[T150838079][WIP] Split the Code Generation for `embedding_backward_split_template.cu` into Smaller Files #1705

q10 · 2023-04-17T17:52:20Z

Summary:

embedding_backward_split_template.cu contains both jinja-template and non-jinja-template code, and some of the templating is unneccessary. Furthermore, the template generates both the vanilla and nobag variants of unweighted into the same source file. This PR moves the non-jinja-template code out of the template, de-duplicates code are unneccessarily templated, and splits the generation of the code to three files per optimizer, one for weighted, unweighted nobag, and unweighted.

Details:

Migrate non-jinja-templated code out of embedding_backward_split_template.cu and into embedding_backward_template_helpers.cuh
De-templatize split_embedding_backward_codegen_{{ optimizer }}_{{ wdesc }}_find_long_segments into split_embedding_backward_codegen_find_long_segments since there is no implementation difference between the optimizers and weighted vs unweighted
Migrate grad_mean_kernel and split_embedding_backward_codegen_find_long_segments into a separate non-template source file to de-duplicate code generation and compilation
Split the code generation of embedding_backward_split_template.cu into 3 files per optimizer, according to weighted, unweighted_nobag, and unweighted

netlify · 2023-04-17T17:52:28Z

✅ Deploy Preview for pytorch-fbgemm-docs canceled.

Name	Link
🔨 Latest commit	`1dba905`
🔍 Latest deploy log	https://app.netlify.com/sites/pytorch-fbgemm-docs/deploys/643f2af218fa4000086dcf64

facebook-github-bot · 2023-04-18T01:42:38Z

@q10 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

…into Smaller Files (pytorch#1705) Summary: `embedding_backward_split_template.cu` contains both jinja-template and non-jinja-template code, and some of the templating is unneccessary. Furthermore, the template generates both the vanilla and `nobag` variants of unweighted into the same source file. This PR moves the non-jinja-template code out of the template, de-duplicates code are unneccessarily templated, and splits the generation of the code to three files per optimizer, one for `weighted`, `unweighted nobag`, and `unweighted`. Details: - Migrate non-jinja-templated code out of `embedding_backward_split_template.cu` and into `embedding_backward_template_helpers.cuh` - De-templatize `split_embedding_backward_codegen_{{ optimizer }}_{{ wdesc }}_find_long_segments` into `split_embedding_backward_codegen_find_long_segments` since there is no implementation difference between the optimizers and weighted vs unweighted - Migrate `grad_mean_kernel` and `split_embedding_backward_codegen_find_long_segments` into a separate non-template source file to de-duplicate code generation and compilation - Split the code generation of `embedding_backward_split_template.cu` into 3 files per optimizer, according to weighted, unweighted_nobag, and unweighted Pull Request resolved: pytorch#1705 Differential Revision: D45073273 Pulled By: q10 fbshipit-source-id: 0ade34cebc3734df7dad3bfbe1da319c975ee61a

facebook-github-bot · 2023-04-18T23:42:57Z

This pull request was exported from Phabricator. Differential Revision: D45073273

facebook-github-bot · 2023-04-19T05:31:29Z

@q10 merged this pull request in 88dc18c.

