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

Merge pull request #321 from mmp/setzero

Browse Source 
More varied support for constant vectors from C++ backend.
This commit is contained in:
Jean-Luc Duprat
2012-07-09 08:57:05 -07:00
parent 950a989744 0d3993fa25
commit 098277b4f0
7 changed files with 330 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files

95
examples/intrinsics/sse4.h
View File

@@ -297,10 +297,18 @@ template <int ALIGN> static FORCEINLINE void __store(__vec4_i1 *p, __vec4_i1 val
_mm_storeu_ps((float *)(&p->v), value.v);
}
static FORCEINLINE __vec4_i1 __smear_i1(__vec4_i1, int v) {
static FORCEINLINE __vec4_i1 __smear_i1(int v) {
return __vec4_i1(v, v, v, v);
}
static FORCEINLINE __vec4_i1 __setzero_i1() {
return __vec4_i1(_mm_setzero_ps());
}
static FORCEINLINE __vec4_i1 __undef_i1() {
return __vec4_i1();
}
///////////////////////////////////////////////////////////////////////////
// int8
@@ -525,10 +533,18 @@ static FORCEINLINE void __insert_element(__vec4_i8 *v, int index, int8_t val) {
((int8_t *)v)[index] = val;
}
static FORCEINLINE __vec4_i8 __smear_i8(__vec4_i8, int8_t v) {
static FORCEINLINE __vec4_i8 __smear_i8(int8_t v) {
return _mm_set1_epi8(v);
}
static FORCEINLINE __vec4_i8 __setzero_i8() {
return _mm_set1_epi8(0);
}
static FORCEINLINE __vec4_i8 __undef_i8() {
return __vec4_i8();
}
static FORCEINLINE __vec4_i8 __broadcast_i8(__vec4_i8 v, int index) {
return _mm_set1_epi8(__extract_element(v, index));
}
@@ -784,10 +800,18 @@ static FORCEINLINE void __insert_element(__vec4_i16 *v, int index, int16_t val)
((int16_t *)v)[index] = val;
}
static FORCEINLINE __vec4_i16 __smear_i16(__vec4_i16, int16_t v) {
static FORCEINLINE __vec4_i16 __smear_i16(int16_t v) {
return _mm_set1_epi16(v);
}
static FORCEINLINE __vec4_i16 __setzero_i16() {
return _mm_set1_epi16(0);
}
static FORCEINLINE __vec4_i16 __undef_i16() {
return __vec4_i16();
}
static FORCEINLINE __vec4_i16 __broadcast_i16(__vec4_i16 v, int index) {
return _mm_set1_epi16(__extract_element(v, index));
}
@@ -1021,10 +1045,18 @@ static FORCEINLINE __vec4_i32 __select(__vec4_i1 mask, __vec4_i32 a, __vec4_i32
_mm_castsi128_ps(a.v), mask.v));
}
static FORCEINLINE __vec4_i32 __smear_i32(__vec4_i32, int32_t v) {
static FORCEINLINE __vec4_i32 __smear_i32(int32_t v) {
return _mm_set1_epi32(v);
}
static FORCEINLINE __vec4_i32 __setzero_i32() {
return _mm_castps_si128(_mm_setzero_ps());
}
static FORCEINLINE __vec4_i32 __undef_i32() {
return __vec4_i32();
}
static FORCEINLINE int32_t __extract_element(__vec4_i32 v, int index) {
return ((int32_t *)&v)[index];
}
@@ -1282,10 +1314,18 @@ static FORCEINLINE __vec4_i64 __select(__vec4_i1 mask, __vec4_i64 a, __vec4_i64
return __vec4_i64(_mm_castpd_si128(r0), _mm_castpd_si128(r1));
}
static FORCEINLINE __vec4_i64 __smear_i64(__vec4_i64, int64_t v) {
static FORCEINLINE __vec4_i64 __smear_i64(int64_t v) {
return __vec4_i64(v, v, v, v);
}
static FORCEINLINE __vec4_i64 __setzero_i64() {
return __vec4_i64(0, 0, 0, 0);
}
static FORCEINLINE __vec4_i64 __undef_i64() {
return __vec4_i64();
}
static FORCEINLINE int64_t __extract_element(__vec4_i64 v, int index) {
return ((int64_t *)&v)[index];
}
@@ -1386,10 +1426,18 @@ static FORCEINLINE __vec4_f __select(__vec4_i1 mask, __vec4_f a, __vec4_f b) {
return _mm_blendv_ps(b.v, a.v, mask.v);
}
static FORCEINLINE __vec4_f __smear_float(__vec4_f, float v) {
static FORCEINLINE __vec4_f __smear_float(float v) {
return _mm_set1_ps(v);
}
static FORCEINLINE __vec4_f __setzero_float() {
return _mm_setzero_ps();
}
static FORCEINLINE __vec4_f __undef_float() {
return __vec4_f();
}
static FORCEINLINE float __extract_element(__vec4_f v, int index) {
return ((float *)&v)[index];
