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

Merge branch 'master' of https://github.com/ncos/ispc into knc-backend-merge

Browse Source 
Conflicts:
	examples/intrinsics/knc.h
This commit is contained in:
Anton Mitrokhin
2014-11-14 11:54:34 +04:00
parent b21043c309 e834703bb7
commit 615d36b236
3 changed files with 236 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
cbackend.cpp
View File

@@ -152,11 +152,7 @@ namespace {
incorporateValue(Aliasee);
}
#if defined(LLVM_3_2) || defined(LLVM_3_3) || defined(LLVM_3_4) || defined(LLVM_3_5)
llvm::SmallVector<std::pair<unsigned, llvm::MDNode*>, 4> MDForInst;
#else // LLVM 3.6+
llvm::SmallVector<std::pair<unsigned, llvm::Value*>, 4> MDForInst;
#endif
// Get types from functions.
for (llvm::Module::const_iterator FI = M.begin(), E = M.end(); FI != E; ++FI) {

231
examples/intrinsics/knc.h
View File

@@ -211,6 +211,23 @@ typedef struct PRE_ALIGN(64) __vec16_i64 {
__m512i v_lo;
} POST_ALIGN(64) __vec16_i64;
static __vec16_i64 zmm2hilo(const __m512i v1, const __m512i v2){
__vec16_i64 v;
v.v_hi = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00,
_mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0),
v2);
v.v_hi = _mm512_mask_permutevar_epi32(v.v_hi, 0x00FF,
_mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1),
v1);
v.v_lo = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00,
_mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1),
v2);
v.v_lo = _mm512_mask_permutevar_epi32(v.v_lo, 0x00FF,
_mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0),
v1);
return v;
}
static void hilo2zmm(const __vec16_i64 &v, __m512i &_v1, __m512i &_v2) {
_v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA,
_mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8),
@@ -331,7 +348,7 @@ inline std::ostream &operator<<(std::ostream &out, const __vec16_i64 &v)
uint32_t *ptr = (uint32_t*)&v;
for (int i=0;i<16;i++) {
uint64_t val = (uint64_t(ptr[i])<<32)+ptr[i+16];
out << (i!=0?",":"") << std::dec << std::setw(8) << val << std::dec;
out << (i!=0?",":"") << std::dec << std::setw(8) << ((int64_t)val) << std::dec;
}
out << "]" << std::flush;
return out;
@@ -964,7 +981,10 @@ template <> FORCEINLINE __vec16_i64 __smear_i64<__vec16_i64>(const int64_t &l)
}
static FORCEINLINE __vec16_i64 __rotate_i64(__vec16_i64 v, int index) {
return __vec16_i64(__shuffle_i32(v.v_lo, index), __shuffle_i32(v.v_hi, index));
__vec16_i32 idx = __smear_i32<__vec16_i32>(index);
__vec16_i32 shuffle = _mm512_and_epi32(_mm512_add_epi32(__ispc_stride1, idx), __smear_i32<__vec16_i32>(0xf));
return __vec16_i64(_mm512_mask_permutevar_epi32(v.v_lo, 0xffff, shuffle, v.v_lo),
_mm512_mask_permutevar_epi32(v.v_hi, 0xffff, shuffle, v.v_hi));
}
static FORCEINLINE __vec16_i64 __shuffle2_i64(__vec16_i64 v0, __vec16_i64 v1, __vec16_i32 index) {
