Add workgroup size attribute to AMDGPU functions in codegen (#4342)

When we did not set the workgroup size, LLVM will use too many registers
for kernel launches with many threads. This resulted in "invalid ISA"
errors. Here we set the maximum workgroup size to the maximum threads
per block from the device API.

Of course, one might look into allowing configurations with fewer
threads at runtime to use more registers.

diff --git a/src/codegen/llvm/codegen_amdgpu.cc b/src/codegen/llvm/codegen_amdgpu.cc
index 19179c3..a1af9bd 100644 (file)
--- a/src/codegen/llvm/codegen_amdgpu.cc
+++ b/src/codegen/llvm/codegen_amdgpu.cc
@@ -36,6 +36,29 @@
 namespace tvm {
 namespace codegen {
 
+namespace {
+
+// calls the device api to get the max threads per block
+static inline int DetectROCMmaxThreadsPerBlock() {
+  TVMContext tvm_ctx;
+  tvm_ctx.device_type = kDLROCM;
+  tvm_ctx.device_id = 0;
+  tvm::runtime::DeviceAPI* api = tvm::runtime::DeviceAPI::Get(tvm_ctx, true);
+  if (api != nullptr) {
+    TVMRetValue val;
+    api->GetAttr(tvm_ctx, tvm::runtime::kExist, &val);
+    if (val.operator int() == 1) {
+      tvm::runtime::DeviceAPI::Get(tvm_ctx)->
+        GetAttr(tvm_ctx, tvm::runtime::kMaxThreadsPerBlock, &val);
+      return val.operator int();
+    }
+  }
+  LOG(WARNING) << "Cannot get maximum number of threads for AMD codegen";
+  return 256;  // see the discussion at PR #4342 for the choice of default
+}
+
+}  // namespace
+
 // AMDGPU code generator.
 class CodeGenAMDGPU : public CodeGenLLVM {
  public:
@@ -43,6 +66,9 @@ class CodeGenAMDGPU : public CodeGenLLVM {
     // add function as void return value
     CodeGenLLVM::AddFunctionInternal(f, true);
     function_->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL);
+    std::ostringstream attr;
+    attr << "1," << DetectROCMmaxThreadsPerBlock();
+    function_->addFnAttr("amdgpu-flat-work-group-size", attr.str());
   }
 
   void VisitStmt_(const Allocate* op) final {