michael@0: /*
michael@0:  *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS. All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: #include <stdlib.h>
michael@0: #include <time.h>
michael@0: 
michael@0: #include "libyuv/cpu_id.h"
michael@0: #include "libyuv/rotate.h"
michael@0: #include "libyuv/row.h"
michael@0: #include "../unit_test/unit_test.h"
michael@0: 
michael@0: namespace libyuv {
michael@0: 
michael@0: static void I420TestRotate(int src_width, int src_height,
michael@0:                            int dst_width, int dst_height,
michael@0:                            libyuv::RotationMode mode,
michael@0:                            int benchmark_iterations) {
michael@0:   if (src_width < 1) {
michael@0:     src_width = 1;
michael@0:   }
michael@0:   if (src_height < 1) {
michael@0:     src_height = 1;
michael@0:   }
michael@0:   if (dst_width < 1) {
michael@0:     dst_width = 1;
michael@0:   }
michael@0:   if (dst_height < 1) {
michael@0:     dst_height = 1;
michael@0:   }
michael@0:   int src_i420_y_size = src_width * src_height;
michael@0:   int src_i420_uv_size = ((src_width + 1) / 2) * ((src_height + 1) / 2);
michael@0:   int src_i420_size = src_i420_y_size + src_i420_uv_size * 2;
michael@0:   align_buffer_64(src_i420, src_i420_size);
michael@0:   for (int i = 0; i < src_i420_size; ++i) {
michael@0:     src_i420[i] = random() & 0xff;
michael@0:   }
michael@0: 
michael@0:   int dst_i420_y_size = dst_width * dst_height;
michael@0:   int dst_i420_uv_size = ((dst_width + 1) / 2) * ((dst_height + 1) / 2);
michael@0:   int dst_i420_size = dst_i420_y_size + dst_i420_uv_size * 2;
michael@0:   align_buffer_64(dst_i420_c, dst_i420_size);
michael@0:   align_buffer_64(dst_i420_opt, dst_i420_size);
michael@0:   memset(dst_i420_c, 2, dst_i420_size);
michael@0:   memset(dst_i420_opt, 3, dst_i420_size);
michael@0: 
michael@0:   MaskCpuFlags(0);  // Disable all CPU optimization.
michael@0:   I420Rotate(src_i420, src_width,
michael@0:              src_i420 + src_i420_y_size, (src_width + 1) / 2,
michael@0:              src_i420 + src_i420_y_size + src_i420_uv_size, (src_width + 1) / 2,
michael@0:              dst_i420_c, dst_width,
michael@0:              dst_i420_c + dst_i420_y_size, (dst_width + 1) / 2,
michael@0:              dst_i420_c + dst_i420_y_size + dst_i420_uv_size,
michael@0:                (dst_width + 1) / 2,
michael@0:              src_width, src_height, mode);
michael@0: 
michael@0:   MaskCpuFlags(-1);  // Enable all CPU optimization.
michael@0:   for (int i = 0; i < benchmark_iterations; ++i) {
michael@0:     I420Rotate(src_i420, src_width,
michael@0:                src_i420 + src_i420_y_size, (src_width + 1) / 2,
michael@0:                src_i420 + src_i420_y_size + src_i420_uv_size,
michael@0:                  (src_width + 1) / 2,
michael@0:                dst_i420_opt, dst_width,
michael@0:                dst_i420_opt + dst_i420_y_size, (dst_width + 1) / 2,
michael@0:                dst_i420_opt + dst_i420_y_size + dst_i420_uv_size,
michael@0:                  (dst_width + 1) / 2,
michael@0:                src_width, src_height, mode);
michael@0:   }
michael@0: 
michael@0:   // Rotation should be exact.
