Merge pull request #293 from YdrMaster/distinct-cuda

issue291 合并 cuda 代码

src/infiniop/ops/swiglu/cuda/swiglu_cuda.cu → src/infiniop/ops/swiglu/nvidia/swiglu_nvidia.cu

-#include "swiglu_cuda.cuh"
-#include "swiglu_cuda_internal.cuh"
+#include "swiglu_nvidia.cuh"
 
-namespace op::swiglu::cuda {
+#include "../../../elementwise/cuda/elementwise_cuda.cuh"
+
+#include "../cuda/kernel.cuh"
+
+namespace op::swiglu::nvidia {
 
 Descriptor::~Descriptor() = default;
 
@@ -42,17 +45,17 @@ infiniStatus_t Descriptor::calculate(
 
     switch (_dtype) {
     case INFINI_DTYPE_F16:
-        return _device_info->calculate<256, SwiGLUOp, half>(_info, workspace, output, inputs, stream);
+        return _device_info->calculate<256, cuda::SwiGLUOp, half>(_info, workspace, output, inputs, stream);
     case INFINI_DTYPE_BF16:
-        return _device_info->calculate<256, SwiGLUOp, __nv_bfloat16>(_info, workspace, output, inputs, stream);
+        return _device_info->calculate<256, cuda::SwiGLUOp, cuda_bfloat16>(_info, workspace, output, inputs, stream);
     case INFINI_DTYPE_F32:
-        return _device_info->calculate<256, SwiGLUOp, float>(_info, workspace, output, inputs, stream);
+        return _device_info->calculate<256, cuda::SwiGLUOp, float>(_info, workspace, output, inputs, stream);
     case INFINI_DTYPE_F64:
-        return _device_info->calculate<256, SwiGLUOp, double>(_info, workspace, output, inputs, stream);
+        return _device_info->calculate<256, cuda::SwiGLUOp, double>(_info, workspace, output, inputs, stream);
     default:
         return INFINI_STATUS_BAD_TENSOR_DTYPE;
     }
 
     return INFINI_STATUS_SUCCESS;
 }
-} // namespace op::swiglu::cuda
+} // namespace op::swiglu::nvidia

src/infiniop/ops/swiglu/cuda/swiglu_cuda.cuh → src/infiniop/ops/swiglu/nvidia/swiglu_nvidia.cuh

@@ -3,6 +3,6 @@
 
 #include "../../../elementwise/cuda/elementwise_cuda_api.cuh"
 
-ELEMENTWISE_DESCRIPTOR(swiglu, cuda)
+ELEMENTWISE_DESCRIPTOR(swiglu, nvidia, cuda)
 
 #endif // __SWIGLU_CUDA_API_H__

src/infiniop/ops/swiglu/operator.cc

@@ -6,13 +6,13 @@
 #include "cpu/swiglu_cpu.h"
 #endif
 #ifdef ENABLE_NVIDIA_API
-#include "cuda/swiglu_cuda.cuh"
+#include "nvidia/swiglu_nvidia.cuh"
 #endif
 #ifdef ENABLE_KUNLUN_API
 #include "kunlun/swiglu_kunlun.h"
 #endif
 #ifdef ENABLE_METAX_API
-#include "maca/swiglu_maca.h"
+#include "metax/swiglu_metax.h"
 #endif
 #ifdef ENABLE_ASCEND_API
 #include "ascend/swiglu_ascend.h"
@@ -40,13 +40,13 @@ __C infiniStatus_t infiniopCreateSwiGLUDescriptor(
         CREATE(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        CREATE(INFINI_DEVICE_NVIDIA, cuda);
+        CREATE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_KUNLUN_API
         CREATE(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
 #ifdef ENABLE_METAX_API
-        CREATE(INFINI_DEVICE_METAX, maca);
+        CREATE(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_CAMBRICON_MLU
     case DevCambriconMlu: {
@@ -83,20 +83,20 @@ __C infiniStatus_t infiniopGetSwiGLUWorkspaceSize(infiniopSwiGLUDescriptor_t des
 #define GET(CASE, NAMESPACE)                                                                  \
     case CASE:                                                                                \
         *size = reinterpret_cast<op::swiglu::NAMESPACE::Descriptor *>(desc)->workspaceSize(); \
-        return INFINI_STATUS_SUCCESS;
+        return INFINI_STATUS_SUCCESS
 