@@ -1605,6 +1625,20 @@ static FORCEINLINE __vec16_f __cast_sitofp(__vec16_f, __vec16_i32 val) {
return _mm512_cvtfxpnt_round_adjustepi32_ps(val, _MM_ROUND_MODE_NEAREST, _MM_EXPADJ_NONE);
}
static FORCEINLINE __vec16_f __cast_sitofp(__vec16_f, __vec16_i64 val) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(val, tmp1, tmp2);
__vec16_f ret;
for (int i = 0; i < 8; i++) {
((float*)&ret)[i] = (float)(((int64_t*)&tmp1)[i]);
}
for (int i = 0; i < 8; i++) {
((float*)&ret)[i + 8] = (float)(((int64_t*)&tmp2)[i]);
}
return ret;
}
static FORCEINLINE __vec16_d __cast_sitofp(__vec16_d, __vec16_i8 val) {
__vec16_i32 vi = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_SINT8, _MM_BROADCAST_16X16, _MM_HINT_NONE);
__vec16_d ret;
@@ -1631,6 +1665,21 @@ static FORCEINLINE __vec16_d __cast_sitofp(__vec16_d, __vec16_i32 val) {
return ret;
}
static FORCEINLINE __vec16_d __cast_sitofp(__vec16_d, __vec16_i64 val) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(val, tmp1, tmp2);
__vec16_d ret;
for (int i = 0; i < 8; i++) {
((double*)&ret.v1)[i] = (double)(((int64_t*)&tmp1)[i]);
}
for (int i = 0; i < 8; i++) {
((double*)&ret.v2)[i] = (double)(((int64_t*)&tmp2)[i]);
}
return ret;
}
static FORCEINLINE __vec16_f __cast_uitofp(__vec16_f, __vec16_i1 v)
{
const __m512 ret = _mm512_setzero_ps();
@@ -1650,6 +1699,39 @@ static FORCEINLINE __vec16_f __cast_uitofp(__vec16_f, __vec16_i32 v) {
return _mm512_cvtfxpnt_round_adjustepu32_ps(v, _MM_FROUND_NO_EXC, _MM_EXPADJ_NONE);
}
static FORCEINLINE __vec16_f __cast_uitofp(__vec16_f, __vec16_i64 val) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(val, tmp1, tmp2);
__vec16_f ret;
// Cycles don't work. I don't know why.
/*
for (int i = 0; i < 8; i++) {
((float*)&ret)[i] = ((float)(((uint64_t*)&tmp1)[i]));
}
for (int i = 0; i < 8; i++) {
((float*)&ret)[i + 8] = ((float)(((uint64_t*)&tmp2)[i]));
}
*/
((float*)&ret)[0] = ((float)(((uint64_t*)&tmp1)[0]));
((float*)&ret)[1] = ((float)(((uint64_t*)&tmp1)[1]));
((float*)&ret)[2] = ((float)(((uint64_t*)&tmp1)[2]));
((float*)&ret)[3] = ((float)(((uint64_t*)&tmp1)[3]));
((float*)&ret)[4] = ((float)(((uint64_t*)&tmp1)[4]));
((float*)&ret)[5] = ((float)(((uint64_t*)&tmp1)[5]));
((float*)&ret)[6] = ((float)(((uint64_t*)&tmp1)[6]));
((float*)&ret)[7] = ((float)(((uint64_t*)&tmp1)[7]));
((float*)&ret)[8] = ((float)(((uint64_t*)&tmp2)[0]));
((float*)&ret)[9] = ((float)(((uint64_t*)&tmp2)[1]));
((float*)&ret)[10] = ((float)(((uint64_t*)&tmp2)[2]));
((float*)&ret)[11] = ((float)(((uint64_t*)&tmp2)[3]));
((float*)&ret)[12] = ((float)(((uint64_t*)&tmp2)[4]));
((float*)&ret)[13] = ((float)(((uint64_t*)&tmp2)[5]));
((float*)&ret)[14] = ((float)(((uint64_t*)&tmp2)[6]));
((float*)&ret)[15] = ((float)(((uint64_t*)&tmp2)[7]));
