Skip to content

GitLab

Unverified Commit 95623d82 authored by PanZezhong1725's avatar PanZezhong1725 Committed by GitHub
Browse files

Merge pull request #328 from YdrMaster/main 

issue/158/refactor: 分离天数和英伟达的算子定义
parents e3022fa8 d9de5133
Hide whitespace changes
Inline Side-by-side
Showing with 216 additions and 34 deletions
src/infiniccl/cuda/infiniccl_cuda.h
View file @ 95623d82
......@@ -4,7 +4,7 @@
#include "../infiniccl_impl.h"
// Windows does not support CUDA
#if defined(ENABLE_NVIDIA_API) && defined(ENABLE_CCL) && !defined(_WIN32)
#if (defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)) && defined(ENABLE_CCL) && !defined(_WIN32)
INFINICCL_DEVICE_API_IMPL(cuda)
#else
INFINICCL_DEVICE_API_NOOP(cuda)
......
src/infiniop/devices/handle.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/cpu_handle.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/nvidia_handle.h"
#endif
#ifdef ENABLE_CAMBRICON_API
......@@ -44,6 +44,9 @@ __C infiniStatus_t infiniopCreateHandle(infiniopHandle_t *handle_ptr) {
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_CAMBRICON_API
CREATE(INFINI_DEVICE_CAMBRICON, bang::cambricon);
#endif
......@@ -81,6 +84,9 @@ __C infiniStatus_t infiniopDestroyHandle(infiniopHandle_t handle) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_CAMBRICON_API
DELETE(INFINI_DEVICE_CAMBRICON, bang::cambricon);
#endif
......
src/infiniop/devices/nvidia/nvidia_common.cu
View file @ 95623d82
#include "nvidia_handle.cuh"
namespace device::nvidia {
namespace device {
Handle::Handle(int device_id)
: InfiniopHandle{INFINI_DEVICE_NVIDIA, device_id},
namespace nvidia {
Handle::Handle(infiniDevice_t device, int device_id)
: InfiniopHandle{device, device_id},
_internal(std::make_shared<Handle::Internal>(device_id)) {}
auto Handle::internal() const -> const std::shared_ptr<Internal> & {
......@@ -71,7 +73,7 @@ cudnnDataType_t getCudnnDtype(infiniDtype_t dt) {
return CUDNN_DATA_INT8;
case INFINI_DTYPE_I32:
return CUDNN_DATA_INT32;
#ifndef ENABLE_ILUVATAR_CUDA_API
#ifndef ENABLE_ILUVATAR_API
case INFINI_DTYPE_I64:
return CUDNN_DATA_INT64;
#endif
......@@ -83,9 +85,23 @@ cudnnDataType_t getCudnnDtype(infiniDtype_t dt) {
}
#endif
infiniStatus_t Handle::create(InfiniopHandle **handle_ptr, int device_id) {
*handle_ptr = new Handle(INFINI_DEVICE_NVIDIA, device_id);
return INFINI_STATUS_SUCCESS;
}
} // namespace nvidia
namespace iluvatar {
Handle::Handle(int device_id)
: nvidia::Handle(INFINI_DEVICE_ILUVATAR, device_id) {}
infiniStatus_t Handle::create(InfiniopHandle **handle_ptr, int device_id) {
*handle_ptr = new Handle(device_id);
return INFINI_STATUS_SUCCESS;
}
} // namespace device::nvidia
} // namespace iluvatar
} // namespace device
src/infiniop/devices/nvidia/nvidia_handle.h
View file @ 95623d82
......@@ -4,13 +4,17 @@
#include "../../handle.h"
#include <memory>
namespace device::nvidia {
namespace device {
namespace nvidia {
struct Handle : public InfiniopHandle {
Handle(int device_id);
class Internal;
auto internal() const -> const std::shared_ptr<Internal> &;
protected:
Handle(infiniDevice_t device, int device_id);
public:
static infiniStatus_t create(InfiniopHandle **handle_ptr, int device_id);
......@@ -18,6 +22,19 @@ private:
std::shared_ptr<Internal> _internal;
};
} // namespace device::nvidia
} // namespace nvidia
namespace iluvatar {
struct Handle : public nvidia::Handle {
Handle(int device_id);
public:
static infiniStatus_t create(InfiniopHandle **handle_ptr, int device_id);
};
} // namespace iluvatar
} // namespace device
#endif // __INFINIOP_CUDA_HANDLE_H__
src/infiniop/devices/nvidia/nvidia_kernel_common.cuh
View file @ 95623d82
......@@ -55,7 +55,7 @@ exp_(const float val) {
return expf(val);
}
#ifndef ENABLE_ILUVATAR_CUDA_API
#ifndef ENABLE_ILUVATAR_API
__forceinline__ __device__ long double
exp_(const long double val) {
return expl(val);
......
