PaddlePaddle
diff --git a/‎cmake/cinn.cmake
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions b/‎cmake/cinn.cmake
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions
diff --git a/‎paddle/cinn/backends/CMakeLists.txt
Copy file name to clipboard
+4Lines changed: 4 additions & 0 deletions b/‎paddle/cinn/backends/CMakeLists.txt
Copy file name to clipboard
+4Lines changed: 4 additions & 0 deletions
diff --git a/‎paddle/cinn/backends/codegen_cuda_host.cc
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion b/‎paddle/cinn/backends/codegen_cuda_host.cc
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion
diff --git a/‎paddle/cinn/backends/codegen_cuda_host.h
Copy file name to clipboard
+19-6Lines changed: 19 additions & 6 deletions b/‎paddle/cinn/backends/codegen_cuda_host.h
Copy file name to clipboard
+19-6Lines changed: 19 additions & 6 deletions
diff --git a/‎paddle/cinn/backends/codegen_device_util.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions b/‎paddle/cinn/backends/codegen_device_util.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions
diff --git a/‎paddle/cinn/backends/codegen_device_util.h
Copy file name to clipboard
+10-4Lines changed: 10 additions & 4 deletions b/‎paddle/cinn/backends/codegen_device_util.h
Copy file name to clipboard
+10-4Lines changed: 10 additions & 4 deletions
diff --git a/‎paddle/cinn/backends/compiler.cc
Copy file name to clipboard
+86-8Lines changed: 86 additions & 8 deletions b/‎paddle/cinn/backends/compiler.cc
Copy file name to clipboard
+86-8Lines changed: 86 additions & 8 deletions
diff --git a/‎paddle/cinn/backends/compiler.h
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions b/‎paddle/cinn/backends/compiler.h
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions
diff --git a/‎paddle/cinn/backends/hip/CMakeLists.txt
Copy file name to clipboard
+3Lines changed: 3 additions & 0 deletions b/‎paddle/cinn/backends/hip/CMakeLists.txt
Copy file name to clipboard
+3Lines changed: 3 additions & 0 deletions
diff --git a/‎paddle/cinn/backends/hip/codegen_hip_dev.cc
Copy file name to clipboard
+35Lines changed: 35 additions & 0 deletions b/‎paddle/cinn/backends/hip/codegen_hip_dev.cc
Copy file name to clipboard
+35Lines changed: 35 additions & 0 deletions
@@ -277,6 +277,8 @@ set(PUBLISH_LIBS ON)
 if(PUBLISH_LIBS)
   set(core_includes
       "${core_includes};paddle/cinn/runtime/cuda/cinn_cuda_runtime_source.cuh")
+  set(core_includes
+      "${core_includes};paddle/cinn/runtime/hip/cinn_hip_runtime_source.h")
   set(core_includes
       "${core_includes};paddle/common/flags.h;paddle/utils/test_macros.h")
   foreach(header ${core_includes})
 
@@ -23,6 +23,10 @@ if(WITH_CUDA)
   list(APPEND srcs cuda_util.cc codegen_cuda_dev.cc)
 endif()
 
+if(WITH_ROCM)
+  add_subdirectory(hip)
+endif()
+
 if(WITH_OPENMP)
   cinn_cc_library(__x86_source_fake_lib SRCS _x86_builtin_source.cc)
 endif()
 
@@ -194,7 +194,7 @@ llvm::Value* CodeGenCudaHost::LowerGPUKernelLauncher(
   return function;
 }
 
-llvm::Value* CodeGenCudaHost::LowerCUDAKernelCall(const ir::Call* call_ir) {
+llvm::Value* CodeGenCudaHost::LowerGPUKernelCall(const ir::Call* call_ir) {
   std::vector<llvm::Value*> ll_function_args;
   std::transform(f_->arg_begin(),
                  f_->arg_end(),
 
