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

Add shift support for knc targets. This is not optimized.

Browse Source
This commit is contained in:
james.brodman
2013-10-28 12:43:42 -04:00
parent 1e80b3b0d7
commit 641d882ea6
1 changed files with 19 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
examples/intrinsics/knc-i1x16.h
View File

@@ -451,6 +451,17 @@ static FORCEINLINE VTYPE __rotate_##NAME(VTYPE v, int index) { \
return ret; \ return ret; \
} \ } \
#define SHIFT(VTYPE, NAME, STYPE) \
static FORCEINLINE VTYPE __shift_##NAME(VTYPE v, int index) { \
VTYPE ret; \
for (int i = 0; i < 16; ++i) { \
int modIndex = i+index; \
STYPE val = ((modIndex >= 0) && (modIndex < 16)) ? v.v[modIndex] : 0; \
ret.v[i] = val; \
} \
return ret; \
} \
/* knc::macro::used */ /* knc::macro::used */
#define SHUFFLES(VTYPE, NAME, STYPE) \ #define SHUFFLES(VTYPE, NAME, STYPE) \
static FORCEINLINE VTYPE __shuffle_##NAME(VTYPE v, __vec16_i32 index) { \ static FORCEINLINE VTYPE __shuffle_##NAME(VTYPE v, __vec16_i32 index) { \
@@ -566,6 +577,7 @@ SETZERO(__vec16_i8, i8)
UNDEF(__vec16_i8, i8) UNDEF(__vec16_i8, i8)
BROADCAST(__vec16_i8, i8, int8_t) BROADCAST(__vec16_i8, i8, int8_t)
ROTATE(__vec16_i8, i8, int8_t) ROTATE(__vec16_i8, i8, int8_t)
SHIFT(__vec16_i8, i8, int8_t)
SHUFFLES(__vec16_i8, i8, int8_t) SHUFFLES(__vec16_i8, i8, int8_t)
LOAD_STORE(__vec16_i8, int8_t) LOAD_STORE(__vec16_i8, int8_t)
@@ -612,6 +624,7 @@ SETZERO(__vec16_i16, i16)
UNDEF(__vec16_i16, i16) UNDEF(__vec16_i16, i16)
BROADCAST(__vec16_i16, i16, int16_t) BROADCAST(__vec16_i16, i16, int16_t)
ROTATE(__vec16_i16, i16, int16_t) ROTATE(__vec16_i16, i16, int16_t)
SHIFT(__vec16_i16, i16, int16_t)
SHUFFLES(__vec16_i16, i16, int16_t) SHUFFLES(__vec16_i16, i16, int16_t)
LOAD_STORE(__vec16_i16, int16_t) LOAD_STORE(__vec16_i16, int16_t)
@@ -688,6 +701,8 @@ static FORCEINLINE __vec16_i32 __rotate_i32(__vec16_i32 v, int index)
return _mm512_mask_permutevar_epi32(v, 0xFFFF, shuffle, v); return _mm512_mask_permutevar_epi32(v, 0xFFFF, shuffle, v);
} }
SHIFT(__vec16_i32, i32, int32_t)
static FORCEINLINE __vec16_i32 __shuffle_i32 (__vec16_i32 v, __vec16_i32 index) static FORCEINLINE __vec16_i32 __shuffle_i32 (__vec16_i32 v, __vec16_i32 index)
{ {
return _mm512_mask_permutevar_epi32(v, 0xFFFF, __and(index, __smear_i32<__vec16_i32>(0xF)), v); return _mm512_mask_permutevar_epi32(v, 0xFFFF, __and(index, __smear_i32<__vec16_i32>(0xF)), v);
@@ -942,6 +957,8 @@ static FORCEINLINE __vec16_i64 __rotate_i64(const __vec16_i64 _v, const int inde
const __vec16_i32 ret_lo = __rotate_i32(v_lo, index); const __vec16_i32 ret_lo = __rotate_i32(v_lo, index);
return CASTI2L(ret_hi, ret_lo); return CASTI2L(ret_hi, ret_lo);
} }
SHIFT(__vec16_i64, i64, int64_t)
static FORCEINLINE __vec16_i64 __shuffle_double(__vec16_i64 _v, const __vec16_i32 index) static FORCEINLINE __vec16_i64 __shuffle_double(__vec16_i64 _v, const __vec16_i32 index)
{ {
CASTL2I(_v, v_hi, v_lo); CASTL2I(_v, v_hi, v_lo);
@@ -1063,6 +1080,7 @@ static FORCEINLINE __vec16_f __rotate_float(__vec16_f _v, int index)
const __vec16_i32 shuffle = _mm512_and_epi32(_mm512_add_epi32(__ispc_stride1, idx), __smear_i32<__vec16_i32>(0xF)); const __vec16_i32 shuffle = _mm512_and_epi32(_mm512_add_epi32(__ispc_stride1, idx), __smear_i32<__vec16_i32>(0xF));
return _mm512_castsi512_ps(_mm512_mask_permutevar_epi32(v, 0xFFFF, shuffle, v)); return _mm512_castsi512_ps(_mm512_mask_permutevar_epi32(v, 0xFFFF, shuffle, v));
} }
SHIFT(__vec16_f, float, float)
static FORCEINLINE __vec16_f __shuffle_float(__vec16_f v, __vec16_i32 index) static FORCEINLINE __vec16_f __shuffle_float(__vec16_f v, __vec16_i32 index)
{ {
return _mm512_castsi512_ps(_mm512_mask_permutevar_epi32(_mm512_castps_si512(v), 0xffff, index, _mm512_castps_si512(v))); return _mm512_castsi512_ps(_mm512_mask_permutevar_epi32(_mm512_castps_si512(v), 0xffff, index, _mm512_castps_si512(v)));
@@ -1333,6 +1351,7 @@ static FORCEINLINE __vec16_d __rotate_double(const __vec16_d _v, const int index
const __vec16_f ret_lo = __rotate_float(v_lo, index); const __vec16_f ret_lo = __rotate_float(v_lo, index);
return CASTF2D(ret_hi, ret_lo); return CASTF2D(ret_hi, ret_lo);
} }
SHIFT(__vec16_d, double, double)
static FORCEINLINE __vec16_d __shuffle_double(__vec16_d _v, const __vec16_i32 index) static FORCEINLINE __vec16_d __shuffle_double(__vec16_d _v, const __vec16_i32 index)
{ {
CASTD2F(_v, v_hi, v_lo); CASTD2F(_v, v_hi, v_lo);