPixels * 4); - align_buffer_64(dst_pixels_opt, kPixels); - align_buffer_64(dst_pixels_c, kPixels); + align_buffer_page_end(src_pixels, kPixels * 4); + align_buffer_page_end(dst_pixels_opt, kPixels); + align_buffer_page_end(dst_pixels_c, kPixels); MemRandomize(src_pixels, kPixels * 4); MemRandomize(dst_pixels_opt, kPixels); @@ -2415,16 +2214,16 @@ TEST_F(LibYUVPlanarTest, TestARGBExtractAlpha) { EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]); } - free_aligned_buffer_64(dst_pixels_c); - free_aligned_buffer_64(dst_pixels_opt); - free_aligned_buffer_64(src_pixels); + free_aligned_buffer_page_end(dst_pixels_c); + free_aligned_buffer_page_end(dst_pixels_opt); + free_aligned_buffer_page_end(src_pixels); } TEST_F(LibYUVPlanarTest, TestARGBCopyYToAlpha) { const int kPixels = benchmark_width_ * benchmark_height_; - align_buffer_64(orig_pixels, kPixels); - align_buffer_64(dst_pixels_opt, kPixels * 4); - align_buffer_64(dst_pixels_c, kPixels * 4); + align_buffer_page_end(orig_pixels, kPixels); + align_buffer_page_end(dst_pixels_opt, kPixels * 4); + align_buffer_page_end(dst_pixels_c, kPixels * 4); MemRandomize(orig_pixels, kPixels); MemRandomize(dst_pixels_opt, kPixels * 4); @@ -2445,9 +2244,9 @@ TEST_F(LibYUVPlanarTest, TestARGBCopyYToAlpha) { EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]); } - free_aligned_buffer_64(dst_pixels_c); - free_aligned_buffer_64(dst_pixels_opt); - free_aligned_buffer_64(orig_pixels); + free_aligned_buffer_page_end(dst_pixels_c); + free_aligned_buffer_page_end(dst_pixels_opt); + free_aligned_buffer_page_end(orig_pixels); } static int TestARGBRect(int width, int height, int benchmark_iterations, @@ -2460,8 +2259,8 @@ static int TestARGBRect(int width, int height, int benchmark_iterations, const int kSize = kStride * height; const uint32 v32 = fastrand() & (bpp == 4 ? 0xffffffff : 0xff); - align_buffer_64(dst_argb_c, kSize + off); - align_buffer_64(dst_argb_opt, kSize + off); + align_buffer_page_end(dst_argb_c, kSize + off); + align_buffer_page_end(dst_argb_opt, kSize + off); MemRandomize(dst_argb_c + off, kSize); memcpy(dst_argb_opt + off, dst_argb_c + off, kSize); @@ -2490,8 +2289,8 @@ static int TestARGBRect(int width, int height, int benchmark_iterations, max_diff = abs_diff; } } - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); return max_diff; } diff --git a/unit_test/rotate_argb_test.cc b/unit_test/rotate_argb_test.cc index 24640800..9c83c356 100644 --- a/unit_test/rotate_argb_test.cc +++ b/unit_test/rotate_argb_test.cc @@ -12,7 +12,6 @@ #include "libyuv/cpu_id.h" #include "libyuv/rotate_argb.h" -#include "libyuv/row.h" #include "../unit_test/unit_test.h" namespace libyuv { @@ -38,15 +37,15 @@ void TestRotateBpp(int src_width, int src_height, } int src_stride_argb = src_width * kBpp; int src_argb_plane_size = src_stride_argb * abs(src_height); - align_buffer_64(src_argb, src_argb_plane_size); + align_buffer_page_end(src_argb, src_argb_plane_size); for (int i = 0; i < src_argb_plane_size; ++i) { src_argb[i] = fastrand() & 0xff; } int dst_stride_argb = dst_width * kBpp; int dst_argb_plane_size = dst_stride_argb * dst_height; - align_buffer_64(dst_argb_c, dst_argb_plane_size); - align_buffer_64(dst_argb_opt, dst_argb_plane_size); + align_buffer_page_end(dst_argb_c, dst_argb_plane_size); + align_buffer_page_end(dst_argb_opt, dst_argb_plane_size); memset(dst_argb_c, 2, dst_argb_plane_size); memset(dst_argb_opt, 3, dst_argb_plane_size); @@ -81,9 +80,9 @@ void TestRotateBpp(int src_width, int src_height, EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]); } - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); - free_aligned_buffer_64(src_argb); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); + free_aligned_buffer_page_end(src_argb); } static void ARGBTestRotate(int src_width, int src_height, diff --git a/unit_test/rotate_test.cc b/unit_test/rotate_test.cc index 1f5b86e9..07e2f73a 100644 --- a/unit_test/rotate_test.cc +++ b/unit_test/rotate_test.cc @@ -12,7 +12,6 @@ #include "libyuv/cpu_id.h" #include "libyuv/rotate.h" -#include "libyuv/row.h" #include "../unit_test/unit_test.h" namespace libyuv { @@ -37,7 +36,7 @@ static void I420TestRotate(int src_width, int src_height, int src_i420_y_size = src_width * Abs(src_height); int src_i420_uv_size = ((src_width + 1) / 2) * ((Abs(src_height) + 1) / 2); int src_i420_size = src_i420_y_size + src_i420_uv_size * 2; - align_buffer_64(src_i420, src_i420_size); + align_buffer_page_end(src_i420, src_i420_size); for (int i = 0; i < src_i420_size; ++i) { src_i420[i] = fastrand() & 0xff; } @@ -45,8 +44,8 @@ static void I420TestRotate(int src_width, int src_height, int dst_i420_y_size = dst_width * dst_height; int dst_i420_uv_size = ((dst_width + 1) / 2) * ((dst_height + 1) / 2); int dst_i420_size = dst_i420_y_size + dst_i420_uv_size * 2; - align_buffer_64(dst_i420_c, dst_i420_size); - align_buffer_64(dst_i420_opt, dst_i420_size); + align_buffer_page_end(dst_i420_c, dst_i420_size); + align_buffer_page_end(dst_i420_opt, dst_i420_size); memset(dst_i420_c, 2, dst_i420_size); memset(dst_i420_opt, 3, dst_i420_size); @@ -78,9 +77,9 @@ static void I420TestRotate(int src_width, int src_height, EXPECT_EQ(dst_i420_c[i], dst_i420_opt[i]); } - free_aligned_buffer_64(dst_i420_c); - free_aligned_buffer_64(dst_i420_opt); - free_aligned_buffer_64(src_i420); + free_aligned_buffer_page_end(dst_i420_c); + free_aligned_buffer_page_end(dst_i420_opt); + free_aligned_buffer_page_end(src_i420); } TEST_F(LibYUVRotateTest, I420Rotate0_Opt) { @@ -163,7 +162,7 @@ static void NV12TestRotate(int src_width, int src_height, int src_nv12_uv_size = ((src_width + 1) / 2) * ((Abs(src_height) + 1) / 2) * 2; int src_nv12_size = src_nv12_y_size + src_nv12_uv_size; - align_buffer_64(src_nv12, src_nv12_size); + align_buffer_page_end(src_nv12, src_nv12_size); for (int i = 0; i < src_nv12_size; ++i) { src_nv12[i] = fastrand() & 0xff; } @@ -171,8 +170,8 @@ static void NV12TestRotate(int src_width, int src_height, int dst_i420_y_size = dst_width * dst_height; int dst_i420_uv_size = ((dst_width + 1) / 2) * ((dst_height + 1) / 2); int dst_i420_size = dst_i420_y_size + dst_i420_uv_size * 