diff options
author | Benjamin Segovia <segovia.benjamin@gmail.com> | 2012-04-12 18:51:11 +0000 |
---|---|---|
committer | Keith Packard <keithp@keithp.com> | 2012-08-10 16:16:18 -0700 |
commit | bfba59eb36af59440dabd2c9ec508eac8a968ab3 (patch) | |
tree | b539e3080ee7cd30af558210daec221f3e7bdec8 /backend/kernels/loop3.cl.ll | |
parent | 99d9b8998677cb870fb0fba9a48a353297821c43 (diff) |
Recompiled kernels with the new stdlib Added definition and use sets per register
Diffstat (limited to 'backend/kernels/loop3.cl.ll')
-rw-r--r-- | backend/kernels/loop3.cl.ll | 62 |
1 files changed, 0 insertions, 62 deletions
diff --git a/backend/kernels/loop3.cl.ll b/backend/kernels/loop3.cl.ll index 378357e9..984165ef 100644 --- a/backend/kernels/loop3.cl.ll +++ b/backend/kernels/loop3.cl.ll @@ -4,68 +4,6 @@ target triple = "ptx32--" %struct.big = type { [10 x i32] } -define ptx_device <2 x float> @_Z3madDv2_fS_S_(<2 x float> %a, <2 x float> %b, <2 x float> %c) nounwind readnone { -entry: - %0 = extractelement <2 x float> %a, i32 0 - %1 = extractelement <2 x float> %b, i32 0 - %2 = extractelement <2 x float> %c, i32 0 - %call = tail call ptx_device float @_Z3madfff(float %0, float %1, float %2) nounwind readnone - %vecinit = insertelement <2 x float> undef, float %call, i32 0 - %3 = extractelement <2 x float> %a, i32 1 - %4 = extractelement <2 x float> %b, i32 1 - %5 = extractelement <2 x float> %c, i32 1 - %call1 = tail call ptx_device float @_Z3madfff(float %3, float %4, float %5) nounwind readnone - %vecinit2 = insertelement <2 x float> %vecinit, float %call1, i32 1 - ret <2 x float> %vecinit2 -} - -declare ptx_device float @_Z3madfff(float, float, float) nounwind readnone - -define ptx_device <3 x float> @_Z3madDv3_fS_S_(<3 x float> %a, <3 x float> %b, <3 x float> %c) nounwind readnone { -entry: - %0 = extractelement <3 x float> %a, i32 0 - %1 = extractelement <3 x float> %b, i32 0 - %2 = extractelement <3 x float> %c, i32 0 - %call = tail call ptx_device float @_Z3madfff(float %0, float %1, float %2) nounwind readnone - %vecinit = insertelement <3 x float> undef, float %call, i32 0 - %3 = extractelement <3 x float> %a, i32 1 - %4 = extractelement <3 x float> %b, i32 1 - %5 = extractelement <3 x float> %c, i32 1 - %call1 = tail call ptx_device float @_Z3madfff(float %3, float %4, float %5) nounwind readnone - %vecinit2 = insertelement <3 x float> %vecinit, float %call1, i32 1 - %6 = extractelement <3 x float> %a, i32 2 - %7 = extractelement <3 x float> %b, i32 2 - %8 = extractelement <3 x float> %c, i32 2 - %call3 = tail call ptx_device float @_Z3madfff(float %6, float %7, float %8) nounwind readnone - %vecinit4 = insertelement <3 x float> %vecinit2, float %call3, i32 2 - ret <3 x float> %vecinit4 -} - -define ptx_device <4 x float> @_Z3madDv4_fS_S_(<4 x float> %a, <4 x float> %b, <4 x float> %c) nounwind readnone { -entry: - %0 = extractelement <4 x float> %a, i32 0 - %1 = extractelement <4 x float> %b, i32 0 - %2 = extractelement <4 x float> %c, i32 0 - %call = tail call ptx_device float @_Z3madfff(float %0, float %1, float %2) nounwind readnone - %vecinit = insertelement <4 x float> undef, float %call, i32 0 - %3 = extractelement <4 x float> %a, i32 1 - %4 = extractelement <4 x float> %b, i32 1 - %5 = extractelement <4 x float> %c, i32 1 - %call1 = tail call ptx_device float @_Z3madfff(float %3, float %4, float %5) nounwind readnone - %vecinit2 = insertelement <4 x float> %vecinit, float %call1, i32 1 - %6 = extractelement <4 x float> %a, i32 2 - %7 = extractelement <4 x float> %b, i32 2 - %8 = extractelement <4 x float> %c, i32 2 - %call3 = tail call ptx_device float @_Z3madfff(float %6, float %7, float %8) nounwind readnone - %vecinit4 = insertelement <4 x float> %vecinit2, float %call3, i32 2 - %9 = extractelement <4 x float> %a, i32 3 - %10 = extractelement <4 x float> %b, i32 3 - %11 = extractelement <4 x float> %c, i32 3 - %call5 = tail call ptx_device float @_Z3madfff(float %9, float %10, float %11) nounwind readnone - %vecinit6 = insertelement <4 x float> %vecinit4, float %call5, i32 3 - ret <4 x float> %vecinit6 -} - define ptx_kernel void @add(i32 addrspace(1)* nocapture %dst, i32 %x, %struct.big* nocapture byval %b) nounwind noinline { entry: %cmp2 = icmp eq i32 %x, 0 |