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

Improve implementaton of 'are both masks equal' check for AVX.

Browse Source 
Previously, we did a vector equal compare and then a movmsk, the
result of which we checked to see if it was on for all lanes.
Because masks are vectors of i32s, under AVX, the vector equal
compare required two 4-wide SSE compares and some shuffling.
Now, we do a movmsk of both masks first and then a scalar
equality comparison of those two values, which seems to generate
overall better code.
This commit is contained in:
Matt Pharr
2011-09-15 06:25:02 -07:00
parent 0848c2cc19
commit 1dedd88132
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ctx.cpp
View File

@@ -704,6 +704,7 @@ FunctionEmitContext::LaneMask(llvm::Value *v) {
llvm::Value * llvm::Value *
FunctionEmitContext::MasksAllEqual(llvm::Value *v1, llvm::Value *v2) { FunctionEmitContext::MasksAllEqual(llvm::Value *v1, llvm::Value *v2) {
#if 0
// Compare the two masks to get a vector of i1s // Compare the two masks to get a vector of i1s
llvm::Value *cmp = CmpInst(llvm::Instruction::ICmp, llvm::CmpInst::ICMP_EQ, llvm::Value *cmp = CmpInst(llvm::Instruction::ICmp, llvm::CmpInst::ICMP_EQ,
v1, v2, "v1==v2"); v1, v2, "v1==v2");
@@ -711,6 +712,12 @@ FunctionEmitContext::MasksAllEqual(llvm::Value *v1, llvm::Value *v2) {
cmp = I1VecToBoolVec(cmp); cmp = I1VecToBoolVec(cmp);
// And see if it's all on // And see if it's all on
return All(cmp); return All(cmp);
#else
llvm::Value *mm1 = LaneMask(v1);
llvm::Value *mm2 = LaneMask(v2);
return CmpInst(llvm::Instruction::ICmp, llvm::CmpInst::ICMP_EQ, mm1, mm2,
"v1==v2");
#endif
} }