@@ -43,11 +43,24 @@ class CodeGenCudaHost : public CodeGenHost {
   }
 
   llvm::Value *Visit(const ir::Call *op) override {
-    if (op->name == runtime::intrinsic::call_cuda_kernel) {
-      return LowerCUDAKernelCall(op);
-    } else {
-      return CodeGenHost::Visit(op);
-    }
+    return common::DefaultDeviceTarget().arch.Match(
+        [&](common::UnknownArch) { return CodeGenHost::Visit(op); },
+        [&](common::X86Arch) { return CodeGenHost::Visit(op); },
+        [&](common::ARMArch) { return CodeGenHost::Visit(op); },
+        [&](common::NVGPUArch) {
+          if (op->name == runtime::intrinsic::call_cuda_kernel) {
+            return LowerGPUKernelCall(op);
+          } else {
+            return CodeGenHost::Visit(op);
+          }
+        },
+        [&](common::HygonDCUArchHIP) {
+          if (op->name == runtime::intrinsic::call_hip_kernel) {
+            return LowerGPUKernelCall(op);
+          } else {
+            return CodeGenHost::Visit(op);
+          }
+        });
   }
 
  private:
@@ -66,7 +79,7 @@ class CodeGenCudaHost : public CodeGenHost {
    */
   llvm::Value *LowerGPUKernelLauncher(const ir::_LoweredFunc_ *func);
 
-  llvm::Value *LowerCUDAKernelCall(const ir::Call *op);
+  llvm::Value *LowerGPUKernelCall(const ir::Call *op);
 };
 
 }  // namespace backends
 
@@ -220,8 +220,9 @@ void detail::CollectBucketStrategyHostFunctionVisitor::ProcessLoweredFunc(
 #endif
       },
       [&](common::HygonDCUArchHIP) {
-        PADDLE_THROW(::common::errors::Unimplemented(
-            "CINN todo: new hardware HygonDCUArchHIP"));
+#ifdef CINN_WITH_HIP
+        shared_mem_bytes = hip::CalculateSharedMemory(func);
+#endif
       });
 
   VLOG(6) << "Add a call node for func_node->name " << func_node->name << "\n"
@@ -241,8 +242,7 @@ void detail::CollectBucketStrategyHostFunctionVisitor::ProcessLoweredFunc(
         call_kernel = runtime::intrinsic::call_cuda_kernel;
       },
       [&](common::HygonDCUArchHIP) {
-        PADDLE_THROW(::common::errors::Unimplemented(
-            "CINN todo: new hardware HygonDCUArchHIP"));
+        call_kernel = runtime::intrinsic::call_hip_kernel;
       });
   ir::Expr call_extern_api =
       ir::Call::Make(Void(),
 
@@ -22,6 +22,9 @@
 #ifdef CINN_WITH_CUDA
 #include "paddle/cinn/backends/codegen_cuda_dev.h"
 #endif
+#ifdef CINN_WITH_HIP
+#include "paddle/cinn/backends/hip/codegen_hip_dev.h"
+#endif
 #include "paddle/cinn/cinn.h"
 #include "paddle/cinn/ir/ir.h"
 #include "paddle/cinn/ir/ir_mutator.h"
@@ -130,8 +133,12 @@ struct CollectHostFunctionVisitor : public ir::IRMutator<> {
 #endif
         },
         [&](common::HygonDCUArchHIP) {
-          PADDLE_THROW(::common::errors::Unimplemented(
-              "CINN todo: new hardware HygonDCUArchHIP"));
+#ifdef CINN_WITH_HIP
+          hip::CodeGenHipDevice codegen_dev(
+              cinn::common::DefaultHygonDcuHipTarget());
+          codegen_dev.Compile(ir::LoweredFunc(func));
+          shared_mem_bytes = codegen_dev.GetDynSharedMemOffset();
+#endif
         });
 
     VLOG(6) << "Add a call node for func->name " << func->name << "\n"
@@ -152,8 +159,7 @@ struct CollectHostFunctionVisitor : public ir::IRMutator<> {
           call_kernel = runtime::intrinsic::call_cuda_kernel;
         },
         [&](common::HygonDCUArchHIP) {
-          PADDLE_THROW(::common::errors::Unimplemented(
-              "CINN todo: new hardware HygonDCUArchHIP"));
+          call_kernel = runtime::intrinsic::call_hip_kernel;
         });
 
