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

'__scatter64_i64' function fixed for icc 13

Browse Source
This commit is contained in:
Anton Mitrokhin
2014-12-11 14:13:16 +04:00
parent dc2e174d4f
commit e4c79418c8
1 changed files with 26 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
examples/intrinsics/knc.h
View File

@@ -3281,6 +3281,28 @@ static FORCEINLINE void __scatter_base_offsets64_i32(uint8_t *_base, uint32_t sc
}
}
static FORCEINLINE void __scatter_base_offsets64_i64(uint8_t *_base, uint32_t scale, __vec16_i64 offsets,
__vec16_i64 value,
__vec16_i1 mask) {
const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN));
__vec16_i1 still_to_do = mask;
while (still_to_do) {
int first_active_lane = _mm_tzcnt_32((int)still_to_do);
const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane];
__vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi,
__smear_i32<__vec16_i32>((int32_t)hi32),
_MM_CMPINT_EQ);
void * base = (void*)((unsigned long)_base + ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN));
_mm512_mask_i32extscatter_epi32(base, match, signed_offsets, value.v_lo, _MM_DOWNCONV_EPI32_NONE, scale, _MM_HINT_NONE);
_mm512_mask_i32extscatter_epi32(base + sizeof(uint32_t), match, signed_offsets, value.v_hi, _MM_DOWNCONV_EPI32_NONE, scale, _MM_HINT_NONE);
still_to_do = _mm512_kxor(match,still_to_do);
}
}
static FORCEINLINE void // TODO
__scatter_base_offsets64_i8(uint8_t *_base, uint32_t scale, __vec16_i64 offsets,
__vec16_i8 value,
@@ -3362,14 +3384,14 @@ static FORCEINLINE void __scatter64_i32(__vec16_i64 ptrs, __vec16_i32 val, __vec
*/
static FORCEINLINE void __scatter64_i64(__vec16_i64 ptrs, __vec16_i64 val, __vec16_i1 mask) {
#if __INTEL_COMPILER < 1500
#warning "__scatter64_i64 is slow due to outdated compiler"
__scatter_base_offsets64_i64(0, 1, ptrs, val, mask);
#else
__vec16_i32 first8ptrs, second8ptrs;
hilo2zmm(ptrs, first8ptrs.v, second8ptrs.v);
__vec16_i32 first8vals, second8vals;
hilo2zmm(val, first8vals.v, second8vals.v);
#if __INTEL_COMPILER < 1500
#warning "__scatter64_i64 is not implemented!"
#else
_mm512_mask_i64extscatter_epi64 (0, mask, first8ptrs, first8vals, _MM_DOWNCONV_EPI64_NONE, 1, _MM_HINT_NONE);
const __mmask8 mask8 = 0x00FF & (mask >> 8);
_mm512_mask_i64extscatter_epi64 (0, mask8, second8ptrs, second8vals, _MM_DOWNCONV_EPI64_NONE, 1, _MM_HINT_NONE);