fix(mge): add param RuntimeArgs to customop kernel on cuda

GitOrigin-RevId: 7ed44c42de

src/CMakeLists.txt

@@ -90,6 +90,15 @@ endif()
 if(MGE_WITH_CUSTOM_OP)
   list(APPEND MGB_INC ${CMAKE_CURRENT_LIST_DIR}/custom/include)
   file(GLOB_RECURSE SOURCES_ custom/impl/*.cpp)
+
+  set(EXCLUDE_PLATFORM_DIR "custom/impl/platform")
+  foreach(CUSOURCE ${SOURCES_})
+    string(FIND ${CUSOURCE} ${EXCLUDE_PLATFORM_DIR} EXCLUDE_DIR_FOUND)
+    if(NOT ${EXCLUDE_DIR_FOUND} EQUAL -1)
+      list(REMOVE_ITEM SOURCES_ ${CUSOURCE})
+    endif()
+  endforeach(CUSOURCE)
+
   list(APPEND SOURCES ${SOURCES_})
 endif()
 

src/custom/impl/op.cpp

@@ -6,6 +6,7 @@
 #include <unordered_set>
 #include "megbrain/custom/op.h"
 #include "megbrain/custom/utils.h"
+#include "megbrain/utils/thin/function.h"
 
 using namespace mgb;
 
@@ -99,40 +100,6 @@ std::string ArgInfo::str() const {
             (arg_info).name().c_str(), static_cast<int>((arg_info).ndim()),           \
             static_cast<int>((real_shape).ndim()))
 
-template <typename T>
-class Function;
-
-template <typename RType, typename... Args>
-class Function<RType(Args...)> {
-public:
-    using Functor = RType (*)(Args...);
-
-    Function() = default;
-    Function(Functor f) : m_f(f) {}
-    Function(const Function& rhs) { m_f = rhs.m_f; }
-
-    RType operator()(Args... args) {
-        custom_assert(m_f != nullptr, "invalid function ptr\n");
-        return m_f(std::forward<Args>(args)...);
-    }
-
-    void operator=(const Function& rhs) {  // not allowed continuous assignment
-        m_f = rhs.m_f;
-    }
-
-    void operator=(const Functor f) { m_f = f; }
-
-private:
-    Functor m_f = nullptr;
-};
-
-template <typename Functions>
-class FuncWithSig : public Functions {
-public:
-    using Functions::operator();
-    using Functions::operator=;
-};
-
 class CustomOpImpl {
     static constexpr uint32_t CURRENT_VERSION = CUSTOM_OP_VERSION;
     const uint32_t m_version;
@@ -143,29 +110,26 @@ class CustomOpImpl {
     std::vector<ArgInfo> m_output_infos;
     ParamInfo m_param_infos;
 
-    using DeviceInfer = FuncWithSig<Function<void(
-            const std::vector<Device>&, const Param&, std::vector<Device>&)>>;
-    using ShapeInfer = FuncWithSig<Function<void(
-            const std::vector<Shape>&, const Param&, std::vector<Shape>&)>>;
-    using DTypeInfer = FuncWithSig<Function<void(
-            const std::vector<DType>&, const Param&, std::vector<DType>&)>>;
-    using FormatInfer = FuncWithSig<Function<void(
-            const std::vector<Format>&, const Param&, std::vector<Format>&)>>;
-    using Preprocess = FuncWithSig<Function<void(
-            const std::vector<Tensor>&, const Param&, std::vector<Tensor>&)>>;
-    using Postprocess = FuncWithSig<Function<void(
-            const std::vector<Tensor>&, const Param&, std::vector<Tensor>&)>>;
-    using Compute = FuncWithSig<Function<void(
-            const std::vector<Tensor>&, const Param&, std::vector<Tensor>&)>>;
+    using DeviceInfer = thin_function<void(
+            const std::vector<Device>&, const Param&, std::vector<Device>&)>;
+    using ShapeInfer = thin_function<void(
+            const std::vector<Shape>&, const Param&, std::vector<Shape>&)>;
+    using DTypeInfer = thin_function<void(
+            const std::vector<DType>&, const Param&, std::vector<DType>&)>;
+    using FormatInfer = thin_function<void(
+            const std::vector<Format>&, const Param&, std::vector<Format>&)>;
+    using Process = thin_function<void(
+            const std::vector<Tensor>&, const Param&, std::vector<Tensor>&,
+            const RuntimeArgs&)>;
 
