added simple SSE support for convolutional layer, very early

liuliu · liuliu · commit 7ac4d634599d · 2014-04-28T02:07:16.000-04:00
diff --git a/lib/.ycm_extra_conf.py b/lib/.ycm_extra_conf.py
@@ -4,6 +4,6 @@
 
 def FlagsForFile(filename):
 	return {
-		'flags' : ['-ffast-math', '-Wall', '-msse2', '-D HAVE_SSE2', '-D HAVE_LIBJPEG', '-D HAVE_LIBPNG', '-D HAVE_GSL', '-D HAVE_FFTW3', '-D HAVE_LIBLINEAR', '-D HAVE_CBLAS', '-D HAVE_AVCODEC', '-D HAVE_AVFORMAT', '-D HAVE_SWSCALE'],
+		'flags' : ['-ffast-math', '-Wall', '-msse3', '-D HAVE_SSE3', '-D HAVE_LIBJPEG', '-D HAVE_LIBPNG', '-D HAVE_GSL', '-D HAVE_FFTW3', '-D HAVE_LIBLINEAR', '-D HAVE_CBLAS', '-D HAVE_AVCODEC', '-D HAVE_AVFORMAT', '-D HAVE_SWSCALE'],
 		'do_cache' : True
 	}
diff --git a/lib/3rdparty/dsfmt/dSFMT-common.h b/lib/3rdparty/dsfmt/dSFMT-common.h
@@ -23,7 +23,7 @@
 
 #include "dSFMT.h"
 
-#if defined(HAVE_SSE2)
+#if defined(HAVE_SSE3)
 #  include <emmintrin.h>
 union X128I_T {
     uint64_t u[2];
@@ -66,7 +66,7 @@ inline static void do_recursion(dw128_t *r, dw128_t *a, dw128_t * b,
     r->s = vec_xor(z, x);
     lung->s = w;
 }
-#elif defined(HAVE_SSE2)
+#elif defined(HAVE_SSE3)
 /**
  * This function represents the recursion formula.
  * @param r output 128-bit
diff --git a/lib/3rdparty/dsfmt/dSFMT-params.h b/lib/3rdparty/dsfmt/dSFMT-params.h
@@ -35,7 +35,7 @@
 #define DSFMT_SR	12
 
 /* for sse2 */
-#if defined(HAVE_SSE2)
+#if defined(HAVE_SSE3)
   #define SSE2_SHUFF 0x1b
 #elif defined(HAVE_ALTIVEC)
   #if defined(__APPLE__)  /* For OSX */
diff --git a/lib/3rdparty/dsfmt/dSFMT.c b/lib/3rdparty/dsfmt/dSFMT.c
@@ -43,7 +43,7 @@ inline static int idxof(int i);
 static void initial_mask(dsfmt_t *dsfmt);
 static void period_certification(dsfmt_t *dsfmt);
 
-#if defined(HAVE_SSE2)
+#if defined(HAVE_SSE3)
 /** 1 in 64bit for sse2 */
 static const union X128I_T sse2_int_one = {{1, 1}};
 /** 2.0 double for sse2 */
@@ -66,7 +66,7 @@ inline static int idxof(int i) {
 }
 #endif
 
-#if defined(HAVE_SSE2)
+#if defined(HAVE_SSE3)
 /**
  * This function converts the double precision floating point numbers which
  * distribute uniformly in the range [1, 2) to those which distribute uniformly
diff --git a/lib/3rdparty/dsfmt/dSFMT.h b/lib/3rdparty/dsfmt/dSFMT.h
@@ -138,7 +138,7 @@ union dW128_T {
     double d[2];
 };
 
-#elif defined(HAVE_SSE2)
+#elif defined(HAVE_SSE3)
 #  include <emmintrin.h>
 
 /** 128-bit data structure */
diff --git a/lib/3rdparty/sfmt/SFMT.c b/lib/3rdparty/sfmt/SFMT.c
@@ -41,7 +41,7 @@ extern "C" {
 /**
  * parameters used by sse2.
  */
-#ifdef HAVE_SSE2
+#ifdef HAVE_SSE3
 static const w128_t sse2_param_mask = {{SFMT_MSK1, SFMT_MSK2,
 					SFMT_MSK3, SFMT_MSK4}};
 #endif
@@ -59,7 +59,7 @@ inline static void swap(w128_t *array, int size);
 
