diff options
author | Cedric Nugteren <web@cedricnugteren.nl> | 2019-01-19 15:44:19 +0100 |
---|---|---|
committer | Cedric Nugteren <web@cedricnugteren.nl> | 2019-01-19 15:44:19 +0100 |
commit | 11f4c7dd936146f9b4f165d8ef69bafa3a33ad26 (patch) | |
tree | bb8a7aa8493e3447b3544b9832cecb678c3087d6 | |
parent | c42e48068bfd51a19f9918f20b43be94c0e3dcf2 (diff) |
Added documentation on the convgemm routine
-rw-r--r-- | README.md | 1 | ||||
-rw-r--r-- | ROADMAP.md | 2 | ||||
-rw-r--r-- | doc/details_conv.md | 22 |
3 files changed, 24 insertions, 1 deletions
@@ -79,6 +79,7 @@ More detailed documentation is available in separate files: * [Testing the library for correctness](doc/testing.md) * [Bindings / wrappers for other languages](doc/bindings.md) * [More details on the GEMM kernel](doc/details_gemm.md) +* [More details on the convolution implementation](doc/details_conv.md) * [Glossary with some terms explained](doc/glossary.md) * [Frequently asked questions (FAQ) and their answers](doc/faq.md) @@ -20,6 +20,6 @@ This file gives an overview of the main features planned for addition to CLBlast | [#228](https://github.com/CNugteren/CLBlast/issues/228) | Mar-Apr '18 | CNugteren | ✔ | Improving performance for Qualcomm Adreno GPUs | | [#270](https://github.com/CNugteren/CLBlast/issues/270) | Oct '18 | CNugteren | ✔ | Implement col2im | | - | ?? | CNugteren | | Add support for OpenCL image buffers | -| [#267](https://github.com/CNugteren/CLBlast/issues/267) | ?? | CNugteren | WIP | Merge im2col and GEMM into a direct kernel | +| [#267](https://github.com/CNugteren/CLBlast/issues/267) | Jan '19 | vbkaisetsu| ✔ | Merge im2col and GEMM into a direct kernel | | [#136](https://github.com/CNugteren/CLBlast/issues/136) | ?? | CNugteren | | Implement xAXPBY and xSET | | [#169](https://github.com/CNugteren/CLBlast/issues/169) | ?? | dividiti | | Problem-specific tuning parameter selection | diff --git a/doc/details_conv.md b/doc/details_conv.md new file mode 100644 index 00000000..65e18e70 --- /dev/null +++ b/doc/details_conv.md @@ -0,0 +1,22 @@ +CLBlast: Details on the CONVGEMM routine +================ + +This document gives a bit more detail on how the CONVGEMM routine is organised and implemented. For other information about CLBlast, see the [main README](../README.md). + + +CONVGEMM: Two approaches +------------- + +CLBlast implements two approaches to batched convolutions using GEMM: through im2col, or stand-alone: + +* `ConvGemmMethod::kWithIm2Col`: running first a batched version of im2col to prepare the data into a temporary buffer, and then running a batched version of GEMM. The implementation is just as the regular im2col and GEMM kernels in CLBlast, but it is implemented as a separate kernel so all the non-needed features can be stripped out and some optimizations can be made. It uses the tuning parameters of the regular im2col and GEMM kernels. + +* `ConvGemmMethod::kSingleKernel`: this is a single kernel approach: it loads the data in such a way that the im2col kernel is no longer needed, i.e. loading the data as the im2col transformation does it. That way it becomes a single kernel and there will be no need for an intermediate large buffer. It uses a separate set of tuning parameters, and can be tuned using the `clblast_tuner_xconvgemm` binary. + + +CONVGEMM: Selecting which approach to use +------------- + +Since CONVGEMM is a relatively new and experimental feature, selection of the approach is hard-coded in [xconvgemm.hpp on line 32](../src/routines/levelx/xconvgemm.hpp:32), but can be changed there in a single place. + +The main drawback of the `ConvGemmMethod::kWithIm2Col` approach is its extra memory usage, but depending on the device and setting, it might be faster compared to the `ConvGemmMethod::kSingleKernel` approach. The latter has as extra advantage that it has its own tuning parameters, so it can be fine-tuned for your specific use-case a bit better than the 2-kernel approach with im2col. |