* using different mechanism for host mapped pinned memory (pytorch#1638) Summary: Pull Request resolved: pytorch#1638 This diff adds another mechanism for allocating the host mapped pinned memory to reduce adverse affect on other processes running on the same host when one process is doing some large allocations. Reviewed By: zyan0, jianyuh Differential Revision: D43950253 fbshipit-source-id: 41a434cb63354509d32e00c851c5f3a2d68be686 * disable use_cpu test (pytorch#1635) Summary: This PR addresses the issue pytorch#1636 akin to https://github.com/pytorch/FBGEMM/blob/8616ed701015f8b9e4c2825ce592b204b4cfaf28/fbgemm_gpu/test/split_table_batched_embeddings_test.py#L1009 Pull Request resolved: pytorch#1635 Reviewed By: shintaro-iwasaki Differential Revision: D44033725 Pulled By: q10 fbshipit-source-id: 49f28fc2f1c20948a42728eebf3defc5195baa5d * Update API interface and reroute backend for exact_rowwise_adagrad FE when using freq based methods (pytorch#1352) Summary: Pull Request resolved: pytorch#1352 1. Update interface to accomadate rowwise_adagrad_with_counter. 2. Route backend for rowwise_adagrad to the new rowwise_adagrad_with_counter when freq based methods (e.g. freq sgd, counter adjusted regularization) are used. Reviewed By: csmiler Differential Revision: D36788395 fbshipit-source-id: 8eb5da8a5c8b52bc1e237af1054aac9f7245c443 * Remove sync point in jagged_dense_elementwise_add_jagged_output backward (pytorch#1642) Summary: Pull Request resolved: pytorch#1642 Remove sync point in jagged_dense_elementwise_add_jagged_output backward Reviewed By: brad-mengchi Differential Revision: D44039901 fbshipit-source-id: 8e7e23e4d9e01359e67e5b166adc57f894a1224d * Add Comprehensive Build Instructions and Isolate CPU and ROCm Builds (pytorch#1639) Summary: - Remove `.post0` suffix from the autogenerated package version - Document the full FBGEMM_GPU OSS build process in a separate Markdown file - Remove installation of packages not needed for ROCm builds - Migrate CPU and ROCm jobs to run on top of Docker containers instead of bare metal instances - Update GitHub workflow configuration to cancel previous jobs for a PR if a new commit is pushed to the PR Pull Request resolved: pytorch#1639 Reviewed By: shintaro-iwasaki Differential Revision: D44076312 Pulled By: q10 fbshipit-source-id: 6b2d083022feb7421b26da2d998678e00c11f283 * include cstdint (pytorch#1640) Summary: fix build with gcc-13 Pull Request resolved: pytorch#1640 Reviewed By: shintaro-iwasaki Differential Revision: D44044422 Pulled By: q10 fbshipit-source-id: 692ec9c34f4aaf726294a2b643fbceabf8159033 * Add support for group size > 54 in group_index_select (pytorch#1611) Summary: Pull Request resolved: pytorch#1611 If group size is larger than 54, internally breaks the group down into smaller groups (each subgroup size is less than or equal to 54). Reviewed By: jianyuh Differential Revision: D43585937 fbshipit-source-id: bf14eeb79881a5737dcf7660e3e0f56d21f7b326 * Implement cache miss emulation in UVM_CACHING (pytorch#1637) Summary: Pull Request resolved: pytorch#1637 Enforce cache misses (even if trace-driven testing doesn't experience cache miss due to limited trace size) so that we can evaluate performance under cache misses. Note that it's not exactly cache misses; enforce access to UVM by overriding lxu_cache_locations -- N / 256 requests. Reviewed By: YuzeDaiMeta Differential Revision: D42194019 fbshipit-source-id: ab04c1cc7a749e84d605cfe4f1687489ceab5725 * Add TensorAccessor with memcheck (pytorch#1602) Summary: Pull Request resolved: pytorch#1602 Illegal memory access is a common problem during GPU kernel execution. The FBGEMM GPU relies on PyTorch's `C10_CUDA_KERNEL_LAUNCH_CHECK()` and the CUDA runtime to detect such problems and throw an error. However, there are a few known issues with this approach. (1) `C10_CUDA_KERNEL_LAUNCH_CHECK()` detects errors on the host. However, due to the non-blocking, asynchronous nature of GPU kernel execution, the error