return ret;
}
static FORCEINLINE __vec16_d __cast_uitofp(__vec16_d, __vec16_i8 val)
{
__vec16_i32 vi = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_UINT8, _MM_BROADCAST_16X16, _MM_HINT_NONE);
@@ -1679,6 +1761,22 @@ static FORCEINLINE __vec16_d __cast_uitofp(__vec16_d, __vec16_i32 val)
return ret;
}
static FORCEINLINE __vec16_d __cast_uitofp(__vec16_d, __vec16_i64 val) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(val, tmp1, tmp2);
__vec16_d ret;
for (int i = 0; i < 8; i++) {
((double*)&ret.v1)[i] = (double)(((uint64_t*)&tmp1)[i]);
}
for (int i = 0; i < 8; i++) {
((double*)&ret.v2)[i] = (double)(((uint64_t*)&tmp2)[i]);
}
return ret;
}
// float/double to signed int
static FORCEINLINE __vec16_i32 __cast_fptosi(__vec16_i32, __vec16_f val) {
return _mm512_cvtfxpnt_round_adjustps_epi32(val, _MM_ROUND_MODE_TOWARD_ZERO, _MM_EXPADJ_NONE);
@@ -1698,6 +1796,18 @@ static FORCEINLINE __vec16_i16 __cast_fptosi(__vec16_i16, __vec16_f val) {
return ret;
}
static FORCEINLINE __vec16_i64 __cast_fptosi(__vec16_i64, __vec16_f val) {
__m512i tmp1;
for (int i = 0; i < 8; i++) {
((int64_t*)&tmp1)[i] = (int64_t)(((float*)&val)[i]);
}
__m512i tmp2;
for (int i = 0; i < 8; i++) {
((int64_t*)&tmp2)[i] = (int64_t)(((float*)&val)[i + 8]);
}
return zmm2hilo(tmp1, tmp2);
}
static FORCEINLINE __vec16_i32 __cast_fptosi(__vec16_i32, __vec16_d val) {
__vec16_i32 tmp = _mm512_cvtfxpnt_roundpd_epi32lo(val.v2, _MM_ROUND_MODE_TOWARD_ZERO);
__vec16_i32 ret_hi8 = _mm512_permute4f128_epi32(tmp, _MM_PERM_BADC);
@@ -1719,7 +1829,17 @@ static FORCEINLINE __vec16_i16 __cast_fptosi(__vec16_i16, __vec16_d val) {
return ret;
}
static FORCEINLINE __vec16_i64 __cast_fptosi(__vec16_i64, __vec16_d val) {
__m512i tmp1;
for (int i = 0; i < 8; i++) {
((int64_t*)&tmp1)[i] = (int64_t)(((double*)&val.v1)[i]);
}
__m512i tmp2;
for (int i = 0; i < 8; i++) {
((int64_t*)&tmp2)[i] = (int64_t)(((double*)&val.v2)[i]);
}
return zmm2hilo(tmp1, tmp2);
}
static FORCEINLINE __vec16_i32 __cast_fptoui(__vec16_i32, __vec16_f val) {
@@ -1740,6 +1860,18 @@ static FORCEINLINE __vec16_i16 __cast_fptoui(__vec16_i16, __vec16_f val) {
return ret;
}
static FORCEINLINE __vec16_i64 __cast_fptoui(__vec16_i64, __vec16_f val) {
__m512i tmp1;
for (int i = 0; i < 8; i++) {
((uint64_t*)&tmp1)[i] = (uint64_t)(((float*)&val)[i]);
}
__m512i tmp2;
for (int i = 0; i < 8; i++) {
((uint64_t*)&tmp2)[i] = (uint64_t)(((float*)&val)[i + 8]);
}
return zmm2hilo(tmp1, tmp2);
}
static FORCEINLINE __vec16_i32 __cast_fptoui(__vec16_i32, __vec16_d val) {
__vec16_i32 tmp = _mm512_cvtfxpnt_roundpd_epu32lo(val.v2, _MM_ROUND_MODE_TOWARD_ZERO);
__vec16_i32 ret_hi8 = _mm512_permute4f128_epi32(tmp, _MM_PERM_BADC);
@@ -1761,7 +1893,17 @@ static FORCEINLINE __vec16_i16 __cast_fptoui(__vec16_i16, __vec16_d val) {
