Add SSE2 support for Windows.

The previous defines controlling SSE2 use assumed GCC. To fix this:
- Import a chunk of defines to typedefs.h from Chrome's build_config.h, primarily to get WEBRTC_ARCH_X86_FAMILY.
- Add a check derived from WebP to define WEBRTC_USE_SSE2.
- Modify cpu_features.cc to work with MSVC. This code is borrowed from chrome/src/base/cpu.cc.
- Change AEC defines to use WEBRTC_USE_SSE2.
- Remove disable_sse2 check from aec.gyp. This is handled by WEBRTC_USE_SSE2.

(Also remove a bit of unused code from aec_core.h)
Review URL: http://webrtc-codereview.appspot.com/95008

git-svn-id: http://webrtc.googlecode.com/svn/trunk@299 4adac7df-926f-26a2-2b94-8c16560cd09d

src/modules/audio_processing/aec/main/source/aec.gyp

@@ -29,20 +29,14 @@
       'sources': [
         '../interface/echo_cancellation.h',
         'echo_cancellation.c',
+        'aec_core.h',
         'aec_core.c',
+        'aec_core_sse2.c',
         'aec_rdft.h',
         'aec_rdft.c',
-        'aec_core.h',
-        'resampler.c',
+        'aec_rdft_sse2.c',
         'resampler.h',
-      ],
-      'conditions': [
-        ['disable_sse2 == 0 and (target_arch == "ia32" or target_arch == "x64")', {
-          'sources': [
-            'aec_core_sse2.c',
-            'aec_rdft_sse2.c',
-          ],
-        }],
+        'resampler.c',
       ],
     },
   ],

src/modules/audio_processing/aec/main/source/aec_core.c

@@ -468,7 +468,7 @@ int WebRtcAec_InitAec(aec_t *aec, int sampFreq)
     WebRtcAec_FilterAdaptation = FilterAdaptation;
     WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppress;
     if (WebRtc_GetCPUInfo(kSSE2)) {
-#if defined(__SSE2__)
+#if defined(WEBRTC_USE_SSE2)
       WebRtcAec_InitAec_SSE2();
 #endif
     }
@@ -561,8 +561,8 @@ void WebRtcAec_ProcessFrame(aec_t *aec, const short *farend,
 }
 
 static void ProcessBlock(aec_t *aec, const short *farend,
-                              const short *nearend, const short *nearendH,
-                              short *output, short *outputH)
+                         const short *nearend, const short *nearendH,
+                         short *output, short *outputH)
 {
     int i;
     float d[PART_LEN], y[PART_LEN], e[PART_LEN], dH[PART_LEN];
@@ -601,7 +601,6 @@ static void ProcessBlock(aec_t *aec, const short *farend,
         }
     }
 
-
     memcpy(fft, aec->xBuf, sizeof(float) * PART_LEN2);
     memcpy(aec->dBuf + PART_LEN, d, sizeof(float) * PART_LEN);
     // For H band

src/modules/audio_processing/aec/main/source/aec_core.h

@@ -16,8 +16,9 @@
 #define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_CORE_H_
 
 #include <stdio.h>
-#include "typedefs.h"
+
 #include "signal_processing_library.h"
+#include "typedefs.h"
 
 //#define G167 // for running G167 tests
 //#define UNCONSTR // time-unconstrained filter
@@ -92,21 +93,13 @@ typedef struct {
     float dMinPow[PART_LEN1];
     float dInitMinPow[PART_LEN1];
     float *noisePow;
-#ifdef FFTW
-    float fftR[PART_LEN2];
-    fftw_complex fftC[PART_LEN2];
-    fftw_plan fftPlan, ifftPlan;
 
-    fftw_complex xfBuf[NR_PART * PART_LEN1];
-    fftw_complex wfBuf[NR_PART * PART_LEN1];
-    fftw_complex sde[PART_LEN1];
-#else
     float xfBuf[2][NR_PART * PART_LEN1]; // farend fft buffer
     float wfBuf[2][NR_PART * PART_LEN1]; // filter fft
     complex_t sde[PART_LEN1]; // cross-psd of nearend and error
     complex_t sxd[PART_LEN1]; // cross-psd of farend and nearend
     complex_t xfwBuf[NR_PART * PART_LEN1]; // farend windowed fft buffer
-#endif
+
     float sx[PART_LEN1], sd[PART_LEN1], se[PART_LEN1]; // far, near and error psd
     float hNs[PART_LEN1];
     float hNlFbMin, hNlFbLocalMin;
@@ -169,8 +162,6 @@ typedef void (*WebRtcAec_FilterFar_t)(aec_t *aec, float yf[2][PART_LEN1]);
 extern WebRtcAec_FilterFar_t WebRtcAec_FilterFar;
 typedef void (*WebRtcAec_ScaleErrorSignal_t)(aec_t *aec, float ef[2][PART_LEN1]);
 extern WebRtcAec_ScaleErrorSignal_t WebRtcAec_ScaleErrorSignal;
-#define IP_LEN PART_LEN // this must be at least ceil(2 + sqrt(PART_LEN))
-#define W_LEN PART_LEN
 typedef void (*WebRtcAec_FilterAdaptation_t)
   (aec_t *aec, float *fft, float ef[2][PART_LEN1]);
 extern WebRtcAec_FilterAdaptation_t WebRtcAec_FilterAdaptation;

src/modules/audio_processing/aec/main/source/aec_core_sse2.c

@@ -12,7 +12,9 @@
  * The core AEC algorithm, SSE2 version of speed-critical functions.
  */
 