is caught on the host at a later point than where the problematic kernel was launched. This can cause the stack trace to be inaccurate and make debugging more difficult. Although the issue can be fixed by running the code with `CUDA_LAUNCH_BLOCKING=1`, this can change the state of the execution and cause Heisenbugs. (2) Not all illegal memory accesses are caught by the runtime. This means that the system may not always throw an error when illegal memory access occurs. (3) Although the runtime throws an error for illegal memory access, it is difficult to pinpoint the specific kernel and memory buffer/address that is causing the problem. For all the aforementioned reasons, we attempt to catch and throw an error as soon as possible in the kernel when illegal memory accesses occur in FBGEMM GPU. We introduce the `FBGEMM_GPU_MEMCHECK` flag to enable memory checking during compile time. We copy PyTorch's `TensorAccessor.h` into the FBGEMM GPU and extend it to check every memory access through the `PackedTensorAccessor`. If an invalid memory access occurs, we throw an error using `CUDA_KERNEL_ASSERT`. The error message includes the name of the tensor and the kernel that caused the problem. If `FBGEMM_GPU_MEMCHECK` is enabled, FBGEMM operators will use `fbgemm::PackedTensorAccessor`. Otherwise, they will use `at::PackedTensorAccessor` `FBGEMM_GPU_MEMCHECK` integration in FBGEMM ops will be done in subsequent diffs Reviewed By: r-barnes Differential Revision: D43421838 fbshipit-source-id: c8ef04970d94bb097cb5f09b42f994db72845167 * Fix compiling with Xcode 14.3 (pytorch#1648) Summary: Pull Request resolved: pytorch#1648 This hack is not needed in Xcode 14.3 anymore, where the clang version is 14.0.3. So change the workaround to only include up to 14.0.2. Reviewed By: MatzeB Differential Revision: D44130421 fbshipit-source-id: 1fb2948567941bdf6ee9487ccfaa9dfb2caf92dd * Add support for building FBGEMM_GPU against Python 3.11 in OSS (pytorch#1646) Summary: - Parallelize the FBGEMM CI builds to build and test static and shared libraries independently instead of in serial - Move the FBGEMM CI builds to run inside Docker containers - Add support for building FBGEMM_GPU against Python 3.11 in OSS - Move all FBGEMM_GPU nightly and release build jobs to run inside `amazonlinux:2023` Docker container - Assuming no build errors or resource starvation, the full OSS build process now runs under 30 minutes. Pull Request resolved: pytorch#1646 Reviewed By: shintaro-iwasaki Differential Revision: D44157228 Pulled By: q10 fbshipit-source-id: 6403ea9955856157785c50837b0b8e4c0cd26d53 * Remove magic numbers from fbgemm/Types.h (pytorch#1629) Summary: Pull Request resolved: pytorch#1629 Replaces magic numbers with constexpr variables Reviewed By: sryap Differential Revision: D43776442 fbshipit-source-id: 5cef7566816f8730f5daa08948ee3260367787aa * added check to avoid div 0 errors in cache report (pytorch#1645) Summary: Pull Request resolved: pytorch#1645 as in title Reviewed By: jianyuh Differential Revision: D44096435 fbshipit-source-id: a7a87a14ffecc2fb6e0be74d199d385357946672 * jagged_dense_bmm operator optimization (pytorch#1643) Summary: Pull Request resolved: pytorch#1643 This diff optimizes the jagged_dense_bmm operator with the following optimizations: * tiling across thread blocks, and use GPU shared memory for thread block * tiling across threads within a thread block, and use registers for each thread Reviewed By: brad-mengchi Differential Revision: D43674845 fbshipit-source-id: 85f0abf89fa958f79636ef59c3070a1c569b73c2 * jagged_dense_bmm: fix ROCm test failures (pytorch#1655) Summary: This patch fixes test failures on AMD GPUs. 1. Remove `__restrict__ `. I don't think it is needed even for CUDA, but it confuses HIPCC. 