Add ARM NEON intrinsic optimizations for SincResampler.

media/base/sinc_resampler.cc

 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 //
 // Input buffer layout, dividing the total buffer into regions (r0_ - r5_):
 //
 // |----------------|-----------------------------------------|----------------|
 //
 //                                   kBlockSize + kKernelSize / 2
 //                   <--------------------------------------------------------->
@@ skipping 18 matching lines @@
 // 5) Goto (2) until all of input is consumed.
 //
 // Note: we're glossing over how the sub-sample handling works with
 // |virtual_source_idx_|, etc.
 
 // MSVC++ requires this to be set before any other includes to get M_PI.
 #define _USE_MATH_DEFINES
 
 #include "media/base/sinc_resampler.h"
 
+#include <cmath>
+
+#include "base/cpu.h"
+#include "base/logging.h"
+#include "build/build_config.h"
+
 #if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
 #include <xmmintrin.h>
 #endif
-#include <cmath>
 
-#include "base/cpu.h"
-#include "base/logging.h"
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+#include <arm_neon.h>
+#endif
 
 namespace media {
 
 enum {
   // The kernel size can be adjusted for quality (higher is better) at the
   // expense of performance.  Must be a multiple of 32.
   // TODO(dalecurtis): Test performance to see if we can jack this up to 64+.
   kKernelSize = 32,
 
   // The number of destination frames generated per processing pass.  Affects
@@ skipping 168 matching lines @@
                               const float* k2,
                               double kernel_interpolation_factor) {
   // Rely on function level static initialization to keep ConvolveProc selection
   // thread safe.
   typedef float (*ConvolveProc)(const float* src, const float* k1,
                                 const float* k2,
                                 double kernel_interpolation_factor);
 #if defined(ARCH_CPU_X86_FAMILY) && defined(__SSE__)
   static const ConvolveProc kConvolveProc =
       base::CPU().has_sse() ? Convolve_SSE : Convolve_C;
+#elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+  static const ConvolveProc kConvolveProc = Convolve_NEON;
 #else
   static const ConvolveProc kConvolveProc = Convolve_C;
 #endif
 
   return kConvolveProc(input_ptr, k1, k2, kernel_interpolation_factor);
 }
 
 float SincResampler::Convolve_C(const float* input_ptr, const float* k1,
                                 const float* k2,
                                 double kernel_interpolation_factor) {
@@ skipping 50 matching lines @@
   // Sum components together.
   float result;
   m_sums2 = _mm_add_ps(_mm_movehl_ps(m_sums1, m_sums1), m_sums1);
   _mm_store_ss(&result, _mm_add_ss(m_sums2, _mm_shuffle_ps(
       m_sums2, m_sums2, 1)));
 
   return result;
 }
 #endif
 
+#if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
+float SincResampler::Convolve_NEON(const float* input_ptr, const float* k1,
+                                   const float* k2,
+                                   double kernel_interpolation_factor) {
+  float32x4_t m_input;
+  float32x4_t m_sums1 = vmovq_n_f32(0);
+  float32x4_t m_sums2 = vmovq_n_f32(0);

[Johann, 2012/09/25 19:02:51]

For some reason it looks like it's assembling with an actual load instruction
(per the last disassembly I saw). Should be able to use VEOR as mentioned by JF.
It might be messy because I don't see one in arm_neon.h that takes float32x4

[DaleCurtis, 2012/09/25 20:30:01]

Done. Old:
 7d0: eddf 2b19   vldr  d18, [pc, #100] ; 838 <_Z13Convolve_NEONPKfS0_S0_d+0x68>
 7d4: eddf 3b1a   vldr  d19, [pc, #104] ; 840 <_Z13Convolve_NEONPKfS0_S0_d+0x70>
 7d8: ef62 61f2   vorr  q11, q9, q9
 7dc: f100 0380   add.w r3, r0, #128  ; 0x80
 7e0: eddd 8b00   vldr  d24, [sp]

New:
 7cc: 2300        movs  r3, #0
 7ce: eddd 8b00   vldr  d24, [sp]
 7d2: eea2 3b90   vdup.32 q9, r3
 7d6: f100 0380   add.w r3, r0, #128  ; 0x80
 7da: ff42 21f2   veor  q9, q9, q9
 7de: ef62 61f2   vorr  q11, q9, q9

The new code seems to be "wasting" time zeroing twice though (vdup.32 and veor),
but I suspect that's still better than two extra loads.

[DaleCurtis, 2012/09/25 21:27:01]

Actually this triggers an uninitialized error and doesn't appear to be any
faster per benchmarks:

Convolve_C took 5642.36ms.
Convolve_NEON(unaligned) took 2421.80ms; which is 2.33x faster than Convolve_C.
Convolve_NEON (aligned) took 2415.05ms; which is 2.34x faster than Convolve_C
and 1.00x faster than Convolve_NEON (unaligned).

So rather than adding a #pragma ignore for the uninitialized variable, I'll just
use vmovq and hope future compilers do something smarter.

+
+  const float* upper = input_ptr + kKernelSize;
+  for (; input_ptr < upper; ) {
+    m_input = vld1q_f32(input_ptr);
+    input_ptr += 4;
+    m_sums1 = vmlaq_f32(m_sums1, m_input, vld1q_f32(k1));
+    k1 += 4;
+    m_sums2 = vmlaq_f32(m_sums2, m_input, vld1q_f32(k2));
+    k2 += 4;
+  }
+
+  // Linearly interpolate the two "convolutions".
+  m_sums1 = vmlaq_f32(
+      vmulq_f32(m_sums1, vmovq_n_f32(1.0 - kernel_interpolation_factor)),
+      m_sums2, vmovq_n_f32(kernel_interpolation_factor));
+
+  // Sum components together.
+  float32x2_t m_half = vadd_f32(vget_high_f32(m_sums1), vget_low_f32(m_sums1));
+  return vget_lane_f32(vpadd_f32(m_half, m_half), 0);
+}
+#endif
+
 }  // namespace media