From f218330e8716c255f1c1b765a5b4060d204348a6 Mon Sep 17 00:00:00 2001
From: zhiboniu <31800336+zhiboniu@users.noreply.github.com>
Date: Thu, 16 Sep 2021 20:30:03 +0800
Subject: [PATCH] fix group_size = floor(C/groups) from ceil(C/groups); add
 groupnorm group divisible check (#35644)

---
 paddle/fluid/operators/group_norm_op.cc | 6 ++++++
 paddle/fluid/operators/group_norm_op.cu | 5 +++--
 paddle/fluid/operators/group_norm_op.h  | 8 ++++----
 3 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/paddle/fluid/operators/group_norm_op.cc b/paddle/fluid/operators/group_norm_op.cc
index e076444626e..2c7fd8f4173 100644
--- a/paddle/fluid/operators/group_norm_op.cc
+++ b/paddle/fluid/operators/group_norm_op.cc
@@ -66,6 +66,12 @@ class GroupNormOp : public framework::OperatorWithKernel {
             "The Attr(groups) of Op(group_norm) must be "
             "greater than or equal to 1. But received: groups is [%s].",
             groups));
+    PADDLE_ENFORCE_EQ(
+        channel_num % groups, 0,
+        platform::errors::InvalidArgument(
+            "Expected number of channels in input to be divisible by "
+            "num_groups, but got input channel is %d and num_groups is %d",
+            channel_num, groups));
 
     if (ctx->HasInput("Scale")) {
       PADDLE_ENFORCE_EQ(
diff --git a/paddle/fluid/operators/group_norm_op.cu b/paddle/fluid/operators/group_norm_op.cu
index f199bfeb944..e029c84090a 100644
--- a/paddle/fluid/operators/group_norm_op.cu
+++ b/paddle/fluid/operators/group_norm_op.cu
@@ -144,7 +144,8 @@ class GroupNormKernel<platform::CUDADeviceContext, T>
     const int C =
         (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
-    const int group_size = (C - 1) / groups + 1;
+    const int group_size = C / groups;
+
     const int W =
         (data_layout == DataLayout::kNCHW ? x_dims[x_dims.size() - 1]
                                           : x_dims[x_dims.size() - 2]);
@@ -314,7 +315,7 @@ class GroupNormGradKernel<platform::CUDADeviceContext, T>
     const int C =
         (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
-    const int group_size = (C - 1) / groups + 1;
+    const int group_size = C / groups;
     const int W =
         (data_layout == DataLayout::kNCHW ? x_dims[x_dims.size() - 1]
                                           : x_dims[x_dims.size() - 2]);
diff --git a/paddle/fluid/operators/group_norm_op.h b/paddle/fluid/operators/group_norm_op.h
index f2388699e26..9cb451235f1 100644
--- a/paddle/fluid/operators/group_norm_op.h
+++ b/paddle/fluid/operators/group_norm_op.h
@@ -52,7 +52,7 @@ class GroupNormKernel : public framework::OpKernel<T> {
     const int C =
         (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
-    const int group_size = (C - 1) / groups + 1;
+    const int group_size = C / groups;
 
     y->mutable_data<T>(ctx.GetPlace());
     mean->mutable_data<T>(ctx.GetPlace());
@@ -100,7 +100,7 @@ class GroupNormKernel : public framework::OpKernel<T> {
             int imid;
             for (imid = 0; imid < imsize - (imsize % M);
                  imid += M, iter_x_data += M) {
-              // TODO(gaoxiang) ：Because AVX/AVX2/AVX512 can not directly used
+              // TODO(gaoxiang): Because AVX/AVX2/AVX512 can not directly used
               // in template class/function, before we complete high
               // performance cpu vector extension, temporarily unrolling
               // loop to get high precision and performance
@@ -138,7 +138,7 @@ class GroupNormKernel : public framework::OpKernel<T> {
             int imid;
             for (imid = 0; imid < imsize - (imsize % M);
                  imid += M, iter_x_data += M * C) {
-              // TODO(gaoxiang) ：Because AVX/AVX2/AVX512 can not directly used
+              // TODO(gaoxiang): Because AVX/AVX2/AVX512 can not directly used
               // in template class/function, before we complete high
               // performance cpu vector extension, temporarily unrolling
               // loop to get high precision and performance
@@ -236,7 +236,7 @@ class GroupNormGradKernel : public framework::OpKernel<T> {
     const int C =
         (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
-    const int group_size = (C - 1) / groups + 1;
+    const int group_size = C / groups;
 
     d_x->mutable_data<T>(ctx.GetPlace());
     math::SetConstant<DeviceContext, T> set_zero;
-- 
GitLab