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..c7add8e3def250af2e2faf34b12258b7999babcf 100644
--- a/media/base/sinc_resampler_unittest.cc
+++ b/media/base/sinc_resampler_unittest.cc
@@ -95,7 +95,6 @@ TEST(SincResamplerTest, Flush) {
 // 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__)
 TEST(SincResamplerTest, Convolve) {
   // Initialize a dummy resampler.
   MockSource mock_source;
@@ -112,21 +111,34 @@ TEST(SincResamplerTest, Convolve) {
   double result = resampler.Convolve_C(
       resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
       resampler.kernel_storage_.get(), kKernelInterpolationFactor);
-  double result2 = resampler.Convolve_SSE(
+
+  double result2 = result;

[Ami GONE FROM CHROMIUM, 2012/09/21 04:41:41]

I don't like this b/c it'll make the test claim to pass (but be a no-op) on
other archs, and will silently regress if defines get changed, etc.

I think it's much safer to 
#else
#error Don't build this on this platform!
at l.123 (etc) and indeed only build this test on appropriate targets.

[DaleCurtis, 2012/09/24 19:54:37]

Done.

+#if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
+  result2 = resampler.Convolve_SSE(
+      resampler.kernel_storage_.get(), 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, 2012/09/21 04:41:41]

Here and elsewhere, I think __ARM_NEON__ is unnecessary given how USE_NEON comes
into being.

[DaleCurtis, 2012/09/24 19:54:37]

Done.

+  result2 = resampler.Convolve_NEON(
       resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
       resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+#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__)
   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)
+  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
 
 // Benchmark for the various Convolve() methods.  Make sure to build with
 // branding=Chrome so that DCHECKs are compiled out when benchmarking.  Original
@@ -139,7 +151,7 @@ TEST(SincResamplerTest, ConvolveBenchmark) {
       base::Bind(&MockSource::ProvideInput, base::Unretained(&mock_source)));
 
   // Retrieve benchmark iterations from command line.
-  int convolve_iterations = 10;
+  int convolve_iterations = 50000000;

[Ami GONE FROM CHROMIUM, 2012/09/21 04:41:41]

what now?

[DaleCurtis, 2012/09/24 19:54:37]

Was just for benchmarking on the try bot, but sadly the trybots don't run the
unit tests :(

   std::string iterations(CommandLine::ForCurrentProcess()->GetSwitchValueASCII(
       kConvolveIterations));
   if (!iterations.empty())
@@ -185,6 +197,33 @@ TEST(SincResamplerTest, ConvolveBenchmark) {
          " 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);
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(__ARM_NEON__) && defined(USE_NEON)

[Ami GONE FROM CHROMIUM, 2012/09/21 04:41:41]

Like with the original SSE CL, IWBN to include benchmark results in the CL
description.

[DaleCurtis, 2012/09/24 19:54:37]

Done.

+  // Benchmark Convolve_NEON() with unaligned input pointer.
+  start = base::TimeTicks::HighResNow();
+  for (int j = 0; j < convolve_iterations; ++j) {
+    resampler.Convolve_NEON(
+        resampler.kernel_storage_.get() + 1, resampler.kernel_storage_.get(),
+        resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  }
+  double total_time_sse_unaligned_ms =

[Ami GONE FROM CHROMIUM, 2012/09/22 02:38:27]

there are a bunch of variables here and below that include "sse" in the name,
probably erroneously.

You could fix that with s/sse/neon/g, but probably better to instead dedup the
analysis code against the version above (and just keep the actual dispatch in
#if's).  For that matter, you could #define your way to a single codepath
entirely:
#if ...SSE...
#define CONVOLVE_OPT Convolve_SSE
#elif ...NEON...
#define CONVOLVE_OPT Convolve_NEON
#else
#error Boom!
#endif
...

[DaleCurtis, 2012/09/24 19:54:37]

Done.

+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf("Convolve_NEON (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);
+
+  // Benchmark Convolve_NEON() with aligned input pointer.
+  start = base::TimeTicks::HighResNow();
+  for (int j = 0; j < convolve_iterations; ++j) {
+    resampler.Convolve_NEON(
+        resampler.kernel_storage_.get(), resampler.kernel_storage_.get(),
+        resampler.kernel_storage_.get(), kKernelInterpolationFactor);
+  }
+  double total_time_sse_aligned_ms =
+      (base::TimeTicks::HighResNow() - start).InMillisecondsF();
+  printf("Convolve_NEON (aligned) took %.2fms; which is %.2fx faster than"
+         " Convolve_C and %.2fx faster than Convolve_SSE (unaligned).\n",

[Ami GONE FROM CHROMIUM, 2012/09/22 02:38:27]

s/SSE/NEON/, I think :)

[DaleCurtis, 2012/09/24 19:54:37]

Done.

+         total_time_sse_aligned_ms, total_time_c_ms / total_time_sse_aligned_ms,
+         total_time_sse_unaligned_ms / total_time_sse_aligned_ms);
 #endif
 }