2. Use `uint32_t` instead of `auto`: old ROCm (including ROCm <= 5.3) does not have `+=` operator for the type of `blockIdx.z`, causing a compilation error. We observed that this issue is fixed in ROCm 5.4.3, but let's use `uint32_t` for now. We should revisit and use `auto` later. See this for details: ROCm/hipamd@86a1634 Pull Request resolved: pytorch#1655 Test Plan: GitHub Actions' AMD CI Reviewed By: q10, brad-mengchi Differential Revision: D44242622 Pulled By: shintaro-iwasaki fbshipit-source-id: c9b88155ebf1ed881b2d03e3be0e8991b4b30174 * Support embedding dim 1024 ~ 2048 (pytorch#1656) Summary: Pull Request resolved: pytorch#1656 wushirong reported the failure on https://fburl.com/code/hae91ra7 . - The embedding config is from f418615450 . - `max_int8_128b_rows` is 10 --> D = 1280 Our embedding dim has grown to 1024 + ? Note that the static shared memory can only go up to 48 KB: > Kernels relying on shared memory allocations over 48 KB per block are architecture-specific, as such they must use dynamic shared memory (rather than statically sized arrays) in https://docs.nvidia.com/cuda/cuda-c-programming-guide/ for ptx shared mem error: ``` [2023-03-21T22:04:33.899-07:00] ptxas error : Entry function '_ZN4nbit60INT8_split_embedding_codegen_forward_weighted_kernel_small_LIiN3c104HalfELm2ELm4ELm4E Lm8ELm16ELb1EEEvN2at27GenericPackedTensorAccessorIhLm1ENS3_17RestrictPtrTraitsElEES6_NS4_IiLm1ES5_iEENS4_IlLm1ES5_iEENS4_IhLm1ES5_iEES7_N10fbgemm_gpu12FixedDiv isorENS4_IT_Lm1ES5_iEESD_llNS4_IfLm1ES5_iEENS4_IT0_Lm2ES5_iEENS4_IhLm2ES5_lEES7_' uses too much shared data (0x10080 bytes, 0xc000 max) ``` Currently we reduce `InputRowsInFlight` to bypass the issue (the static shared memory used in the kernel is ``` typedef uint4 AllBuffers[WarpsPerBlock][OutputRowsPerThread][InputRowsInFlight][NumUint4LoadsPerRow]; __shared__ AllBuffers buffers; ``` Long term, we can change the static shared memory to dynamic shared memory, and increase the shared memory size to be 64 KB +. Reviewed By: wushirong Differential Revision: D44270081 fbshipit-source-id: 367ae838ea073dfe58d859ea3c0e6c7190beca6a * Containerize the remaining FBGEMM_GPU CI jobs (pytorch#1658) Summary: - Containerize the remaining FBGEMM_GPU CI jobs - Add Conda cleanups to make PyTorch and CUDA installs more reliable - Update post-install checks for PyTorch to work with ROCm - Update the CI to continue running on jobs that fail on just a few variants - Use PIP to install PyTorch GPU nightly as the nightly packages show up in PIP more reliably than in Conda Pull Request resolved: pytorch#1658 Reviewed By: shintaro-iwasaki Differential Revision: D44306708 Pulled By: q10 fbshipit-source-id: 5f0862f18eca7151759d9983aa97849222539d7d * Add tbe_input_combine_with_length for GPU (pytorch#1647) Summary: Pull Request resolved: pytorch#1647 Implement `tbe_input_combine_with_length` for GPU. The operator takes 3 lists of tensors (`indices`, `lengths`, and `per_sample_weights`) and concatenates each one into a single tensor. Implicit type casting is also performed if the input types are different from the output types. `indices` and `lengths` tensors can be of type `int32_t` or `int64_t`. The outputs for `indices` concatenation and `lengths` concatenation are fixed to `int32_t`. `per_sample_weights` must be `float`. Reviewed By: bangshengtang Differential Revision: D44076452 fbshipit-source-id: f6ce8628e7345093bb55835f9523870c2914516f * jagged_jagged_bmm operator optimization (pytorch#1644) Summary: Pull Request resolved: pytorch#1644 This diff optimizes the jagged_jagged_bmm operator using tiling across thread blocks and GPU shared memory. Reviewed By: brad-mengchi Differential Revision: D44029528 fbshipit-source-id: fa5cd5a26893f935427bce5efb7dfcc731c3f47d * Specify device to emulate_cache_miss kernel (pytorch#1660) Summary: Pull Request resolved: pytorch#1660 When enabled emulate cache miss, it caused illegal memory access, if we're using more than one GPU. It turns out that previous diff didn't specify device within emulate_cache_miss kernel. This diff fixes it. In addition, cleaned up a bit (e.g., no need to used index_t based kernel launch for emulate_cache_miss