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

More AVX fixes

Browse Source 
Fix RNG state initialization for 16-wide targets
Fix a number of bugs in reduce_add builtin implementations for AVX.
Fix some tests that had incorrect expected results for the 16-wide
  case.
This commit is contained in:
Matt Pharr
2011-09-06 15:53:11 -07:00
parent c76ef7b174
commit 4f451bd041
8 changed files with 24 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
builtins-avx-x2.ll
View File

@@ -232,8 +232,8 @@ define internal float @__reduce_add_float(<16 x float>) nounwind readonly always
%v1 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %va, <8 x float> %vb)
%v2 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %v1, <8 x float> %v1)
%v3 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %v2, <8 x float> %v2)
%scalar1 = extractelement <8 x float> %v2, i32 0
%scalar2 = extractelement <8 x float> %v2, i32 4
%scalar1 = extractelement <8 x float> %v3, i32 0
%scalar2 = extractelement <8 x float> %v3, i32 4
%sum = fadd float %scalar1, %scalar2
ret float %sum
}
@@ -316,7 +316,9 @@ define internal double @__reduce_add_double(<16 x double>) nounwind readonly alw
%sum0 = call <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double> %vab, <4 x double> %vcd)
%sum1 = call <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double> %sum0, <4 x double> %sum0)
%sum = extractelement <4 x double> %sum1, i32 0
%final0 = extractelement <4 x double> %sum1, i32 0
%final1 = extractelement <4 x double> %sum1, i32 2
%sum = fadd double %final0, %final1
ret double %sum
}