     DeviceInfer infer_output_device_func;
     ShapeInfer infer_output_shape_func;
     DTypeInfer infer_output_dtype_func;
     FormatInfer infer_output_format_func;
 
-    std::unordered_map<std::string, Compute> compute_funcs;
-    std::unordered_map<std::string, Preprocess> preprocess_funcs;
-    std::unordered_map<std::string, Postprocess> postprocess_funcs;
+    std::unordered_map<std::string, Process> compute_funcs;
+    std::unordered_map<std::string, Process> preprocess_funcs;
+    std::unordered_map<std::string, Process> postprocess_funcs;
 
 public:
     CustomOpImpl(const std::string&, uint32_t version);
@@ -215,7 +179,8 @@ CustomOpImpl::CustomOpImpl(const std::string& op_type, uint32_t version)
 
     for (const auto& device : Device::legal_devices()) {
         compute_funcs[device] = [](const std::vector<Tensor>&, const Param&,
-                                   std::vector<Tensor>& outputs) -> void {
+                                   std::vector<Tensor>& outputs,
+                                   const RuntimeArgs&) -> void {
             auto device = outputs[0].device();
             mgb_assert(
                     false,
@@ -224,9 +189,11 @@ CustomOpImpl::CustomOpImpl(const std::string& op_type, uint32_t version)
                     device.str().c_str());
         };
         preprocess_funcs[device] = [](const std::vector<Tensor>&, const Param&,
-                                      std::vector<Tensor>&) -> void { return; };
+                                      std::vector<Tensor>&,
+                                      const RuntimeArgs&) -> void { return; };
         postprocess_funcs[device] = [](const std::vector<Tensor>&, const Param&,
-                                       std::vector<Tensor>&) -> void { return; };
+                                       std::vector<Tensor>&,
+                                       const RuntimeArgs&) -> void { return; };
     }
     m_param_infos.set_tag(op_type);
 }
@@ -256,33 +223,78 @@ CustomOp& CustomOp::set_format_infer(FormatInferFuncPtr func) {
     return *this;
 }
 
-CustomOp& CustomOp::set_preprocess(PreprocessFuncPtr func) {
+CustomOp& CustomOp::set_preprocess(ProcessFuncPtrWithoutRuntimeArgs func) {
+    set_preprocess("x86", func);
+    return *this;
+}
+
+CustomOp& CustomOp::set_preprocess(
+        const std::string& device, ProcessFuncPtrWithoutRuntimeArgs func) {
+    auto wrap_func = [func](const std::vector<Tensor>& input, const Param& param,
+                            std::vector<Tensor>& output, const RuntimeArgs&) -> void {
+        return func(input, param, output);
+    };
+
+    OpImplRef(m_impl.get())->preprocess_funcs[device] = wrap_func;
+    return *this;
+}
+
+CustomOp& CustomOp::set_preprocess(ProcessFuncPtr func) {
     set_preprocess("x86", func);
     return *this;
 }
 
-CustomOp& CustomOp::set_preprocess(const std::string& device, PreprocessFuncPtr func) {
+CustomOp& CustomOp::set_preprocess(const std::string& device, ProcessFuncPtr func) {
     OpImplRef(m_impl.get())->preprocess_funcs[device] = func;
     return *this;
 }
 
