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

start fixing gather/scatter functions (INT_MIN fix)

Browse Source
This commit is contained in:
Anton Mitrokhin
2014-09-18 10:20:19 +04:00
parent 8b8e313dc6
commit 1a2979aa7f
1 changed files with 93 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
examples/intrinsics/knc.h
View File

@@ -963,6 +963,71 @@ template <> FORCEINLINE void __store<128>(__vec16_i64 *p, __vec16_i64 v) {
__store<64>(p, v);
}
/*! gather vector of 64-bit ints from addresses pointing to uniform ints
(iw) WARNING: THIS CODE ONLY WORKS FOR GATHERS FROM ARRAYS OF
***UNIFORM*** INT64's/POINTERS. (problem is that ispc doesn't
expose whether it's from array of uniform or array of varying
poitners, so in here there's no way to tell - only thing we can do
is pick one...
*/
static FORCEINLINE __vec16_i64
__gather_base_offsets32_i64(uint8_t *base, uint32_t scale, __vec16_i32 offsets,
__vec16_i1 mask) {
__vec16_i64 ret;
ret.v_lo = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets,
base, _MM_UPCONV_EPI32_NONE, scale,
_MM_HINT_NONE);
ret.v_hi = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets,
base+4, _MM_UPCONV_EPI32_NONE, scale,
_MM_HINT_NONE);
return ret;
}
/*! gather vector of 64-bit ints from addresses pointing to uniform ints
(iw) WARNING: THIS CODE ONLY WORKS FOR GATHERS FROM ARRAYS OF
***UNIFORM*** INT64's/POINTERS. (problem is that ispc doesn't
expose whether it's from array of uniform or array of varying
poitners, so in here there's no way to tell - only thing we can do
is pick one...
*/
static FORCEINLINE __vec16_i64
__gather64_i64(__vec16_i64 addr,
__vec16_i1 mask)
{//TODO
__vec16_i64 ret;
// There is no gather instruction with 64-bit offsets in KNC.
// We have to manually iterate over the upper 32 bits ;-)
__vec16_i1 still_to_do = mask;
__m512i offsets = addr.v_lo;
while (still_to_do) {
int first_active_lane = _mm_tzcnt_32((int)still_to_do);
const uint32_t &hi32 = ((uint*)&addr.v_hi)[first_active_lane];
__vec16_i1 match = _mm512_mask_cmp_epi32_mask(still_to_do,addr.v_hi,
__smear_i32<__vec16_i32>((int32_t)hi32),
_MM_CMPINT_EQ);
const uint8_t * base = (const uint8_t*)((uint64_t)hi32 << 32);
ret.v_lo = _mm512_mask_i32extgather_epi32(ret.v_lo, match, offsets,
base, _MM_UPCONV_EPI32_NONE, 1,
_MM_HINT_NONE);
ret.v_hi = _mm512_mask_i32extgather_epi32(ret.v_hi, match, offsets,
base+4, _MM_UPCONV_EPI32_NONE, 1,
_MM_HINT_NONE);
still_to_do = _mm512_kxor(match, still_to_do);
}
return ret;
}
///////////////////////////////////////////////////////////////////////////
// float
///////////////////////////////////////////////////////////////////////////
@@ -1945,6 +2010,7 @@ __gather64_float(__vec16_i64 addr, __vec16_i1 mask)
return ret;
}
/*! gather with 64-bit offsets.
\todo add optimization that falls back to 32-bit offset gather if
@@ -2056,6 +2122,33 @@ __scatter_base_offsets64_i32(uint8_t *_base, uint32_t scale, __vec16_i64 offsets
}
}
static FORCEINLINE void // TODO
__scatter_base_offsets64_i8(uint8_t *_base, uint32_t scale, __vec16_i64 offsets,
__vec16_i8 value,
__vec16_i1 mask) {
__vec16_i1 still_to_do = mask;
__vec16_i32 tmp = _mm512_extload_epi32(&value, _MM_UPCONV_EPI32_SINT8,
_MM_BROADCAST32_NONE, _MM_HINT_NONE);
// PING;
// _mm512_mask_extstore_epi32(p, mask, tmp, _MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE);
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));
_mm512_mask_i32extscatter_epi32(base, match, offsets.v_lo,
tmp,
_MM_DOWNCONV_EPI32_SINT8, scale,
_MM_HINT_NONE);
still_to_do = _mm512_kxor(match,still_to_do);
}
}
static FORCEINLINE __vec16_i32