github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-06-28 16:41:01 +03:00
Code Issues Packages Projects Releases Wiki Activity

Special case for Shuffle test added as well as comments for respectives implementations

Browse Source
This commit is contained in:
apostolos 2021-10-26 11:48:33 +03:00
parent 1eb3b19f63
commit bf54aae779
3 changed files with 47 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
src/util/arch/ppc64el/simd_utils.h
View File

@ -462,6 +462,9 @@ char testbit128(m128 val, unsigned int n) {
static really_inline
m128 pshufb_m128(m128 a, m128 b) {
/* On Intel, if bit 0x80 is set, then result is zero, otherwise which the lane it is &0xf.
In NEON or PPC, if >=16, then the result is zero, otherwise it is that lane.
below is the version that is converted from Intel to PPC. */
uint8x16_t mask =(uint8x16_t)vec_cmpge((uint8x16_t)b, (uint8x16_t)vec_splats((uint8_t)0x80));
uint8x16_t res = vec_perm ((uint8x16_t)a, (uint8x16_t)a, (uint8x16_t)b);
return (m128) vec_sel((uint8x16_t)res, (uint8x16_t)zeroes128(), (uint8x16_t)mask);

3
src/util/supervector/arch/ppc64el/impl.cpp
View File

@ -603,6 +603,9 @@ template<>
template<>
really_inline SuperVector<16> SuperVector<16>::pshufb<false>(SuperVector<16> b)
{
/* On Intel, if bit 0x80 is set, then result is zero, otherwise which the lane it is &0xf.
In NEON or PPC, if >=16, then the result is zero, otherwise it is that lane.
below is the version that is converted from Intel to PPC. */
uint8x16_t mask =(uint8x16_t)vec_cmpge((uint8x16_t)b.u.v128[0], (uint8x16_t)vec_splats((uint8_t)0x80));
uint8x16_t res = vec_perm ((uint8x16_t)u.v128[0], (uint8x16_t)u.v128[0], (uint8x16_t)b.u.v128[0]);
return (m128) vec_sel((uint8x16_t)res, (uint8x16_t)vec_splat_s8(0), (uint8x16_t)mask);

45
unit/internal/simd_utils.cpp
View File

@ -845,26 +845,63 @@ TEST(SimdUtilsTest, pshufb_m128) {
srand (time(NULL));
u8 vec[16];
for (int i=0; i<16; i++) {
vec[i] = rand() % 100 + 1;
vec[i] = rand() % 1000 + 1;
}
u8 vec2[16];
for (int i=0; i<16; i++) {
vec2[i]=i + (rand() % 15 + 0);
vec2[i]=i + (rand() % 100 + 0);
}
/* On Intel, if bit 0x80 is set, then result is zero, otherwise which the lane it is &0xf.
In NEON or PPC, if >=16, then the result is zero, otherwise it is that lane.
Thus bellow we have to check thah case to NEON or PPC. */
/*Insure that vec2 has at least 1 or more 0x80*/
u8 vec3[16] = {0};
vec3[15] = 0x80;
for (int i=0; i<15; i++) {
int l = rand() % 1000 + 0;
if (l % 16 ==0){
vec3[i]= 0x80;
} else{
vec3[i]= vec2[i];
}
}
/*
printf("vec3: ");
for(int i=15; i>=0; i--) { printf("%02x, ", vec3[i]); }
printf("\n");
*/
/*Test Special Case*/
m128 v1 = loadu128(vec);
m128 v2 = loadu128(vec2);
m128 v2 = loadu128(vec3);
m128 vres = pshufb_m128(v1, v2);
u8 res[16];
storeu128(res, vres);
for (int i=0; i<16; i++) {
if(vec3[i] & 0x80){
ASSERT_EQ(res[i], 0);
}else{
ASSERT_EQ(vec[vec3[i] % 16 ], res[i]);
}
}
/*Test Other Cases*/
v1 = loadu128(vec);
v2 = loadu128(vec2);
vres = pshufb_m128(v1, v2);
storeu128(res, vres);
for (int i=0; i<16; i++) {
if(vec2[i] & 0x80){
ASSERT_EQ(res[i], 0);
}else{
ASSERT_EQ(vec[vec2[i] % 16 ], res[i]);
}
}
}
}