2; - align_buffer_64(dst_i420_c, dst_i420_size); - align_buffer_64(dst_i420_opt, dst_i420_size); + align_buffer_page_end(dst_i420_c, dst_i420_size); + align_buffer_page_end(dst_i420_opt, dst_i420_size); memset(dst_i420_c, 2, dst_i420_size); memset(dst_i420_opt, 3, dst_i420_size); @@ -201,9 +200,9 @@ static void NV12TestRotate(int src_width, int src_height, EXPECT_EQ(dst_i420_c[i], dst_i420_opt[i]); } - free_aligned_buffer_64(dst_i420_c); - free_aligned_buffer_64(dst_i420_opt); - free_aligned_buffer_64(src_nv12); + free_aligned_buffer_page_end(dst_i420_c); + free_aligned_buffer_page_end(dst_i420_opt); + free_aligned_buffer_page_end(src_nv12); } TEST_F(LibYUVRotateTest, NV12Rotate0_Opt) { diff --git a/unit_test/scale_argb_test.cc b/unit_test/scale_argb_test.cc index d470ff1d..f99782f7 100644 --- a/unit_test/scale_argb_test.cc +++ b/unit_test/scale_argb_test.cc @@ -13,7 +13,6 @@ #include "libyuv/convert_argb.h" #include "libyuv/cpu_id.h" -#include "libyuv/row.h" // For align_buffer_64 #include "libyuv/scale_argb.h" #include "libyuv/video_common.h" #include "../unit_test/unit_test.h" @@ -156,7 +155,7 @@ static int ARGBClipTestFilter(int src_width, int src_height, (Abs(src_height) + b * 2) * 4; int src_stride_argb = (b * 2 + Abs(src_width)) * 4; - align_buffer_64(src_argb, src_argb_plane_size); + align_buffer_page_end(src_argb, src_argb_plane_size); if (!src_argb) { printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n"); return 0; @@ -173,8 +172,8 @@ static int ARGBClipTestFilter(int src_width, int src_height, } } - align_buffer_64(dst_argb_c, dst_argb_plane_size); - align_buffer_64(dst_argb_opt, dst_argb_plane_size); + align_buffer_page_end(dst_argb_c, dst_argb_plane_size); + align_buffer_page_end(dst_argb_opt, dst_argb_plane_size); if (!dst_argb_c || !dst_argb_opt) { printf("Skipped. Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n"); return 0; @@ -216,9 +215,9 @@ static int ARGBClipTestFilter(int src_width, int src_height, } } - free_aligned_buffer_64(dst_argb_c); - free_aligned_buffer_64(dst_argb_opt); - free_aligned_buffer_64(src_argb); + free_aligned_buffer_page_end(dst_argb_c); + free_aligned_buffer_page_end(dst_argb_opt); + free_aligned_buffer_page_end(src_argb); return max_diff; } @@ -322,8 +321,7 @@ int YUVToARGBScaleReference2(const uint8* src_y, int src_stride_y, int clip_x, int clip_y, int clip_width, int clip_height, enum FilterMode filtering) { - - uint8* argb_buffer = (uint8*)malloc(src_width * src_height * 4); + uint8* argb_buffer = static_cast<uint8*>(malloc(src_width * src_height * 4)); int r; I420ToARGB(src_y, src_stride_y, src_u, src_stride_u, diff --git a/unit_test/unit_test.h b/unit_test/unit_test.h index 2d62b1e7..51e938df 100644 --- a/unit_test/unit_test.h +++ b/unit_test/unit_test.h @@ -91,6 +91,16 @@ static inline double get_time() { } #endif +#ifndef SIMD_ALIGNED +#if defined(_MSC_VER) && !defined(__CLR_VER) +#define SIMD_ALIGNED(var) __declspec(align(16)) var +#elif defined(__GNUC__) && !defined(__pnacl__) +#define SIMD_ALIGNED(var) var __attribute__((aligned(16))) +#else +#define SIMD_ALIGNED(var) var +#endif +#endif + extern unsigned int fastrand_seed; inline int fastrand() { fastrand_seed = fastrand_seed * 214013u + 2531011u; |