Issue/445 沐曦平台添加macaSDK支持 (#468)

* initial add mc support for meta

* add command description for maca compilation

* rebase metax maca support to main

* issue/445 - clang format code on ubuntu

* issue//445 - change config from use_mc to use-mc and format code

README.md

@@ -97,6 +97,7 @@ python scripts/install.py [XMAKE_CONFIG_FLAGS]
 | `--ascend-npu=[y\|n]`    | 是否编译昇腾 NPU 接口实现         | n
 | `--cambricon-mlu=[y\|n]` | 是否编译寒武纪 MLU 接口实现       | n
 | `--metax-gpu=[y\|n]`     | 是否编译沐曦 GPU 接口实现         | n
+| `--use-mc=[y\|n]`        | 是否沐曦 GPU 接口实现使用maca SDK | n
 | `--moore-gpu=[y\|n]`     | 是否编译摩尔线程 GPU 接口实现     | n
 | `--iluvatar-gpu=[y\|n]`  | 是否编译沐曦 GPU 接口实现         | n
 | `--hygon-dcu=[y\|n]`     | 是否编译海光 DCU 接口实现         | n

src/infiniop/devices/metax/metax_common.h

 #include "../../../utils.h"
 #include "../pool.h"
 #include "metax_handle.h"
+#include "metax_ht2mc.h"
+
+#ifdef ENABLE_METAX_MC_API
+#include <mcblas/mcblas.h>
+#include <mcdnn/mcdnn.h>
+#else
 #include <hcblas/hcblas.h>
 #include <hcdnn/hcdnn.h>
+#endif
 #include <memory>
 
 #define CHECK_MCBLAS(API) CHECK_INTERNAL(API, HCBLAS_STATUS_SUCCESS)

src/infiniop/devices/metax/metax_ht2mc.h

+#ifdef ENABLE_METAX_MC_API
+#define hcblasHandle_t mcblasHandle_t
+#define hcdnnHandle_t mcdnnHandle_t
+#define hcStream_t mcStream_t
+#define hcdnnDataType_t mcdnnDataType_t
+#define hpcc_bfloat16 maca_bfloat16
+#define hpcc_bfloat162 maca_bfloat162
+#define __hpcc_bfloat16 __maca_bfloat16
+#define __hpcc_bfloat16 __maca_bfloat16
+#define hcError_t mcError_t
+#define hcSuccess mcSuccess
+#define hcDeviceProp_t mcDeviceProp_t
+#define hcGetDeviceProperties mcGetDeviceProperties
+#define HCBLAS_STATUS_SUCCESS MCBLAS_STATUS_SUCCESS
+#define HCDNN_STATUS_SUCCESS MCDNN_STATUS_SUCCESS
+#define hcblasCreate mcblasCreate
+#define hcblasSetStream mcblasSetStream
+#define hcGetDeviceCount mcGetDeviceCount
+#define hcSetDevice mcSetDevice
+#define hcDeviceSynchronize mcDeviceSynchronize
+#define hcStreamCreate mcStreamCreate
+#define hcStreamDestroy mcStreamDestroy
+#define hcStreamSynchronize mcStreamSynchronize
+#define hcdnnCreate mcdnnCreate
+#define hcdnnSetStream mcdnnSetStream
+#define hcEvent_t mcEvent_t
+#define hcStreamWaitEvent mcStreamWaitEvent
+#define HCDNN_DATA_HALF MCDNN_DATA_HALF
+#define HCDNN_DATA_FLOAT MCDNN_DATA_FLOAT
+#define HCDNN_DATA_DOUBLE MCDNN_DATA_DOUBLE
+#define HCDNN_DATA_BFLOAT16 MCDNN_DATA_BFLOAT16
+#define HCDNN_DATA_INT8 MCDNN_DATA_INT8
+#define HCDNN_DATA_INT32 MCDNN_DATA_INT32
+#define HCDNN_DATA_INT64 MCDNN_DATA_INT64
+#define HCDNN_DATA_UINT8 MCDNN_DATA_UINT8
+#define hcEventCreate mcEventCreate
+#define hcEventRecord mcEventRecord
+#define hcEventQuery mcEventQuery
+#define hcEventSynchronize mcEventSynchronize
+#define hcEventDestroy mcEventDestroy
+#define hcMalloc mcMalloc
+#define hpccDataType macaDataType
+#define hcblasComputeType_t mcblasComputeType_t
+#define hcMallocHost mcMallocHost
+#define hcFree mcFree
+#define hcFreeHost mcFreeHost
+#define hcMemcpyKind mcMemcpyKind
+#define hcMemcpyHostToDevice mcMemcpyHostToDevice
+#define hcMemcpyDeviceToHost mcMemcpyDeviceToHost
+#define hcMemcpyDeviceToDevice mcMemcpyDeviceToDevice
+#define hcMemcpyHostToHost mcMemcpyHostToHost
+#define hcMemcpyDefault mcMemcpyDefault
+#define hcMemcpy mcMemcpy
+#define hcMemcpyAsync mcMemcpyAsync
+#define hcMallocAsync mcMallocAsync
+#define hcFreeAsync mcFreeAsync
+#define HPCC_R_16F MACA_R_16F
+#define HPCC_R_16BF MACA_R_16BF
+#define HPCC_R_32F MACA_R_32F
+#define HCBLAS_COMPUTE_32F MCBLAS_COMPUTE_32F
+#define HCBLAS_COMPUTE_32F_FAST_TF32 MCBLAS_COMPUTE_32F_FAST_TF32
+#define HCBLAS_OP_N MCBLAS_OP_N
+#define HCBLAS_OP_T MCBLAS_OP_T
+#define HCBLAS_GEMM_DEFAULT_TENSOR_OP MCBLAS_GEMM_DEFAULT_TENSOR_OP
+#define hcblasGemmStridedBatchedEx mcblasGemmStridedBatchedEx
+#define hcLaunchKernel mcLaunchKernel
+#endif

src/infiniop/ops/causal_softmax/metax/causal_softmax_metax.maca

 #include "../../../devices/metax/metax_common.h"
 #include "causal_softmax_metax.h"
 
