From 115c4592c08080112ea736e4e87676a1456331d7 Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Fri, 1 Apr 2022 14:55:05 +0800
Subject: [PATCH] fix(dnn/opencl): fix opencl elemwise tuning issue

GitOrigin-RevId: 317640547d262cbfec90f79786c60a872253f0b8
---
 dnn/test/common/benchmarker.h | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/dnn/test/common/benchmarker.h b/dnn/test/common/benchmarker.h
index dfa60c3ad..957ffc24e 100644
--- a/dnn/test/common/benchmarker.h
+++ b/dnn/test/common/benchmarker.h
@@ -158,7 +158,7 @@ float BenchmarkerBase<Opr, T>::exec(TensorLayoutArray layouts) {
     opr->param() = m_param;
     auto user_layouts = layouts;
     m_proxy->deduce_layout(opr, layouts);
-    for (size_t i = 0; i < layouts.size(); ++i)
+    for (size_t i = 0; i < layouts.size(); ++i) {
         if (user_layouts[i].ndim > 0) {
             auto run = [&]() {
                 ASSERT_TRUE(layouts[i].eq_shape(user_layouts[i]))
@@ -169,13 +169,14 @@ float BenchmarkerBase<Opr, T>::exec(TensorLayoutArray layouts) {
             };
             run();
         }
+    }
     auto allocate = [&layouts](Handle* handle) {
         TensorNDArray tensors(layouts.size());
         auto trans_func = [handle](const TensorLayout& layout) {
             auto span = layout.span();
             TensorND res;
             res.reset_ptr(
-                    static_cast<uint8_t*>(megdnn_malloc(handle, span.dist_byte())) +
+                    static_cast<uint8_t*>(megdnn_malloc(handle, span.dist_byte())) -
                     span.low_byte);
             res.layout = layout;
             return res;
@@ -244,7 +245,9 @@ float BenchmarkerBase<Opr, T>::exec(TensorLayoutArray layouts) {
     }
     auto free = [](Handle* handle, TensorNDArray& tensors) {
         std::for_each(tensors.begin(), tensors.end(), [handle](const TensorND& tensor) {
-            megdnn_free(handle, tensor.raw_ptr());
+            megdnn_free(
+                    handle, static_cast<dt_byte*>(tensor.raw_ptr()) +
+                                    tensor.layout.span().low_byte);
         });
     };
     free(m_handle, tensors_cur);
@@ -283,7 +286,7 @@ float BenchmarkerBase<Opr, T>::exect(const TensorValueArray& testcase_in) {
             auto span = layout.span();
             TensorND res;
             res.reset_ptr(
-                    static_cast<uint8_t*>(megdnn_malloc(handle, span.dist_byte())) +
+                    static_cast<uint8_t*>(megdnn_malloc(handle, span.dist_byte())) -
                     span.low_byte);
             res.layout = layout;
             return res;
@@ -341,7 +344,9 @@ float BenchmarkerBase<Opr, T>::exect(const TensorValueArray& testcase_in) {
     }
     auto free = [](Handle* handle, TensorNDArray& tensors) {
         std::for_each(tensors.begin(), tensors.end(), [handle](const TensorND& tensor) {
-            megdnn_free(handle, tensor.raw_ptr());
+            megdnn_free(
+                    handle, static_cast<dt_byte*>(tensor.raw_ptr()) +
+                                    tensor.layout.span().low_byte);
         });
     };
     free(m_handle, tensors_cur);
-- 
GitLab