     auto call_extern_api =
 
@@ -16,20 +16,25 @@
 
 #include <sys/stat.h>
 #include <fstream>
-
+#include "paddle/cinn/backends/codegen_cuda_host.h"
+#include "paddle/cinn/backends/codegen_device_util.h"
 #include "paddle/cinn/backends/llvm/runtime_symbol_registry.h"
 #include "paddle/cinn/common/context.h"
 #include "paddle/cinn/hlir/framework/graph_compiler_util.h"
 #include "paddle/cinn/ir/ir_printer.h"
+#include "paddle/cinn/runtime/backend_api.h"
 #ifdef CINN_WITH_CUDA
 #include "paddle/cinn/backends/codegen_cuda_dev.h"
-#include "paddle/cinn/backends/codegen_cuda_host.h"
-#include "paddle/cinn/backends/codegen_device_util.h"
 #include "paddle/cinn/backends/nvrtc/nvrtc_util.h"
 #include "paddle/cinn/runtime/cuda/cuda_module.h"
 #include "paddle/cinn/runtime/cuda/cuda_util.h"
 #include "paddle/cinn/runtime/flags.h"
 #endif
+#ifdef CINN_WITH_HIP
+#include "paddle/cinn/backends/hip/codegen_hip_dev.h"
+#include "paddle/cinn/backends/hip/compiler_hip.h"
+#include "paddle/cinn/runtime/hip/hip_module.h"
+#endif
 #include "paddle/cinn/adt/adt.h"
 
 PD_DECLARE_string(cinn_source_code_save_path);
@@ -270,8 +275,17 @@ std::string Compiler::GetSourceCode(const ir::Module& module) {
 #endif
       },
       [&](common::HygonDCUArchHIP) -> std::string {
-        PADDLE_THROW(::common::errors::Unimplemented(
-            "CINN todo: new hardware HygonDCUArchHIP"));
+#ifdef CINN_WITH_HIP
+        auto _host_module_device_module_ =
+            SplitDeviceAndHostModule(module);  // NOLINT
+        auto& host_module = std::get<0>(_host_module_device_module_);
+        auto& device_module = std::get<1>(_host_module_device_module_);
+        hip::CodeGenHipDevice codegen(target_);
+        auto source_code = codegen.Compile(device_module);
+        return source_code;
+#else
+        CINN_NOT_IMPLEMENTED
+#endif
       });
 }
 
@@ -308,7 +322,7 @@ void Compiler::RegisterDeviceModuleSymbol() {
       [&](common::X86Arch) { return; },
       [&](common::ARMArch) { return; },
       [&](common::NVGPUArch) { RegisterCudaModuleSymbol(); },
-      [&](common::HygonDCUArchHIP) { CINN_NOT_IMPLEMENTED; });
+      [&](common::HygonDCUArchHIP) { RegisterHipModuleSymbol(); });
 }
 
 void Compiler::RegisterCudaModuleSymbol() {
@@ -338,6 +352,39 @@ void Compiler::RegisterCudaModuleSymbol() {
 #endif
 }
 
+void Compiler::RegisterHipModuleSymbol() {
+#ifdef CINN_WITH_HIP
+  hiprtc::Compiler compiler;
+  std::string source_code =
+      hip::CodeGenHipDevice::GetSourceHeader() + device_fn_code_;
+  std::string hsaco = compiler(source_code);
+  PADDLE_ENFORCE_EQ(
+      !hsaco.empty(),
+      true,
+      phi::errors::Fatal("Compile hsaco failed from source code:\n%s",
+                         source_code));
+  using runtime::hip::HIPModule;
+  hip_module_.reset(new HIPModule(hsaco));
+  // get device id
+  using cinn::runtime::BackendAPI;
+  int device_id = BackendAPI::get_backend(target_)->get_device();
+  // register kernel
+  RuntimeSymbols symbols;
+  for (const auto& kernel_fn_name : device_fn_name_) {
+    auto fn_kernel = hip_module_->GetFunction(device_id, kernel_fn_name);
+    PADDLE_ENFORCE_NOT_NULL(
+        fn_kernel,
+        phi::errors::Fatal("HIP GetFunction Error: get valid kernel."));
+    fn_ptr_.push_back(reinterpret_cast<void*>(fn_kernel));
+    symbols.RegisterVar(kernel_fn_name + "_ptr_",
+                        reinterpret_cast<void*>(fn_kernel));
+  }
+  engine_->RegisterModuleRuntimeSymbols(std::move(symbols));
+#else
+  CINN_NOT_IMPLEMENTED
+#endif
+}
+
 void Compiler::CompileCudaModule(const Module& module,
                                  const std::string& code) {
 #ifdef CINN_WITH_CUDA
@@ -379,8 +426,39 @@ void Compiler::CompileCudaModule(const Module& module,
 }
 