-#if defined(__SSE2__)
+#include "typedefs.h"
+
+#if defined(WEBRTC_USE_SSE2)
 #include <emmintrin.h>
 #include <math.h>
 
@@ -210,14 +212,6 @@ static void FilterAdaptationSSE2(aec_t *aec, float *fft, float ef[2][PART_LEN1])
   }
 }
 
-#ifdef _MSC_VER /* visual c++ */
-# define ALIGN16_BEG __declspec(align(16))
-# define ALIGN16_END
-#else /* gcc or icc */
-# define ALIGN16_BEG
-# define ALIGN16_END __attribute__((aligned(16)))
-#endif
-
 static __m128 mm_pow_ps(__m128 a, __m128 b)
 {
   // a^b = exp2(b * log2(a))
@@ -432,4 +426,4 @@ void WebRtcAec_InitAec_SSE2(void) {
   WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressSSE2;
 }
 
-#endif   //__SSE2__
+#endif   // WEBRTC_USE_SSE2

src/modules/audio_processing/aec/main/source/aec_rdft.c

@@ -19,10 +19,12 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include "aec_rdft.h"
+
 #include <math.h>
 
-#include "aec_rdft.h"
 #include "system_wrappers/interface/cpu_features_wrapper.h"
+#include "typedefs.h"
 
 // constants shared by all paths (C, SSE2).
 float rdft_w[64];
@@ -571,7 +573,7 @@ void aec_rdft_init(void) {
   rftfsub_128 = rftfsub_128_C;
   rftbsub_128 = rftbsub_128_C;
   if (WebRtc_GetCPUInfo(kSSE2)) {
-#if defined(__SSE2__)
+#if defined(WEBRTC_USE_SSE2)
     aec_rdft_init_sse2();
 #endif
   }

src/modules/audio_processing/aec/main/source/aec_rdft.h

@@ -8,6 +8,9 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_RDFT_H_
+#define WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_RDFT_H_
+
 #ifdef _MSC_VER /* visual c++ */
 # define ALIGN16_BEG __declspec(align(16))
 # define ALIGN16_END
@@ -40,3 +43,5 @@ void aec_rdft_init(void);
 void aec_rdft_init_sse2(void);
 void aec_rdft_forward_128(float *a);
 void aec_rdft_inverse_128(float *a);
+
+#endif  // WEBRTC_MODULES_AUDIO_PROCESSING_AEC_MAIN_SOURCE_AEC_RDFT_H_

src/modules/audio_processing/aec/main/source/aec_rdft_sse2.c

@@ -8,7 +8,9 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#if defined(__SSE2__)
+#include "typedefs.h"
+
+#if defined(WEBRTC_USE_SSE2)
 #include <emmintrin.h>
 
 #include "aec_rdft.h"
@@ -261,4 +263,4 @@ void aec_rdft_init_sse2(void) {
   rftbsub_128 = rftbsub_128_SSE2;
 }
 
-#endif  // __SSE2__
+#endif  // WEBRTC_USE_SS2

src/system_wrappers/source/cpu_features.cc

@@ -8,17 +8,29 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+// Parts of this file derived from Chromium's base/cpu.cc.
+
 #include "cpu_features_wrapper.h"
 
+#include "typedefs.h"
+
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+#if defined(_MSC_VER)
+#include <intrin.h>
+#endif
+#endif
+
 // No CPU feature is available => straight C path.
 int GetCPUInfoNoASM(CPUFeature feature) {
   (void)feature;
   return 0;
 }
 
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+#ifndef _MSC_VER
 // Intrinsic for "cpuid".
 #if defined(__pic__) && defined(__i386__)
-static inline void cpuid(int cpu_info[4], int info_type) {
+static inline void __cpuid(int cpu_info[4], int info_type) {
   __asm__ volatile (
     "mov %%ebx, %%edi\n"
     "cpuid\n"
@@ -26,20 +38,22 @@ static inline void cpuid(int cpu_info[4], int info_type) {
     : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
     : "a"(info_type));
 }
-#elif defined(__i386__) || defined(__x86_64__)
-static inline void cpuid(int cpu_info[4], int info_type) {
+#else
+static inline void __cpuid(int cpu_info[4], int info_type) {
   __asm__ volatile (
     "cpuid\n"
     : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
     : "a"(info_type));
 }
 #endif
+#endif  // _MSC_VER
+#endif  // WEBRTC_ARCH_X86_FAMILY
 
-#if defined(__i386__) || defined(__x86_64__)
+#if defined(WEBRTC_ARCH_X86_FAMILY)
 // Actual feature detection for x86.
 static int GetCPUInfo(CPUFeature feature) {
   int cpu_info[4];
-  cpuid(cpu_info, 1);
+  __cpuid(cpu_info, 1);
   if (feature == kSSE2) {
     return 0 != (cpu_info[3] & 0x04000000);
   }

src/typedefs.h

@@ -8,21 +8,23 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-/*
- *
- * This file contains type definitions used in all WebRtc APIs.
- * 
- */
+// This file contains platform-specific typedefs and defines.
 
