[TOPI] Conv2d Added and Optimized for Intel HD Graphics #1290
Conversation
Laurawly
changed the title
|
@yidawang Why not intel_gpu? seems intel_graphics is a bit too long for abbreviation. |
|
@tqchen I feel that "Intel GPU" is a coined name, Intel normally calls it "Intel Graphics". And, Intel plans to launch their first version of discrete GPU, which will be the real "Intel GPU", in 2020. So I think it is helpful to reduce the confusion from the beginning. |
|
It would be nice if later we can parameterize schedules, like conv2d for x86. |
|
How about 'Iris'? Intel brand its high end graphics under the name "iris plus". |
|
We will rewrite all performance bottleneck operators (e.g. conv2d) in autotuning fashion after merging the auto tuner. For now it is good to follow @kevinthesun 's comment . |
| s[tensor].bind(xi, thread_x) | ||
| return xi, thread_z, thread_y, thread_x | ||
|
|
||
| @conv2d_alter_layout.register(["intel_gpu"]) |
There was a problem hiding this comment.
consider making the keys consistent as per @yidawang 's suggestion
| @generic.schedule_conv2d_nchw.register(["intel_gpu"]) | ||
| def schedule_conv2d_nchw(outs): | ||
| """Schedule for conv2d_nchw for Intel GPU | ||
| Parameters |
There was a problem hiding this comment.
there needs to be a new line between Parameters and usage.
See https://docs.tvm.ai/contribute/document.html#document-python
|
@Laurawly please act on the comments and let us bring the changes in |
|
@tqchen So we should change all intel_gpu to intel_graphics including the target naming in tvm right? If not I'd get a "no such target" error. |
include/tvm/build_module.h
Outdated
| /*! \return A target for Intel GPU */ | ||
| EXPORT Target intel_gpu(const std::vector<std::string>& options = | ||
| /*! \return A target for Intel Graphics */ | ||
| EXPORT Target intel_graphics(const std::vector<std::string>& options = | ||
| std::vector<std::string>()); |
|
@yidawang please review again and approve or request more changes |
|
@tqchen I agree with @kevinthesun and @merrymercy about parameterizing the schedules like https://github.com/dmlc/tvm/blob/master/topi/python/topi/x86/conv2d.py#L15 But I think we can do it in a separated PR later. Other than that, LGTM |
|
@yidawang please explicit approve, see https://docs.tvm.ai/contribute/code_review.html#approve-and-request-changes-explicitly |
Optimize convolutions specifically for Intel Graphics which compared with other opencl based backends, adds specific intel_sub_group operations. And we support consistent data packing (thanks to @yzhliu's implementation ) for end-to-end networks. Note that currently conv2d is most optimized for workloads for resnet. We compared with state-of-the-art deep learning libraries for intel GPUs. Following is some end-to-end benchmark for Intel® Processor Graphics Gen9 using opt_level 3.