-CustomOp& CustomOp::set_postprocess(PostprocessFuncPtr func) {
+CustomOp& CustomOp::set_postprocess(ProcessFuncPtrWithoutRuntimeArgs func) {
     set_postprocess("x86", func);
     return *this;
 }
 
 CustomOp& CustomOp::set_postprocess(
-        const std::string& device, PostprocessFuncPtr func) {
+        const std::string& device, ProcessFuncPtrWithoutRuntimeArgs func) {
+    auto wrap_func = [func](const std::vector<Tensor>& input, const Param& param,
+                            std::vector<Tensor>& output,
+                            const RuntimeArgs&) -> void { func(input, param, output); };
+
+    OpImplRef(m_impl.get())->postprocess_funcs[device] = wrap_func;
+    return *this;
+}
+
+CustomOp& CustomOp::set_postprocess(ProcessFuncPtr func) {
+    set_postprocess("x86", func);
+    return *this;
+}
+
+CustomOp& CustomOp::set_postprocess(const std::string& device, ProcessFuncPtr func) {
     OpImplRef(m_impl.get())->postprocess_funcs[device] = func;
     return *this;
 }
 
-CustomOp& CustomOp::set_compute(ComputeFuncPtr func) {
+CustomOp& CustomOp::set_compute(ProcessFuncPtrWithoutRuntimeArgs func) {
+    set_compute("x86", func);
+    return *this;
+}
+
+CustomOp& CustomOp::set_compute(
+        const std::string& device, ProcessFuncPtrWithoutRuntimeArgs func) {
+    auto wrap_func = [func](const std::vector<Tensor>& input, const Param& param,
+                            std::vector<Tensor>& output,
+                            const RuntimeArgs&) -> void { func(input, param, output); };
+
+    OpImplRef(m_impl.get())->compute_funcs[device] = wrap_func;
+    return *this;
+}
+
+CustomOp& CustomOp::set_compute(ProcessFuncPtr func) {
     set_compute("x86", func);
     return *this;
 }
 
-CustomOp& CustomOp::set_compute(const std::string& device, ComputeFuncPtr func) {
+CustomOp& CustomOp::set_compute(const std::string& device, ProcessFuncPtr func) {
     OpImplRef(m_impl.get())->compute_funcs[device] = func;
     return *this;
 }
@@ -513,23 +525,28 @@ void CustomOp::compute(
         return;
     }
 
-    std::string device = outputs[0].device().str();
+    Device device = outputs[0].device();
+    std::string device_str = device.str();
     for (size_t i = 1; i < outputs.size(); ++i) {
         mgb_assert(
-                outputs[i].device().str() == device,
+                outputs[i].device().str() == device_str,
                 "all output tensors should have the same device attribute");
     }
 
     // need to add other input/output check
-    mgb_assert(Device::is_legal(device), "unsupported device type: %s", device.c_str());
+    mgb_assert(
+            Device::is_legal(device_str), "unsupported device type: %s",
+            device_str.c_str());
+
+    auto preprocess_func = OpImplRef(m_impl.get())->preprocess_funcs[device_str];
+    auto forward_func = OpImplRef(m_impl.get())->compute_funcs[device_str];
+    auto postprocess_func = OpImplRef(m_impl.get())->postprocess_funcs[device_str];
 
-    auto preprocess_func = OpImplRef(m_impl.get())->preprocess_funcs[device];
-    auto forward_func = OpImplRef(m_impl.get())->compute_funcs[device];
-    auto postprocess_func = OpImplRef(m_impl.get())->postprocess_funcs[device];
+    RuntimeArgs rt_args(device);
 