     switch (desc->device_type) {
 #ifdef ENABLE_CPU_API
-        GET(INFINI_DEVICE_CPU, cpu)
+        GET(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        GET(INFINI_DEVICE_NVIDIA, cuda)
+        GET(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_KUNLUN_API
-        GET(INFINI_DEVICE_KUNLUN, kunlun)
+        GET(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
 #ifdef ENABLE_METAX_API
-        GET(INFINI_DEVICE_METAX, maca);
+        GET(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_CAMBRICON_MLU
     case DevCambriconMlu: {
@@ -104,12 +104,7 @@ __C infiniStatus_t infiniopGetSwiGLUWorkspaceSize(infiniopSwiGLUDescriptor_t des
     }
 #endif
 #ifdef ENABLE_ASCEND_API
-        GET(INFINI_DEVICE_ASCEND, ascend)
-#endif
-#ifdef ENABLE_METAX_GPU
-    case DevMetaxGpu: {
-        return macaGetSwiGLUWorkspaceSize((SwiGLUMacaDescriptor_t)desc, size);
-    }
+        GET(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_MTHREADS_GPU
     case DevMthreadsGpu: {
@@ -143,13 +138,13 @@ __C infiniStatus_t infiniopSwiGLU(
         CALCULATE(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        CALCULATE(INFINI_DEVICE_NVIDIA, cuda);
+        CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_KUNLUN_API
         CALCULATE(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
 #ifdef ENABLE_METAX_API
-        CALCULATE(INFINI_DEVICE_METAX, maca);
+        CALCULATE(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_CAMBRICON_MLU
     case DevCambriconMlu: {
@@ -189,13 +184,13 @@ infiniopDestroySwiGLUDescriptor(infiniopSwiGLUDescriptor_t desc) {
         DELETE(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        DELETE(INFINI_DEVICE_NVIDIA, cuda);
+        DELETE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_KUNLUN_API
         DELETE(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
 #ifdef ENABLE_METAX_API
-        DELETE(INFINI_DEVICE_METAX, maca);
+        DELETE(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_CAMBRICON_MLU
     case DevCambriconMlu: {

src/infiniop/reduce/cuda/reduce.cuh

 #ifndef __INFINIOP_REDUCE_CUDA_H__
 #define __INFINIOP_REDUCE_CUDA_H__
 
-#include <cub/block/block_reduce.cuh>
-
 /*
  * Device functions for reduction operations on CUDA.
  *
  * Note: Only local result on thread 0 is guranteed to be correct.
  *       A manual broadcast is needed for other threads.
+ *
+ * Important Note: This is a device-independent header file containing reduce kernels
+ *                 for all cuda-supporting platforms. Include device-specific headers
+ *                 (such as <cub/block/block_reduce.cuh> for nvidia) in your source file
+ *                 and then include this file for proper usage.
  */
 namespace op::common_cuda::reduce_op {
 

