ref: b93e4a149c1bb77e65f4e2c66249553d2a9c4428
parent: ed9006038966d8ab059964e9bd4918e8d5180512
author: Jean-Marc Valin <jmvalin@amazon.com>
date: Mon Nov 20 21:13:06 EST 2023
Start enabling AVX2 silk_inner_product_FLP() Not yet with rtcd
--- a/Makefile.am
+++ b/Makefile.am
@@ -39,7 +39,10 @@
if HAVE_SSE4_1
SILK_SOURCES += $(SILK_SOURCES_SSE4_1)
endif
+if HAVE_AVX2
+SILK_SOURCES += $(SILK_SOURCES_FLOAT_AVX2)
endif
+endif
if DISABLE_FLOAT_API
else
@@ -427,6 +430,7 @@
if HAVE_AVX2
AVX2_OBJ = $(CELT_SOURCES_AVX2:.c=.lo) \
$(SILK_SOURCES_AVX2:.c=.lo) \
+ $(SILK_SOURCES_FLOAT_AVX2:.c=.lo) \
$(DNN_SOURCES_AVX2:.c=.lo)
$(AVX2_OBJ): CFLAGS += $(OPUS_X86_AVX2_CFLAGS)
endif
--- a/silk/float/SigProc_FLP.h
+++ b/silk/float/SigProc_FLP.h
@@ -30,6 +30,7 @@
#include "SigProc_FIX.h"
#include "float_cast.h"
+#include "main.h"
#include <math.h>
#ifdef __cplusplus
@@ -124,11 +125,16 @@
);
/* inner product of two silk_float arrays, with result as double */
-double silk_inner_product_FLP(
+double silk_inner_product_FLP_c(
const silk_float *data1,
const silk_float *data2,
opus_int dataSize
);
+
+#ifndef OVERRIDE_inner_product_FLP
+#define silk_inner_product_FLP(data1, data2, dataSize) silk_inner_product_FLP_c(data1, data2, dataSize)
+#endif
+
/* sum of squares of a silk_float array, with result as double */
double silk_energy_FLP(
--- a/silk/float/inner_product_FLP.c
+++ b/silk/float/inner_product_FLP.c
@@ -32,7 +32,7 @@
#include "SigProc_FLP.h"
/* inner product of two silk_float arrays, with result as double */
-double silk_inner_product_FLP(
+double silk_inner_product_FLP_c(
const silk_float *data1,
const silk_float *data2,
opus_int dataSize
--- a/silk/float/x86/inner_product_FLP_avx2.c
+++ b/silk/float/x86/inner_product_FLP_avx2.c
@@ -35,7 +35,7 @@
/* inner product of two silk_float arrays, with result as double */
-double silk_inner_product_FLP(
+double silk_inner_product_FLP_avx2(
const silk_float *data1,
const silk_float *data2,
opus_int dataSize
--- a/silk/x86/main_sse.h
+++ b/silk/x86/main_sse.h
@@ -269,5 +269,22 @@
# endif
+double silk_inner_product_FLP_avx2(
+ const silk_float *data1,
+ const silk_float *data2,
+ opus_int dataSize
+);
+
+#if defined (OPUS_X86_PRESUME_AVX2)
+
+#define OVERRIDE_inner_product_FLP
+#define silk_inner_product_FLP(data1, data2, dataSize) silk_inner_product_FLP_avx2(data1, data2, dataSize)
+
+#elif defined(OPUS_HAVE_RTCD) && defined(OPUS_X86_MAY_HAVE_AVX2)
+
+/*#define OVERRIDE_inner_product_FLP*/
+
+#endif
+
# endif
#endif
--- a/silk_sources.mk
+++ b/silk_sources.mk
@@ -159,3 +159,6 @@
silk/float/scale_vector_FLP.c \
silk/float/schur_FLP.c \
silk/float/sort_FLP.c
+
+SILK_SOURCES_FLOAT_AVX2 = \
+silk/float/x86/inner_product_FLP_avx2.c
--
⑨