[PATCH] D92628: [HIP] Fix bug in driver about wavefront size

Yaxun Liu via Phabricator via cfe-commits cfe-commits at lists.llvm.org
Thu Dec 3 19:01:42 PST 2020


yaxunl created this revision.
yaxunl added reviewers: tra, arsenm.
Herald added subscribers: kerbowa, nhaehnle, jvesely.
yaxunl requested review of this revision.
Herald added a subscriber: wdng.

The static variable causes it only initialized once and take
the same value for different GPU archs, whereas they
may be different for different GPU archs, e.g. when
there are both gfx900 and gfx1010.

Removing `static` fixes that.


https://reviews.llvm.org/D92628

Files:
  clang/lib/Driver/ToolChains/AMDGPU.cpp
  clang/test/Driver/hip-wavefront-size.hip


Index: clang/test/Driver/hip-wavefront-size.hip
===================================================================
--- /dev/null
+++ clang/test/Driver/hip-wavefront-size.hip
@@ -0,0 +1,21 @@
+// REQUIRES: clang-driver,amdgpu-registered-target
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN:   --cuda-gpu-arch=gfx900 \
+// RUN:   --rocm-path=%S/Inputs/rocm --cuda-device-only %s \
+// RUN:   2>&1 | FileCheck %s --check-prefixes=WAVE64
+// WAVE64: "-mlink-builtin-bitcode" "{{.*}}oclc_wavefrontsize64_on.bc"{{.*}} "-target-cpu" "gfx900"
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN:   --cuda-gpu-arch=gfx1010 \
+// RUN:   --rocm-path=%S/Inputs/rocm --cuda-device-only %s \
+// RUN:   2>&1 | FileCheck %s --check-prefixes=WAVE32
+// WAVE32: "-mlink-builtin-bitcode" "{{.*}}oclc_wavefrontsize64_off.bc"{{.*}} "-target-cpu" "gfx1010"
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN:   --cuda-gpu-arch=gfx1010 \
+// RUN:   --cuda-gpu-arch=gfx900 \
+// RUN:   --rocm-path=%S/Inputs/rocm --cuda-device-only %s \
+// RUN:   2>&1 | FileCheck %s --check-prefixes=BOTH
+// BOTH-DAG: "-mlink-builtin-bitcode" "{{.*}}oclc_wavefrontsize64_on.bc"{{.*}} "-target-cpu" "gfx900"
+// BOTH-DAG: "-mlink-builtin-bitcode" "{{.*}}oclc_wavefrontsize64_off.bc"{{.*}} "-target-cpu" "gfx1010"
Index: clang/lib/Driver/ToolChains/AMDGPU.cpp
===================================================================
--- clang/lib/Driver/ToolChains/AMDGPU.cpp
+++ clang/lib/Driver/ToolChains/AMDGPU.cpp
@@ -499,7 +499,7 @@
 bool AMDGPUToolChain::isWave64(const llvm::opt::ArgList &DriverArgs,
                                llvm::AMDGPU::GPUKind Kind) {
   const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind);
-  static bool HasWave32 = (ArchAttr & llvm::AMDGPU::FEATURE_WAVE32);
+  bool HasWave32 = (ArchAttr & llvm::AMDGPU::FEATURE_WAVE32);
 
   return !HasWave32 || DriverArgs.hasFlag(
     options::OPT_mwavefrontsize64, options::OPT_mno_wavefrontsize64, false);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D92628.309439.patch
Type: text/x-patch
Size: 1978 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20201204/4307832f/attachment.bin>


More information about the cfe-commits mailing list