-    preprocess_func(inputs, param, outputs);
-    forward_func(inputs, param, outputs);
-    postprocess_func(outputs, param, outputs);
+    preprocess_func(inputs, param, outputs, rt_args);
+    forward_func(inputs, param, outputs, rt_args);
+    postprocess_func(outputs, param, outputs, rt_args);
     assert_outputs_size_right(outputs);
 }
 

src/custom/impl/platform/custom_cuda.cpp

@@ -0,0 +1,21 @@
+#include "megbrain/common.h"
+
+#include "megbrain/comp_node_env.h"
+#include "megbrain/custom/data_adaptor.h"
+#include "megbrain/custom/platform/custom_cuda.h"
+
+using namespace mgb;
+
+namespace custom {
+
+const CudaRuntimeArgs get_cuda_runtime_args(const RuntimeArgs& rt_args) {
+    mgb_assert(
+            rt_args.device().enumv() == DeviceEnum::cuda,
+            "devive type should be cuda.");
+    const CompNodeEnv& env =
+            CompNodeEnv::from_comp_node(to_builtin<CompNode, Device>(rt_args.device()));
+    const CompNodeEnv::CudaEnv& cuda_env = env.cuda_env();
+    return {cuda_env.device, cuda_env.stream};
+}
+
+}  // namespace custom

src/custom/include/megbrain/custom/op.h

@@ -36,6 +36,18 @@ class MGE_WIN_DECLSPEC_FUC ArgInfo {
     std::string str() const;
 };
 
+class CudaRuntimeArgs;
+
+class MGE_WIN_DECLSPEC_FUC RuntimeArgs {
+    Device m_device;
+
+public:
+    RuntimeArgs() = default;
+    RuntimeArgs(Device device) : m_device(device){};
+
+    const Device& device() const { return m_device; }
+};
+
 class MGE_WIN_DECLSPEC_FUC CustomOp {
     std::unique_ptr<void, void_deleter> m_impl;
 
@@ -51,11 +63,10 @@ public:
             void (*)(const std::vector<DType>&, const Param&, std::vector<DType>&);
     using FormatInferFuncPtr =
             void (*)(const std::vector<Format>&, const Param&, std::vector<Format>&);
-    using PreprocessFuncPtr =
-            void (*)(const std::vector<Tensor>&, const Param&, std::vector<Tensor>&);
-    using PostprocessFuncPtr =
-            void (*)(const std::vector<Tensor>&, const Param&, std::vector<Tensor>&);
-    using ComputeFuncPtr =
+    using ProcessFuncPtr = void (*)(
+            const std::vector<Tensor>&, const Param&, std::vector<Tensor>&,
+            const RuntimeArgs&);
+    using ProcessFuncPtrWithoutRuntimeArgs =
             void (*)(const std::vector<Tensor>&, const Param&, std::vector<Tensor>&);
 
     // write for forward
@@ -63,12 +74,24 @@ public:
     CustomOp& set_shape_infer(ShapeInferFuncPtr func);
     CustomOp& set_dtype_infer(DTypeInferFuncPtr func);
     CustomOp& set_format_infer(FormatInferFuncPtr func);
-    CustomOp& set_preprocess(PreprocessFuncPtr func);
-    CustomOp& set_preprocess(const std::string& device, PreprocessFuncPtr func);
-    CustomOp& set_postprocess(PostprocessFuncPtr func);
-    CustomOp& set_postprocess(const std::string& device, PostprocessFuncPtr func);
-    CustomOp& set_compute(ComputeFuncPtr func);
-    CustomOp& set_compute(const std::string& device, ComputeFuncPtr func);
+    //! set process function with RuntimeArgs e.g. cuda
+    CustomOp& set_preprocess(ProcessFuncPtr func);
+    CustomOp& set_preprocess(const std::string& device, ProcessFuncPtr func);
+    CustomOp& set_postprocess(ProcessFuncPtr func);
+    CustomOp& set_postprocess(const std::string& device, ProcessFuncPtr func);
+    CustomOp& set_compute(ProcessFuncPtr func);
+    CustomOp& set_compute(const std::string& device, ProcessFuncPtr func);
+
+    //! set process function without RuntimeArgs e.g. cpu
+    CustomOp& set_preprocess(ProcessFuncPtrWithoutRuntimeArgs func);
+    CustomOp& set_preprocess(
+            const std::string& device, ProcessFuncPtrWithoutRuntimeArgs func);
+    CustomOp& set_postprocess(ProcessFuncPtrWithoutRuntimeArgs func);
+    CustomOp& set_postprocess(
+            const std::string& device, ProcessFuncPtrWithoutRuntimeArgs func);
+    CustomOp& set_compute(ProcessFuncPtrWithoutRuntimeArgs func);
+    CustomOp& set_compute(
+            const std::string& device, ProcessFuncPtrWithoutRuntimeArgs func);
 
