media/base/vector_math.cc

   1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include "media/base/vector_math.h"
   6 #include "media/base/vector_math_testing.h"
   7
   8 #include "base/cpu.h"
   9 #include "base/logging.h"
  10 #include "build/build_config.h"
  11
  12 #if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
  13 #include <arm_neon.h>
  14 #endif
  15
  16 namespace media {
  17 namespace vector_math {
  18
  19 // If we know the minimum architecture at compile time, avoid CPU detection.
  20 // Force NaCl code to use C routines since (at present) nothing there uses these
  21 // methods and plumbing the -msse built library is non-trivial.
  22 #if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL)
  23 #if defined(__SSE__)
  24 #define FMAC_FUNC FMAC_SSE
  25 #define FMUL_FUNC FMUL_SSE
  26 void Initialize() {}
  27 #else
  28 // X86 CPU detection required.  Functions will be set by Initialize().
  29 // TODO(dalecurtis): Once Chrome moves to an SSE baseline this can be removed.
  30 #define FMAC_FUNC g_fmac_proc_
  31 #define FMUL_FUNC g_fmul_proc_
  32
  33 typedef void (*MathProc)(const float src[], float scale, int len, float dest[]);
  34 static MathProc g_fmac_proc_ = NULL;
  35 static MathProc g_fmul_proc_ = NULL;
  36
  37 void Initialize() {
  38   CHECK(!g_fmac_proc_);
  39   CHECK(!g_fmul_proc_);
  40   const bool kUseSSE = base::CPU().has_sse();
  41   g_fmac_proc_ = kUseSSE ? FMAC_SSE : FMAC_C;
  42   g_fmul_proc_ = kUseSSE ? FMUL_SSE : FMUL_C;
  43 }
  44 #endif
  45 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
  46 #define FMAC_FUNC FMAC_NEON
  47 #define FMUL_FUNC FMUL_NEON
  48 void Initialize() {}
  49 #else
  50 // Unknown architecture.
  51 #define FMAC_FUNC FMAC_C
  52 #define FMUL_FUNC FMUL_C
  53 void Initialize() {}
  54 #endif
  55
  56 void FMAC(const float src[], float scale, int len, float dest[]) {
  57   // Ensure |src| and |dest| are 16-byte aligned.
  58   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1));
  59   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1));
  60   return FMAC_FUNC(src, scale, len, dest);
  61 }
  62
  63 void FMAC_C(const float src[], float scale, int len, float dest[]) {
  64   for (int i = 0; i < len; ++i)
  65     dest[i] += src[i] * scale;
  66 }
  67
  68 void FMUL(const float src[], float scale, int len, float dest[]) {
  69   // Ensure |src| and |dest| are 16-byte aligned.
  70   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1));
  71   DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1));
  72   return FMUL_FUNC(src, scale, len, dest);
  73 }
  74
  75 void FMUL_C(const float src[], float scale, int len, float dest[]) {
  76   for (int i = 0; i < len; ++i)
  77     dest[i] = src[i] * scale;
  78 }
  79
  80 #if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
  81 void FMAC_NEON(const float src[], float scale, int len, float dest[]) {
  82   const int rem = len % 4;
  83   const int last_index = len - rem;
  84   float32x4_t m_scale = vmovq_n_f32(scale);
  85   for (int i = 0; i < last_index; i += 4) {
  86     vst1q_f32(dest + i, vmlaq_f32(
  87         vld1q_f32(dest + i), vld1q_f32(src + i), m_scale));
  88   }
  89
  90   // Handle any remaining values that wouldn't fit in an NEON pass.
  91   for (int i = last_index; i < len; ++i)
  92     dest[i] += src[i] * scale;
  93 }
  94
  95 void FMUL_NEON(const float src[], float scale, int len, float dest[]) {
  96   const int rem = len % 4;
  97   const int last_index = len - rem;
  98   float32x4_t m_scale = vmovq_n_f32(scale);
  99   for (int i = 0; i < last_index; i += 4)
 100     vst1q_f32(dest + i, vmulq_f32(vld1q_f32(src + i), m_scale));
 101
 102   // Handle any remaining values that wouldn't fit in an NEON pass.
 103   for (int i = last_index; i < len; ++i)
 104     dest[i] = src[i] * scale;
 105 }
 106 #endif
 107
 108 }  // namespace vector_math
 109 }  // namespace media