src/infiniop/reduce/maca/reduce.h

-#ifndef __INFINIOP_REDUCE_MACA_H__
-#define __INFINIOP_REDUCE_MACA_H__
-
-#include <hccub/block/block_reduce.cuh>
-
-/*
- * Device functions for reduction operations on MACA.
- *
- * Note: Only local result on thread 0 is guranteed to be correct.
- *       A manual broadcast is needed for other threads.
- */
-namespace op::common_maca::reduce_op {
-
-// Sum(x^2) on contiguous data of length count
-template <unsigned int BLOCK_SIZE, typename Tdata, typename Tcompute>
-__device__ __forceinline__ Tcompute sumSquared(const Tdata *data_ptr, size_t count) {
-    Tcompute ss = 0;
-
-    // Each thread computes its partial sum
-    for (size_t i = threadIdx.x; i < count; i += BLOCK_SIZE) {
-        ss += Tcompute(data_ptr[i]) * Tcompute(data_ptr[i]);
-    }
-
-    // Use CUB block-level reduction
-    using BlockReduce = cub::BlockReduce<Tcompute, BLOCK_SIZE>;
-    __shared__ typename BlockReduce::TempStorage temp_storage;
-
-    return BlockReduce(temp_storage).Sum(ss);
-}
-
-// Sum(x) on contiguous data of length count
-template <unsigned int BLOCK_SIZE, typename Tdata, typename Tcompute>
-__device__ __forceinline__ Tcompute sum(const Tdata *data_ptr, size_t count) {
-    Tcompute s = 0;
-
-    for (size_t i = threadIdx.x; i < count; i += BLOCK_SIZE) {
-        s += Tcompute(data_ptr[i]);
-    }
-
-    using BlockReduce = cub::BlockReduce<Tcompute, BLOCK_SIZE>;
-    __shared__ typename BlockReduce::TempStorage temp_storage;
-
-    return BlockReduce(temp_storage).Sum(s);
-}
-
-// Max(x) on contiguous data of length count
-template <unsigned int BLOCK_SIZE, typename Tdata>
-__device__ __forceinline__ Tdata max(const Tdata *data_ptr, size_t count) {
-    Tdata max_ = data_ptr[0];
-
-    for (size_t i = threadIdx.x; i < count; i += BLOCK_SIZE) {
-        max_ = cub::Max()(max_, data_ptr[i]);
-    }
-
-    using BlockReduce = cub::BlockReduce<Tdata, BLOCK_SIZE>;
-    __shared__ typename BlockReduce::TempStorage temp_storage;
-
-    return BlockReduce(temp_storage).Reduce(max_, cub::Max(), BLOCK_SIZE);
-}
-
-} // namespace op::common_maca::reduce_op
-
-#endif

xmake.lua

@@ -108,7 +108,7 @@ option_end()
 
 if has_config("metax-gpu") then
     add_defines("ENABLE_METAX_API")
-    includes("xmake/maca.lua")
+    includes("xmake/metax.lua")
 end
 
 -- 摩尔线程
@@ -174,7 +174,7 @@ target("infini-utils")
         add_cxflags("-fPIC", "-Wno-unknown-pragmas")
         if has_config("omp") then
             add_cxflags("-fopenmp")
-            add_ldflags("-fopenmp")
+            add_ldflags("-fopenmp", {force = true})
         end
     end
 

xmake/cuda.lua

@@ -46,7 +46,7 @@ target("infiniop-cuda")
     add_cuflags("-Xcompiler=-Wno-error=deprecated-declarations")
 
     set_languages("cxx17")
-    add_files("../src/infiniop/devices/cuda/*.cu", "../src/infiniop/ops/*/cuda/*.cu", "../build/ninetoothed/*.c")
+    add_files("../src/infiniop/devices/cuda/*.cu", "../src/infiniop/ops/*/cuda/*.cu", "../src/infiniop/ops/*/nvidia/*.cu", "../build/ninetoothed/*.c")
 target_end()
 
 target("infinirt-cuda")

xmake/maca.lua → xmake/metax.lua

@@ -34,8 +34,8 @@ target("infiniop-metax")
     set_languages("cxx17")
     set_warnings("all", "error")
     add_cxflags("-lstdc++", "-fPIC", "-Wno-defaulted-function-deleted", "-Wno-strict-aliasing")
-    add_files("../src/infiniop/devices/maca/*.cc", "../src/infiniop/ops/*/maca/*.cc")
-    add_files("../src/infiniop/ops/*/maca/*.maca", {rule = "maca"})
+    add_files("../src/infiniop/devices/maca/*.cc", "../src/infiniop/ops/*/metax/*.cc")
+    add_files("../src/infiniop/ops/*/metax/*.maca", {rule = "maca"})
 target_end()
 
 target("infinirt-metax")
@@ -61,5 +61,5 @@ target("infiniccl-metax")
         add_files("../src/infiniccl/maca/*.cc")
     end
     set_languages("cxx17")
-    
+
 target_end()