 #if defined(HAVE_ALTIVEC)
   #include "SFMT-alti.h"
-#elif defined(HAVE_SSE2)
+#elif defined(HAVE_SSE3)
   #include "SFMT-sse2.h"
 #endif
 
@@ -77,7 +77,7 @@ inline static int idxof(int i) {
 }
 #endif
 
-#if (!defined(HAVE_ALTIVEC)) && (!defined(HAVE_SSE2))
+#if (!defined(HAVE_ALTIVEC)) && (!defined(HAVE_SSE3))
 /**
  * This function fills the user-specified array with pseudorandom
  * integers.
@@ -228,7 +228,7 @@ int sfmt_get_min_array_size64(sfmt_t * sfmt) {
     return SFMT_N64;
 }
 
-#if !defined(HAVE_SSE2) && !defined(HAVE_ALTIVEC)
+#if !defined(HAVE_SSE3) && !defined(HAVE_ALTIVEC)
 /**
  * This function fills the internal state array with pseudorandom
  * integers.
diff --git a/lib/3rdparty/sfmt/SFMT.h b/lib/3rdparty/sfmt/SFMT.h
@@ -79,7 +79,7 @@ union W128_T {
     uint32_t u[4];
     uint64_t u64[2];
 };
-#elif defined(HAVE_SSE2)
+#elif defined(HAVE_SSE3)
   #include <emmintrin.h>
 
 /** 128-bit data structure */
diff --git a/lib/ccv.h b/lib/ccv.h
@@ -14,8 +14,10 @@
 #include <string.h>
 #include <float.h>
 #include <math.h>
-#ifdef HAVE_SSE2
-#include <xmmintrin.h>
+#ifndef __CUDACC__
+#ifdef HAVE_SSE3
+#include <x86intrin.h>
+#endif
 #endif
 #include <assert.h>
 #include <alloca.h>
diff --git a/lib/ccv_basic.c b/lib/ccv_basic.c
@@ -221,7 +221,7 @@ static void _ccv_atan2(float* x, float* y, float* angle, float* mag, int len)
 {
 	int i = 0;
 	float scale = (float)(180.0 / CCV_PI);
-#ifdef HAVE_SSE2
+#ifdef HAVE_SSE3
 #ifndef _WIN32
 	union { int i; float fl; } iabsmask; iabsmask.i = 0x7fffffff;
 	__m128 eps = _mm_set1_ps((float)1e-6), absmask = _mm_set1_ps(iabsmask.fl);
diff --git a/lib/ccv_convnet.c b/lib/ccv_convnet.c
@@ -152,14 +152,35 @@ static void _ccv_convnet_convolutional_forward_propagate(ccv_convnet_layer_t* la
 				float* w = layer_w + comx * ch_per_partition + comy;
 				float* apz = ap + ccv_max(j * strides - border, 0) * ch;
 				// when we have border, we simply do zero padding
+#if HAVE_SSE3
+				__m128 v4 = _mm_setzero_ps();
+#endif
 				for (y = 0; y < maxy; y++)
 				{
 					for (x = 0; x < maxx; x++)
-						for (c = 0; c < ch_per_partition; c++)
+					{
+						c = 0;
+#if HAVE_SSE3
+						for (; c < ch_per_partition - 4; c += 4)
+						{
+							__m128 w4 = _mm_loadu_ps(w + x * ch_per_partition + c);
+							__m128 apz4 = _mm_loadu_ps(apz + x * ch + c);
+							v4 = _mm_add_ps(_mm_mul_ps(w4, apz4), v4);
+						}
+#endif
+						for (; c < ch_per_partition; c++)
 							v += w[x * ch_per_partition + c] * apz[x * ch + c];
+					}
 					w += kernel_cols * ch_per_partition;
 					apz += a->cols * ch;
 				}
+#if HAVE_SSE3
+				v4 = _mm_hadd_ps(v4, v4);
+				v4 = _mm_hadd_ps(v4, v4);
+				float pv;
+				_mm_store_ss(&pv, v4);
+				v += pv;
+#endif
 				bp[j * count] = ccv_max(0, v); // ReLU
 			}
 			bp += db->cols * count;
diff --git a/lib/configure b/lib/configure
@@ -3296,10 +3296,10 @@ fi
 done
 
