issue/87:增加沐曦平台的handle

src/infiniop/devices/handle.cc

@@ -17,6 +17,9 @@
 #ifdef ENABLE_KUNLUN_API
 #include "kunlun/kunlun_handle.h"
 #endif
+#ifdef ENABLE_MACA_API
+#include "maca/maca_handle.h"
+#endif
 
 __C infiniStatus_t infiniopCreateHandle(infiniopHandle_t *handle_ptr) {
     if (handle_ptr == nullptr) {
@@ -53,6 +56,9 @@ __C infiniStatus_t infiniopCreateHandle(infiniopHandle_t *handle_ptr) {
         return createKunlunHandle((infiniopKunlunHandle_t *)handle_ptr);
     }
 #endif
+#ifdef ENABLE_MACA_API
+        CREATE(INFINI_DEVICE_METAX, maca);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
@@ -89,6 +95,9 @@ __C infiniStatus_t infiniopDestroyHandle(infiniopHandle_t handle) {
     case INFINI_DEVICE_KUNLUN: {
         return destroyKunlunHandle((infiniopKunlunHandle_t)handle);
     }
+#endif
+#ifdef ENABLE_MACA_API
+        DELETE(INFINI_DEVICE_METAX, maca);
 #endif
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;

src/infiniop/devices/maca/common_maca.h

+#include "../../../utils.h"
+#include "../pool.h"
+#include "maca_handle.h"
+#include <hcblas/hcblas.h>
+#include <hcdnn/hcdnn.h>
+#include <memory>
+
+#define CHECK_MCBLAS(API) CHECK_INTERNAL(API, HCBLAS_STATUS_SUCCESS)
+#define CHECK_MCDNN(API) CHECK_INTERNAL(API, HCDNN_STATUS_SUCCESS)
+
+namespace device::maca {
+
+class Handle::Internal {
+    Pool<hcblasHandle_t> mcblas_handles;
+    Pool<hcdnnHandle_t> mcdnn_handles;
+
+    template <typename T>
+    using Fn = std::function<infiniStatus_t(T)>;
+
+public:
+    infiniStatus_t use_mcblas(hcStream_t stream, const Fn<hcblasHandle_t> &f) const;
+    infiniStatus_t use_mcdnn(hcStream_t stream, const Fn<hcdnnHandle_t> &f) const;
+};
+
+hcdnnDataType_t getHcdnnDtype(infiniDtype_t dt);
+
+} // namespace device::maca

src/infiniop/devices/maca/maca_handle.cc

+#include "common_maca.h"
+
+namespace device::maca {
+Handle::Handle(infiniDevice_t device, int device_id)
+    : InfiniopHandle{device, device_id},
+      _internal(std::make_shared<Handle::Internal>()) {}
+
+auto Handle::internal() const -> const std::shared_ptr<Internal> & {
+    return _internal;
+}
+
+infiniStatus_t Handle::Internal::use_mcblas(hcStream_t stream, const Fn<hcblasHandle_t> &f) const {
+    auto handle = mcblas_handles.pop();
+    if (!handle) {
+        CHECK_MCBLAS(hcblasCreate(&(*handle)));
+    }
+    CHECK_MCBLAS(hcblasSetStream(*handle, stream));
+    CHECK_STATUS(f(*handle));
+    mcblas_handles.push(std::move(*handle));
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t Handle::Internal::use_mcdnn(hcStream_t stream, const Fn<hcdnnHandle_t> &f) const {
+    auto handle = mcdnn_handles.pop();
+    if (!handle) {
+        CHECK_MCDNN(hcdnnCreate(&(*handle)));
+    }
+    CHECK_MCDNN(hcdnnSetStream(*handle, stream));
+    CHECK_STATUS(f(*handle));
+    mcdnn_handles.push(std::move(*handle));
+    return INFINI_STATUS_SUCCESS;
+}
+
+hcdnnDataType_t getHcdnnDtype(infiniDtype_t dt) {
+    switch (dt) {
+    case INFINI_DTYPE_F16:
+        return HCDNN_DATA_HALF;
+    case INFINI_DTYPE_F32:
+        return HCDNN_DATA_FLOAT;
+    case INFINI_DTYPE_F64:
+        return HCDNN_DATA_DOUBLE;
+    case INFINI_DTYPE_BF16:
+        return HCDNN_DATA_BFLOAT16;
+    case INFINI_DTYPE_I8:
+        return HCDNN_DATA_INT8;
+    case INFINI_DTYPE_I32:
+        return HCDNN_DATA_INT32;
+    case INFINI_DTYPE_I64:
+        return HCDNN_DATA_INT64;
+    case INFINI_DTYPE_U8:
+        return HCDNN_DATA_UINT8;
+    default:
+        return HCDNN_DATA_FLOAT;
+    }
+}
+
+infiniStatus_t Handle::create(InfiniopHandle **handle_ptr, int device_id) {
+    *handle_ptr = new Handle(INFINI_DEVICE_METAX, device_id);
+    return INFINI_STATUS_SUCCESS;
+}
+
+} // namespace device::maca

