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

modified all gather/scatter instructions in 'knc-i1x16.h'

Browse Source
This commit is contained in:
Anton Mitrokhin
2014-08-12 17:02:33 +04:00
parent 9c9c77d2db
commit 6b5b547e2f
1 changed files with 17 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
examples/intrinsics/knc-i1x16.h
View File

@@ -53,7 +53,7 @@
#endif
#define KNC 1
#if 0
#if 1
extern "C"
{
int printf(const unsigned char *, ...);
@@ -2164,6 +2164,7 @@ static FORCEINLINE __vec16_i8 __gather_base_offsets32_i8(uint8_t *base, uint32_t
static FORCEINLINE __vec16_i8 __gather_base_offsets64_i8(uint8_t *_base, uint32_t scale, __vec16_i64 _offsets, __vec16_i1 mask)
{
const __vec16_i64 offsets = _offsets.cvt2hilo();
const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT32_MIN));
__vec16_i1 still_to_do = mask;
__vec16_i32 tmp;
while (still_to_do) {
@@ -2174,8 +2175,8 @@ static FORCEINLINE __vec16_i8 __gather_base_offsets64_i8(uint8_t *_base, uint32_
_MM_CMPINT_EQ);
void * base = (void*)((unsigned long)_base +
((scale*(unsigned long)hi32) << 32));
tmp = _mm512_mask_i32extgather_epi32(tmp, match, offsets.v_lo, base,
((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT32_MIN));
tmp = _mm512_mask_i32extgather_epi32(tmp, match, signed_offsets, base,
_MM_UPCONV_EPI32_SINT8, scale,
_MM_HINT_NONE);
still_to_do = _mm512_kxor(match,still_to_do);
@@ -2199,6 +2200,7 @@ static FORCEINLINE __vec16_i32 __gather_base_offsets32_i32(uint8_t *base, uint32
static FORCEINLINE __vec16_i32 __gather_base_offsets64_i32(uint8_t *_base, uint32_t scale, __vec16_i64 _offsets, __vec16_i1 mask)
{
const __vec16_i64 offsets = _offsets.cvt2hilo();
const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT32_MIN));
// 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;
@@ -2209,10 +2211,10 @@ static FORCEINLINE __vec16_i32 __gather_base_offsets64_i32(uint8_t *_base, uint3
__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));
ret = _mm512_mask_i32extgather_epi32(ret, match, offsets.v_lo, base,
((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT32_MIN));
ret = _mm512_mask_i32extgather_epi32(ret, match, signed_offsets, base,
_MM_UPCONV_EPI32_NONE, scale,
_MM_HINT_NONE);
still_to_do = _mm512_kxor(match, still_to_do);
@@ -2232,6 +2234,7 @@ static FORCEINLINE __vec16_f __gather_base_offsets32_float(uint8_t *base, uint32
static FORCEINLINE __vec16_f __gather_base_offsets64_float(uint8_t *_base, uint32_t scale, __vec16_i64 _offsets, __vec16_i1 mask)
{
const __vec16_i64 offsets = _offsets.cvt2hilo();
const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT32_MIN));
// 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;
@@ -2244,8 +2247,8 @@ static FORCEINLINE __vec16_f __gather_base_offsets64_float(uint8_t *_base, uint3
_MM_CMPINT_EQ);
void * base = (void*)((unsigned long)_base +
((scale*(unsigned long)hi32) << 32));
ret = _mm512_mask_i32extgather_ps(ret, match, offsets.v_lo, base,
((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT32_MIN));
ret = _mm512_mask_i32extgather_ps(ret, match, signed_offsets, base,
_MM_UPCONV_PS_NONE, scale,
_MM_HINT_NONE);
still_to_do = _mm512_kxor(match, still_to_do);
@@ -2341,7 +2344,8 @@ static FORCEINLINE void __scatter_base_offsets32_i32(uint8_t *b, uint32_t scale,
static FORCEINLINE void __scatter_base_offsets64_i32(uint8_t *_base, uint32_t scale, __vec16_i64 _offsets, __vec16_i32 value, __vec16_i1 mask)
{
const __vec16_i64 offsets = _offsets.cvt2hilo();
const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT32_MIN));
__vec16_i1 still_to_do = mask;
while (still_to_do) {
int first_active_lane = _mm_tzcnt_32((int)still_to_do);
@@ -2351,8 +2355,8 @@ static FORCEINLINE void __scatter_base_offsets64_i32(uint8_t *_base, uint32_t sc
_MM_CMPINT_EQ);
void * base = (void*)((unsigned long)_base +
((scale*(unsigned long)hi32) << 32));
_mm512_mask_i32extscatter_epi32(base, match, offsets.v_lo,
((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT32_MIN));
_mm512_mask_i32extscatter_epi32(base, match, signed_offsets,
value,
_MM_DOWNCONV_EPI32_NONE, scale,
_MM_HINT_NONE);
@@ -2373,7 +2377,7 @@ static FORCEINLINE void __scatter_base_offsets64_float(uint8_t *_base, uint32_t
{
const __vec16_i64 offsets = _offsets.cvt2hilo();
const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT32_MIN));
__vec16_i1 still_to_do = mask;
while (still_to_do) {
int first_active_lane = _mm_tzcnt_32((int)still_to_do);
@@ -2383,7 +2387,7 @@ static FORCEINLINE void __scatter_base_offsets64_float(uint8_t *_base, uint32_t
_MM_CMPINT_EQ);
void * base = (void*)((unsigned long)_base +
((scale*(unsigned long)hi32) << 32) - INT32_MIN);
((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT32_MIN));
_mm512_mask_i32extscatter_ps(base, match, signed_offsets,
value,