-
Notifications
You must be signed in to change notification settings - Fork 5.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add the first implememtation of fusion_group op #19621
Conversation
…kernel using nvrtc. test=develop
test=develop
test=develop
…ad_per_thread. test=develop
test=develop
test=develop
93928ab
to
cfc91e1
Compare
test=develop
…up op. test=develop
test=develop
test=develop
2cbdb48
to
336179a
Compare
336179a
to
a1fd12e
Compare
…e dump when failing to call the CUDA driver API.
test=develop
platform::DeviceCodePool::Instance().Get(place, func_name); | ||
VLOG(3) << "func_name: " << func_name; | ||
|
||
if (type == 0) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
type==elementwise_relu
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
elementwise_add/mul
之类的,认为是elementwise类型的二元操作,relu/sigmoid
之类的,认为是elementwise类型的一元操作。所以elementwise_relu不合适。- type属性,目前是留个空间,后面可能会支持更多类型的计算模式。目前type对op的计算kernel实现没有作用,主要是用来决定InferShape如何检查各个输入Tensor的dims,以及推导输出Tensor的dims。
先用int吧,后面和pass里面统一一下,考虑用一个枚举类型。但是因为现在只支持这一种类型,所以还没想好。
} | ||
|
||
extern "C" __global__ | ||
void elementwise_cuda_kernel_0(size_t n, float *x, float* y, float* z) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
改成elementwise_relu_cuda_kernel比较容易懂
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
如上,我觉得elementwise_relu
不合适。
* Add the dynamic load of nvrtc, and support runtime compiling of CUDA kernel using nvrtc. test=develop * Call CUDA driver api to launch the kernel compiled by nvrtc. test=develop * Disable for mac and windows. test=develop * Refine the codes to support manually specified num_threads and workload_per_thread. test=develop * Refine the CUDA kernel to support large dims. test=develop * Add DeviceCodePool to manage all device codes. * Add the first implementation fusion_group op. * Add unit-test for fusion_group op. * Add the check of result. * Add the check of nvrtc in unit-test. test=develop * Add comment to explain the inputs, outputs and features of fusion_group op. test=develop * Disable fusion_group op for mac and windows. test=develop * Make the compiling of device code return status instead of hanging up. test=develop * Add the check of whether there is CUDA driver library, and do not core dump when failing to call the CUDA driver API. * Unify fusion_group_op's input and output names. test=develop * Add the check of CUDA driver library in unittest. test=develop
test=develop
1d0c120
to
f040c70
Compare
test=develop
f040c70
to
cf97946
Compare
} | ||
|
||
std::string func_name = ctx.Attr<std::string>("func_name"); | ||
platform::DeviceCode* dev_code = |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
这个dev_code需要每次compute的时候都去查找吗?网络创建完后就不会变化了吧,如果比较多的话查找会不会很慢
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
每次compute都要去pool里面拿到,就是访问map,相比program其他那些map访问来说,这个应该算是少的,开销应该还好。如果后面发现这个有性能问题,可以再来改进,比如直接把DeviceCode的指针保存到op信息里面。
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM
* Add the first implememtation of fusion_group op PaddlePaddle#19621 (#3) * Add the dynamic load of nvrtc, and support runtime compiling of CUDA kernel using nvrtc. test=develop * Call CUDA driver api to launch the kernel compiled by nvrtc. test=develop * Disable for mac and windows. test=develop * Refine the codes to support manually specified num_threads and workload_per_thread. test=develop * Refine the CUDA kernel to support large dims. test=develop * Add DeviceCodePool to manage all device codes. * Add the first implementation fusion_group op. * Add unit-test for fusion_group op. * Add the check of result. * Add the check of nvrtc in unit-test. test=develop * Add comment to explain the inputs, outputs and features of fusion_group op. test=develop * Disable fusion_group op for mac and windows. test=develop * Make the compiling of device code return status instead of hanging up. test=develop * Add the check of whether there is CUDA driver library, and do not core dump when failing to call the CUDA driver API. * Unify fusion_group_op's input and output names. test=develop * Add the check of CUDA driver library in unittest. test=develop * Enable generating code for a given subgraph. PaddlePaddle#21126 (#4) * Enable generating code for a given subgraph. * Support sorting the subgraph. * Remove the rearange of expressions because we use the sorted subgraph directly. * Enable generating code for a subgraph which is composed of grad ops. * Use expression information to check the accuracy in unittest. * Separate load and store from computation expressions. test=develop * Improve the loading statements in generated codes. test=develop * Remove unused arguments from formal list. test=develop * Enable the detection of subgraph of grad