changed codestyle to 2 spaces in 'knc.h'
This commit is contained in:
@@ -1862,6 +1862,7 @@ template <int ALIGN> static FORCEINLINE __vec16_i8 __load(const __vec16_i8 *p) {
|
|||||||
template <int ALIGN> static FORCEINLINE void __store(__vec16_i8 *p, __vec16_i8 v) {
|
template <int ALIGN> static FORCEINLINE void __store(__vec16_i8 *p, __vec16_i8 v) {
|
||||||
*p = v;
|
*p = v;
|
||||||
}
|
}
|
||||||
|
|
||||||
static FORCEINLINE void
|
static FORCEINLINE void
|
||||||
__scatter_base_offsets32_i8(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
|
__scatter_base_offsets32_i8(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
|
||||||
__vec16_i8 val, __vec16_i1 mask)
|
__vec16_i8 val, __vec16_i1 mask)
|
||||||
@@ -2184,8 +2185,7 @@ __scatter_base_offsets32_float(void *base, uint32_t scale, __vec16_i32 offsets,
|
|||||||
// packed load/store
|
// packed load/store
|
||||||
///////////////////////////////////////////////////////////////////////////
|
///////////////////////////////////////////////////////////////////////////
|
||||||
|
|
||||||
static FORCEINLINE int32_t __packed_load_active(uint32_t *p, __vec16_i32 *val,
|
static FORCEINLINE int32_t __packed_load_active(uint32_t *p, __vec16_i32 *val, __vec16_i1 mask) {
|
||||||
__vec16_i1 mask) {
|
|
||||||
__vec16_i32 v = __load<64>(val);
|
__vec16_i32 v = __load<64>(val);
|
||||||
v = _mm512_mask_extloadunpacklo_epi32(v, mask, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
|
v = _mm512_mask_extloadunpacklo_epi32(v, mask, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
|
||||||
v = _mm512_mask_extloadunpackhi_epi32(v, mask, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
|
v = _mm512_mask_extloadunpackhi_epi32(v, mask, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE);
|
||||||
@@ -2193,8 +2193,7 @@ static FORCEINLINE int32_t __packed_load_active(uint32_t *p, __vec16_i32 *val,
|
|||||||
return _mm_countbits_32(uint32_t(mask));
|
return _mm_countbits_32(uint32_t(mask));
|
||||||
}
|
}
|
||||||
|
|
||||||
static FORCEINLINE int32_t __packed_store_active(uint32_t *p, __vec16_i32 val,
|
static FORCEINLINE int32_t __packed_store_active(uint32_t *p, __vec16_i32 val, __vec16_i1 mask) {
|
||||||
__vec16_i1 mask) {
|
|
||||||
_mm512_mask_extpackstorelo_epi32(p, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
|
_mm512_mask_extpackstorelo_epi32(p, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
|
||||||
_mm512_mask_extpackstorehi_epi32((uint8_t*)p+64, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
|
_mm512_mask_extpackstorehi_epi32((uint8_t*)p+64, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE);
|
||||||
return _mm_countbits_32(uint32_t(mask));
|
return _mm_countbits_32(uint32_t(mask));
|
||||||
|
|||||||
Reference in New Issue
Block a user