Merge pull request #322 from YdrMaster/main

issue/291/style: 所有 maca 改为 metax

src/infiniop/ops/rms_norm/metax/rms_norm_metax.maca

-#include "../../../devices/maca/common_maca.h"
+#include "../../../devices/metax/metax_common.h"
 #include "rms_norm_metax.cuh"
 
-#include "../../../devices/maca/maca_kernel_common.h"
+#include "../../../devices/metax/metax_kernel_common.h"
 #include <cub/block/block_reduce.cuh>
 
 #include "../../../reduce/cuda/reduce.cuh"
@@ -9,7 +9,7 @@
 #include "../cuda/kernel.cuh"
 
 template <unsigned int BLOCK_SIZE, typename Tcompute, typename Tdata, typename Tweight>
-INFINIOP_MACA_KERNEL rmsnormKernel(
+INFINIOP_METAX_KERNEL rmsnormKernel(
     Tdata *__restrict__ y,
     ptrdiff_t stride_y,
     const Tdata *__restrict__ x,
@@ -20,10 +20,10 @@ INFINIOP_MACA_KERNEL rmsnormKernel(
     rmsnormBlock<BLOCK_SIZE, Tcompute>(y, stride_y, x, stride_x, w, dim, epsilon);
 }
 
-namespace op::rms_norm::maca {
+namespace op::rms_norm::metax {
 
 struct Descriptor::Opaque {
-    std::shared_ptr<device::maca::Handle::Internal> internal;
+    std::shared_ptr<device::metax::Handle::Internal> internal;
 };
 
 Descriptor::~Descriptor() {
@@ -47,7 +47,7 @@ infiniStatus_t Descriptor::create(
     }
 
     *desc_ptr = new Descriptor(
-        new Opaque{reinterpret_cast<device::maca::Handle *>(handle)->internal()},
+        new Opaque{reinterpret_cast<device::metax::Handle *>(handle)->internal()},
         std::move(info),
         0,
         handle->device, handle->device_id);
@@ -62,10 +62,10 @@ infiniStatus_t launchKernel(
     const void *x, ptrdiff_t stride_x,
     const void *w, infiniDtype_t wtype,
     float epsilon,
-    hcStream_t maca_stream) {
+    hcStream_t stream) {
 
 #define LAUNCH_KERNEL(Tdata, Tweight, Tcompute)                                                      \
-    rmsnormKernel<BLOCK_SIZE, Tcompute, Tdata, Tweight><<<batch_size, BLOCK_SIZE, 0, maca_stream>>>( \
+    rmsnormKernel<BLOCK_SIZE, Tcompute, Tdata, Tweight><<<batch_size, BLOCK_SIZE, 0, stream>>>( \
         reinterpret_cast<Tdata *>(y),                                                                \
         stride_y,                                                                                    \
         reinterpret_cast<const Tdata *>(x),                                                          \
@@ -96,7 +96,7 @@ infiniStatus_t launchKernel(
 infiniStatus_t Descriptor::calculate(
     void *workspace, size_t workspace_size,
     void *y, const void *x, const void *w,
-    void *stream) const {
+    void *stream_) const {
 
     if (workspace_size < _workspace_size) {
         return INFINI_STATUS_INSUFFICIENT_WORKSPACE;
@@ -106,14 +106,14 @@ infiniStatus_t Descriptor::calculate(
     auto stride_y = _info.y_strides[0];
     auto dim = _info.dim();
     uint32_t batch_size = static_cast<uint32_t>(_info.shape[0]);
-    auto maca_stream = reinterpret_cast<hcStream_t>(stream);
+    auto stream = reinterpret_cast<hcStream_t>(stream_);
 
     // launch kernel with different block sizes
-    if (_opaque->internal->maxThreadsPerBlock() == MACA_BLOCK_SIZE_1024) {
-        CHECK_STATUS(launchKernel<MACA_BLOCK_SIZE_1024>(batch_size, dim, y, _info.atype, stride_y, x, stride_x, w, _info.wtype, _info.epsilon, maca_stream));
+    if (_opaque->internal->maxThreadsPerBlock() == METAX_BLOCK_SIZE_1024) {
+        CHECK_STATUS(launchKernel<METAX_BLOCK_SIZE_1024>(batch_size, dim, y, _info.atype, stride_y, x, stride_x, w, _info.wtype, _info.epsilon, stream));
     } else {
         return INFINI_STATUS_DEVICE_ARCHITECTURE_NOT_SUPPORTED;
     }
     return INFINI_STATUS_SUCCESS;
 }
-} // namespace op::rms_norm::maca
+} // namespace op::rms_norm::metax

