78 lines
3 KiB
Diff
78 lines
3 KiB
Diff
diff --git a/aom_dsp/x86/synonyms.h b/aom_dsp/x86/synonyms.h
|
|
--- a/aom_dsp/x86/synonyms.h
|
|
+++ b/aom_dsp/x86/synonyms.h
|
|
@@ -41,23 +41,34 @@ static inline __m128i xx_loadl_64(const
|
|
static inline __m128i xx_load_128(const void *a) {
|
|
return _mm_load_si128((const __m128i *)a);
|
|
}
|
|
|
|
static inline __m128i xx_loadu_128(const void *a) {
|
|
return _mm_loadu_si128((const __m128i *)a);
|
|
}
|
|
|
|
+// _mm_loadu_si64 has been introduced in GCC 9, reimplement the function
|
|
+// manually on older compilers.
|
|
+#if !defined(__clang__) && __GNUC_MAJOR__ < 9
|
|
+static inline __m128i xx_loadu_2x64(const void *hi, const void *lo) {
|
|
+ __m64 hi_, lo_;
|
|
+ memcpy(&hi_, hi, sizeof(hi_));
|
|
+ memcpy(&lo_, lo, sizeof(lo_));
|
|
+ return _mm_set_epi64(hi_, lo_);
|
|
+}
|
|
+#else
|
|
// Load 64 bits from each of hi and low, and pack into an SSE register
|
|
// Since directly loading as `int64_t`s and using _mm_set_epi64 may violate
|
|
// the strict aliasing rule, this takes a different approach
|
|
static inline __m128i xx_loadu_2x64(const void *hi, const void *lo) {
|
|
return _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)lo),
|
|
_mm_loadl_epi64((const __m128i *)hi));
|
|
}
|
|
+#endif
|
|
|
|
static inline void xx_storel_32(void *const a, const __m128i v) {
|
|
const int val = _mm_cvtsi128_si32(v);
|
|
memcpy(a, &val, sizeof(val));
|
|
}
|
|
|
|
static inline void xx_storel_64(void *const a, const __m128i v) {
|
|
_mm_storel_epi64((__m128i *)a, v);
|
|
diff --git a/aom_dsp/x86/synonyms_avx2.h b/aom_dsp/x86/synonyms_avx2.h
|
|
--- a/aom_dsp/x86/synonyms_avx2.h
|
|
+++ b/aom_dsp/x86/synonyms_avx2.h
|
|
@@ -71,21 +71,36 @@ static inline __m256i yy_loadu_4x64(cons
|
|
__m128d v23 = _mm_loadh_pd(v2, (const double *)e3);
|
|
// Note this can be replaced with
|
|
// `_mm256_castpd_si256(_mm256_set_m128d(v23, v01))` if immintrin.h contains
|
|
// _mm256_set_m128d() with all supported compilers. This version is used to
|
|
// match the behavior with yy_set_m128i().
|
|
return yy_set_m128i(_mm_castpd_si128(v23), _mm_castpd_si128(v01));
|
|
}
|
|
|
|
+#define GCC_VERSION (__GNUC__ * 10000 \
|
|
+ + __GNUC_MINOR__ * 100 \
|
|
+ + __GNUC_PATCHLEVEL__)
|
|
+
|
|
+// _mm256_loadu2_m128i has been introduced in GCC 10.1
|
|
+#if !defined(__clang__) && GCC_VERSION < 101000
|
|
+static inline __m256i yy_loadu2_128(const void *hi, const void *lo) {
|
|
+ __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
|
|
+ __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
|
|
+ return _mm256_set_m128i(mhi, mlo);
|
|
+}
|
|
+#else
|
|
static inline __m256i yy_loadu2_128(const void *hi, const void *lo) {
|
|
__m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
|
|
__m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
|
|
return yy_set_m128i(mhi, mlo);
|
|
}
|
|
+#endif
|
|
+
|
|
+#undef GCC_VERSION
|
|
|
|
static inline void yy_storeu2_128(void *hi, void *lo, const __m256i a) {
|
|
_mm_storeu_si128((__m128i *)hi, _mm256_extracti128_si256(a, 1));
|
|
_mm_storeu_si128((__m128i *)lo, _mm256_castsi256_si128(a));
|
|
}
|
|
|
|
static inline __m256i yy_roundn_epu16(__m256i v_val_w, int bits) {
|
|
const __m256i v_s_w = _mm256_srli_epi16(v_val_w, bits - 1);
|