ops. * Generate code for detected subgraph in fusion_group_pass. * Add an option in BuildStrategy to enable fusion_group_pass and add unittest. test=develop * Fix a bug when checking whether the shape of all inputs are the same. * Add debug information. * Remove subgraph_detector from inference/analysis to the common framework/ir directory. (#5) test=develop * Call subgraph_detector in fusion_group pass. test=develop * Disable fusion_group when WITH_GPU is OFF. test=develop * Refine all PADDLE_ENFORCE message. test=develop * Fix the case that some inputs are not defined in grad ops, and set op_role for fused op. test=develop * add backward gradient computation for op argsort (PaddlePaddle#22203) * add backward gradient computation for op argsort test=developo * use pre-commit test=develop * fix the bug of profile update (PaddlePaddle#22207) * fix the bug of profile update test=develop * add NotImplementedError for multi optimizers (PaddlePaddle#22181) * add NotImplementedError for multi optimizers used on multi-places . test=develop * assert error only if num_devices>1. test=develop * set test_optimizer_in_control_flow in CMakeLists for using multi-GPU.test=develop * support fluid-lite subgraph run resnet test=develop (PaddlePaddle#22191) - 添加了fluid-lite子图方式运行resnet的单测 - 修改了依赖Lite的git commit id * fix bug fot test_dygraph_mnist_fp16.py, test=develop (PaddlePaddle#22222) * Check dygraph weight name (PaddlePaddle#22140) * add parameter check; test=develop * change parameter name checker in dygraph guard; test=develop * fix test layers error; test=develop * revert some code to develop; test=develop * fix exampel error; test=develop * fix comment error; test=develop * fix comment error; test=develop * only import used test case and function(PaddlePaddle#22208) Co-authored-by: FlyingQianMM <245467267@qq.com> Co-authored-by: wangchaochaohu <wangchao66@baidu.com> Co-authored-by: liym27 <33742067+liym27@users.noreply.github.com> Co-authored-by: Wilber <jiweibo1028@outlook.com> Co-authored-by: zhongpu <2013000149@qq.com> Co-authored-by: hong <43953930+phlrain@users.noreply.github.com> Co-authored-by: Zhang Ting <709968123@qq.com>
* Add the first implememtation of fusion_group op #19621 (#3) * Add the dynamic load of nvrtc, and support runtime compiling of CUDA kernel using nvrtc. test=develop * Call CUDA driver api to launch the kernel compiled by nvrtc. test=develop * Disable for mac and windows. test=develop * Refine the codes to support manually specified num_threads and workload_per_thread. test=develop * Refine the CUDA kernel to support large dims. test=develop * Add DeviceCodePool to manage all device codes. * Add the first implementation fusion_group op. * Add unit-test for fusion_group op. * Add the check of result. * Add the check of nvrtc in unit-test. test=develop * Add comment to explain the inputs, outputs and features of fusion_group op. test=develop * Disable fusion_group op for mac and windows. test=develop * Make the compiling of device code return status instead of hanging up. test=develop * Add the check of whether there is CUDA driver library, and do not core dump when failing to call the CUDA driver API. * Unify fusion_group_op's input and output names. test=develop * Add the check of CUDA driver library in unittest. test=develop * Enable generating code for a given subgraph. #21126 (#4) * Enable generating code for a given subgraph. * Support sorting the subgraph. * Remove the rearange of expressions because we use the sorted subgraph directly. * Enable generating code for a subgraph which is composed of grad ops. * Use expression information to check the accuracy in unittest. * Separate load and store from computation expressions. test=develop * Improve the loading statements in generated codes. test=develop * Remove unused arguments from formal list. test=develop * Enable the detection of subgraph of grad ops. * Generate code for detected subgraph in fusion_group_pass. * Add an option in BuildStrategy to enable fusion_group_pass and add unittest. test=develop * Fix a bug when checking whether the shape of all inputs are the same. * Add debug information. * Remove subgraph_detector from inference/analysis to the common framework/ir directory. (#5) test=develop * Call subgraph_detector in fusion_group pass. test=develop * Disable fusion_group when WITH_GPU is OFF. test=develop * Refine all PADDLE_ENFORCE message. test=develop * Fix the case that some inputs are not defined in grad ops, and set op_role for fused op. test=develop * Follow review comments. test=develop
这个PR的工作:
DeviceCode
的Compile()
,若调用CUDA driver函数或者NVRTC函数失败时,不会直接退出,而是会返回false;Launch()
调用CUDA driver函数失败依旧直接退出。DeviceCodePool
的实现,管理所有生成的核函数。Launch()
接口。