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

Add reduce_add() for int8 and int16 types.

Browse Source 
This maps to specialized instructions (e.g. PSADBW) when available.
This commit is contained in:
Matt Pharr
2013-07-25 09:11:39 -07:00
parent 2d063925a1
commit b6df447b55
24 changed files with 464 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
examples/intrinsics/generic-16.h
View File

@@ -1162,19 +1162,20 @@ REDUCE_ADD(double, __vec16_d, __reduce_add_double)
REDUCE_MINMAX(double, __vec16_d, __reduce_min_double, <)
REDUCE_MINMAX(double, __vec16_d, __reduce_max_double, >)
REDUCE_ADD(uint32_t, __vec16_i32, __reduce_add_int32)
REDUCE_ADD(int16_t, __vec16_i8, __reduce_add_int8)
REDUCE_ADD(int32_t, __vec16_i16, __reduce_add_int16)
REDUCE_ADD(int64_t, __vec16_i32, __reduce_add_int32)
REDUCE_MINMAX(int32_t, __vec16_i32, __reduce_min_int32, <)
REDUCE_MINMAX(int32_t, __vec16_i32, __reduce_max_int32, >)
REDUCE_ADD(uint32_t, __vec16_i32, __reduce_add_uint32)
REDUCE_MINMAX(uint32_t, __vec16_i32, __reduce_min_uint32, <)
REDUCE_MINMAX(uint32_t, __vec16_i32, __reduce_max_uint32, >)
REDUCE_ADD(uint64_t, __vec16_i64, __reduce_add_int64)
REDUCE_ADD(int64_t, __vec16_i64, __reduce_add_int64)
REDUCE_MINMAX(int64_t, __vec16_i64, __reduce_min_int64, <)
REDUCE_MINMAX(int64_t, __vec16_i64, __reduce_max_int64, >)
REDUCE_ADD(uint64_t, __vec16_i64, __reduce_add_uint64)
REDUCE_MINMAX(uint64_t, __vec16_i64, __reduce_min_uint64, <)
REDUCE_MINMAX(uint64_t, __vec16_i64, __reduce_max_uint64, >)

9
examples/intrinsics/generic-32.h
View File

@@ -1231,19 +1231,20 @@ REDUCE_ADD(double, __vec32_d, __reduce_add_double)
REDUCE_MINMAX(double, __vec32_d, __reduce_min_double, <)
REDUCE_MINMAX(double, __vec32_d, __reduce_max_double, >)
REDUCE_ADD(uint32_t, __vec32_i32, __reduce_add_int32)
REDUCE_ADD(int16_t, __vec16_i8, __reduce_add_int8)
REDUCE_ADD(int32_t, __vec16_i16, __reduce_add_int16)
REDUCE_ADD(int64_t, __vec32_i32, __reduce_add_int32)
REDUCE_MINMAX(int32_t, __vec32_i32, __reduce_min_int32, <)
REDUCE_MINMAX(int32_t, __vec32_i32, __reduce_max_int32, >)
REDUCE_ADD(uint32_t, __vec32_i32, __reduce_add_uint32)
REDUCE_MINMAX(uint32_t, __vec32_i32, __reduce_min_uint32, <)
REDUCE_MINMAX(uint32_t, __vec32_i32, __reduce_max_uint32, >)
REDUCE_ADD(uint64_t, __vec32_i64, __reduce_add_int64)
REDUCE_ADD(int64_t, __vec32_i64, __reduce_add_int64)
REDUCE_MINMAX(int64_t, __vec32_i64, __reduce_min_int64, <)
REDUCE_MINMAX(int64_t, __vec32_i64, __reduce_max_int64, >)
REDUCE_ADD(uint64_t, __vec32_i64, __reduce_add_uint64)
REDUCE_MINMAX(uint64_t, __vec32_i64, __reduce_min_uint64, <)
REDUCE_MINMAX(uint64_t, __vec32_i64, __reduce_max_uint64, >)