src/infiniop/devices/maca/maca_handle.h

+#ifndef __INFINIOP_MACA_HANDLE_H__
+#define __INFINIOP_MACA_HANDLE_H__
+
+#include "../../handle.h"
+#include <memory>
+
+namespace device::maca {
+struct Handle : public InfiniopHandle {
+    class Internal;
+    auto internal() const -> const std::shared_ptr<Internal> &;
+
+public:
+    static infiniStatus_t create(InfiniopHandle **handle_ptr, int device_id);
+
+protected:
+    Handle(infiniDevice_t device, int device_id);
+
+private:
+    std::shared_ptr<Internal> _internal;
+};
+
+} // namespace device::maca
+
+#endif // __INFINIOP_MACA_HANDLE_H__

xmake.lua

@@ -75,6 +75,7 @@ option_end()
 
 if has_config("metax-gpu") then
     add_defines("ENABLE_MACA_API")
+    includes("xmake/maca.lua")
 end
 
 -- 摩尔线程
@@ -169,7 +170,7 @@ target("infiniop")
         add_deps("infiniop-ascend")
     end
     if has_config("metax-gpu") then
-        add_deps("metax-gpu")
+        add_deps("infiniop-metax")
     end
     if has_config("kunlun-xpu") then
         add_deps("infiniop-kunlun")

xmake/maca.lua

+
+local MACA_ROOT = os.getenv("MACA_PATH") or os.getenv("MACA_HOME") or os.getenv("MACA_ROOT")
+
+add_includedirs(MACA_ROOT .. "/include")
+add_linkdirs(MACA_ROOT .. "/lib")
+add_links("libhcdnn.so")
+add_links("libhcblas.so")
+add_links("libhcruntime.so")
+
+rule("maca")
+    set_extensions(".maca")
+
+    on_load(function (target)
+        target:add("includedirs", "include")
+    end)
+
+    on_build_file(function (target, sourcefile)
+        local objectfile = target:objectfile(sourcefile)
+        os.mkdir(path.directory(objectfile))
+        local htcc = path.join(MACA_ROOT, "htgpu_llvm/bin/htcc")
+        local includedirs = table.concat(target:get("includedirs"), " ")
+
+        local args = { "-x", "hpcc", "-c", sourcefile, "-o", objectfile, "-I" .. MACA_ROOT .. "/include", "-O3", "-fPIC", "-Werror", "-std=c++17"}
+
+        for _, includedir in ipairs(target:get("includedirs")) do
+            table.insert(args, "-I" .. includedir)
+        end
+
+        os.execv(htcc, args)
+        table.insert(target:objectfiles(), objectfile)
+    end)
+rule_end()
+
+target("infiniop-metax")
+    set_kind("static")
+    on_install(function (target) end)
+    set_languages("cxx17")
+    add_files("../src/infiniop/devices/maca/*.cc", "../src/infiniop/ops/*/maca/*.cc")
+    add_files("../src/infiniop/ops/*/maca/*.maca", {rule = "maca"})
+    add_cxflags("-lstdc++ -Werror -fPIC")
+target_end()