ref: 59dc75fa9713d6543bbb85fe83cb56555513a4de
parent: 1186fb8ea4f977f5d1d0816a7bc8384978966304
author: Timothy B. Terriberry <tterribe@xiph.org>
date: Thu Feb 22 00:48:18 EST 2024
Rework 32-bit SSE loads yet again. The existing code in vec_avx.h produced warning: dereferencing type-punned pointer will break strict-aliasing rules with gcc 6.4.0. We already had a macro to work around this within the rules of the C standard, but trying to use that here does not get optimized into a single MOVD like we were hoping. Replacing it with memcpy() instead does get optimized correctly, but requires switching from a macro to an inline function in order to be able to declare a local variable and return a value. We already have such an inline function in NSQ_del_dec_avx2.c, so hoist that out and use it everywhere, and then convert vec_avx.h to use it also.
--- a/celt/x86/x86cpu.h
+++ b/celt/x86/x86cpu.h
@@ -60,18 +60,33 @@
int opus_select_arch(void);
# endif
+# if defined(OPUS_X86_MAY_HAVE_SSE2)
+# include "opus_defines.h"
+
/*MOVD should not impose any alignment restrictions, but the C standard does,
and UBSan will report errors if we actually make unaligned accesses.
Use this to work around those restrictions (which should hopefully all get
- optimized to a single MOVD instruction).*/
-#define OP_LOADU_EPI32(x) \
- (int)((*(unsigned char *)(x) | *((unsigned char *)(x) + 1) << 8U |\
- *((unsigned char *)(x) + 2) << 16U | (opus_uint32)*((unsigned char *)(x) + 3) << 24U))
+ optimized to a single MOVD instruction).
+ GCC implemented _mm_loadu_si32() since GCC 11; HOWEVER, there is a bug!
+ https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99754 */
+# if !defined(_MSC_VER) && !OPUS_GNUC_PREREQ(11,3) && !(defined(__clang__) && (__clang_major__ >= 8))
+# include <string.h>
+# include <emmintrin.h>
-#define OP_CVTEPI8_EPI32_M32(x) \
- (_mm_cvtepi8_epi32(_mm_cvtsi32_si128(OP_LOADU_EPI32(x))))
+# define _mm_loadu_si32 WORKAROUND_mm_loadu_si32
+static inline __m128i WORKAROUND_mm_loadu_si32(void const* mem_addr) {+ int val;
+ memcpy(&val, mem_addr, sizeof(val));
+ return _mm_cvtsi32_si128(val);
+}
+# endif
-#define OP_CVTEPI16_EPI32_M64(x) \
+# define OP_CVTEPI8_EPI32_M32(x) \
+ (_mm_cvtepi8_epi32(_mm_loadu_si32(x)))
+
+# define OP_CVTEPI16_EPI32_M64(x) \
(_mm_cvtepi16_epi32(_mm_loadl_epi64((__m128i *)(void*)(x))))
+
+# endif
#endif
--- a/dnn/vec_avx.h
+++ b/dnn/vec_avx.h
@@ -34,8 +34,8 @@
#include <immintrin.h>
#include <math.h>
+#include "celt/x86/x86cpu.h"
-
#define MAX_INPUTS (2048)
#define USE_SU_BIAS
@@ -196,6 +196,14 @@
#define _mm256_storeu_si256(dst, src) mm256_storeu_si256(dst, src)
+static inline mm256i_emu mm256_broadcastd_epi32(__m128i x) {+ mm256i_emu ret;
+ ret.hi = ret.lo = _mm_shuffle_epi32(x, 0);
+ return ret;
+}
+#define _mm256_broadcastd_epi32(x) mm256_broadcastd_epi32(x)
+
+
static inline mm256i_emu mm256_set1_epi32(int x) {mm256i_emu ret;
ret.lo = _mm_set1_epi32(x);
@@ -786,19 +794,19 @@
{__m256i vxj;
__m256i vw;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[*idx++]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[*idx++]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[*idx++]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[*idx++]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[*idx++]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[*idx++]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[*idx++]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[*idx++]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
@@ -808,9 +816,7 @@
{__m256i vxj;
__m256i vw;
- int pos;
- pos = (*idx++);
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[pos]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[*idx++]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
@@ -837,19 +843,19 @@
{__m256i vxj;
__m256i vw;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[j]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[j]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[j+4]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[j+4]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[j+8]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[j+8]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[j+12]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[j+12]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
@@ -859,7 +865,7 @@
{__m256i vxj;
__m256i vw;
- vxj = _mm256_set1_epi32(*(int*)(void*)&x[j]);
+ vxj = _mm256_broadcastd_epi32(_mm_loadu_si32(&x[j]));
vw = _mm256_loadu_si256((const __m256i *)(void*)w);
vy0 = opus_mm256_dpbusds_epi32(vy0, vxj, vw);
w += 32;
--- a/silk/x86/NSQ_del_dec_avx2.c
+++ b/silk/x86/NSQ_del_dec_avx2.c
@@ -86,18 +86,6 @@
}
#endif
-/*
- * GCC implemented _mm_loadu_si32() since GCC 11; HOWEVER, there is a bug!
- * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99754
- */
-#if !defined(_MSC_VER) && !OPUS_GNUC_PREREQ(11,3) && !(defined(__clang__) && (__clang_major__ >= 8))
-#define _mm_loadu_si32 WORKAROUND_mm_loadu_si32
-static inline __m128i WORKAROUND_mm_loadu_si32(void const* mem_addr)
-{- return _mm_cvtsi32_si128(OP_LOADU_EPI32(mem_addr));
-}
-#endif
-
static OPUS_INLINE __m128i silk_cvtepi64_epi32_high(__m256i num)
{return _mm256_castsi256_si128(_mm256_permutevar8x32_epi32(num, _mm256_set_epi32(0, 0, 0, 0, 7, 5, 3, 1)));
--
⑨