From 1d74eb9d99b8fdc5126dc912bbfb51c534930cc2 Mon Sep 17 00:00:00 2001
From: liuruilong <liuruilong@baidu.com>
Date: Mon, 15 Oct 2018 15:06:24 +0800
Subject: [PATCH]  init image when init kernel

---
 src/framework/cl/cl_image.h                       |  1 +
 .../kernel/cl/conv_add_bn_relu_kernel.cpp         | 15 +++++++++++----
 src/operators/kernel/cl/conv_add_kernel.cpp       |  3 +++
 src/operators/kernel/cl/conv_kernel.cpp           |  5 ++---
 src/operators/kernel/cl/depthwise_conv_kernel.cpp |  1 +
 5 files changed, 18 insertions(+), 7 deletions(-)

diff --git a/src/framework/cl/cl_image.h b/src/framework/cl/cl_image.h
index 1c5da6146d..b8645a4298 100644
--- a/src/framework/cl/cl_image.h
+++ b/src/framework/cl/cl_image.h
@@ -61,6 +61,7 @@ class CLImage {
       PADDLE_MOBILE_THROW_EXCEPTION(
           " empty image tensor data shouldn't have value");
     }
+    DLOG << " init empty image ";
     InitCLImage(context, nullptr, dim);
     initialized_ = true;
   }
diff --git a/src/operators/kernel/cl/conv_add_bn_relu_kernel.cpp b/src/operators/kernel/cl/conv_add_bn_relu_kernel.cpp
index 29b13b6abc..6b16039f15 100644
--- a/src/operators/kernel/cl/conv_add_bn_relu_kernel.cpp
+++ b/src/operators/kernel/cl/conv_add_bn_relu_kernel.cpp
@@ -24,9 +24,16 @@ namespace operators {
 template <>
 bool ConvAddBNReluKernel<GPU_CL, float>::Init(
     FusionConvAddBNReluParam<GPU_CL> *param) {
+  PADDLE_MOBILE_ENFORCE(
+          param->Filter()->dims()[2] == param->Filter()->dims()[3] &&
+          param->Paddings()[0] == param->Paddings()[1],
+          "need equal");
+
+  param->Filter()->InitCLImage(cl_helper_.CLContext());
+  param->Bias()->InitCLImage(cl_helper_.CLContext());
+
   //  const CL *mean = param->InputMean();
   const framework::CLImage *mean = param->InputMean();
-
   const framework::CLImage *variance = param->InputVariance();
   const framework::CLImage *scale = param->InputScale();
   const framework::CLImage *bias = param->InputBias();
@@ -52,9 +59,6 @@ bool ConvAddBNReluKernel<GPU_CL, float>::Init(
     new_bias_ptr[i] = bias_ptr[i] - mean_ptr[i] * inv_std_ptr[i] * scale_ptr[i];
   }
 
-  delete[](new_scale_ptr);
-  delete[](new_bias_ptr);
-
   framework::CLImage *new_scale = new framework::CLImage();
 
   new_scale->SetTensorData(new_scale_ptr, variance->dims());
@@ -68,6 +72,9 @@ bool ConvAddBNReluKernel<GPU_CL, float>::Init(
   param->SetNewScale(new_scale);
   param->SetNewBias(new_bias);
 
+  delete[](new_scale_ptr);
+  delete[](new_bias_ptr);
+
   PADDLE_MOBILE_ENFORCE(
       param->Filter()->dims()[2] == param->Filter()->dims()[3] &&
           param->Paddings()[0] == param->Paddings()[1],
diff --git a/src/operators/kernel/cl/conv_add_kernel.cpp b/src/operators/kernel/cl/conv_add_kernel.cpp
index b5fd82c47a..ac36e70ed7 100644
--- a/src/operators/kernel/cl/conv_add_kernel.cpp
+++ b/src/operators/kernel/cl/conv_add_kernel.cpp
@@ -25,6 +25,9 @@ bool ConvAddKernel<GPU_CL, float>::Init(FusionConvAddParam<GPU_CL> *param) {
       param->Filter()->dims()[2] == param->Filter()->dims()[3] &&
           param->Paddings()[0] == param->Paddings()[1],
       "need equal");
+  param->Filter()->InitCLImage(cl_helper_.CLContext());
+  param->Bias()->InitCLImage(cl_helper_.CLContext());
+
   int offset = static_cast<int>(param->Filter()->dims()[2]) / 2 -
                static_cast<int>(param->Paddings()[1]);
   param->SetOffset(offset);
diff --git a/src/operators/kernel/cl/conv_kernel.cpp b/src/operators/kernel/cl/conv_kernel.cpp
index 8386144685..63c1776f83 100644
--- a/src/operators/kernel/cl/conv_kernel.cpp
+++ b/src/operators/kernel/cl/conv_kernel.cpp
@@ -26,19 +26,18 @@ bool ConvKernel<GPU_CL, float>::Init(ConvParam<GPU_CL> *param) {
           param->Paddings()[0] == param->Paddings()[1],
       "need equal");
 
+  param->Filter()->InitCLImage(cl_helper_.CLContext());
+
   int offset = static_cast<int>(param->Filter()->dims()[2]) / 2 -
                static_cast<int>(param->Paddings()[1]);
   param->SetOffset(offset);
 
   DLOG << " init helper: " << &cl_helper_;
   DLOG << " conv kernel add kernel ~ ";
-
   DLOG << " width of one block: " << param->Filter()->WidthOfOneBlock();
   DLOG << " height of one block: " << param->Filter()->HeightOfOneBlock();
-
   DLOG << " filter dims: " << param->Filter()->dims();
 
-
   if (param->Filter()->WidthOfOneBlock() == 1 &&
       param->Filter()->HeightOfOneBlock() == 1) {
 
diff --git a/src/operators/kernel/cl/depthwise_conv_kernel.cpp b/src/operators/kernel/cl/depthwise_conv_kernel.cpp
index 99b5a714d6..dcb95b4c6a 100644
--- a/src/operators/kernel/cl/depthwise_conv_kernel.cpp
+++ b/src/operators/kernel/cl/depthwise_conv_kernel.cpp
@@ -27,6 +27,7 @@ bool DepthwiseConvKernel<GPU_CL, float>::Init(ConvParam<GPU_CL> *param) {
       param->Filter()->dims()[2] == param->Filter()->dims()[3] &&
           param->Paddings()[0] == param->Paddings()[1],
       "need equal");
+  param->Filter()->InitCLImage(cl_helper_.CLContext());
   int offset = static_cast<int>(param->Filter()->dims()[2]) / 2 -
                static_cast<int>(param->Paddings()[1]);
   param->SetOffset(offset);
-- 
GitLab