-#include <hccub/block/block_reduce.cuh>
+#ifdef ENABLE_METAX_MC_API
+    #include <cub/block/block_reduce.cuh>
+#else
+    #include <hccub/block/block_reduce.cuh>
+#endif
 #include "../../../devices/metax/metax_kernel_common.h"
 
 #include "../../../reduce/cuda/reduce.cuh"

src/infiniop/ops/random_sample/metax/random_sample_kernel.h

 #include "../../../devices/metax/metax_kernel_common.h"
 #include "infinicore.h"
-#include <hccub/device/device_radix_sort.cuh>
-#include <hccub/device/device_reduce.cuh>
-#include <hccub/device/device_scan.cuh>
+
+#ifdef ENABLE_METAX_MC_API
+#include <cub/device/device_radix_sort.cuh>
+#include <cub/device/device_reduce.cuh>
+#include <cub/device/device_scan.cuh>
+#else
+#include <cub/device/device_radix_sort.cuh>
+#include <cub/device/device_reduce.cuh>
+#include <cub/device/device_scan.cuh>
+#endif
 
 namespace op::random_sample::metax {
 

src/infiniop/ops/random_sample/metax/random_sample_metax.h

 #ifndef __RANDOM_SAMPLE_METAX_H__
 #define __RANDOM_SAMPLE_METAX_H__
-
 #include "../random_sample.h"
 
 DESCRIPTOR(metax)

src/infiniop/ops/random_sample/metax/random_sample_metax.maca

+#include "../../../devices/metax/metax_ht2mc.h"
 #include "../../../devices/metax/metax_common.h"
 #include "../../../devices/metax/metax_handle.h"
 #include "../info.h"

src/infiniop/ops/softplus/operator.cc

@@ -8,9 +8,6 @@
 #if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
 #include "nvidia/softplus_nvidia.cuh"
 #endif
-#ifdef ENABLE_METAX_API
-#include "metax/softplus_metax.h"
-#endif
 
 __C infiniStatus_t infiniopCreateSoftplusDescriptor(
     infiniopHandle_t handle,

src/infinirt/metax/infinirt_metax.cc

 #include "infinirt_metax.h"
 #include "../../utils.h"
+#ifdef ENABLE_METAX_MC_API
+#include <mcr/mc_runtime.h>
+#include <mcr/mc_runtime_api.h>
+#else
 #include <hcr/hc_runtime.h>
 #include <hcr/hc_runtime_api.h>
+#endif
 
 #define CHECK_MACART(RT_API) CHECK_INTERNAL(RT_API, hcSuccess)
 

src/infinirt/metax/infinirt_metax.h

 #ifndef __INFINIRT_MACA_H__
 #define __INFINIRT_MACA_H__
+#include "../../infiniop/devices/metax/metax_ht2mc.h"
 #include "../infinirt_impl.h"
 
 namespace infinirt::metax {

xmake.lua

@@ -109,8 +109,17 @@ option("metax-gpu")
     set_description("Whether to compile implementations for MetaX GPU")
 option_end()
 
+option("use-mc")
+    set_default(false)
+    set_showmenu(true)
+    set_description("Use MC version")
+option_end()
+
 if has_config("metax-gpu") then
     add_defines("ENABLE_METAX_API")
+    if has_config("use-mc") then
+        add_defines("ENABLE_METAX_MC_API")
+    end
     includes("xmake/metax.lua")
 end
 

xmake/metax.lua

@@ -2,7 +2,11 @@
 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("hcdnn", "hcblas", "hcruntime")
+if has_config("use-mc") then
+    add_links("mcdnn", "mcblas", "mcruntime")
+else
+    add_links("hcdnn", "hcblas", "hcruntime")
+end
 
 rule("maca")
     set_extensions(".maca")
@@ -14,11 +18,16 @@ rule("maca")
     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 args
+        local htcc
+        if has_config("use-mc") then
+            htcc = path.join(MACA_ROOT, "mxgpu_llvm/bin/mxcc")
+            args = { "-x", "maca", "-c", sourcefile, "-o", objectfile, "-I" .. MACA_ROOT .. "/include", "-O3", "-fPIC", "-Werror", "-std=c++17"}
+        else
+            htcc = path.join(MACA_ROOT, "htgpu_llvm/bin/htcc")
+            args = { "-x", "hpcc", "-c", sourcefile, "-o", objectfile, "-I" .. MACA_ROOT .. "/include", "-O3", "-fPIC", "-Werror", "-std=c++17"}
+        end
         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
@@ -66,7 +75,11 @@ target("infiniccl-metax")
         add_cxflags("-fPIC")
     end
     if has_config("ccl") then
+        if has_config("use-mc") then
+            add_links("libmccl.so")
+        else
             add_links("libhccl.so")
+        end
         add_files("../src/infiniccl/metax/*.cc")
     end
     set_languages("cxx17")