[llvm] r295134 - [AMDGPU] Fix MaxWorkGroupsPerCU for large workgroups

[Stanislav Mekhanoshin via llvm-commits]

Author: rampitec
Date: Tue Feb 14 19:03:59 2017
New Revision: 295134

URL: http://llvm.org/viewvc/llvm-project?rev=295134&view=rev
Log:
[AMDGPU] Fix MaxWorkGroupsPerCU for large workgroups

This patch corrects the maximum workgroups per CU if we have big
workgroups (more than 128). This calculation contributes to the
occupancy calculation in respect to LDS size.

Differential Revision: https://reviews.llvm.org/D29974

Modified:
    llvm/trunk/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
    llvm/trunk/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll

Modified: llvm/trunk/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp?rev=295134&r1=295133&r2=295134&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp (original)
+++ llvm/trunk/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp Tue Feb 14 19:03:59 2017
@@ -151,7 +151,11 @@ unsigned getMaxWorkGroupsPerCU(const Fea
                                unsigned FlatWorkGroupSize) {
   if (!Features.test(FeatureGCN))
     return 8;
-  return getWavesPerWorkGroup(Features, FlatWorkGroupSize) == 1 ? 40 : 16;
+  unsigned N = getWavesPerWorkGroup(Features, FlatWorkGroupSize);
+  if (N == 1)
+    return 40;
+  N = 40 / N;
+  return std::min(N, 16u);
 }
 
 unsigned getMaxWavesPerCU(const FeatureBitset &Features) {

Modified: llvm/trunk/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll?rev=295134&r1=295133&r2=295134&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll (original)
+++ llvm/trunk/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll Tue Feb 14 19:03:59 2017
@@ -69,7 +69,8 @@ entry:
 }
 
 ; ALL-LABEL: @occupancy_0(
-; ALL: alloca [5 x i32]
+; CI-NOT: alloca [5 x i32]
+; SI: alloca [5 x i32]
 define void @occupancy_0(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) #3 {
 entry:
   %stack = alloca [5 x i32], align 4
@@ -91,7 +92,8 @@ entry:
 }
 
 ; ALL-LABEL: @occupancy_max(
-; ALL: alloca [5 x i32]
+; CI-NOT: alloca [5 x i32]
+; SI: alloca [5 x i32]
 define void @occupancy_max(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) #4 {
 entry:
   %stack = alloca [5 x i32], align 4