!1175 LayerNormGrad fix & codex

Merge pull request !1175 from chenweifeng/layernorm

mindspore/ccsrc/kernel/gpu/arrays/unsorted_segment_sum_gpu_kernel.cc

@@ -37,6 +37,5 @@ MS_REG_GPU_KERNEL_TWO(
   UnsortedSegmentSum,
   KernelAttr().AddInputAttr(kNumberTypeInt32).AddInputAttr(kNumberTypeInt64).AddOutputAttr(kNumberTypeInt32),
   UnsortedSegmentSumGpuKernel, int, int64_t)
-
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/gpu/data/dataset_utils.cc

@@ -18,7 +18,6 @@
 
 namespace mindspore {
 namespace kernel {
-
 size_t UnitSizeInBytes(const mindspore::TypeId &t) {
   size_t bytes = 0;
   switch (t) {
@@ -65,6 +64,5 @@ int ElementNums(const std::vector<int> &shape) {
 
   return nums;
 }
-
 }  // namespace kernel
 }  // namespace mindspore

mindspore/ccsrc/kernel/gpu/data/dataset_utils.h

@@ -21,7 +21,6 @@
 #include "ir/dtype/type.h"
 namespace mindspore {
 namespace kernel {
-
 size_t UnitSizeInBytes(const mindspore::TypeId &t);
 int ElementNums(const std::vector<int> &shape);
 }  // namespace kernel

mindspore/ccsrc/kernel/gpu/math/broadcast_gpu_kernel.h

@@ -27,7 +27,6 @@
 #include "kernel/gpu/kernel_constants.h"
 namespace mindspore {
 namespace kernel {
-
 template <typename T, typename S>
 class BroadcastOpGpuKernel : public GpuKernel {
  public:
@@ -70,14 +69,14 @@ class BroadcastOpGpuKernel : public GpuKernel {
       output_num_ *= shape3[i];
     }
     int offset = shape3.size() - shape1.size();
-    for (size_t i = 0; i < shape1.size(); i++) {
+    for (size_t j = 0; j < shape1.size(); j++) {
-      lhs_shape_[i + offset] = shape1[i];
+      lhs_shape_[j + offset] = shape1[j];
-      input1_num_ *= shape1[i];
+      input1_num_ *= shape1[j];
     }
     offset = shape3.size() - shape2.size();
-    for (size_t i = 0; i < shape2.size(); i++) {
+    for (size_t k = 0; k < shape2.size(); k++) {
-      rhs_shape_[i + offset] = shape2[i];
+      rhs_shape_[k + offset] = shape2[k];
-      input2_num_ *= shape2[i];
+      input2_num_ *= shape2[k];
     }
 
     InitSizeLists();

mindspore/ccsrc/kernel/gpu/math/broadcast_grad_gpu_kernel.h

@@ -1,4 +1,3 @@
-
 /**
  * Copyright 2020 Huawei Technologies Co., Ltd
  *
@@ -28,7 +27,6 @@
 #include "kernel/gpu/kernel_constants.h"
 namespace mindspore {
 namespace kernel {
-
 template <typename T>
 class BroadcastOpGradGpuKernel : public GpuKernel {
  public:

mindspore/ccsrc/kernel/gpu/nn/layer_norm_grad_gpu_kernel.h

@@ -36,8 +36,8 @@ class LayerNormGradGpuKernel : public GpuKernel {
 
   bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
               const std::vector<AddressPtr> &outputs, uintptr_t stream_ptr) override {
-    auto dy = GetDeviceAddress<T>(inputs, 0);
+    auto x = GetDeviceAddress<T>(inputs, 0);
-    auto x = GetDeviceAddress<T>(inputs, 1);
+    auto dy = GetDeviceAddress<T>(inputs, 1);
     auto var = GetDeviceAddress<T>(inputs, 2);
     auto mean = GetDeviceAddress<T>(inputs, 3);
     auto gamma = GetDeviceAddress<T>(inputs, 4);

mindspore/ccsrc/kernel/gpu/nn/rmsprop_gpu_kernel.cc

@@ -44,6 +44,5 @@ MS_REG_GPU_KERNEL_ONE(ApplyCenteredRMSProp,
                         .AddInputAttr(kNumberTypeFloat32)
                         .AddOutputAttr(kNumberTypeFloat32),
                       RMSPropGpuKernel, float)
-
 }  // namespace kernel
 }  // namespace mindspore

tests/st/ops/gpu/test_layer_norm_grad_op.py

@@ -80,7 +80,7 @@ def test_layernormgrad0():
     gamma_ms = Tensor(gamma_np)
 
     net = LayerNormGradNet(begin_norm_axis, begin_params_axis)
-    dx_ms, dg_ms, db_ms = net(dy_ms, x_ms, var_ms, mean_ms, gamma_ms)
+    dx_ms, dg_ms, db_ms = net(x_ms, dy_ms, var_ms, mean_ms, gamma_ms)
 
     assert np.allclose(dx_ms.asnumpy(), dx_np, rtol=1e-6, atol=1e-6)
     assert np.allclose(dg_ms.asnumpy(), dg_np, rtol=1e-6, atol=1e-3)
@@ -107,7 +107,7 @@ def test_layernormgrad1():
     gamma_ms = Tensor(gamma_np)
 
     net = LayerNormGradNet(begin_norm_axis, begin_params_axis)
-    dx_ms, dg_ms, db_ms = net(dy_ms, x_ms, var_ms, mean_ms, gamma_ms)
+    dx_ms, dg_ms, db_ms = net(x_ms, dy_ms, var_ms, mean_ms, gamma_ms)
 
     assert np.allclose(dx_ms.asnumpy(), dx_np, rtol=1e-6, atol=1e-6)
     assert np.allclose(dg_ms.asnumpy(), dg_np, rtol=1e-6, atol=1e-3)
@@ -133,8 +133,8 @@ def test_layernormgrad2():
     gamma_ms = Tensor(gamma_np)
 
     net = LayerNormGradNet(begin_norm_axis, begin_params_axis)
-    dx_ms, dg_ms, db_ms = net(dy_ms, x_ms, var_ms, mean_ms, gamma_ms)
+    dx_ms, dg_ms, db_ms = net(x_ms, dy_ms, var_ms, mean_ms, gamma_ms)
 
     assert np.allclose(dx_ms.asnumpy(), dx_np, rtol=1e-6, atol=1e-6)
     assert np.allclose(dg_ms.asnumpy(), dg_np, rtol=1e-6, atol=1e-3)
-    assert np.allclose(db_ms.asnumpy(), db_np, rtol=1e-6, atol=1e-3)
+    assert np.allclose(db_ms.asnumpy(), db_np, rtol=1e-6, atol=1e-3)