From cd5d97e34b3b79dbd12ba9baf4523934589137ba Mon Sep 17 00:00:00 2001
From: liu zhengxi <380185688@qq.com>
Date: Fri, 15 Nov 2019 10:49:17 +0800
Subject: [PATCH] fix the cuda bilinear and nearest precision, test=develop
 (#2426)

fix the cuda bilinear and nearest precision caused by data type conversion.
---
 lite/kernels/cuda/bilinear_interp_compute.cu     | 4 ++--
 lite/kernels/cuda/nearest_interp_compute.cu      | 4 ++--
 lite/kernels/cuda/nearest_interp_compute_test.cc | 4 +++-
 3 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/lite/kernels/cuda/bilinear_interp_compute.cu b/lite/kernels/cuda/bilinear_interp_compute.cu
index 4f8b5dc8f8..00b1457938 100644
--- a/lite/kernels/cuda/bilinear_interp_compute.cu
+++ b/lite/kernels/cuda/bilinear_interp_compute.cu
@@ -28,8 +28,8 @@ inline std::vector<int> get_new_shape(
   for (size_t i = 0; i < list_new_shape_tensor.size(); ++i) {
     auto tensor = list_new_shape_tensor[i];
     lite::Tensor temp;
-    auto temp_data = temp.mutable_data<int32_t>();
-    auto tensor_data = tensor->data<int32_t>();
+    auto temp_data = temp.mutable_data<float>();
+    auto tensor_data = tensor->data<float>();
     cudaMemcpy(temp_data,
                tensor_data,
                tensor->dims().production() * sizeof(float),
diff --git a/lite/kernels/cuda/nearest_interp_compute.cu b/lite/kernels/cuda/nearest_interp_compute.cu
index 89d9c66ccd..adae034a1d 100644
--- a/lite/kernels/cuda/nearest_interp_compute.cu
+++ b/lite/kernels/cuda/nearest_interp_compute.cu
@@ -28,8 +28,8 @@ inline std::vector<int> get_new_shape(
   for (size_t i = 0; i < list_new_shape_tensor.size(); ++i) {
     auto tensor = list_new_shape_tensor[i];
     lite::Tensor temp;
-    auto temp_data = temp.mutable_data<int32_t>();
-    auto tensor_data = tensor->data<int32_t>();
+    auto temp_data = temp.mutable_data<float>();
+    auto tensor_data = tensor->data<float>();
     cudaMemcpy(temp_data,
                tensor_data,
                tensor->dims().production() * sizeof(float),
diff --git a/lite/kernels/cuda/nearest_interp_compute_test.cc b/lite/kernels/cuda/nearest_interp_compute_test.cc
index 60e30e55de..ad2ef9294e 100644
--- a/lite/kernels/cuda/nearest_interp_compute_test.cc
+++ b/lite/kernels/cuda/nearest_interp_compute_test.cc
@@ -202,6 +202,7 @@ TEST(nearest_interp, update) {
   float* size_tensor1_ref_data = size_tensor_ref[1].mutable_data<float>();
   float* input_scale_ref_data = input_scale_ref.mutable_data<float>();
   float* osz_ref_data = osz_ref.mutable_data<float>();
+  float* out_ref_data = out_ref.mutable_data<float>();
 
   for (int i = 0; i < x_cpu.numel(); ++i) {
     x_cpu_data[i] = i + 5.0;
@@ -247,8 +248,9 @@ TEST(nearest_interp, update) {
 
   CopySync<TARGET(kCUDA)>(
       out_cpu_data, out_data, sizeof(float) * out.numel(), IoDirection::DtoH);
+  NearestInterpRef(&x_ref, &out_ref, false);
   for (int i = 0; i < out.numel(); i++) {
-    LOG(INFO) << out_cpu_data[i];
+    EXPECT_NEAR(out_cpu_data[i], out_ref_data[i], 1e-5);
   }
 }
 
-- 
GitLab