9
examples/intrinsics/generic-64.h
View File

@@ -1364,19 +1364,20 @@ REDUCE_ADD(double, __vec64_d, __reduce_add_double)
REDUCE_MINMAX(double, __vec64_d, __reduce_min_double, <)
REDUCE_MINMAX(double, __vec64_d, __reduce_max_double, >)
REDUCE_ADD(uint32_t, __vec64_i32, __reduce_add_int32)
REDUCE_ADD(int16_t, __vec16_i8, __reduce_add_int8)
REDUCE_ADD(int32_t, __vec16_i16, __reduce_add_int16)
REDUCE_ADD(int64_t, __vec64_i32, __reduce_add_int32)
REDUCE_MINMAX(int32_t, __vec64_i32, __reduce_min_int32, <)
REDUCE_MINMAX(int32_t, __vec64_i32, __reduce_max_int32, >)
REDUCE_ADD(uint32_t, __vec64_i32, __reduce_add_uint32)
REDUCE_MINMAX(uint32_t, __vec64_i32, __reduce_min_uint32, <)
REDUCE_MINMAX(uint32_t, __vec64_i32, __reduce_max_uint32, >)
REDUCE_ADD(uint64_t, __vec64_i64, __reduce_add_int64)
REDUCE_ADD(int64_t, __vec64_i64, __reduce_add_int64)
REDUCE_MINMAX(int64_t, __vec64_i64, __reduce_min_int64, <)
REDUCE_MINMAX(int64_t, __vec64_i64, __reduce_max_int64, >)
REDUCE_ADD(uint64_t, __vec64_i64, __reduce_add_uint64)
REDUCE_MINMAX(uint64_t, __vec64_i64, __reduce_min_uint64, <)
REDUCE_MINMAX(uint64_t, __vec64_i64, __reduce_max_uint64, >)

16
examples/intrinsics/knc.h
View File

@@ -1511,6 +1511,22 @@ static FORCEINLINE int64_t __count_trailing_zeros_i64(const __vec1_i64 mask) {
// reductions
///////////////////////////////////////////////////////////////////////////
static FORCEINLINE int16_t __reduce_add_i8(__vec16_i8 v) {
// TODO: improve this!
int16_t ret = 0;
for (int i = 0; i < 16; ++i)
ret += v.v[i];
return ret;
}
static FORCEINLINE int32_t __reduce_add_i16(__vec16_i16 v) {
// TODO: improve this!
int32_t ret = 0;
for (int i = 0; i < 16; ++i)
ret += v.v[i];
return ret;
}
static FORCEINLINE uint32_t __reduce_add_i32(__vec16_i32 v) {
return _mm512_reduce_add_epi32(v);
}

3
examples/intrinsics/knc2x.h
View File

@@ -1607,6 +1607,9 @@ static FORCEINLINE int64_t __count_leading_zeros_i64(uint64_t v) {
///////////////////////////////////////////////////////////////////////////
// reductions
REDUCE_ADD(int16_t, __vec32_i8, __reduce_add_int8)
REDUCE_ADD(int32_t, __vec32_i16, __reduce_add_int16)
static FORCEINLINE float __reduce_add_float(__vec32_f v) {
return _mm512_reduce_add_ps(v.v1) + _mm512_reduce_add_ps(v.v2);
}

16
examples/intrinsics/sse4.h
View File

@@ -2528,6 +2528,22 @@ static FORCEINLINE int64_t __count_leading_zeros_i64(uint64_t v) {
///////////////////////////////////////////////////////////////////////////
// reductions
static FORCEINLINE int16_t __reduce_add_int8(__vec4_i8 v) {
// TODO: improve
int16_t ret = 0;
for (int i = 0; i < 4; ++i)
ret += v.v[i];
return ret;
}
static FORCEINLINE int32_t __reduce_add_int16(__vec4_i16 v) {
// TODO: improve
int32_t ret = 0;
for (int i = 0; i < 4; ++i)
ret += v.v[i];
return ret;
}
static FORCEINLINE float __reduce_add_float(__vec4_f v) {
float r = bits_as_float(_mm_extract_ps(v.v, 0));
r += bits_as_float(_mm_extract_ps(v.v, 1));