src/infiniop/ops/add/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/add_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/add_nvidia.cuh"
#endif
#ifdef ENABLE_METAX_API
......@@ -36,6 +36,9 @@ __C infiniStatus_t infiniopCreateAddDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -61,6 +64,9 @@ __C infiniStatus_t infiniopGetAddWorkspaceSize(infiniopAddDescriptor_t desc, siz
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
GET(INFINI_DEVICE_METAX, metax);
#endif
......@@ -94,6 +100,9 @@ __C infiniStatus_t infiniopAdd(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -121,6 +130,9 @@ infiniopDestroyAddDescriptor(infiniopAddDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DELETE(INFINI_DEVICE_METAX, metax);
#endif
......
src/infiniop/ops/causal_softmax/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/causal_softmax_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/causal_softmax_nvidia.cuh"
#endif
#ifdef ENABLE_METAX_API
......@@ -36,6 +36,9 @@ __C infiniStatus_t infiniopCreateCausalSoftmaxDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia)
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax)
#endif
......@@ -71,6 +74,9 @@ __C infiniStatus_t infiniopGetCausalSoftmaxWorkspaceSize(infiniopCausalSoftmaxDe
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia)
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
GET(INFINI_DEVICE_METAX, metax)
#endif
......@@ -112,6 +118,9 @@ __C infiniStatus_t infiniopCausalSoftmax(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia)
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax)
#endif
......@@ -147,6 +156,9 @@ __C infiniStatus_t infiniopDestroyCausalSoftmaxDescriptor(infiniopCausalSoftmaxD
#ifdef ENABLE_NVIDIA_API
DESTROY(INFINI_DEVICE_NVIDIA, nvidia)
#endif
#ifdef ENABLE_ILUVATAR_API
DESTROY(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DESTROY(INFINI_DEVICE_METAX, metax)
#endif
......
src/infiniop/ops/clip/cuda/kernel.cuh
View file @ 95623d82
......@@ -10,7 +10,7 @@ public:
template <typename T>
__device__ __forceinline__ T operator()(const T &x, const T &min_val, const T &max_val) const {
if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, cuda_bfloat162>) {
#ifndef ENABLE_ILUVATAR_CUDA_API
#ifndef ENABLE_ILUVATAR_API
return __hmax2(__hmin2(x, max_val), min_val);
#else
return {std::clamp(x.x, min_val.x, max_val.x), std::clamp(x.y, min_val.y, max_val.y)};
......
src/infiniop/ops/clip/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/clip_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/clip_nvidia.cuh"
#endif
#ifdef ENABLE_METAX_API
......@@ -36,6 +36,9 @@ __C infiniStatus_t infiniopCreateClipDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -61,6 +64,9 @@ __C infiniStatus_t infiniopGetClipWorkspaceSize(infiniopClipDescriptor_t desc, s
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia)
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
GET(INFINI_DEVICE_METAX, metax)
#endif
......@@ -94,6 +100,9 @@ __C infiniStatus_t infiniopClip(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -121,6 +130,9 @@ infiniopDestroyClipDescriptor(infiniopClipDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DELETE(INFINI_DEVICE_METAX, metax);
#endif
......
src/infiniop/ops/conv/nvidia/conv_nvidia.cu
View file @ 95623d82
......@@ -345,7 +345,7 @@ public:
}
return utils::Result<Opaque>(std::move(opaque));
#else
return INFINI_STATUS_SUCCESS;
return INFINI_STATUS_NOT_IMPLEMENTED;
#endif
}
};
......@@ -380,7 +380,7 @@ infiniStatus_t Descriptor::create(
auto conv_info = result.take();
auto opaque_result = Opaque::create(handle->internal(), conv_info, dtype);
CHECK_RESULT(opaque_result);
auto opaque = new Opaque(std::move(opaque_result.take()));
auto opaque = new Opaque(opaque_result.take());
*desc_ptr = new Descriptor(
dtype,
......@@ -391,7 +391,7 @@ infiniStatus_t Descriptor::create(
handle->device_id);
return INFINI_STATUS_SUCCESS;
#else
return INFINI_STATUS_SUCCESS;
return INFINI_STATUS_NOT_IMPLEMENTED;
#endif
}
......@@ -450,7 +450,7 @@ infiniStatus_t Descriptor::calculate(
return INFINI_STATUS_SUCCESS;
#else
return INFINI_STATUS_SUCCESS;
return INFINI_STATUS_NOT_IMPLEMENTED;
#endif
}
} // namespace op::conv::nvidia
src/infiniop/ops/conv/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/conv_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/conv_nvidia.cuh"
#endif
......@@ -39,6 +39,10 @@ __C __export infiniStatus_t infiniopCreateConvDescriptor(infiniopHandle_t handle
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
default:
return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
}
......@@ -63,6 +67,10 @@ infiniopGetConvWorkspaceSize(
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
default:
return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
}
......@@ -95,6 +103,9 @@ __C infiniStatus_t infiniopConv(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
default:
return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
......@@ -116,6 +127,10 @@ infiniopDestroyConvDescriptor(infiniopConvDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
default:
return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
}
......
