issue/291/refactor: 改造 rms_norm、rope、swiglu

Signed-off-by: YdrMaster <ydrml@hotmail.com>

src/infiniop/ops/rms_norm/cuda/rms_norm_kernel.cuh → src/infiniop/ops/rms_norm/cuda/kernel.cuh

 #ifndef __RMS_NORM_CUDA_KERNEL_H__
 #define __RMS_NORM_CUDA_KERNEL_H__
 
-#include "../../../devices/cuda/cuda_kernel_common.cuh"
-
-#include <cub/block/block_reduce.cuh>
-
-#include "../../../reduce/cuda/reduce.cuh"
-
 template <unsigned int BLOCK_SIZE, typename Tdata, typename Tweight, typename Tcompute>
 INFINIOP_CUDA_KERNEL rmsnormBlock(
     Tdata *__restrict__ y,

src/infiniop/ops/rms_norm/cuda/rms_norm_cuda.cu → src/infiniop/ops/rms_norm/nvidia/rms_norm_nvidia.cu

 #include "../../../devices/cuda/cuda_common.cuh"
-#include "rms_norm_cuda.cuh"
-#include "rms_norm_kernel.cuh"
+#include "rms_norm_nvidia.cuh"
 
-namespace op::rms_norm::cuda {
+#include "../../../devices/cuda/cuda_kernel_common.cuh"
+#include <cub/block/block_reduce.cuh>
+
+#include "../../../reduce/cuda/reduce.cuh"
+
+#include "../cuda/kernel.cuh"
+
+namespace op::rms_norm::nvidia {
 
 struct Descriptor::Opaque {
     std::shared_ptr<device::cuda::Handle::Internal> internal;

src/infiniop/ops/rms_norm/cuda/rms_norm_cuda.cuh → src/infiniop/ops/rms_norm/nvidia/rms_norm_nvidia.cuh

@@ -3,6 +3,6 @@
 
 #include "../rms_norm.h"
 
-DESCRIPTOR(cuda)
+DESCRIPTOR(nvidia)
 
 #endif

src/infiniop/ops/rms_norm/operator.cc

@@ -6,7 +6,7 @@
 #include "cpu/rms_norm_cpu.h"
 #endif
 #ifdef ENABLE_NVIDIA_API
-#include "cuda/rms_norm_cuda.cuh"
+#include "nvidia/rms_norm_nvidia.cuh"
 #endif
 #ifdef ENABLE_ASCEND_API
 #include "ascend/rms_norm_aclnn.h"
@@ -37,17 +37,17 @@ __C infiniStatus_t infiniopCreateRMSNormDescriptor(
             y_desc,                                                             \
             x_desc,                                                             \
             w_desc,                                                             \
-            epsilon);
+            epsilon)
 
     switch (handle->device) {
 #ifdef ENABLE_CPU_API
-        CREATE(INFINI_DEVICE_CPU, cpu)
+        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)
+        CREATE(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
 #ifdef ENABLE_CAMBRICON_MLU
     case DevCambriconMlu: {
@@ -55,13 +55,13 @@ __C infiniStatus_t infiniopCreateRMSNormDescriptor(
     }
 #endif
 #ifdef ENABLE_ASCEND_API
-        CREATE(INFINI_DEVICE_ASCEND, ascend)
+        CREATE(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        CREATE(INFINI_DEVICE_METAX, maca)
+        CREATE(INFINI_DEVICE_METAX, maca);
 #endif
 #ifdef ENABLE_MOORE_API
-        CREATE(INFINI_DEVICE_MOORE, musa)
+        CREATE(INFINI_DEVICE_MOORE, musa);
 #endif
     }
 
@@ -75,17 +75,17 @@ __C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t d
 #define GET(CASE, NAMESPACE)                                                                    \
     case CASE:                                                                                  \
         *size = reinterpret_cast<op::rms_norm::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_CAMBRICON_MLU
     case DevCambriconMlu: {
@@ -93,13 +93,13 @@ __C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t d
     }
 #endif
 #ifdef ENABLE_ASCEND_API
-        GET(INFINI_DEVICE_ASCEND, ascend)
+        GET(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        GET(INFINI_DEVICE_METAX, maca)
+        GET(INFINI_DEVICE_METAX, maca);
 #endif
 #ifdef ENABLE_MOORE_API
-        GET(INFINI_DEVICE_MOORE, musa)
+        GET(INFINI_DEVICE_MOORE, musa);
 #endif
     }
 
@@ -114,17 +114,17 @@ __C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *works
 #define CALCULATE(CASE, NAMESPACE)                                                       \
     case CASE:                                                                           \
         return reinterpret_cast<op::rms_norm::NAMESPACE::Descriptor *>(desc)->calculate( \
-            workspace, workspace_size, y, x, w, stream);
+            workspace, workspace_size, y, x, w, stream)
 
     switch (desc->device_type) {
 #ifdef ENABLE_CPU_API
-        CALCULATE(INFINI_DEVICE_CPU, cpu)
+        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)
+        CALCULATE(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
 #ifdef ENABLE_CAMBRICON_MLU
     case DevCambriconMlu: {
@@ -132,13 +132,13 @@ __C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *works
     }
 #endif
 #ifdef ENABLE_ASCEND_API
-        CALCULATE(INFINI_DEVICE_ASCEND, ascend)
+        CALCULATE(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        CALCULATE(INFINI_DEVICE_METAX, maca)
+        CALCULATE(INFINI_DEVICE_METAX, maca);
 #endif
 #ifdef ENABLE_MOORE_API
-        CALCULATE(INFINI_DEVICE_MOORE, musa)
+        CALCULATE(INFINI_DEVICE_MOORE, musa);
 #endif
     }
 
