aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add scatter_int16

Browse Source
This commit is contained in:
Anton Mitrokhin
2014-12-24 19:15:14 +03:00
parent 515c5f76e0
commit 83b65ef534
1 changed files with 13 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
examples/intrinsics/knc.h
View File

@@ -3019,6 +3019,18 @@ __scatter_base_offsets32_i8(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
} }
static FORCEINLINE void
__scatter_base_offsets32_i16(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
__vec16_i16 val, __vec16_i1 mask)
{
__vec16_i32 tmp = _mm512_extload_epi32(&val,_MM_UPCONV_EPI32_SINT16,
_MM_BROADCAST32_NONE, _MM_HINT_NONE);
_mm512_mask_i32extscatter_epi32(b, mask, offsets, tmp,
_MM_DOWNCONV_EPI32_SINT16, scale,
_MM_HINT_NONE);
}
static FORCEINLINE void __masked_store_i16(void *p, const __vec16_i16 &val, __vec16_i1 mask) { static FORCEINLINE void __masked_store_i16(void *p, const __vec16_i16 &val, __vec16_i1 mask) {
__vec16_i32 tmp = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_SINT16, _MM_BROADCAST32_NONE, _MM_HINT_NONE); __vec16_i32 tmp = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_SINT16, _MM_BROADCAST32_NONE, _MM_HINT_NONE);
#ifdef ISPC_FORCE_ALIGNED_MEMORY #ifdef ISPC_FORCE_ALIGNED_MEMORY
@@ -3485,16 +3497,6 @@ static FORCEINLINE void __scatter64_float(__vec16_i64 ptrs, __vec16_f val, __vec
const __mmask8 mask_hi = 0x00FF & (mask >> 8); const __mmask8 mask_hi = 0x00FF & (mask >> 8);
_mm512_mask_i64scatter_pslo (0, mask_hi, second8ptrs, _mm512_permute4f128_ps(val.v, _MM_PERM_DCDC), 1); _mm512_mask_i64scatter_pslo (0, mask_hi, second8ptrs, _mm512_permute4f128_ps(val.v, _MM_PERM_DCDC), 1);
} }
/*
static FORCEINLINE void __scatter64_double(__vec16_i64 ptrs, __vec16_d val, __vec16_i1 mask) {
}
static FORCEINLINE void __scatter64_i8(__vec16_i64 ptrs, __vec16_i8 val, __vec16_i1 mask) {
}
static FORCEINLINE void __scatter64_i16(__vec16_i64 ptrs, __vec16_i16 val, __vec16_i1 mask) {
}
*/
static FORCEINLINE void __scatter64_i32(__vec16_i64 ptrs, __vec16_i32 val, __vec16_i1 mask) { static FORCEINLINE void __scatter64_i32(__vec16_i64 ptrs, __vec16_i32 val, __vec16_i1 mask) {
__vec16_i32 first8ptrs, second8ptrs; __vec16_i32 first8ptrs, second8ptrs;