michael@0:   for (int i = 0; i < dst_i420_size; ++i) {
michael@0:     EXPECT_EQ(dst_i420_c[i], dst_i420_opt[i]);
michael@0:   }
michael@0: 
michael@0:   free_aligned_buffer_64(dst_i420_c);
michael@0:   free_aligned_buffer_64(dst_i420_opt);
michael@0:   free_aligned_buffer_64(src_i420);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate0) {
michael@0:   I420TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_width_, benchmark_height_,
michael@0:                  kRotate0, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate90) {
michael@0:   I420TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_height_, benchmark_width_,
michael@0:                  kRotate90, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate180) {
michael@0:   I420TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_width_, benchmark_height_,
michael@0:                  kRotate180, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate270) {
michael@0:   I420TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_height_, benchmark_width_,
michael@0:                  kRotate270, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate0_Odd) {
michael@0:   I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  kRotate0, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate90_Odd) {
michael@0:   I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_height_ - 1, benchmark_width_ - 3,
michael@0:                  kRotate90, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate180_Odd) {
michael@0:   I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  kRotate180, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, I420Rotate270_Odd) {
michael@0:   I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_height_ - 1, benchmark_width_ - 3,
michael@0:                  kRotate270, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: static void NV12TestRotate(int src_width, int src_height,
michael@0:                            int dst_width, int dst_height,
michael@0:                            libyuv::RotationMode mode,
michael@0:                            int benchmark_iterations) {
michael@0:   if (src_width < 1) {
michael@0:     src_width = 1;
michael@0:   }
michael@0:   if (src_height < 1) {
michael@0:     src_height = 1;
michael@0:   }
michael@0:   if (dst_width < 1) {
michael@0:     dst_width = 1;
michael@0:   }
michael@0:   if (dst_height < 1) {
michael@0:     dst_height = 1;
michael@0:   }
michael@0:   int src_nv12_y_size = src_width * src_height;
michael@0:   int src_nv12_uv_size = ((src_width + 1) / 2) * ((src_height + 1) / 2) * 2;
michael@0:   int src_nv12_size = src_nv12_y_size + src_nv12_uv_size;
michael@0:   align_buffer_64(src_nv12, src_nv12_size);
michael@0:   for (int i = 0; i < src_nv12_size; ++i) {
michael@0:     src_nv12[i] = random() & 0xff;
michael@0:   }
michael@0: 
michael@0:   int dst_i420_y_size = dst_width * dst_height;
michael@0:   int dst_i420_uv_size = ((dst_width + 1) / 2) * ((dst_height + 1) / 2);
michael@0:   int dst_i420_size = dst_i420_y_size + dst_i420_uv_size * 2;
michael@0:   align_buffer_64(dst_i420_c, dst_i420_size);
michael@0:   align_buffer_64(dst_i420_opt, dst_i420_size);
michael@0:   memset(dst_i420_c, 2, dst_i420_size);
michael@0:   memset(dst_i420_opt, 3, dst_i420_size);
michael@0: 
michael@0:   MaskCpuFlags(0);  // Disable all CPU optimization.
michael@0:   NV12ToI420Rotate(src_nv12, src_width,
michael@0:                    src_nv12 + src_nv12_y_size, (src_width + 1) & ~1,
michael@0:                    dst_i420_c, dst_width,
michael@0:                    dst_i420_c + dst_i420_y_size, (dst_width + 1) / 2,
michael@0:                    dst_i420_c + dst_i420_y_size + dst_i420_uv_size,
michael@0:                      (dst_width + 1) / 2,
michael@0:                    src_width, src_height, mode);
michael@0: 
michael@0:   MaskCpuFlags(-1);  // Enable all CPU optimization.
michael@0:   for (int i = 0; i < benchmark_iterations; ++i) {
michael@0:     NV12ToI420Rotate(src_nv12, src_width,
michael@0:                      src_nv12 + src_nv12_y_size, (src_width + 1) & ~1,
michael@0:                      dst_i420_opt, dst_width,
michael@0:                      dst_i420_opt + dst_i420_y_size, (dst_width + 1) / 2,
michael@0:                      dst_i420_opt + dst_i420_y_size + dst_i420_uv_size,
michael@0:                        (dst_width + 1) / 2,
michael@0:                      src_width, src_height, mode);
michael@0:   }
michael@0: 
michael@0:   // Rotation should be exact.
michael@0:   for (int i = 0; i < dst_i420_size; ++i) {
michael@0:     EXPECT_EQ(dst_i420_c[i], dst_i420_opt[i]);
michael@0:   }
michael@0: 
michael@0:   free_aligned_buffer_64(dst_i420_c);
michael@0:   free_aligned_buffer_64(dst_i420_opt);
michael@0:   free_aligned_buffer_64(src_nv12);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate0) {
michael@0:   NV12TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_width_, benchmark_height_,
michael@0:                  kRotate0, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate90) {
michael@0:   NV12TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_height_, benchmark_width_,
michael@0:                  kRotate90, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate180) {
michael@0:   NV12TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_width_, benchmark_height_,
michael@0:                  kRotate180, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate270) {
michael@0:   NV12TestRotate(benchmark_width_, benchmark_height_,
michael@0:                  benchmark_height_, benchmark_width_,
michael@0:                  kRotate270, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate0_Odd) {
michael@0:   NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  kRotate0, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate90_Odd) {
michael@0:   NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_height_ - 1, benchmark_width_ - 3,
michael@0:                  kRotate90, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate180_Odd) {
michael@0:   NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  kRotate180, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: TEST_F(libyuvTest, NV12Rotate270_Odd) {
michael@0:   NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
michael@0:                  benchmark_height_ - 1, benchmark_width_ - 3,
michael@0:                  kRotate270, benchmark_iterations_);
michael@0: }
michael@0: 
michael@0: }  // namespace libyuv