-/* Reserved words definitions */
+#ifndef WEBRTC_TYPEDEFS_H_
+#define WEBRTC_TYPEDEFS_H_
+
+// Reserved words definitions
 #define WEBRTC_EXTERN extern
 #define G_CONST const
 #define WEBRTC_INLINE extern __inline
 
-#ifndef WEBRTC_TYPEDEFS_H
-#define WEBRTC_TYPEDEFS_H
-
-/* Define WebRtc preprocessor identifiers based on the current build platform */
+// Define WebRTC preprocessor identifiers based on the current build platform.
+// TODO(ajm): Clean these up. We can probably remove everything in this block.
+//   - TARGET_MAC_INTEL and TARGET_MAC aren't used anywhere.
+//   - In the few places where TARGET_PC is used, it should be replaced by
+//     something more specific.
+//   - Do we really support PowerPC? Probably not. Remove WEBRTC_MAC_INTEL
+//     from build/common.gypi as well.
 #if defined(WIN32)
     // Windows & Windows Mobile
     #if !defined(WEBRTC_TARGET_PC)
@@ -33,11 +35,11 @@
     #if defined(__LITTLE_ENDIAN__ ) //TODO: is this used?
         #if !defined(WEBRTC_TARGET_MAC_INTEL)
             #define WEBRTC_TARGET_MAC_INTEL
-        #endif  
+        #endif
     #else
         #if !defined(WEBRTC_TARGET_MAC)
             #define WEBRTC_TARGET_MAC
-        #endif  
+        #endif
     #endif
 #else
     // Linux etc.
@@ -46,6 +48,40 @@
     #endif
 #endif
 
+// Derived from Chromium's build/build_config.h
+// Processor architecture detection.  For more info on what's defined, see:
+//   http://msdn.microsoft.com/en-us/library/b0084kay.aspx
+//   http://www.agner.org/optimize/calling_conventions.pdf
+//   or with gcc, run: "echo | gcc -E -dM -"
+// TODO(ajm): replace WEBRTC_LITTLE_ENDIAN with WEBRTC_ARCH_LITTLE_ENDIAN?
+#if defined(_M_X64) || defined(__x86_64__)
+#define WEBRTC_ARCH_X86_FAMILY
+#define WEBRTC_ARCH_X86_64
+#define WEBRTC_ARCH_64_BITS
+#define WEBRTC_ARCH_LITTLE_ENDIAN
+#elif defined(_M_IX86) || defined(__i386__)
+#define WEBRTC_ARCH_X86_FAMILY
+#define WEBRTC_ARCH_X86
+#define WEBRTC_ARCH_32_BITS
+#define WEBRTC_ARCH_LITTLE_ENDIAN
+#elif defined(__ARMEL__)
+// TODO(ajm): Chromium uses the two commented defines. Should we switch?
+#define WEBRTC_ARCH_ARM
+//#define WEBRTC_ARCH_ARM_FAMILY
+//#define WEBRTC_ARCH_ARMEL
+#define WEBRTC_ARCH_32_BITS
+#define WEBRTC_ARCH_LITTLE_ENDIAN
+#else
+#error Please add support for your architecture in typedefs.h
+#endif
+
+// TODO(ajm): SSE2 is disabled on Windows for the moment, because AEC
+// optimization is broken. Enable it as soon as AEC is fixed.
+//#if defined(__SSE2__) || defined(_MSC_VER)
+#if defined(__SSE2__)
+#define WEBRTC_USE_SSE2
+#endif
+
 #if defined(WEBRTC_TARGET_PC)
 
 #if !defined(_MSC_VER)
@@ -79,7 +115,7 @@
     typedef char                WebRtc_Word8;
     typedef uint8_t             WebRtc_UWord8;
 
-    /* Define endian for the platform */
+    // Define endian for the platform
     #define WEBRTC_LITTLE_ENDIAN
 
 #elif defined(WEBRTC_TARGET_MAC_INTEL)
@@ -94,14 +130,11 @@
     typedef uint16_t            WebRtc_UWord16;
     typedef uint8_t             WebRtc_UWord8;
 
-    /* Define endian for the platform */
+    // Define endian for the platform
     #define WEBRTC_LITTLE_ENDIAN
 
 #else
-
-    #error "No platform defined for WebRtc type definitions (webrtc_typedefs.h)"
-
+    #error "No platform defined for WebRTC type definitions (typedefs.h)"
 #endif
 
-
-#endif // WEBRTC_TYPEDEFS_H
+#endif  // WEBRTC_TYPEDEFS_H_