kernel, as lxu_cache_locations is always with int32_t. Reviewed By: sryap, YuzeDaiMeta Differential Revision: D44340131 fbshipit-source-id: d99ba2364e9030cbca6c1166e578d24d99646bb1 * Add C++17 Support to FBGEMM and FBGEMM_GPU OSS builds (pytorch#1652) Summary: - Add C++17 support for the entire FBGEMM_GPU build - Add C++17 support for the entire FBGEMM build - Update FBGEMM tests and benchmarks to be C++17-compatible - Make FBGEMM builds output more logging - Cherry-pick code changes from D43776442 v4 now that C++17 is fully supported Pull Request resolved: pytorch#1652 Reviewed By: shintaro-iwasaki Differential Revision: D44287321 Pulled By: q10 fbshipit-source-id: 4bf2bcf66d528939865d42b6deafc470bee55d17 * Prune CPU/GPU TBE optimizer codegen (pytorch#1659) Summary: Pull Request resolved: pytorch#1659 This diff aims to reduce the build time and libary size of `//deeplearning/fbgemm/fbgemm_gpu/codegen:embedding_ops`. The diff modifies the build target to generate and compile only the necessary files. This is based on the fact that CPU and GPU do not support all optimizers in `SplitTBE`. (Before this diff, all optimizers were generated and compiled for both CPU and GPU.) The following is the list of supported optimizers |OptimType|Generated optimizer|Supported on CPU|Supported on GPU| |EXACT_ADAGRAD|adagrad|x|x| |EXACT_ROWWISE_ADAGRAD|rowwise_adagrad_with_counter|x|x| ||rowwise_adagrad|x|x| |EXACT_ROWWISE_WEIGHTED_ADAGRAD|rowwise_weighted_adagrad|x|x| |EXACT_SGD|sgd|x|x| |SGD|approx_sgd|x|x| |ROWWISE_ADAGRAD|approx_rowwise_adagrad_with_counter|x|| ||approx_rowwise_adagrad|x|| |ADAM|adam||x| |LAMB|lamb||x| |LARS_SGD|lars_sgd||x| |PARTIAL_ROWWISE_ADAM|partial_rowwise_adam||x| |PARTIAL_ROWWISE_LAMB|partial_rowwise_lamb||x| |-|rowwise_adagrad_with_weight_decay||| |-|approx_rowwise_adagrad_with_weight_decay||| Note: x = supported Reviewed By: jianyuh Differential Revision: D44326540 fbshipit-source-id: 02413256b4a675f13ada8e8820820cb5112cb405 * Fix the Documentation Build Job (pytorch#1673) Summary: - Rewrite the documentation builds job to use the build infrastructure tooling - Rename workflow files for consistency Pull Request resolved: pytorch#1673 Reviewed By: shintaro-iwasaki Differential Revision: D44472660 Pulled By: q10 fbshipit-source-id: 60434c1f7098b7efa8c750133bb22f14fc98d5dc * Back out "Prune CPU/GPU TBE optimizer codegen" (pytorch#1675) Summary: Pull Request resolved: pytorch#1675 Original commit changeset: 02413256b4a6 Original Phabricator Diff: D44326540 Reviewed By: q10, jianyuh Differential Revision: D44475251 fbshipit-source-id: 5be66944a833e03a2737fc6d1baaa5c351455b2c * Prepare bounds_check_indices for VBE (pytorch#1633) Summary: Pull Request resolved: pytorch#1633 Prepare `bounds_check_indices` for variable batch size TBE (VBE). - Update the frontend API to accept VBE args - Update the backend logic to process VBE data Reviewed By: jianyuh Differential Revision: D43253703 fbshipit-source-id: 2870f0c41a96265650281a9b6362d4e6dc48009b * Move pruning/index_remapping support to embedding inplace update files (pytorch#1667) Summary: Pull Request resolved: pytorch#1667 As title. This diff moves pruning/index_remapping support to embedding inplace update files. Reviewed By: jianyuh Differential Revision: D44409419 fbshipit-source-id: 93fc91d83502eb95cb0feca2a8a03b003c336078 * jagged_softmax forward optimization (pytorch#1661) Summary: Pull Request resolved: pytorch#1661 This diff optimizes jagged_softmax forward with more efficient reduction from cub library. Reviewed By: brad-mengchi Differential Revision: D44161021 fbshipit-source-id: bf2e059d14ef4d7ad311edac65155a463ba653ff * jagged_softmax backward optimization (pytorch#1662) Summary: Pull Request resolved: pytorch#1662 This diff optimizes jagged_softmax backward with more efficient reduction from cub library Reviewed By: brad-mengchi Differential Revision: D44205819 fbshipit-source-id: cd1d7a886d6ba68201dc1ad782c2e8cde7ff706b * multi-gpu all_to_one