optimization of max_pool3d forward #45820
Conversation
|
|
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
I think it is better to list the perf value with a table. |
| if (ele < | ||
| input_data_cur[(d * input_height + h) * input_width + w]) { | ||
| max_index = (d * input_height + h) * input_width + w; | ||
| ele = input_data_cur[max_index]; |
There was a problem hiding this comment.
Formula (d * input_height + h) * input_width + w has appeard twice, would it be better to assign this formula to a local data, example below:
T1 cur_data = input_data_cur[(d * input_height + h) * input_width + w];
ele = ele < cur_data] ? cur_data: ele;
There was a problem hiding this comment.
Because max_index should be assigned only when the condition ele < input_data_cur[(d * input_height + h) * input_width + w] is true and it would be used at the end of funciton for return mask_data, so I should assign the value to max_index inside the if statement and therefore formula (d * input_height + h) * input_width + w appears twice.
There was a problem hiding this comment.
Will time cost of formula (d * input_height + h) * input_width + w shrink with code below ?
for (int w = wstart; w < wend; ++w) {
max_index = (d * input_height + h) * input_width + w;
if (ele < input_data[max_index]) {
ele = input_data[max_index];
}
}
There was a problem hiding this comment.
The max_index cannot be calculated correctly with this codes, and consequently the mask_data return in the function would be incorrect.
Thanks for your advice, I have modified description to present the perf value in the form of table. |
| int blocks = (nthreads + thread_num - 1) / thread_num; | ||
| dim3 threads(thread_num, 1); | ||
| dim3 grid(blocks, 1); | ||
| int thread_x = 32; |
There was a problem hiding this comment.
@fengxiaoshuai hi, this PR deletes code for NV_JETSON about threads config, however the thread_num is 256 in all which is under the limitation for NV_JETSON.
Can this PR be merged?
There was a problem hiding this comment.
jetson系列中比较低端的GPU寄存器很少，有时候会出现资源不足，kernel launch失败，这个QA后台有CE监控
PR types
Performance optimization
PR changes
OPs
Describe