}
@@ -1518,10 +1566,18 @@ static FORCEINLINE __vec4_d __select(__vec4_i1 mask, __vec4_d a, __vec4_d b) {
return __vec4_d(r0, r1);
}
static FORCEINLINE __vec4_d __smear_double(__vec4_d, double v) {
static FORCEINLINE __vec4_d __smear_double(double v) {
return __vec4_d(_mm_set1_pd(v), _mm_set1_pd(v));
}
static FORCEINLINE __vec4_d __setzero_double() {
return __vec4_d(_mm_setzero_pd(), _mm_setzero_pd());
}
static FORCEINLINE __vec4_d __undef_double() {
return __vec4_d();
}
static FORCEINLINE double __extract_element(__vec4_d v, int index) {
return ((double *)&v)[index];
}
@@ -1618,13 +1674,11 @@ static FORCEINLINE __vec4_i16 __cast_sext(__vec4_i16, __vec4_i8 val) {
}
static FORCEINLINE __vec4_i8 __cast_sext(__vec4_i8, __vec4_i1 v) {
return __select(v, __smear_i8(__vec4_i8(), 0xff),
__smear_i8(__vec4_i8(), 0));
return __select(v, __smear_i8(0xff), __setzero_i8());
}
static FORCEINLINE __vec4_i16 __cast_sext(__vec4_i16, __vec4_i1 v) {
return __select(v, __smear_i16(__vec4_i16(), 0xffff),
__smear_i16(__vec4_i16(), 0));
return __select(v, __smear_i16(0xffff), __setzero_i16());
}
static FORCEINLINE __vec4_i32 __cast_sext(__vec4_i32, __vec4_i1 v) {
@@ -1684,12 +1738,11 @@ static FORCEINLINE __vec4_i16 __cast_zext(__vec4_i16, __vec4_i8 val) {
}
static FORCEINLINE __vec4_i8 __cast_zext(__vec4_i8, __vec4_i1 v) {
return __select(v, __smear_i8(__vec4_i8(), 1), __smear_i8(__vec4_i8(), 0));
return __select(v, __smear_i8(1), __setzero_i8());
}
static FORCEINLINE __vec4_i16 __cast_zext(__vec4_i16, __vec4_i1 v) {
return __select(v, __smear_i16(__vec4_i16(), 1),
__smear_i16(__vec4_i16(), 0));
return __select(v, __smear_i16(1), __setzero_i16());
}
static FORCEINLINE __vec4_i32 __cast_zext(__vec4_i32, __vec4_i1 v) {
@@ -1697,7 +1750,7 @@ static FORCEINLINE __vec4_i32 __cast_zext(__vec4_i32, __vec4_i1 v) {
}
static FORCEINLINE __vec4_i64 __cast_zext(__vec4_i64, __vec4_i1 v) {
return __select(v, __smear_i64(__vec4_i64(), 1), __smear_i64(__vec4_i64(), 0));
return __select(v, __smear_i64(1), __setzero_i64());
}
// truncations
@@ -1857,11 +1910,11 @@ static FORCEINLINE __vec4_d __cast_uitofp(__vec4_d, __vec4_i64 val) {
}
static FORCEINLINE __vec4_f __cast_uitofp(__vec4_f, __vec4_i1 v) {
return __select(v, __smear_float(__vec4_f(), 1.), __smear_float(__vec4_f(), 0.));
return __select(v, __smear_float(1.), __setzero_float());
}
static FORCEINLINE __vec4_d __cast_uitofp(__vec4_d, __vec4_i1 v) {
return __select(v, __smear_double(__vec4_d(), 1.), __smear_double(__vec4_d(), 0.));
return __select(v, __smear_double(1.), __setzero_double());
}
// float/double to signed int
@@ -2796,8 +2849,8 @@ lGatherBaseOffsets32(RetVec, RetScalar, unsigned char *p, __vec4_i32 offsets,
RetScalar r[4];
#if 1
// "Fast gather" trick...
offsets = __select(mask, offsets, __smear_i32(__vec4_i32(), 0));
constOffset = __select(mask, constOffset, __smear_i32(__vec4_i32(), 0));
offsets = __select(mask, offsets, __setzero_i32());
constOffset = __select(mask, constOffset, __setzero_i32());
int offset = scale * _mm_extract_epi32(offsets.v, 0) + _mm_extract_epi32(constOffset.v, 0);
RetScalar *ptr = (RetScalar *)(p + offset);
@@ -2854,8 +2907,8 @@ lGatherBaseOffsets64(RetVec, RetScalar, unsigned char *p, __vec4_i64 offsets,
RetScalar r[4];
#if 1
// "Fast gather" trick...
offsets = __select(mask, offsets, __smear_i64(__vec4_i64(), 0));
constOffset = __select(mask, constOffset, __smear_i64(__vec4_i64(), 0));
offsets = __select(mask, offsets, __setzero_i64());
constOffset = __select(mask, constOffset, __setzero_i64());
int64_t offset = scale * _mm_extract_epi64(offsets.v[0], 0) + _mm_extract_epi64(constOffset.v[0], 0);
RetScalar *ptr = (RetScalar *)(p + offset);