未验证 提交 e9989afd 编写于 作者: X xiaogang 提交者: GitHub

fix: remove winograd extend_workspace to run (#3168)

上级 ba54e537
develop _release/v2.6.2 github/fork/AnBaolei1984/baolei/bitmain github/fork/Cambricon/develop github/fork/DannyIsFunny/Android5 github/fork/DannyIsFunny/Update_MemOpt github/fork/DannyIsFunny/fix_pow github/fork/DannyIsFunny/fix_v26_windows github/fork/LDOUBLEV/ocr github/fork/Leonardo-Ding/dwh_dev github/fork/MaxwellDing/develop github/fork/MyPandaShaoxiang/int8 github/fork/PaddleLite-EB/merge1.4 github/fork/Wangzheee/matrix_nms_op github/fork/cathwong/patch-1 github/fork/cclauss/patch-1 github/fork/chenjiaoAngel/cherry_pic github/fork/chenjiaoAngel/conv_dw_5x5 github/fork/chenjiaoAngel/conv_dw_5x5s2 github/fork/edimetia3d/arm_update_elementwise_op github/fork/edimetia3d/host_deformable_conv github/fork/edimetia3d/matrix_nms_host github/fork/edimetia3d/update_pow_op github/fork/edimetia3d/update_yolo_box github/fork/haozech/develop github/fork/haozech/infershape_chz github/fork/haozech/parl-develop github/fork/jackzhang235/develop github/fork/jiansowa/jiansowa/img_nna github/fork/jiweibo/stream_manage github/fork/juncaipeng/add_cast github/fork/qili93/update_sup_model_v26 github/fork/qjing666/develop github/fork/sunsetlh/sunsetlh/xpu_multi_test github/fork/weihaoji/whj_27 github/fork/weihaoji/xpu_res2net_fusion github/fork/weihaoji/xpu_weihaoji_dev github/fork/xiebaiyuan/fix_leak_opencl github/fork/xiebaiyuan/opencl_depthwised1 github/fork/xiebaiyuan/opencl_softmax github/fork/yanghongtian/yanghongtian/add_ascend310_target_place github/fork/yiicy/computelib github/fork/yongqiangma/bm_card github/fork/yongqiangma/calib github/fork/yongqiangma/pass github/fork/yongqiangma/pool github/fork/yongqiangma/priorbox github/fork/yongqiangma/shape github/fork/yongqiangma/trans github/fork/yongqiangma/trans2 github/fork/yongqiangma/workspace github/fork/ysh329/add-cl-kernel-member-for-opencl github/fork/ysh329/cherry-pick-precision-profiler-enhance github/fork/ysh329/fix-opencl-concat github/fork/ysh329/support-int64-copy-from-to-cpu github/fork/zhaoyang-star/enable_prifile_in_tiny_publish github/fork/zhaoyang-star/fix_openc_demo github/fork/zhaoyang-star/patch-1 github/fork/zhupengyang/opt release/v2.6 release/v2.6.0 release/v2.7 revert-4368-hongming/test_v26 v2.7-beta v2.6.3-beta2 v2.6.3-beta1 v2.6.2 v2.6.1 v2.6.0
1 合并请求!3210[Opencl] fix opencl bug
......@@ -69,10 +69,8 @@ void WinogradConv<PRECISION(kFloat), PRECISION(kFloat)>::ReInitWhenNeeded() {
(tile_block * ((ic + 3) / 4 + (oc + 3) / 4) * 4 * wino_iw * wino_iw +
8 * wino_iw * wino_iw) *
threads;
ctx.ExtendWorkspace((temp_size + new_input_size) * sizeof(float));
workspace_size_ = (temp_size + new_input_size) * sizeof(float);
weights_.Resize({1, 1, 1, wino_iw * wino_iw * oc_pad * ic_pad});
ctx.ExtendWorkspace((temp_size + new_input_size) * sizeof(float));
void* trans_tmp_ptr = malloc(sizeof(float) * wino_iw * wino_iw * oc * ic);
auto weights_data_ = weights_.mutable_data<float>();
if (!choose_small_) {
......@@ -96,6 +94,7 @@ template <>
void WinogradConv<PRECISION(kFloat), PRECISION(kFloat)>::Run() {
auto& param = this->Param<param_t>();
auto& ctx = this->ctx_->template As<ARMContext>();
ctx.ExtendWorkspace(workspace_size_);
const auto* i_data = param.x->data<float>();
const auto* w_data = weights_.data<float>();
const auto* b_data = param.bias ? param.bias->data<float>() : nullptr;
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册
反馈
建议
客服 返回
顶部