return ret;
}
static FORCEINLINE __vec16_i64 __cast_fptoui(__vec16_i64, __vec16_d val) {
__m512i tmp1;
for (int i = 0; i < 8; i++) {
((uint64_t*)&tmp1)[i] = (uint64_t)(((double*)&val.v1)[i]);
}
__m512i tmp2;
for (int i = 0; i < 8; i++) {
((uint64_t*)&tmp2)[i] = (uint64_t)(((double*)&val.v2)[i]);
}
return zmm2hilo(tmp1, tmp2);
}
@@ -1920,6 +2062,22 @@ CMP_OP(__vec16_i8, i8, int8_t, __signed_greater_than, >)
SELECT(__vec16_i8)
static FORCEINLINE int8_t __extract_element(__vec16_i8 v, uint32_t index) {
return v[index];
}
static FORCEINLINE void __insert_element(__vec16_i8 *v, uint32_t index, int8_t val) {
((int32_t *)v)[index] = val;
}
static FORCEINLINE __vec16_i8 __broadcast_i8(__vec16_i8 v, int index) {
int32_t val = __extract_element(v, index & 0xf);
__vec16_i32 tmp = _mm512_set1_epi32(val);
__vec16_i8 ret;
_mm512_extstore_epi32(&ret, tmp, _MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE);
return ret;
}
static FORCEINLINE __vec16_i1 __not_equal_i8(__vec16_i8 a, __vec16_i8 b) {
__vec16_i32 tmp_a = _mm512_extload_epi32(&a, _MM_UPCONV_EPI32_SINT8, _MM_BROADCAST32_NONE, _MM_HINT_NONE);
__vec16_i32 tmp_b = _mm512_extload_epi32(&b, _MM_UPCONV_EPI32_SINT8, _MM_BROADCAST32_NONE, _MM_HINT_NONE);
@@ -2014,6 +2172,22 @@ CMP_OP(__vec16_i16, i16, int16_t, __signed_greater_than, >)
SELECT(__vec16_i16)
static FORCEINLINE int16_t __extract_element(__vec16_i16 v, uint32_t index) {
return v[index];
}
static FORCEINLINE void __insert_element(__vec16_i16 *v, uint32_t index, int16_t val) {
((int16_t *)v)[index] = val;
}
static FORCEINLINE __vec16_i16 __broadcast_i16(__vec16_i16 v, int index) {
int32_t val = __extract_element(v, index & 0xf);
__vec16_i32 tmp = _mm512_set1_epi32(val);
__vec16_i16 ret;
_mm512_extstore_epi32(&ret, tmp, _MM_DOWNCONV_EPI32_SINT16,_MM_HINT_NONE);
return ret;
}
static FORCEINLINE __vec16_i1 __not_equal_i16(__vec16_i16 a, __vec16_i16 b) {
__vec16_i32 tmp_a = _mm512_extload_epi32(&a, _MM_UPCONV_EPI32_SINT16, _MM_BROADCAST32_NONE, _MM_HINT_NONE);
__vec16_i32 tmp_b = _mm512_extload_epi32(&b, _MM_UPCONV_EPI32_SINT16, _MM_BROADCAST32_NONE, _MM_HINT_NONE);
@@ -2221,7 +2395,7 @@ static FORCEINLINE int64_t __count_trailing_zeros_i64(const __vec1_i64 mask) {
// reductions
///////////////////////////////////////////////////////////////////////////
static FORCEINLINE int16_t __reduce_add_int8(__vec16_i8 v) {
static FORCEINLINE int8_t __reduce_add_int8(__vec16_i8 v) {
// TODO: improve this!
int16_t ret = 0;
for (int i = 0; i < 16; ++i)
@@ -2229,7 +2403,7 @@ static FORCEINLINE int16_t __reduce_add_int8(__vec16_i8 v) {
return ret;
}
static FORCEINLINE int32_t __reduce_add_int16(__vec16_i16 v) {
static FORCEINLINE int16_t __reduce_add_int16(__vec16_i16 v) {
// TODO: improve this!
