Align diopiStd to device with torch version>=2.0

dipu/scripts/autogen_diopi_wrapper/autogen_diopi_wrapper.py

@@ -1203,10 +1203,25 @@ def main():
                 continue
 
         # filter torch version
-        in_torch_vers = merged_fun_config.get("torch_ver", None)
+        supported_torch_ver_list = merged_fun_config.get("torch_ver", None)
         cur_torch_ver = merged_fun_config.get("current_torch_ver", None)
-        if in_torch_vers is not None and cur_torch_ver not in in_torch_vers:
-            continue
+
+        if supported_torch_ver_list != None:
+            exclude_torch_ver_list = []
+            include_torch_ver_list = []
+            all_include = False
+            for supported_torch_ver in supported_torch_ver_list:
+                if supported_torch_ver.startswith("-"):
+                    exclude_torch_ver_list.append(supported_torch_ver[1:])
+                elif supported_torch_ver == "all":
+                    all_include = True
+                else:
+                    include_torch_ver_list.append(supported_torch_ver)
+
+            if (cur_torch_ver in exclude_torch_ver_list) or (
+                all_include == False and (cur_torch_ver not in include_torch_ver_list)
+            ):
+                continue
 
         fun_code, register_code = functions_code_gen(merged_fun_config)
 

dipu/scripts/autogen_diopi_wrapper/diopi_functions.yaml

@@ -838,32 +838,40 @@
   custom_code_at_the_beginning: |
     c10::DimVector output_shape = infer_reduce_op_shape(self.sizes(), dim.value_or(c10::DimVector()), keepdim);
     auto out = nodispatch::empty(output_shape, self.options());
-    bool unbiased = correction.value_or(1) == 1;
+    ::diopiScalar_t correctionDiopiScalar;
+    const ::diopiScalar_t* correctionDiopiScalarPtr = nullptr;
+    if (correction.has_value()) {
+        correctionDiopiScalar = dipu::diopi_helper::toDiopiScalar(correction.value());
+        correctionDiopiScalarPtr = &correctionDiopiScalar;
+    }
     ::diopiSize_t diopi_size = toDiopiSize(dim);
-  interface: diopiStd(ctx, out, self, diopi_size, unbiased);
+  interface: diopiStd(ctx, out, self, diopi_size, correctionDiopiScalarPtr);
 
 - schema: "std.correction_out(Tensor self, int[1]? dim=None, *, int? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)"
   torch_ver: ["20000",]
   custom_code_at_the_beginning: |
     ::diopiSize_t diopi_size = toDiopiSize(dim);
-    bool unbiased = correction.value_or(1) == 1;
-  interface: diopiStd(ctx, out, self, diopi_size, unbiased);
+    ::diopiScalar_t correctionDiopiScalar;
+    const ::diopiScalar_t* correctionDiopiScalarPtr = nullptr;
+    if (correction.has_value()) {
+        correctionDiopiScalar = dipu::diopi_helper::toDiopiScalar(correction.value());
+        correctionDiopiScalarPtr = &correctionDiopiScalar;
+    }
+  interface: diopiStd(ctx, out, self, diopi_size, correctionDiopiScalarPtr);
 
 - schema: "std.correction(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False) -> Tensor"
-  torch_ver: ["20100", "20101", "20202"]
+  torch_ver: [all, "-20000"]
   custom_code_at_the_beginning: |
     c10::DimVector output_shape = infer_reduce_op_shape(self.sizes(), dim.value_or(c10::DimVector()), keepdim);
     auto out = nodispatch::empty(output_shape, self.options());
-    bool unbiased = correction.value_or(1).toLong() == 1;
     ::diopiSize_t diopi_size = toDiopiSize(dim);
-  interface: diopiStd(ctx, out, self, diopi_size, unbiased);
+  interface: diopiStd(ctx, out, self, diopi_size, correctionDiopiScalarPtr);
 
 - schema: "std.correction_out(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)"
-  torch_ver: ["20100", "20101", "20202"]
+  torch_ver: [all, "-20000"]
   custom_code_at_the_beginning: |
     ::diopiSize_t diopi_size = toDiopiSize(dim);
-    bool unbiased = correction.value_or(1).toLong() == 1;
-  interface: diopiStd(ctx, out, self, diopi_size, unbiased);
+  interface: diopiStd(ctx, out, self, diopi_size, correctionDiopiScalarPtr)
 
 - schema: "linear_backward(Tensor input, Tensor grad_output, Tensor weight, bool[3] output_mask) -> (Tensor grad_input, Tensor grad_weight, Tensor grad_bias)"
   device: [all, -cuda, -muxi, -ascend]

dipu/tests/python/unittests/test_mean_std.py

@@ -1,7 +1,7 @@
 # Copyright (c) 2023, DeepLink.
 import torch
 import torch_dipu
-from torch_dipu.testing._internal.common_utils import TestCase, run_tests
+from torch_dipu.testing._internal.common_utils import TestCase, run_tests, skipOn
 
 
 class TestMeanStd(TestCase):
@@ -61,6 +61,17 @@ def test_std(self):
             )
         )
 
+    @skipOn("MLU", "camb does not support this type")
+    def test_std_correction(self):
+        self.assertTrue(
+            torch.allclose(
+                torch.std(self.a, dim=-1, correction=20).cpu(),
+                torch.std(self.a.cpu(), dim=-1, correction=20),
+                atol=1e-3,
+                rtol=1e-3,
+            )
+        )
+
 
 if __name__ == "__main__":
     run_tests()

dipu/tests/pytorch_config_cuda.py

@@ -56,8 +56,6 @@
     "TestReductionsDIPU": {
         "test_ref_large_input_1D",
         "test_ref_large_input_64bit_indexing",
-        # will fail because diopiStd not align with torch.std, will fix later
-        "test_warn_invalid_degrees_of_freedom",
     },
 }