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. iOS lies about 22 // its architecture, so we also need to exclude it here. 23 #if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_NACL) && !defined(OS_IOS) 24 #if defined(__SSE__) 25 #define FMAC_FUNC FMAC_SSE 26 #define FMUL_FUNC FMUL_SSE 27 void Initialize() {} 28 #else 29 // X86 CPU detection required. Functions will be set by Initialize(). 30 // TODO(dalecurtis): Once Chrome moves to an SSE baseline this can be removed. 31 #define FMAC_FUNC g_fmac_proc_ 32 #define FMUL_FUNC g_fmul_proc_ 33 34 typedef void (*MathProc)(const float src[], float scale, int len, float dest[]); 35 static MathProc g_fmac_proc_ = NULL; 36 static MathProc g_fmul_proc_ = NULL; 37 38 void Initialize() { 39 CHECK(!g_fmac_proc_); 40 CHECK(!g_fmul_proc_); 41 const bool kUseSSE = base::CPU().has_sse(); 42 g_fmac_proc_ = kUseSSE ? FMAC_SSE : FMAC_C; 43 g_fmul_proc_ = kUseSSE ? FMUL_SSE : FMUL_C; 44 } 45 #endif 46 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON) 47 #define FMAC_FUNC FMAC_NEON 48 #define FMUL_FUNC FMUL_NEON 49 void Initialize() {} 50 #else 51 // Unknown architecture. 52 #define FMAC_FUNC FMAC_C 53 #define FMUL_FUNC FMUL_C 54 void Initialize() {} 55 #endif 56 57 void FMAC(const float src[], float scale, int len, float dest[]) { 58 // Ensure |src| and |dest| are 16-byte aligned. 59 DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1)); 60 DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1)); 61 return FMAC_FUNC(src, scale, len, dest); 62 } 63 64 void FMAC_C(const float src[], float scale, int len, float dest[]) { 65 for (int i = 0; i < len; ++i) 66 dest[i] += src[i] * scale; 67 } 68 69 void FMUL(const float src[], float scale, int len, float dest[]) { 70 // Ensure |src| and |dest| are 16-byte aligned. 71 DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(src) & (kRequiredAlignment - 1)); 72 DCHECK_EQ(0u, reinterpret_cast<uintptr_t>(dest) & (kRequiredAlignment - 1)); 73 return FMUL_FUNC(src, scale, len, dest); 74 } 75 76 void FMUL_C(const float src[], float scale, int len, float dest[]) { 77 for (int i = 0; i < len; ++i) 78 dest[i] = src[i] * scale; 79 } 80 81 #if defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON) 82 void FMAC_NEON(const float src[], float scale, int len, float dest[]) { 83 const int rem = len % 4; 84 const int last_index = len - rem; 85 float32x4_t m_scale = vmovq_n_f32(scale); 86 for (int i = 0; i < last_index; i += 4) { 87 vst1q_f32(dest + i, vmlaq_f32( 88 vld1q_f32(dest + i), vld1q_f32(src + i), m_scale)); 89 } 90 91 // Handle any remaining values that wouldn't fit in an NEON pass. 92 for (int i = last_index; i < len; ++i) 93 dest[i] += src[i] * scale; 94 } 95 96 void FMUL_NEON(const float src[], float scale, int len, float dest[]) { 97 const int rem = len % 4; 98 const int last_index = len - rem; 99 float32x4_t m_scale = vmovq_n_f32(scale); 100 for (int i = 0; i < last_index; i += 4) 101 vst1q_f32(dest + i, vmulq_f32(vld1q_f32(src + i), m_scale)); 102 103 // Handle any remaining values that wouldn't fit in an NEON pass. 104 for (int i = last_index; i < len; ++i) 105 dest[i] = src[i] * scale; 106 } 107 #endif 108 109 } // namespace vector_math 110 } // namespace media 111