 void Compiler::CompileHipModule(const Module& module, const std::string& code) {
-  PADDLE_THROW(::common::errors::Unimplemented(
-      "CINN todo: new hardware HygonDCUArchHIP"));
+#ifdef CINN_WITH_HIP
+  auto _host_module_device_module_ =
+      SplitDeviceAndHostModule(module);  // NOLINT
+  auto& host_module = std::get<0>(_host_module_device_module_);
+  auto& device_module = std::get<1>(_host_module_device_module_);
+  VLOG(3) << "[HIP] host module:\n" << host_module;
+  VLOG(3) << "[HIP] device module:\n" << device_module;
+  std::string source_code;
+  if (!FLAGS_cinn_debug_custom_code_path.empty()) {
+    std::string file_path = FLAGS_cinn_debug_custom_code_path;
+    source_code = GetFileContent(file_path);
+  } else if (code.empty()) {
+    hip::CodeGenHipDevice codegen(target_);
+    source_code = codegen.Compile(device_module);
+  } else {
+    source_code = code;
+  }
+  PADDLE_ENFORCE_EQ(
+      !source_code.empty(),
+      true,
+      phi::errors::Fatal("Compile HIP code failed from device module:\n%s",
+                         device_module));
+  VLOG(3) << "[HIP]:\n" << source_code;
+  SourceCodePrint::GetInstance()->write(source_code);
+  device_fn_code_ += source_code;
+  for (auto& fn : device_module.functions()) {
+    std::string kernel_fn_name = fn->name;
+    device_fn_name_.emplace_back(kernel_fn_name);
+  }
+  engine_->Link<CodeGenCUDA_Host>(host_module);
+#else
+  CINN_NOT_IMPLEMENTED
+#endif
 }
 
 void Compiler::CompileX86Module(const Module& module) {
 
@@ -29,6 +29,9 @@
 #ifdef CINN_WITH_CUDA
 #include "paddle/cinn/runtime/cuda/cuda_module.h"
 #endif
+#ifdef CINN_WITH_HIP
+#include "paddle/cinn/runtime/hip/hip_module.h"
+#endif
 
 namespace cinn {
 namespace backends {
@@ -131,6 +134,8 @@ class Compiler final {
 
   void RegisterCudaModuleSymbol();
 
+  void RegisterHipModuleSymbol();
+
   void CompileCudaModule(const ir::Module& module,
                          const std::string& code = "");
 
@@ -154,6 +159,9 @@ class Compiler final {
 #ifdef CINN_WITH_CUDA
   std::unique_ptr<runtime::cuda::CUDAModule> cuda_module_;
 #endif
+#ifdef CINN_WITH_HIP
+  std::unique_ptr<runtime::hip::HIPModule> hip_module_;
+#endif
 };
 
 }  // namespace backends
 
@@ -0,0 +1,3 @@
+core_gather_headers()
+
+gather_srcs(cinnapi_src SRCS codegen_hip_dev.cc compiler_hip.cc)
@@ -0,0 +1,35 @@
+// Copyright (c) 2024 CINN Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/cinn/backends/hip/codegen_hip_dev.h"
+
+namespace cinn {
+namespace backends {
+namespace hip {
+
+const std::string CodeGenHipDevice::source_header_ =  // NOLINT
+    R"(#include "cinn_hip_runtime_source.h"
+)";
+
+const std::string &CodeGenHipDevice::GetSourceHeader() {
+  return source_header_;
+}
+
+CodeGenHipDevice::CodeGenHipDevice(Target target) : CodeGenGpuDev(target) {}
+
+void CodeGenHipDevice::PrintIncludes() { str_ += GetSourceHeader(); }
+
+}  // namespace hip
+}  // namespace backends
+}  // namespace cinn
Original file line number	Diff line number	Diff line change
`@@ -194,7 +194,7 @@ llvm::Value* CodeGenCudaHost::LowerGPUKernelLauncher(`
`194`	`194`	`return function;`
`195`	`195`	`}`
`196`	`196`
`197`		`-llvm::Value* CodeGenCudaHost::LowerCUDAKernelCall(const ir::Call* call_ir) {`
	`197`	`+llvm::Value* CodeGenCudaHost::LowerGPUKernelCall(const ir::Call* call_ir) {`
`198`	`198`	`std::vector<llvm::Value*> ll_function_args;`
`199`	`199`	`std::transform(f_->arg_begin(),`
`200`	`200`	`f_->arg_end(),`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+core_gather_headers()`
	`2`	`+`
	`3`	`+gather_srcs(cinnapi_src SRCS codegen_hip_dev.cc compiler_hip.cc)`