@@ -152,17 +152,17 @@ __C infiniStatus_t infiniopDestroyRMSNormDescriptor(infiniopRMSNormDescriptor_t
 #define DESTROY(CASE, NAMESPACE)                                              \
     case CASE:                                                                \
         delete reinterpret_cast<op::rms_norm::NAMESPACE::Descriptor *>(desc); \
-        return INFINI_STATUS_SUCCESS;
+        return INFINI_STATUS_SUCCESS
 
     switch (desc->device_type) {
 #ifdef ENABLE_CPU_API
-        DESTROY(INFINI_DEVICE_CPU, cpu)
+        DESTROY(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        DESTROY(INFINI_DEVICE_NVIDIA, cuda)
+        DESTROY(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_KUNLUN_API
-        DESTROY(INFINI_DEVICE_KUNLUN, kunlun)
+        DESTROY(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
 #ifdef ENABLE_CAMBRICON_MLU
     case DevCambriconMlu: {
@@ -170,13 +170,13 @@ __C infiniStatus_t infiniopDestroyRMSNormDescriptor(infiniopRMSNormDescriptor_t
     }
 #endif
 #ifdef ENABLE_ASCEND_API
-        DESTROY(INFINI_DEVICE_ASCEND, ascend)
+        DESTROY(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        DESTROY(INFINI_DEVICE_METAX, maca)
+        DESTROY(INFINI_DEVICE_METAX, maca);
 #endif
 #ifdef ENABLE_MOORE_API
-        DESTROY(INFINI_DEVICE_MOORE, musa)
+        DESTROY(INFINI_DEVICE_MOORE, musa);
 #endif
     }
 

src/infiniop/ops/rope/cuda/rope_cuda_kernel.cuh → src/infiniop/ops/rope/cuda/kernel.cuh

 #ifndef __INFINIOP_ROPE_CUDA_KERNEL_CUH__
 #define __INFINIOP_ROPE_CUDA_KERNEL_CUH__
 
-#include "../../../devices/cuda/cuda_kernel_common.cuh"
-
 template <typename Tdata, typename Tindex, typename Tangle>
 INFINIOP_CUDA_KERNEL ropeThreadPerItem(
     Tdata *y_,

src/infiniop/ops/rope/cuda/rope_cuda.cu → src/infiniop/ops/rope/nvidia/rope_nvidia.cu

 #include "../../../devices/cuda/cuda_common.cuh"
-#include "rope_cuda.cuh"
-#include "rope_cuda_kernel.cuh"
+#include "rope_nvidia.cuh"
 
-namespace op::rope::cuda {
+#include "../../../devices/cuda/cuda_kernel_common.cuh"
+
+#include "../cuda/kernel.cuh"
+
+namespace op::rope::nvidia {
 
 struct Descriptor::Opaque {
     std::shared_ptr<device::cuda::Handle::Internal> internal;

src/infiniop/ops/rope/cuda/rope_cuda.cuh → src/infiniop/ops/rope/nvidia/rope_nvidia.cuh

@@ -3,6 +3,6 @@
 
 #include "../rope.h"
 
-DESCRIPTOR(cuda)
+DESCRIPTOR(nvidia)
 
 #endif // __INFINIOP_ROPE_CUDA_H__

src/infiniop/ops/rope/operator.cc

@@ -6,7 +6,7 @@
 #include "cpu/rope_cpu.h"
 #endif
 #ifdef ENABLE_NVIDIA_API
-#include "cuda/rope_cuda.cuh"
+#include "nvidia/rope_nvidia.cuh"
 #endif
 #ifdef ENABLE_ASCEND_API
 #include "ascend/rope_ascend.h"
@@ -40,7 +40,7 @@ __C infiniStatus_t infiniopCreateRoPEDescriptor(
         CREATE(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        CREATE(INFINI_DEVICE_NVIDIA, cuda);
+        CREATE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_METAX_API
         CREATE(INFINI_DEVICE_METAX, maca);
@@ -81,7 +81,7 @@ __C infiniStatus_t infiniopGetRoPEWorkspaceSize(infiniopRoPEDescriptor_t desc,
         GET(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        GET(INFINI_DEVICE_NVIDIA, cuda);
+        GET(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_METAX_API
         GET(INFINI_DEVICE_METAX, maca);
@@ -132,7 +132,7 @@ __C infiniStatus_t infiniopRoPE(
         CALCULATE(INFINI_DEVICE_CPU, cpu);
 #endif
 #ifdef ENABLE_NVIDIA_API
-        CALCULATE(INFINI_DEVICE_NVIDIA, cuda);
+        CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
 #ifdef ENABLE_METAX_API
         CALCULATE(INFINI_DEVICE_METAX, maca);
@@ -178,7 +178,7 @@ infiniopDestroyRoPEDescriptor(infiniopRoPEDescriptor_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_METAX_API
         DELETE(INFINI_DEVICE_METAX, maca);

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"
+#include "../cuda/kernel.cuh"
 
-namespace op::swiglu::cuda {
+namespace op::swiglu::nvidia {
 
 Descriptor::~Descriptor() = default;
 
@@ -42,13 +42,13 @@ 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, __nv_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;
     }

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, cuda)
+ELEMENTWISE_DESCRIPTOR(swiglu, nvidia, cuda)
 
 #endif // __SWIGLU_CUDA_API_H__

src/infiniop/ops/swiglu/operator.cc

@@ -6,7 +6,7 @@
 #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"
@@ -40,7 +40,7 @@ __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);
@@ -83,17 +83,17 @@ __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);
@@ -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: {
@@ -143,7 +138,7 @@ __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);
@@ -189,7 +184,7 @@ 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);