src/infiniop/ops/rms_norm/operator.cc

@@ -58,7 +58,7 @@ __C infiniStatus_t infiniopCreateRMSNormDescriptor(
         CREATE(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        CREATE(INFINI_DEVICE_METAX, maca);
+        CREATE(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_MOORE_API
         CREATE(INFINI_DEVICE_MOORE, musa);
@@ -96,7 +96,7 @@ __C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t d
         GET(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        GET(INFINI_DEVICE_METAX, maca);
+        GET(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_MOORE_API
         GET(INFINI_DEVICE_MOORE, musa);
@@ -135,7 +135,7 @@ __C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *works
         CALCULATE(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        CALCULATE(INFINI_DEVICE_METAX, maca);
+        CALCULATE(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_MOORE_API
         CALCULATE(INFINI_DEVICE_MOORE, musa);
@@ -173,7 +173,7 @@ __C infiniStatus_t infiniopDestroyRMSNormDescriptor(infiniopRMSNormDescriptor_t
         DESTROY(INFINI_DEVICE_ASCEND, ascend);
 #endif
 #ifdef ENABLE_METAX_API
-        DESTROY(INFINI_DEVICE_METAX, maca);
+        DESTROY(INFINI_DEVICE_METAX, metax);
 #endif
 #ifdef ENABLE_MOORE_API
         DESTROY(INFINI_DEVICE_MOORE, musa);

src/infiniop/ops/rope/metax/rope_metax.h

-#ifndef __INFINIOP_ROPE_MACA_H__
-#define __INFINIOP_ROPE_MACA_H__
+#ifndef __INFINIOP_ROPE_METAX_H__
+#define __INFINIOP_ROPE_METAX_H__
 
 #include "../rope.h"
 
 DESCRIPTOR(metax)
 
-#endif // __INFINIOP_ROPE_MACA_H__
+#endif // __INFINIOP_ROPE_METAX_H__

src/infiniop/ops/rope/metax/rope_metax.maca

-#include "../../../devices/maca/common_maca.h"
+#include "../../../devices/metax/metax_common.h"
 #include "rope_metax.h"
 
-#include "../../../devices/maca/maca_kernel_common.h"
+#include "../../../devices/metax/metax_kernel_common.h"
 
 #include "../cuda/kernel.cuh"
 
 template <typename Tdata, typename Tindex, typename Tangle>
-INFINIOP_MACA_KERNEL ropeThreadPerItemKernel(
+INFINIOP_METAX_KERNEL ropeThreadPerItemKernel(
     Tdata *y_,
     const Tdata *x_,
     const Tindex *__restrict__ pos_ids,
@@ -28,7 +28,7 @@ INFINIOP_MACA_KERNEL ropeThreadPerItemKernel(
 namespace op::rope::metax {
 
 struct Descriptor::Opaque {
-    std::shared_ptr<device::maca::Handle::Internal> internal;
+    std::shared_ptr<device::metax::Handle::Internal> internal;
 };
 
 Descriptor::~Descriptor() {
@@ -44,7 +44,7 @@ infiniStatus_t Descriptor::create(
     infiniopTensorDescriptor_t sin_desc,
     infiniopTensorDescriptor_t cos_desc) {
 
-    auto handle = reinterpret_cast<device::maca::Handle *>(handle_);
+    auto handle = reinterpret_cast<device::metax::Handle *>(handle_);
 
     auto info = RoPEInfo::createRoPEInfo(y_desc, x_desc, pos_desc, sin_desc, cos_desc);
     CHECK_RESULT(info);
@@ -53,7 +53,7 @@ infiniStatus_t Descriptor::create(
     *desc_ptr = new Descriptor(
         info.take(),
         0,
-        new Opaque{reinterpret_cast<device::maca::Handle *>(handle)->internal()},
+        new Opaque{reinterpret_cast<device::metax::Handle *>(handle)->internal()},
         handle->device,
         handle->device_id);
 
@@ -141,4 +141,4 @@ infiniStatus_t Descriptor::calculate(
 #undef ROPE_TYPE
 #undef CALCULATE_ROPE
 
-} // namespace op::rope::maca
+} // namespace op::rope::metax

src/infiniop/ops/swiglu/metax/swiglu_metax.h

-#ifndef __SWIGLU_MACA_API_H__
-#define __SWIGLU_MACA_API_H__
+#ifndef __SWIGLU_METAX_API_H__
+#define __SWIGLU_METAX_API_H__
 
-#include "../../../elementwise/maca/elementwise_maca_api.h"
+#include "../../../elementwise/metax/elementwise_metax_api.h"
 
-ELEMENTWISE_DESCRIPTOR(swiglu, metax, maca)
+ELEMENTWISE_DESCRIPTOR(swiglu, metax, metax)
 
-#endif // __SWIGLU_MACA_API_H__
+#endif // __SWIGLU_METAX_API_H__

src/infiniop/ops/swiglu/metax/swiglu_metax.maca

 #include "swiglu_metax.h"
 
-#include "../../../elementwise/maca/elementwise_maca.h"
+#include "../../../elementwise/metax/elementwise_metax.h"
 
 #include "../cuda/kernel.cuh"
 
@@ -14,7 +14,7 @@ infiniStatus_t Descriptor::create(
     infiniopTensorDescriptor_t out_desc,
     std::vector<infiniopTensorDescriptor_t> input_desc_vec) {
 
-    auto handle = reinterpret_cast<device::maca::Handle *>(handle_);
+    auto handle = reinterpret_cast<device::metax::Handle *>(handle_);
     auto dtype = out_desc->dtype();
 
     const auto &up_desc = input_desc_vec.at(0);
@@ -26,8 +26,8 @@ infiniStatus_t Descriptor::create(
     CHECK_DTYPE(dtype, INFINI_DTYPE_F16, INFINI_DTYPE_BF16, INFINI_DTYPE_F32, INFINI_DTYPE_F64);
     CHECK_SAME_SHAPE(out_shape, up_shape, gate_shape);
 
-    // create MACA elementwise descriptor
-    CREATE_ELEMENTWISE_MACA_DESCRIPTOR(handle, dtype, out_desc, input_desc_vec)
+    // create METAX elementwise descriptor
+    CREATE_ELEMENTWISE_METAX_DESCRIPTOR(handle, dtype, out_desc, input_desc_vec)
 
     return INFINI_STATUS_SUCCESS;
 }

src/infinirt/infinirt.cc

@@ -5,7 +5,7 @@
 #include "cpu/infinirt_cpu.h"
 #include "cuda/infinirt_cuda.cuh"
 #include "kunlun/infinirt_kunlun.h"
-#include "maca/infinirt_maca.h"
+#include "metax/infinirt_metax.h"
 #include "musa/infinirt_musa.h"
 
 thread_local infiniDevice_t CURRENT_DEVICE_TYPE = INFINI_DEVICE_CPU;
@@ -62,7 +62,7 @@ __C infiniStatus_t infinirtGetDevice(infiniDevice_t *device_ptr, int *device_id_
             _status = infinirt::ascend::API PARAMS;                    \
             break;                                                     \
         case INFINI_DEVICE_METAX:                                      \
-            _status = infinirt::maca::API PARAMS;                      \
+            _status = infinirt::metax::API PARAMS;                     \
             break;                                                     \
         case INFINI_DEVICE_MOORE:                                      \
             _status = infinirt::musa::API PARAMS;                      \

src/infinirt/maca/infinirt_maca.cc → src/infinirt/metax/infinirt_metax.cc

-#include "infinirt_maca.h"
+#include "infinirt_metax.h"
 #include "../../utils.h"
 #include <hcr/hc_runtime.h>
 #include <hcr/hc_runtime_api.h>
 
 #define CHECK_MACART(RT_API) CHECK_INTERNAL(RT_API, hcSuccess)
 
-namespace infinirt::maca {
+namespace infinirt::metax {
 infiniStatus_t getDeviceCount(int *count) {
     CHECK_MACART(hcGetDeviceCount(count));
     return INFINI_STATUS_SUCCESS;
@@ -124,4 +124,4 @@ infiniStatus_t freeAsync(void *ptr, infinirtStream_t stream) {
     CHECK_MACART(hcFreeAsync(ptr, (hcStream_t)stream));
     return INFINI_STATUS_SUCCESS;
 }
-} // namespace infinirt::maca
+} // namespace infinirt::metax

src/infinirt/maca/infinirt_maca.h → src/infinirt/metax/infinirt_metax.h

@@ -2,12 +2,12 @@
 #define __INFINIRT_MACA_H__
 #include "../infinirt_impl.h"
 
-namespace infinirt::maca {
+namespace infinirt::metax {
 #ifdef ENABLE_METAX_API
 INFINIRT_DEVICE_API_IMPL
 #else
 INFINIRT_DEVICE_API_NOOP
 #endif
-} // namespace infinirt::maca
+} // namespace infinirt::metax
 
 #endif // __INFINIRT_MACA_H__

xmake/metax.lua

@@ -34,7 +34,7 @@ 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/*/metax/*.cc")
+    add_files("../src/infiniop/devices/metax/*.cc", "../src/infiniop/ops/*/metax/*.cc")
     add_files("../src/infiniop/ops/*/metax/*.maca", {rule = "maca"})
 target_end()
 
@@ -45,7 +45,7 @@ target("infinirt-metax")
     add_deps("infini-utils")
     set_warnings("all", "error")
     add_cxflags("-lstdc++ -fPIC")
-    add_files("../src/infinirt/maca/*.cc")
+    add_files("../src/infinirt/metax/*.cc")
 target_end()
 
 target("infiniccl-metax")
@@ -58,7 +58,7 @@ target("infiniccl-metax")
     end
     if has_config("ccl") then
         add_links("libhccl.so")
-        add_files("../src/infiniccl/maca/*.cc")
+        add_files("../src/infiniccl/metax/*.cc")
     end
     set_languages("cxx17")