     CustomOp& set_description(const std::string& op_desc);
     CustomOp& add_input(

src/custom/include/megbrain/custom/platform/custom_cuda.h

@@ -0,0 +1,25 @@
+#pragma once
+
+#include "megbrain/custom/op.h"
+
+#include <cuda_runtime_api.h>
+
+namespace custom {
+
+class CudaRuntimeArgs {
+private:
+    int m_device;
+    cudaStream_t m_stream;
+
+public:
+    CudaRuntimeArgs(int device, cudaStream_t stream)
+            : m_device(device), m_stream(stream) {}
+
+    int device() const { return m_device; }
+
+    cudaStream_t stream() const { return m_stream; }
+};
+
+const CudaRuntimeArgs get_cuda_runtime_args(const RuntimeArgs& rt_args);
+
+}  // namespace custom

src/custom/test/op.cpp

@@ -119,6 +119,34 @@ void gpu_kernel(
     ASSERT_TRUE(params["device"] == "cuda");
 }
 
+void cpu_kernel_with_runtime_args(
+        const std::vector<Tensor>& inputs, const Param& params,
+        std::vector<Tensor>& outputs, const RuntimeArgs& args) {
+    (void)inputs;
+    (void)params;
+    (void)outputs;
+    (void)args;
+#if OP_TEST_LOG
+    std::cout << "Checking CPU Forward - " << params["device"].as<std::string>()
+              << std::endl;
+#endif
+    ASSERT_TRUE(params["device"] == "x86");
+}
+
+void gpu_kernel_with_runtime_args(
+        const std::vector<Tensor>& inputs, const Param& params,
+        std::vector<Tensor>& outputs, const RuntimeArgs& args) {
+    (void)inputs;
+    (void)params;
+    (void)outputs;
+    (void)args;
+#if OP_TEST_LOG
+    std::cout << "Checking GPU Forward - " << params["device"].as<std::string>()
+              << std::endl;
+#endif
+    ASSERT_TRUE(params["device"] == "cuda");
+}
+
 TEST(TestCustomOp, TestCustomOpFuncSetter) {
 #if MGB_CUDA
     CustomOp test("TestOp", CUSTOM_OP_VERSION);
@@ -179,6 +207,7 @@ TEST(TestCustomOp, TestCustomOpFuncSetter) {
     ASSERT_TRUE(iformats[0].is_default());
     ASSERT_TRUE(iformats[1].is_default());
 
+    test.set_compute(cpu_kernel_with_runtime_args);
     test.set_compute(cpu_kernel);
     DeviceTensorND cdev_itensor0(CompNode::load("cpux"), {3, 2}, dtype::Int32{});
     DeviceTensorND cdev_itensor1(CompNode::load("cpux"), {3, 2}, dtype::Float32{});
@@ -192,6 +221,7 @@ TEST(TestCustomOp, TestCustomOpFuncSetter) {
     param["device"] = "x86";
     test.compute(cinputs, param, coutputs);
 
+    test.set_compute("cuda", gpu_kernel_with_runtime_args);
     test.set_compute("cuda", gpu_kernel);
     DeviceTensorND gdev_itensor0(CompNode::load("gpux"), {3, 2}, dtype::Int32{});
     DeviceTensorND gdev_itensor1(CompNode::load("gpux"), {3, 2}, dtype::Float32{});