[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

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295134 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index 02e47afeb91c4214d771e3331378f10b43e67ca6..0a1ab73d8dcfda545f321fc05d041e346dbb3cb0 100644 (file)
--- a/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -151,7 +151,11 @@ unsigned getMaxWorkGroupsPerCU(const FeatureBitset &Features,
                                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) {

diff --git a/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll b/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll
index 71d21b120f0fa8f9ff3ec29bde131ff77d40d5cf..93b33736b2a044f76288ebc0794c1db4cf49f4e1 100644 (file)
--- a/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll
+++ b/test/CodeGen/AMDGPU/large-work-group-promote-alloca.ll
@@ -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