Zmz/matmul12 (DeepLink-org#869)

* support matmul

impl/ascend/convert_config.yaml

@@ -283,6 +283,9 @@
 - diopiHardswishBackward:
     dtype: (float64)->float32
 
+- diopiMatmul:
+    dtype: (float64)->float32
+
 - diopiAtan:
     dtype: (uint8, int8, int32, int16, int64, bool)->float32
 

impl/ascend/device_configs.py

@@ -431,7 +431,7 @@
             args=[
                 {
                     "ins": ['input'],
-                    "shape": [Skip((128, 49, 128)),Skip((5,)),Skip((128, 4, 49, 32)),Skip((2, 1, 3136, 3136)),Skip((2, 784, 64)),Skip((2, 16, 8, 64)),Skip((2, 31, 6, 40, 512)),],
+                    "shape": [Skip((2, 31, 6, 40, 512)),],
                 },
             ]
         ),

impl/ascend_npu/ascend_config.yaml

@@ -200,6 +200,7 @@ ascend:
 - diopiScatterScalar
 - diopiScatterInpScalar
 ascend_npu:
+- diopiMatmul
 - diopiCastDtype
 - diopiCopyInp
 - diopiCat

impl/ascend_npu/diopi_impl/matmul.cpp

@@ -0,0 +1,18 @@
+/**
+ * @file
+ * @author DeepLink
+ * @copyright  (c) 2023, DeepLink.
+ */
+
+#include "helper.hpp"
+#include "op_plugin/AclOpsInterface.h"
+
+namespace OP_IMPL_NS {
+
+diopiError_t diopiMatmul(diopiContextHandle_t ctx, diopiTensorHandle_t out, diopiConstTensorHandle_t input, diopiConstTensorHandle_t other) {
+    BEGIN_CALL_ACL_OP(input, out, other);
+    acl_op::matmul_out(inputAt, otherAt, outAt);
+    END_CALL_ACL_OP();
+}
+
+}  // namespace OP_IMPL_NS

impl/ascend_npu/torch_npu/csrc/DIOPIAdapter.cpp

@@ -2895,7 +2895,24 @@ at::Tensor viewStorage(const at::Tensor input, const c10::IntArrayRef sizes, con
             if (st != -1) st *= sizes[i - 1];
         }
     }
-    return fromPreAllocated(input.data_ptr() + storageOffset * input.itemsize(), sizes, stridesVec, input.options());
+
+    // when shape[0]=-1, fill data
+    std::vector<int64_t> sizeVec(sizes.size(), 1);
+    std::copy(sizes.begin(), sizes.end(), sizeVec.begin());
+    if (!sizes.empty() && sizes[0] == -1) {
+        bool flag = true;
+        for (auto i : sizes) {
+            if (!flag && i < 0) {
+                TORCH_CHECK(false, "more than one -1, sizes=", sizes);
+            }
+            if (i < 0) {
+                flag = false;
+            }
+        }
+        int count = std::accumulate(sizeVec.begin() + 1, sizeVec.end(), 1, std::multiplies<int>());
+        sizeVec[0] = input.numel() / count;
+    }
+    return fromPreAllocated(input.data_ptr() + storageOffset * input.itemsize(), sizeVec, stridesVec, input.options());
 }
 
 c10::List<c10::optional<at::Tensor>> castIntIndicesToLongIndices(const c10::List<c10::optional<at::Tensor>>& indices) {
@@ -3057,7 +3074,11 @@ at::Tensor wrapper__transpose(const at::Tensor& self, int64_t dim0, int64_t dim1
 }
 
 at::Scalar wrapper___local_scalar_dense(const at::Tensor& self) { return at_npu::native::NPUNativeFunctions::_local_scalar_dense(self); }
+at::Tensor& wrapper_out_mm_out(const at::Tensor& self, const at::Tensor& mat2, at::Tensor& out) { return acl_op::mm_out(self, mat2, out); }
 
+at::Tensor& wrapper_source_Tensor_set_(at::Tensor& self, const at::Tensor& source) { return at_npu::native::NPUNativeFunctions::set_(self, source); }
+at::Tensor& wrapper_out_bmm_out(const at::Tensor& self, const at::Tensor& mat2, at::Tensor& out) { return acl_op::bmm_out(self, mat2, out); }
+at::Tensor wrapper__dot(const at::Tensor& self, const at::Tensor& tensor) { return acl_op::dot(self, tensor); }
 }  // namespace
 
 namespace at {
@@ -3092,6 +3113,11 @@ TORCH_LIBRARY_IMPL(aten, XLA, m) {
     m.impl("repeat", TORCH_FN(wrapper__repeat));
     m.impl("transpose.int", TORCH_FN(wrapper__transpose));
     m.impl("_local_scalar_dense", TORCH_FN(wrapper___local_scalar_dense));
+    m.impl("cat", TORCH_FN(wrapper__cat));
+    m.impl("mm.out", TORCH_FN(wrapper_out_mm_out));
+    m.impl("set_.source_Tensor", TORCH_FN(wrapper_source_Tensor_set_));
+    m.impl("dot", TORCH_FN(wrapper__dot));
+    m.impl("bmm.out", TORCH_FN(wrapper_out_bmm_out));
 };
 
 TORCH_LIBRARY_IMPL(_, XLA, m) { m.fallback(torch::CppFunction::makeFromBoxedFunction<&ascend_diopi_fallback>()); }