[GPU/OpenCL] Updated the SwiGLU, Reshape and Concat Layers with latest GPU pipeline changes @open sesame 10/04 17:23 #2745
Conversation
niket-agarwal
commented
Oct 4, 2024
niket-agarwal
commented
…es with OpenCL ops Added naive version of OpenCL implementation for Transpose function using blas Incorporated kernels for ops used. Added unit tests for transpose about different axes. Signed-off-by: Niket Agarwal <[email protected]>
Updated the swiglu, reshape, and concat layers with the new shared_ptr flow.
Replaced clCreateKernel with registerClKernel for all these layers.
Self evaluation:
Build test: [X]Passed [ ]Failed [ ]Skipped
Run test: [X]Passed [ ]Failed [ ]Skipped
Signed-off-by: Niket Agarwal <[email protected]>
|
📝 TAOS-CI Version: 1.5.20200925. Thank you for submitting PR #2745. Please a submit 1commit/1PR (one commit per one PR) policy to get comments quickly from reviewers. Your PR must pass all verificiation processes of cibot before starting a review process from reviewers. If you are new member to join this project, please read manuals in documentation folder and wiki page. In order to monitor a progress status of your PR in more detail, visit http://ci.nnstreamer.ai/. |
niket-agarwal
changed the title
jijoongmoon
changed the title
There was a problem hiding this comment.
@niket-agarwal, 💯 All CI checkers are successfully verified. Thanks.
| do { | ||
| result = context.clCreateKernel(copy_cl_kernel_, context.LayerKernel::COPY, | ||
| ReshapeLayerCl::kernel_copy); | ||
| if (!result) { | ||
| ClContext::SharedPtrClKernel kernel_copy_ptr = | ||
| cl_context_ref.registerClKernel(copy_cl_kernel_fp16_, "copy_cl_fp16"); | ||
| if (!kernel_copy_ptr) { | ||
| break; | ||
| } | ||
|
|
There was a problem hiding this comment.
What about making this part as a separate function and letting the function to be called when the clLayer is registered? This opinion comes from #2723. Could you please think about this and share your opinion?
There was a problem hiding this comment.
About this comment, we may need additional discussions and can apply it in another PR. As I understood, this PR aims to update the kernels to work with the recent commit. You can ignore this comment for this PR now :)
| @@ -148,15 +149,13 @@ class ConcatLayerCl : public Layer { | |||
| * @param[in] input1_height represents the height of the input tensor | |||
| * @param[in] input1_width represents the width of the input tensor A | |||
| * @param[in] input2_width represents the width of the input tensor X | |||
| * @param[in] context RunLayerContext reference | |||
| */ | |||
| void concat_cl_axis3_fp16(const __fp16 *matAdata, const __fp16 *vecXdata, | |||
There was a problem hiding this comment.
Please add ENABLE_FP16 condition to this header.
| void concat_cl_axis3_fp16(const __fp16 *matAdata, const __fp16 *vecXdata, | |
| #ifdef ENABLE_FP16 | |
| void concat_cl_axis3_fp16(const __fp16 *matAdata, const __fp16 *vecXdata, |
Besides this function prototypes, all concat_cl_axis*_fp16 should be imported only when ENABLE_FP16 is defined.
Please consider this conditional compilation for cpp file as well.
|
Please rebase it so that it's in sync with PR #2738 (merged). Make sure to add the kernel strings in |