 
-ac_fn_c_check_header_mongrel "$LINENO" "xmmintrin.h" "ac_cv_header_xmmintrin_h" "$ac_includes_default"
-if test "x$ac_cv_header_xmmintrin_h" = xyes; then :
-  DEFINE_MACROS="$DEFINE_MACROS-D HAVE_SSE2 "
- MKCFLAGS="$MKCFLAGS-msse2 "
+ac_fn_c_check_header_mongrel "$LINENO" "x86intrin.h" "ac_cv_header_x86intrin_h" "$ac_includes_default"
+if test "x$ac_cv_header_x86intrin_h" = xyes; then :
+  DEFINE_MACROS="$DEFINE_MACROS-D HAVE_SSE3 "
+ MKCFLAGS="$MKCFLAGS-msse3 "
 
 fi
 
diff --git a/lib/configure.ac b/lib/configure.ac
@@ -12,8 +12,8 @@ AC_CHECK_PROG(CC, clang, clang)
 AC_CHECK_PROG(NVCC, nvcc, nvcc)
 
 # check for sse2 support, libpng, libjpeg, fftw3, liblinear, cblas, Accelerate framework, avformat, avcodec, swscale
-AC_CHECK_HEADER(xmmintrin.h,
-				[AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D HAVE_SSE2 "]) AC_SUBST(MKCFLAGS, ["$MKCFLAGS-msse2 "])])
+AC_CHECK_HEADER(x86intrin.h,
+				[AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D HAVE_SSE3 "]) AC_SUBST(MKCFLAGS, ["$MKCFLAGS-msse3 "])])
 AC_CHECK_HEADER(png.h,
 				[AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D HAVE_LIBPNG "]) AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-lpng "])])
 AC_CHECK_HEADER(jpeglib.h,
diff --git a/lib/cuda/.ycm_extra_conf.py b/lib/cuda/.ycm_extra_conf.py
@@ -6,7 +6,7 @@
 	'-ffast-math',
 	'-Wall',
 	'-msse2',
-	'-D HAVE_SSE2',
+	'-D HAVE_SSE3',
 	'-D HAVE_LIBJPEG',
 	'-D HAVE_LIBPNG',
 	'-D HAVE_GSL',

Original file line number	Diff line number	Diff line change
`@@ -4,6 +4,6 @@`
`4`	`4`
`5`	`5`	`def FlagsForFile(filename):`
`6`	`6`	`return {`
`7`		`- 'flags' : ['-ffast-math', '-Wall', '-msse2', '-D HAVE_SSE2', '-D HAVE_LIBJPEG', '-D HAVE_LIBPNG', '-D HAVE_GSL', '-D HAVE_FFTW3', '-D HAVE_LIBLINEAR', '-D HAVE_CBLAS', '-D HAVE_AVCODEC', '-D HAVE_AVFORMAT', '-D HAVE_SWSCALE'],`
	`7`	`+ 'flags' : ['-ffast-math', '-Wall', '-msse3', '-D HAVE_SSE3', '-D HAVE_LIBJPEG', '-D HAVE_LIBPNG', '-D HAVE_GSL', '-D HAVE_FFTW3', '-D HAVE_LIBLINEAR', '-D HAVE_CBLAS', '-D HAVE_AVCODEC', '-D HAVE_AVFORMAT', '-D HAVE_SWSCALE'],`
`8`	`8`	`'do_cache' : True`
`9`	`9`	`}`
Original file line number	Diff line number	Diff line change
`@@ -221,7 +221,7 @@ static void _ccv_atan2(float* x, float* y, float* angle, float* mag, int len)`
`221`	`221`	`{`
`222`	`222`	`int i = 0;`
`223`	`223`	`float scale = (float)(180.0 / CCV_PI);`
`224`		`-#ifdef HAVE_SSE2`
	`224`	`+#ifdef HAVE_SSE3`
`225`	`225`	`#ifndef _WIN32`
`226`	`226`	`union { int i; float fl; } iabsmask; iabsmask.i = 0x7fffffff;`
`227`	`227`	`__m128 eps = _mm_set1_ps((float)1e-6), absmask = _mm_set1_ps(iabsmask.fl);`