issue/32: 实现摩尔线程rms_norm算子

src/infiniop/devices/musa/common_musa.h

@@ -10,18 +10,36 @@
 #define CHECK_MUBLAS(API) CHECK_INTERNAL(API, MUBLAS_STATUS_SUCCESS)
 #define CHECK_MUDNN(API) CHECK_INTERNAL((int)API, (int)::musa::dnn::Status::SUCCESS)
 
+#define INFINIOP_MUSA_KERNEL __global__ void
+#define MUSA_BLOCK_SIZE_1024 1024
+
 namespace device::musa {
 
 class Handle::Internal {
     Pool<std::unique_ptr<mublasHandle_t>> mublas_handles;
     Pool<std::unique_ptr<::musa::dnn::Handle>> mudnn_handles;
 
+    int _warp_size,
+        _max_threads_per_block,
+        _block_size[3],
+        _grid_size[3];
+
     template <typename T>
     using Fn = std::function<infiniStatus_t(T)>;
 
 public:
+    Internal(int);
     infiniStatus_t useMublas(musaStream_t stream, const Fn<mublasHandle_t> &f) const;
     infiniStatus_t useMudnn(musaStream_t stream, const Fn<::musa::dnn::Handle &> &f) const;
+
+    int warpSize() const;
+    int maxThreadsPerBlock() const;
+    int blockSizeX() const;
+    int blockSizeY() const;
+    int blockSizeZ() const;
+    int gridSizeX() const;
+    int gridSizeY() const;
+    int gridSizeZ() const;
 };
 
 } // namespace device::musa

src/infiniop/devices/musa/musa_handle.cc

@@ -3,7 +3,7 @@
 namespace device::musa {
 Handle::Handle(infiniDevice_t device, int device_id)
     : InfiniopHandle{device, device_id},
-      _internal(std::make_shared<Handle::Internal>()) {}
+      _internal(std::make_shared<Handle::Internal>(device_id)) {}
 
 Handle::Handle(int device_id) : Handle(INFINI_DEVICE_MOORE, device_id) {}
 
@@ -11,6 +11,19 @@ auto Handle::internal() const -> const std::shared_ptr<Internal> & {
     return _internal;
 }
 
+Handle::Internal::Internal(int device_id) {
+    musaDeviceProp prop;
+    musaGetDeviceProperties(&prop, device_id);
+    _warp_size = prop.warpSize;
+    _max_threads_per_block = prop.maxThreadsPerBlock;
+    _block_size[0] = prop.maxThreadsDim[0];
+    _block_size[1] = prop.maxThreadsDim[1];
+    _block_size[2] = prop.maxThreadsDim[2];
+    _grid_size[0] = prop.maxGridSize[0];
+    _grid_size[1] = prop.maxGridSize[1];
+    _grid_size[2] = prop.maxGridSize[2];
+}
+
 infiniStatus_t Handle::Internal::useMublas(musaStream_t stream, const Fn<mublasHandle_t> &f) const {
     std::unique_ptr<mublasHandle_t> handle;
     auto opt_handle = mublas_handles.pop();
@@ -40,6 +53,15 @@ infiniStatus_t Handle::Internal::useMudnn(musaStream_t stream, const Fn<::musa::
     return INFINI_STATUS_SUCCESS;
 }
 
+int Handle::Internal::warpSize() const { return _warp_size; }
+int Handle::Internal::maxThreadsPerBlock() const { return _max_threads_per_block; }
+int Handle::Internal::blockSizeX() const { return _block_size[0]; }
+int Handle::Internal::blockSizeY() const { return _block_size[1]; }
+int Handle::Internal::blockSizeZ() const { return _block_size[2]; }
+int Handle::Internal::gridSizeX() const { return _grid_size[0]; }
+int Handle::Internal::gridSizeY() const { return _grid_size[1]; }
+int Handle::Internal::gridSizeZ() const { return _grid_size[2]; }
+
 infiniStatus_t Handle::create(InfiniopHandle **handle_ptr, int device_id) {
     *handle_ptr = new Handle(INFINI_DEVICE_MOORE, device_id);
     return INFINI_STATUS_SUCCESS;

src/infiniop/ops/rms_norm/musa/rms_norm_kernel.h

+#ifndef __RMS_NORM_MUSA_KERNEL_H__
+#define __RMS_NORM_MUSA_KERNEL_H__
+
+#include "../../../reduce/musa/reduce.h"
+
+template <unsigned int BLOCK_SIZE, typename Tdata, typename Tweight, typename Tcompute>
+INFINIOP_MUSA_KERNEL rmsnormBlock(
+    Tdata *__restrict__ y,
+    ptrdiff_t stride_y,
+    const Tdata *__restrict__ x,
+    ptrdiff_t stride_x,
+    const Tweight *__restrict__ w,
+    size_t dim,
+    float epsilon) {
+    // Each block takes care of a row of continuous data of length dim
+    // Each thread deals with every block_size element in the row
+    auto y_ptr = y + blockIdx.x * stride_y;
+    auto x_ptr = x + blockIdx.x * stride_x;
+    auto w_ptr = w;
+
+    // Block-reduce sum of x^2
+    Tcompute ss = op::common_musa::reduce_op::sumSquared<BLOCK_SIZE, Tdata, Tcompute>(x_ptr, dim);
+
+    // Thread_0 computes RMS=1/sqrt(ss/dim+epsilon) and stores in shared memory
+    __shared__ Tcompute rms;
+    if (threadIdx.x == 0) {
+        rms = Tdata(rsqrtf(ss / Tcompute(dim) + epsilon));
+    }
+    __syncthreads();
+
+    for (size_t i = threadIdx.x; i < dim; i += BLOCK_SIZE) {
+        y_ptr[i] = Tdata(Tcompute(x_ptr[i]) * Tcompute(w_ptr[i]) * rms);
+    }
+}
+
+#endif

src/infiniop/ops/rms_norm/musa/rms_norm_musa.h

+#ifndef __RMS_NORM_MUSA_H__
+#define __RMS_NORM_MUSA_H__
+
+#include "../rms_norm.h"
+
+DESCRIPTOR(musa)
+
+#endif

src/infiniop/ops/rms_norm/musa/rms_norm_musa.mu

+#include "../../../devices/musa/common_musa.h"
+#include "rms_norm_kernel.h"
+#include "rms_norm_musa.h"
+
+namespace op::rms_norm::musa {
+
+struct Descriptor::Opaque {
+    std::shared_ptr<device::musa::Handle::Internal> internal;
+};
+
+Descriptor::~Descriptor() {
+    delete _opaque;
+}
+
+infiniStatus_t Descriptor::create(
+    infiniopHandle_t handle,
+    Descriptor **desc_ptr,
+    infiniopTensorDescriptor_t y_desc,
+    infiniopTensorDescriptor_t x_desc,
+    infiniopTensorDescriptor_t w_desc,
+    float epsilon) {
+    auto result = RMSNormInfo::create(y_desc, x_desc, w_desc, epsilon);
+    CHECK_RESULT(result);
+    auto info = result.take();
+
+    // only support contiguous last dimension
+    if (info.x_strides[1] != 1 || info.y_strides[1] != 1) {
+        return INFINI_STATUS_BAD_TENSOR_STRIDES;
+    }
+
+    *desc_ptr = new Descriptor(
+        new Opaque{reinterpret_cast<device::musa::Handle *>(handle)->internal()},
+        std::move(info),
+        0,
+        handle->device, handle->device_id);
+    return INFINI_STATUS_SUCCESS;
+}
+
+// launch kernel with different data types
+template <unsigned int BLOCK_SIZE>
+infiniStatus_t launchKernel(
+    uint32_t batch_size, size_t dim,
+    void *y, infiniDtype_t atype, ptrdiff_t stride_y,
+    const void *x, ptrdiff_t stride_x,
+    const void *w, infiniDtype_t wtype,
+    float epsilon,
+    musaStream_t musa_stream) {
+
+#define LAUNCH_KERNEL(Tdata, Tweight, Tcompute)                                                     \
+    rmsnormBlock<BLOCK_SIZE, Tdata, Tweight, Tcompute><<<batch_size, BLOCK_SIZE, 0, musa_stream>>>( \
+        reinterpret_cast<Tdata *>(y),                                                               \
+        stride_y,                                                                                   \
+        reinterpret_cast<const Tdata *>(x),                                                         \
+        stride_x,                                                                                   \
+        reinterpret_cast<const Tweight *>(w),                                                       \
+        dim,                                                                                        \
+        epsilon)
+
+    if (atype == INFINI_DTYPE_F16 && wtype == INFINI_DTYPE_F16) {
+        LAUNCH_KERNEL(half, half, float);
+    } else if (atype == INFINI_DTYPE_F16 && wtype == INFINI_DTYPE_F32) {
+        LAUNCH_KERNEL(half, float, float);
+    } else if (atype == INFINI_DTYPE_F32 && wtype == INFINI_DTYPE_F32) {
+        LAUNCH_KERNEL(float, float, float);
+    } else {
+        return INFINI_STATUS_BAD_TENSOR_DTYPE;
+    }
+
+#undef LAUNCH_KERNEL
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t Descriptor::calculate(
+    void *workspace, size_t workspace_size,
+    void *y, const void *x, const void *w,
+    void *stream) const {
+
+    if (workspace_size < _workspace_size) {
+        return INFINI_STATUS_INSUFFICIENT_WORKSPACE;
+    }
+
+    auto stride_x = _info.x_strides[0];
+    auto stride_y = _info.y_strides[0];
+    auto dim = _info.dim();
+    uint32_t batch_size = static_cast<uint32_t>(_info.shape[0]);
+    auto musa_stream = reinterpret_cast<musaStream_t>(stream);
+
+    // launch kernel with different block sizes
+    if (_opaque->internal->maxThreadsPerBlock() == MUSA_BLOCK_SIZE_1024) {
+        CHECK_STATUS(launchKernel<MUSA_BLOCK_SIZE_1024>(batch_size, dim, y, _info.atype, stride_y, x, stride_x, w, _info.wtype, _info.epsilon, musa_stream));
+    } else {
+        return INFINI_STATUS_DEVICE_ARCHITECTURE_NOT_SUPPORTED;
+    }
+    return INFINI_STATUS_SUCCESS;
+}
+} // namespace op::rms_norm::musa

src/infiniop/ops/rms_norm/operator.cc

@@ -11,6 +11,9 @@
 #ifdef ENABLE_ASCEND_API
 #include "ascend/rms_norm_aclnn.h"
 #endif
+#ifdef ENABLE_MOORE_API
+#include "musa/rms_norm_musa.h"
+#endif
 #ifdef ENABLE_KUNLUN_API
 #include "kunlun/rms_norm_kunlun.h"
 #endif
@@ -56,10 +59,8 @@ __C infiniStatus_t infiniopCreateRMSNormDescriptor(
         return macaCreateRMSNormDescriptor((MacaHandle_t)handle, (RMSNormMacaDescriptor_t *)desc_ptr, y_desc, x_desc, w_desc, epsilon);
     }
 #endif
-#ifdef ENABLE_MTHREADS_GPU
-    case DevMthreadsGpu: {
-        return musaCreateRMSNormDescriptor((MusaHandle_t)handle, (RMSNormMusaDescriptor_t *)desc_ptr, y_desc, x_desc, w_desc, epsilon);
-    }
+#ifdef ENABLE_MOORE_API
+        CREATE(INFINI_DEVICE_MOORE, musa)
 #endif
     }
 
@@ -98,10 +99,8 @@ __C infiniStatus_t infiniopGetRMSNormWorkspaceSize(infiniopRMSNormDescriptor_t d
         return macaGetRMSNormWorkspaceSize((RMSNormMacaDescriptor_t)desc, size);
     }
 #endif
-#ifdef ENABLE_MTHREADS_GPU
-    case DevMthreadsGpu: {
-        return musaGetRMSNormWorkspaceSize((RMSNormMusaDescriptor_t)desc, size);
-    }
+#ifdef ENABLE_MOORE_API
+        GET(INFINI_DEVICE_MOORE, musa)
 #endif
     }
 
@@ -141,10 +140,8 @@ __C infiniStatus_t infiniopRMSNorm(infiniopRMSNormDescriptor_t desc, void *works
         return macaRMSNorm((RMSNormMacaDescriptor_t)desc, workspace, workspace_size, y, x, w, stream);
     }
 #endif
-#ifdef ENABLE_MTHREADS_GPU
-    case DevMthreadsGpu: {
-        return musaRMSNorm((RMSNormMusaDescriptor_t)desc, workspace, workspace_size, y, x, w, stream);
-    }
+#ifdef ENABLE_MOORE_API
+        CALCULATE(INFINI_DEVICE_MOORE, musa)
 #endif
     }
 
@@ -183,10 +180,8 @@ __C infiniStatus_t infiniopDestroyRMSNormDescriptor(infiniopRMSNormDescriptor_t
         return macaDestroyRMSNormDescriptor((RMSNormMacaDescriptor_t)desc);
     }
 #endif
-#ifdef ENABLE_MTHREADS_GPU
-    case DevMthreadsGpu: {
-        return musaDestroyRMSNormDescriptor((RMSNormMusaDescriptor_t)desc);
-    }
+#ifdef ENABLE_MOORE_API
+        DESTROY(INFINI_DEVICE_MOORE, musa)
 #endif
     }
 

src/infiniop/reduce/musa/reduce.h

+#ifndef __INFINIOP_REDUCE_MUSA_H__
+#define __INFINIOP_REDUCE_MUSA_H__
+
+#include <cub/block/block_reduce.cuh>
+
+namespace op::common_musa::reduce_op {
+
+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] * 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);
+}
+
+} // namespace op::common_musa::reduce_op
+
+#endif

src/infinirt/infinirt.cc

@@ -4,6 +4,8 @@
 #include "bang/infinirt_bang.h"
 #include "cpu/infinirt_cpu.h"
 #include "cuda/infinirt_cuda.cuh"
+#include "maca/infinirt_maca.h"
+#include "musa/infinirt_musa.h"
 
 thread_local infiniDevice_t CURRENT_DEVICE_TYPE = INFINI_DEVICE_CPU;
 thread_local int CURRENT_DEVICE_ID = 0;
@@ -58,6 +60,12 @@ __C infiniStatus_t infinirtGetDevice(infiniDevice_t *device_ptr, int *device_id_
         case INFINI_DEVICE_ASCEND:                          \
             _status = infinirt::ascend::API PARAMS;         \
             break;                                          \
+        case INFINI_DEVICE_METAX:                           \
+            _status = infinirt::maca::API PARAMS;           \
+            break;                                          \
+        case INFINI_DEVICE_MOORE:                           \
+            _status = infinirt::musa::API PARAMS;           \
+            break;                                          \
         default:                                            \
             return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED; \
         }                                                   \