src/infiniop/ops/gemm/nvidia/gemm_nvidia.cu
View file @ 95623d82
......@@ -43,7 +43,7 @@ infiniStatus_t Descriptor::calculate(
void *stream) const {
cudaDataType a_type, b_type, c_type;
#ifdef ENABLE_ILUVATAR_CUDA_API
#ifdef ENABLE_ILUVATAR_API
cudaDataType compute_type;
#else
cublasComputeType_t compute_type;
......@@ -52,7 +52,7 @@ infiniStatus_t Descriptor::calculate(
switch (_dtype) {
case INFINI_DTYPE_F16:
a_type = b_type = c_type = CUDA_R_16F;
#ifdef ENABLE_ILUVATAR_CUDA_API
#ifdef ENABLE_ILUVATAR_API
compute_type = CUDA_R_32F;
#else
compute_type = CUBLAS_COMPUTE_32F;
......@@ -60,7 +60,7 @@ infiniStatus_t Descriptor::calculate(
break;
case INFINI_DTYPE_BF16:
a_type = b_type = c_type = CUDA_R_16BF;
#ifdef ENABLE_ILUVATAR_CUDA_API
#ifdef ENABLE_ILUVATAR_API
compute_type = CUDA_R_32F;
#else
compute_type = CUBLAS_COMPUTE_32F;
......@@ -68,7 +68,7 @@ infiniStatus_t Descriptor::calculate(
break;
case INFINI_DTYPE_F32:
a_type = b_type = c_type = CUDA_R_32F;
#if defined ENABLE_ILUVATAR_CUDA_API
#if defined ENABLE_ILUVATAR_API
compute_type = CUDA_R_32F;
#elif defined ENABLE_SUGON_CUDA_API
compute_type = CUBLAS_COMPUTE_32F;
......
src/infiniop/ops/gemm/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/gemm_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/gemm_nvidia.cuh"
#endif
#ifdef ENABLE_CAMBRICON_API
......@@ -48,6 +48,9 @@ __C infiniStatus_t infiniopCreateGemmDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_CAMBRICON_API
CREATE(INFINI_DEVICE_CAMBRICON, bang);
#endif
......@@ -90,6 +93,9 @@ infiniopGetGemmWorkspaceSize(
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_CAMBRICON_API
GET(INFINI_DEVICE_CAMBRICON, bang);
#endif
......@@ -139,6 +145,9 @@ __C infiniStatus_t infiniopGemm(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_CAMBRICON_API
CALCULATE(INFINI_DEVICE_CAMBRICON, bang);
#endif
......@@ -178,6 +187,9 @@ infiniopDestroyGemmDescriptor(infiniopGemmDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_CAMBRICON_API
DELETE(INFINI_DEVICE_CAMBRICON, bang);
#endif
......
src/infiniop/ops/mul/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/mul_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/mul_nvidia.cuh"
#endif
#ifdef ENABLE_METAX_API
......@@ -36,6 +36,9 @@ __C infiniStatus_t infiniopCreateMulDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -61,9 +64,13 @@ __C infiniStatus_t infiniopGetMulWorkspaceSize(infiniopMulDescriptor_t desc, siz
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
GET(INFINI_DEVICE_METAX, metax);
#endif
default:
return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
}
......@@ -94,6 +101,9 @@ __C infiniStatus_t infiniopMul(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -121,6 +131,9 @@ infiniopDestroyMulDescriptor(infiniopMulDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DELETE(INFINI_DEVICE_METAX, metax);
#endif
......
src/infiniop/ops/random_sample/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/random_sample_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/random_sample_nvidia.cuh"
#endif
#ifdef ENABLE_METAX_API
......@@ -38,6 +38,9 @@ infiniopCreateRandomSampleDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -71,6 +74,9 @@ __C infiniStatus_t infiniopGetRandomSampleWorkspaceSize(
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
GET(INFINI_DEVICE_METAX, metax);
#endif
......@@ -114,6 +120,9 @@ __C infiniStatus_t infiniopRandomSample(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -144,6 +153,9 @@ __C infiniStatus_t infiniopDestroyRandomSampleDescriptor(
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DELETE(INFINI_DEVICE_METAX, metax);
#endif
......