int32_t ret = 0;
for (int i = 0; i < 16; ++i)
@@ -2257,6 +2431,51 @@ static FORCEINLINE uint32_t __reduce_max_uint32(__vec16_i32 v) {
return _mm512_reduce_max_epu32(v);
}
static FORCEINLINE int64_t __reduce_add_int64(__vec16_i64 v) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(v, tmp1, tmp2);
int64_t res1 = _mm512_reduce_add_epi64(tmp1);
int64_t res2 = _mm512_reduce_add_epi64(tmp2);
return res1 + res2;
}
static FORCEINLINE int64_t __reduce_min_int64(__vec16_i64 v) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(v, tmp1, tmp2);
int64_t res1 = _mm512_reduce_min_epi64(tmp1);
int64_t res2 = _mm512_reduce_min_epi64(tmp2);
return (res1 < res2) ? res1 : res2;
}
static FORCEINLINE int64_t __reduce_max_int64(__vec16_i64 v) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(v, tmp1, tmp2);
int64_t res1 = _mm512_reduce_max_epi64(tmp1);
int64_t res2 = _mm512_reduce_max_epi64(tmp2);
return (res1 > res2) ? res1 : res2;
}
static FORCEINLINE uint64_t __reduce_min_uint64(__vec16_i64 v) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(v, tmp1, tmp2);
uint64_t res1 = _mm512_reduce_min_epu64(tmp1);
uint64_t res2 = _mm512_reduce_min_epu64(tmp2);
return (res1 < res2) ? res1 : res2;
}
static FORCEINLINE uint64_t __reduce_max_uint64(__vec16_i64 v) {
__m512i tmp1;
__m512i tmp2;
hilo2zmm(v, tmp1, tmp2);
uint64_t res1 = _mm512_reduce_max_epu64(tmp1);
uint64_t res2 = _mm512_reduce_max_epu64(tmp2);
return (res1 > res2) ? res1 : res2;
}
static FORCEINLINE float __reduce_add_float(__vec16_f v) {
return _mm512_reduce_add_ps(v);
}

22
opt.cpp
View File

@@ -185,11 +185,8 @@ lCopyMetadata(llvm::Value *vto, const llvm::Instruction *from) {
if (!to)
return;
#if defined(LLVM_3_2) || defined(LLVM_3_3) || defined(LLVM_3_4) || defined(LLVM_3_5)
llvm::SmallVector<std::pair<unsigned int, llvm::MDNode *>, 8> metadata;
#else // LLVM 3.6+
llvm::SmallVector<std::pair<unsigned int, llvm::Value *>, 8> metadata;
#endif
from->getAllMetadata(metadata);
for (unsigned int i = 0; i < metadata.size(); ++i)
to->setMetadata(metadata[i].first, metadata[i].second);
@@ -219,19 +216,12 @@ lCopyMetadata(llvm::Value *vto, const llvm::Instruction *from) {
*/
static bool
lGetSourcePosFromMetadata(const llvm::Instruction *inst, SourcePos *pos) {
#if defined(LLVM_3_2) || defined(LLVM_3_3) || defined(LLVM_3_4) || defined(LLVM_3_5)
llvm::MDNode *filename = inst->getMetadata("filename");
llvm::MDNode *first_line = inst->getMetadata("first_line");
llvm::MDNode *first_column = inst->getMetadata("first_column");
llvm::MDNode *last_line = inst->getMetadata("last_line");
llvm::MDNode *last_column = inst->getMetadata("last_column");
#else // LLVM 3.6+
llvm::MDNode *filename = inst->getMDNode("filename");
llvm::MDNode *first_line = inst->getMDNode("first_line");
llvm::MDNode *first_column = inst->getMDNode("first_column");
llvm::MDNode *last_line = inst->getMDNode("last_line");
llvm::MDNode *last_column = inst->getMDNode("last_column");
#endif
if (!filename || !first_line || !first_column || !last_line || !last_column)
return false;
@@ -1210,7 +1200,11 @@ InstructionSimplifyPass::simplifyBoolVec(llvm::Value *value) {
zext->getOperand(0)->getType() == LLVMTypes::Int1VectorType)
return zext->getOperand(0);
}
/*
// This optimization has discernable benefit on the perf
// suite on latest LLVM versions.
// On 3.4+ (maybe even older), it can result in illegal
// operations, so it's being disabled.
llvm::ICmpInst *icmp = llvm::dyn_cast<llvm::ICmpInst>(value);
if (icmp != NULL) {
// icmp(ne, {sext,zext}(foo), zeroinitializer) -> foo
@@ -1226,7 +1220,9 @@ InstructionSimplifyPass::simplifyBoolVec(llvm::Value *value) {
return zext->getOperand(0);
}
}
}
*/
return NULL;
}