improvements (pytorch#1674) Summary: Pull Request resolved: pytorch#1674 improved multi-gpu all_to_one with: 1. new intermediate hop selection taking advantage of distinct NVLinks 2. overlapping of intermediate hop transfers with each-other and with direct-peer transfers Reviewed By: doehyun Differential Revision: D44285941 fbshipit-source-id: 0202083f04388b5ba60b8155809433f334993ef4 * Extract and export weights offsets/placements initialization functions (pytorch#1669) Summary: Pull Request resolved: pytorch#1669 Extract portions initializing the weights_placements/offsets tensors into separate functions and jit.export them. SplitState is converted to a NamedTuple since we can't jit.script a dataclass that also holds an enum. Reviewed By: houseroad Differential Revision: D44338256 fbshipit-source-id: e1c12e5956f7217d51cd190958c3764d220e521d * Fix the ROCm Test Job (pytorch#1668) Summary: - Clean up the ROCm test job and re-enable ROCm testing on the rocm instances. - Update the build scripts framework to build FBGEMM_GPU against the correct hardware target that it is intended to be tested on. One thing that was discovered was that if FBGEMM_GPU was built with `PYTORCH_ROCM_ARCH=gfx90a` but run on `gfx908` target, the tests will fail with a segfault. While the failure is expected, the segfault can be unfriendly and confusing for users. - Enable correct compilation of `merge_pooled_embeddings` operator under ROCm - Fix existing code in `jagged_tensor_ops` from PR pytorch#1661 and pytorch#1662 that break its compilation under ROCm 5.3 Pull Request resolved: pytorch#1668 Reviewed By: shintaro-iwasaki Differential Revision: D44453594 Pulled By: q10 fbshipit-source-id: 2030cd0e00c6ff9694c2783dfd62c31cf5543da2 * Use exported functions instead of calling initialize_weights in weights loading (pytorch#1676) Summary: Pull Request resolved: pytorch#1676 Export a function to reset the embedding specs by target location Reviewed By: RoshanPAN, houseroad Differential Revision: D44338258 fbshipit-source-id: 502733e9f3a164450a02656d2822492fbf69f994 * Extract index remappings array initialization and jit.export it (pytorch#1670) Summary: Pull Request resolved: pytorch#1670 ATT Reviewed By: RoshanPAN, houseroad Differential Revision: D44338257 fbshipit-source-id: c091666c7a4d294c283f5e3774d0494089fc3478 * Disable COUNTER in FBGEMM test (pytorch#1683) Summary: Pull Request resolved: pytorch#1683 Disable FBGEMM test on COUNTER mode temporarily. Reviewed By: sryap Differential Revision: D44589052 fbshipit-source-id: f2af6f9e3cce75d4c599c4708055e5f52ac705e2 * update hipify_torch and remove manual mapping of C10 macros (pytorch#1682) Summary: Pull Request resolved: pytorch#1682 Reviewed By: shintaro-iwasaki Differential Revision: D44599348 Pulled By: q10 fbshipit-source-id: 8f968a7c21b09358eac070a35ee15d5b767ea94c * Install NVIDIA Drivers on Instances Missing the Drivers (pytorch#1684) Summary: - Use the pytorch/test-infra action ot install NVIDIA drivers properly if the instance is missing the drivers Pull Request resolved: pytorch#1684 Reviewed By: shintaro-iwasaki Differential Revision: D44603925 Pulled By: q10 fbshipit-source-id: 712bdf5c2af67c5a6f540567abcc47ed892912c1 * Clean up the linting job (pytorch#1686) Summary: Sumary: - Clean up the linting job to use the build scripts infrastructure - Delete the Conda prefix directory before creating a new environment, if it exists Pull Request resolved: pytorch#1686 Reviewed By: shintaro-iwasaki Differential Revision: D44646234 Pulled By: q10 fbshipit-source-id: d754efeadffb265c9e55bc302606fc1e60ef8b51 * reduce_to_one (pytorch#1571) Summary: Pull Request resolved: pytorch#1571 reduce_to_one for row-wise sharding in inference Similar approach to all_to_one but without having the source waiting for target to be ready for potential WAR and WAW dependency violation because in this reduce_to_one implementation we create a new destination tensor. Reviewed By: xing-liu, jianyuh