src/infiniop/ops/rearrange/operator.cc
View file @ 95623d82
......@@ -8,8 +8,7 @@
#ifdef ENABLE_ASCEND_API
#include "ascend/rearrange_ascend.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/rearrange_nvidia.cuh"
#endif
#ifdef ENABLE_METAX_API
......@@ -42,6 +41,9 @@ __C infiniStatus_t infiniopCreateRearrangeDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -75,6 +77,9 @@ __C infiniStatus_t infiniopRearrange(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -102,10 +107,12 @@ __C infiniStatus_t infiniopDestroyRearrangeDescriptor(
#ifdef ENABLE_ASCEND_API
DELETE(INFINI_DEVICE_ASCEND, ascend);
#endif
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DELETE(INFINI_DEVICE_METAX, metax);
#endif
......
src/infiniop/ops/rms_norm/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/rms_norm_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/rms_norm_nvidia.cuh"
#endif
#ifdef ENABLE_ASCEND_API
......@@ -46,6 +46,9 @@ __C infiniStatus_t infiniopCreateRMSNormDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
CREATE(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......@@ -84,6 +87,9 @@ __C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t d
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
GET(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......@@ -123,6 +129,9 @@ __C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *works
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
CALCULATE(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......@@ -161,6 +170,9 @@ __C infiniStatus_t infiniopDestroyRMSNormDescriptor(infiniopRMSNormDescriptor_t
#ifdef ENABLE_NVIDIA_API
DESTROY(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DESTROY(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
DESTROY(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......
src/infiniop/ops/rope/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/rope_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/rope_nvidia.cuh"
#endif
#ifdef ENABLE_ASCEND_API
......@@ -42,6 +42,9 @@ __C infiniStatus_t infiniopCreateRoPEDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -83,6 +86,9 @@ __C infiniStatus_t infiniopGetRoPEWorkspaceSize(infiniopRoPEDescriptor_t desc,
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
GET(INFINI_DEVICE_METAX, metax);
#endif
......@@ -134,6 +140,9 @@ __C infiniStatus_t infiniopRoPE(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -180,6 +189,9 @@ infiniopDestroyRoPEDescriptor(infiniopRoPEDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DELETE(INFINI_DEVICE_METAX, metax);
#endif
......
src/infiniop/ops/sub/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/sub_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/sub_nvidia.cuh"
#endif
#ifdef ENABLE_METAX_API
......@@ -36,6 +36,9 @@ __C infiniStatus_t infiniopCreateSubDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CREATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -61,6 +64,9 @@ __C infiniStatus_t infiniopGetSubWorkspaceSize(infiniopSubDescriptor_t desc, siz
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
GET(INFINI_DEVICE_METAX, metax);
#endif
......@@ -94,6 +100,9 @@ __C infiniStatus_t infiniopSub(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
CALCULATE(INFINI_DEVICE_METAX, metax);
#endif
......@@ -121,6 +130,9 @@ infiniopDestroySubDescriptor(infiniopSubDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_METAX_API
DELETE(INFINI_DEVICE_METAX, metax);
#endif
......
src/infiniop/ops/swiglu/operator.cc
View file @ 95623d82
......@@ -5,7 +5,7 @@
#ifdef ENABLE_CPU_API
#include "cpu/swiglu_cpu.h"
#endif
#ifdef ENABLE_NVIDIA_API
#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
#include "nvidia/swiglu_nvidia.cuh"
#endif
#ifdef ENABLE_KUNLUN_API
......@@ -42,6 +42,9 @@ __C infiniStatus_t infiniopCreateSwiGLUDescriptor(
#ifdef ENABLE_NVIDIA_API
CREATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
CREATE(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......@@ -92,6 +95,9 @@ __C infiniStatus_t infiniopGetSwiGLUWorkspaceSize(infiniopSwiGLUDescriptor_t des
#ifdef ENABLE_NVIDIA_API
GET(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
GET(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
GET(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......@@ -140,6 +146,9 @@ __C infiniStatus_t infiniopSwiGLU(
#ifdef ENABLE_NVIDIA_API
CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
CALCULATE(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......@@ -186,6 +195,9 @@ infiniopDestroySwiGLUDescriptor(infiniopSwiGLUDescriptor_t desc) {
#ifdef ENABLE_NVIDIA_API
DELETE(INFINI_DEVICE_NVIDIA, nvidia);
#endif
#ifdef ENABLE_ILUVATAR_API
DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
#endif
#ifdef ENABLE_KUNLUN_API
DELETE(INFINI_DEVICE_KUNLUN, kunlun);
#endif
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment