Merge pull request #350 from InfiniTensor/p800-add

issue/349 P800 add

src/infiniop/ops/add/kunlun/add_kunlun.h

+#ifndef __ADD_KUNLUN_API_H__
+#define __ADD_KUNLUN_API_H__
+
+#include "../../../elementwise/kunlun/elementwise_kunlun_api.h"
+
+ELEMENTWISE_DESCRIPTOR(add, kunlun)
+
+#endif // __ADD_KUNLUN_API_H__

src/infiniop/ops/add/kunlun/add_kunlun.xpu

+#include "../../../elementwise/kunlun/elementwise_kunlun.h"
+#include "add_kunlun.h"
+#include "kernel.h"
+
+namespace op::elementwise::kunlun {
+
+using AddOp = op::add::kunlun::AddOp;
+
+INSTANTIATE_ELEMENTWISE_KERNEL(AddOp::num_inputs, AddOp, float);
+INSTANTIATE_ELEMENTWISE_KERNEL(AddOp::num_inputs, AddOp, half);
+INSTANTIATE_ELEMENTWISE_KERNEL(AddOp::num_inputs, AddOp, bfloat16_t);
+} // namespace op::elementwise::kunlun
+
+namespace op::add::kunlun {
+
+Descriptor::~Descriptor() = default;
+
+infiniStatus_t Descriptor::create(
+    infiniopHandle_t handle_,
+    Descriptor **desc_ptr,
+    infiniopTensorDescriptor_t out_desc,
+    std::vector<infiniopTensorDescriptor_t> input_desc_vec) {
+
+    auto handle = reinterpret_cast<device::kunlun::Handle *>(handle_);
+    auto dtype = out_desc->dtype();
+
+    const auto &a_desc = input_desc_vec.at(0);
+    const auto &b_desc = input_desc_vec.at(1);
+    const auto &c_shape = out_desc->shape();
+    const auto &a_shape = a_desc->shape();
+    const auto &b_shape = b_desc->shape();
+
+    CHECK_DTYPE(dtype, INFINI_DTYPE_F16, INFINI_DTYPE_F32, INFINI_DTYPE_BF16);
+
+    CHECK_SAME_SHAPE(c_shape, a_shape, b_shape);
+
+    // create KUNLUN elementwise descriptor
+    CREATE_ELEMENTWISE_KUNLUN_DESCRIPTOR(handle, dtype, out_desc, input_desc_vec)
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t Descriptor::calculate(
+    void *workspace,
+    size_t workspace_size,
+    void *output,
+    std::vector<const void *> inputs,
+    void *stream) const {
+
+    if (workspace_size < _workspace_size) {
+        return INFINI_STATUS_INSUFFICIENT_WORKSPACE;
+    }
+
+    switch (_dtype) {
+    case INFINI_DTYPE_F16:
+        return _device_info->calculate<8, AddOp, half>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_BF16:
+        return _device_info->calculate<8, AddOp, bfloat16_t>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_F32:
+        return _device_info->calculate<8, AddOp, float>(_info, workspace, output, inputs, stream);
+    default:
+        return INFINI_STATUS_BAD_TENSOR_DTYPE;
+    }
+
+    return INFINI_STATUS_SUCCESS;
+}
+} // namespace op::add::kunlun

src/infiniop/ops/add/kunlun/kernel.h

+#ifndef __ADD_KUNLUN_KERNEL_H__
+#define __ADD_KUNLUN_KERNEL_H__
+
+namespace op::add::kunlun {
+
+typedef struct AddOp {
+public:
+    static constexpr int num_inputs = 2;
+    template <typename T>
+    inline __device__ T operator()(const T *inputs) const {
+        T a = inputs[0];
+        T b = inputs[1];
+        return a + b;
+    }
+    // bfloat16 特化版本（使用 float 计算精度）
+    inline __device__ bfloat16_t operator()(const bfloat16_t *inputs) const {
+        float a_f = __bfloat162float(inputs[0]);
+        float b_f = __bfloat162float(inputs[1]);
+        return __float2bfloat16(a_f + b_f);
+    }
+} AddOp;
+} // namespace op::add::kunlun
+
+#endif // __ADD_KUNLUN_KERNEL_H__

src/infiniop/ops/add/operator.cc

@@ -11,6 +11,9 @@
 #ifdef ENABLE_METAX_API
 #include "metax/add_metax.h"
 #endif
+#ifdef ENABLE_KUNLUN_API
+#include "kunlun/add_kunlun.h"
+#endif
 
 __C infiniStatus_t infiniopCreateAddDescriptor(
     infiniopHandle_t handle,
@@ -42,6 +45,9 @@ __C infiniStatus_t infiniopCreateAddDescriptor(
 #ifdef ENABLE_METAX_API
         CREATE(INFINI_DEVICE_METAX, metax);
 #endif
+#ifdef ENABLE_KUNLUN_API
+        CREATE(INFINI_DEVICE_KUNLUN, kunlun);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
@@ -69,6 +75,9 @@ __C infiniStatus_t infiniopGetAddWorkspaceSize(infiniopAddDescriptor_t desc, siz
 #endif
 #ifdef ENABLE_METAX_API
         GET(INFINI_DEVICE_METAX, metax);
+#endif
+#ifdef ENABLE_KUNLUN_API
+        GET(INFINI_DEVICE_KUNLUN, kunlun);
 #endif
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
@@ -106,6 +115,9 @@ __C infiniStatus_t infiniopAdd(
 #ifdef ENABLE_METAX_API
         CALCULATE(INFINI_DEVICE_METAX, metax);
 #endif
+#ifdef ENABLE_KUNLUN_API
+        CALCULATE(INFINI_DEVICE_KUNLUN, kunlun);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
@@ -136,6 +148,9 @@ infiniopDestroyAddDescriptor(infiniopAddDescriptor_t desc) {
 #ifdef ENABLE_METAX_API
         DELETE(INFINI_DEVICE_METAX, metax);
 #endif
+#ifdef ENABLE_KUNLUN_API
+        DELETE(INFINI_DEVICE_KUNLUN, kunlun);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;