Add ARM NEON intrinsic optimizations for SincResampler.

media/base/sinc_resampler_unittest.cc

Index: media/base/sinc_resampler_unittest.cc
diff --git a/media/base/sinc_resampler_unittest.cc b/media/base/sinc_resampler_unittest.cc
index 77a963e2ca5db9db11d1d08416abb804aac8caed..cae1d422d3eb4b8f7f6f8fa29ec56adb81ebc3fa 100644
--- a/media/base/sinc_resampler_unittest.cc
+++ b/media/base/sinc_resampler_unittest.cc
@@ -12,7 +12,9 @@
 #include "base/command_line.h"
 #include "base/logging.h"
 #include "base/string_number_conversions.h"
+#include "base/stringize_macros.h"
 #include "base/time.h"
+#include "build/build_config.h"
 #include "media/base/sinc_resampler.h"
 #include "testing/gmock/include/gmock/gmock.h"
 #include "testing/gtest/include/gtest/gtest.h"
@@ -89,13 +91,14 @@ TEST(SincResamplerTest, Flush) {
       .Times(1).WillOnce(ClearBuffer());
   resampler.Resample(resampled_destination.get(), resampler.ChunkSize() / 2);
   for (int i = 0; i < resampler.ChunkSize() / 2; ++i)
-    ASSERT_EQ(resampled_destination[i], 0);
+    ASSERT_FLOAT_EQ(resampled_destination[i], 0);
 }
 
 // Ensure various optimized Convolve() methods return the same value.  Only run
 // this test if other optimized methods exist, otherwise the default Convolve()
 // will be tested by the parameterized SincResampler tests below.
-#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+#if (defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)) || \
+    (defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON))
 TEST(SincResamplerTest, Convolve) {
   // Initialize a dummy resampler.
   MockSource mock_source;
@@ -112,18 +115,32 @@ TEST(SincResamplerTest, Convolve) {
   double result = resampler.Convolve_C(
       resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
       resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
   double result2 = resampler.Convolve_SSE(
       resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
       resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+  double result2 = resampler.Convolve_NEON(
+      resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
+      resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+#else
+#error This test should only be compiled when SSE or NEON is available.

[Ami GONE FROM CHROMIUM, 2012/09/24 20:04:16]

This'd be a lot clearer if the #if test was reversed, since then the endif could
be real close to the #if.

[Ami GONE FROM CHROMIUM, 2012/09/24 20:04:16]

I wonder if all our bots have one or the other.  I guess we'll find out!

[DaleCurtis, 2012/09/24 20:13:50]

I don't follow. Can you elaborate? You mean #if !neon && !sse #error #elif...
#elif #endif? If so, I don't think that's cleaner.

[DaleCurtis, 2012/09/24 20:13:50]

The #if || check @100 prevents any issues here :)

+#endif
   EXPECT_NEAR(result2, result, kEpsilon);
 
-  // Test Convolve_SSE() w/ unaligned input pointer.
+  // Test Convolve() w/ unaligned input pointer.
   result = resampler.Convolve_C(
       resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
       resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)

[Ami GONE FROM CHROMIUM, 2012/09/24 20:04:16]

you could avoid repeating these by defining an OPTIMIZATION_TYPE macro (SSE or
NEON) as part of the #error-triggering #if, and then use that here.

[DaleCurtis, 2012/09/24 20:22:27]

Split out into a common section for clarity.

   result2 = resampler.Convolve_SSE(
       resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
       resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(__ARM_NEON__) && defined(USE_NEON)

[Ami GONE FROM CHROMIUM, 2012/09/24 20:04:16]

drop arm_neon

[DaleCurtis, 2012/09/24 20:22:27]

Done.

+  result2 = resampler.Convolve_NEON(
+      resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
+      resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+#endif
   EXPECT_NEAR(result2, result, kEpsilon);
 }
 #endif
@@ -158,33 +175,44 @@ TEST(SincResamplerTest, ConvolveBenchmark) {
       (base::TimeTicks::HighResNow() - start).InMillisecondsF();
   printf("Convolve_C took %.2fms.\n", total_time_c_ms);
 
+#if (defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)) || \
+    (defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON))
 #if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
-  // Benchmark Convolve_SSE() with unaligned input pointer.
+#define CONVOLVE_FUNC Convolve_SSE
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#define CONVOLVE_FUNC Convolve_NEON

[Ami GONE FROM CHROMIUM, 2012/09/24 20:04:16]

Could as well go in the first #if, to save repetition?

[DaleCurtis, 2012/09/24 20:13:50]

I don't see how, the first if is a || check. We don't want to error out if SSE
and NEON are not available, just skip the extra benchmark.

+#else
+#error This benchmark should only be compiled when SSE or NEON is available.
+#endif
+  // Benchmark with unaligned input pointer.
   start = base::TimeTicks::HighResNow();
   for (int j = 0; j < convolve_iterations; ++j) {
-    resampler.Convolve_SSE(
+    resampler.CONVOLVE_FUNC(
         resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
         resampler.kernel_storage_.get(), kKernelInterpolationFactor);
   }
-  double total_time_sse_unaligned_ms =
+  double total_time_optimized_unaligned_ms =
       (base::TimeTicks::HighResNow() - start).InMillisecondsF();
-  printf("Convolve_SSE (unaligned) took %.2fms; which is %.2fx faster than"
-         " Convolve_C.\n", total_time_sse_unaligned_ms,
-         total_time_c_ms / total_time_sse_unaligned_ms);
+  printf(STRINGIZE(CONVOLVE_FUNC) "(unaligned) took %.2fms; which is %.2fx "
+         "faster than Convolve_C.\n", total_time_optimized_unaligned_ms,
+         total_time_c_ms / total_time_optimized_unaligned_ms);
 
-  // Benchmark Convolve_SSE() with aligned input pointer.
+  // Benchmark with aligned input pointer.
   start = base::TimeTicks::HighResNow();
   for (int j = 0; j < convolve_iterations; ++j) {
-    resampler.Convolve_SSE(
+    resampler.CONVOLVE_FUNC(
         resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
         resampler.kernel_storage_.get(), kKernelInterpolationFactor);
   }
-  double total_time_sse_aligned_ms =
+  double total_time_optimized_aligned_ms =
       (base::TimeTicks::HighResNow() - start).InMillisecondsF();
-  printf("Convolve_SSE (aligned) took %.2fms; which is %.2fx faster than"
-         " Convolve_C and %.2fx faster than Convolve_SSE (unaligned).\n",
-         total_time_sse_aligned_ms, total_time_c_ms / total_time_sse_aligned_ms,
-         total_time_sse_unaligned_ms / total_time_sse_aligned_ms);
+  printf(STRINGIZE(CONVOLVE_FUNC) " (aligned) took %.2fms; which is %.2fx "
+         "faster than Convolve_C and %.2fx faster than "
+         STRINGIZE(CONVOLVE_FUNC) " (unaligned).\n",
+         total_time_optimized_aligned_ms,
+         total_time_c_ms / total_time_optimized_aligned_ms,
+         total_time_optimized_unaligned_ms / total_time_optimized_aligned_ms);
+#undef CONVOLVE_FUNC
 #endif
 }