Differential Revision: D34263436 fbshipit-source-id: 7b1630b395311cfd6fef124113436f87f51a6fba * Reorganize the build scripts (pytorch#1685) Summary: Pull Request resolved: pytorch#1685 Reviewed By: r-barnes, shintaro-iwasaki Differential Revision: D44654808 Pulled By: q10 fbshipit-source-id: a58987b4a3970139bba72db8cecc89c0256fba76 * Prune CPU/GPU TBE optimizer codegen (pytorch#1678) Summary: Pull Request resolved: pytorch#1678 This diff aims to reduce the build time and libary size of `//deeplearning/fbgemm/fbgemm_gpu/codegen:embedding_ops`. [1/2] Update `lookup_invoker` to enable the function invoker based on `has_cpu_support` and `has_gpu_support` [2/2] Update the code generation part The diff modifies the build target to generate and compile only the necessary files. This is based on the fact that CPU and GPU do not support all optimizers in `SplitTBE`. (Before this diff, all optimizers were generated and compiled for both CPU and GPU.) The following is the list of supported optimizers |OptimType|Generated optimizer|Supported on CPU|Supported on GPU| |EXACT_ADAGRAD|adagrad|x|x| |EXACT_ROWWISE_ADAGRAD|rowwise_adagrad_with_counter|x|x| ||rowwise_adagrad|x|x| |EXACT_ROWWISE_WEIGHTED_ADAGRAD|rowwise_weighted_adagrad|x|x| |EXACT_SGD|sgd|x|x| |SGD|approx_sgd|x|x| |ROWWISE_ADAGRAD|approx_rowwise_adagrad_with_counter|x|| ||approx_rowwise_adagrad|x|| |ADAM|adam||x| |LAMB|lamb||x| |LARS_SGD|lars_sgd||x| |PARTIAL_ROWWISE_ADAM|partial_rowwise_adam||x| |PARTIAL_ROWWISE_LAMB|partial_rowwise_lamb||x| |-|rowwise_adagrad_with_weight_decay||| |-|approx_rowwise_adagrad_with_weight_decay||| Reviewed By: q10 Differential Revision: D44484764 fbshipit-source-id: f04710e66498bdcbdad619d48411c2403316901c * thread tiling for jagged_jagged_bmm (pytorch#1691) Summary: Pull Request resolved: pytorch#1691 This diff adds thread tiling optimization in jagged_jagged_bmm operator, where each thread will process a tile of elements instead of one. The implementation is similar to the one applied to jagged_dense_bmm: D43674845. Reviewed By: brad-mengchi Differential Revision: D44764339 fbshipit-source-id: ca4cf257bac755ab97754fdc6605072cfbfb1c4d * tune the tile sizes for jagged_dense_bmm (pytorch#1692) Summary: Pull Request resolved: pytorch#1692 Tune the tile sizes based on the input tensor size. If M > N, then use larger tile size in M dimension, otherwise use larger tile size in N dimension. Reviewed By: brad-mengchi Differential Revision: D44791699 fbshipit-source-id: 348a66089d781e9fef141b63d7a56e6dfa5da905 * Populate supported optims to match OSS Pytorch state dict (pytorch#1632) Summary: Pull Request resolved: pytorch#1632 ATT. Reviewed By: jianyuh Differential Revision: D43887969 fbshipit-source-id: 048ff61a925113b29c547abf20d7acdc4a50b8d7 * Build Scripts and README Improvements (pytorch#1695) Summary: - Update build scripts to print out cc, c++, and nvcc preprocessor defines - Print out all undefined symbols in the output library after build to inspect whether or not templates have been un-instantiated - Handle the case where `TORCH_CUDA_ARCH_LIST` is pre-defined in the environment - Clean up the FBGEMM_GPU READMEs to consolidate all FBGEMM_GPU build instructions into `docs/BuildInstructions.md` - Fix the build badges for FBGEMM and FBGEMM_GPU - Add Slack contact information to the READMEs - Remove deprecated GitHub workflows and build scripts in favor of the new scripts, which cover all the functionality of the old scripts Pull Request resolved: pytorch#1695 Reviewed By: shintaro-iwasaki Differential Revision: D44901368 Pulled By: q10 fbshipit-source-id: bef6045347c905a051970e4e5f8630175e0f5ef6 * Add Documentation to Work Around GCC 12 Regressions (pytorch#1697) Summary: Pull Request resolved: pytorch#1697 Reviewed By: shintaro-iwasaki Differential Revision: D44935915 Pulled By: q10 fbshipit-source-id: e1bdd4ebff18bd9708208a5b659ef9a93ebc866a * Fix build instructions (pytorch#1701) Summary: This change fixes a missing step (cd) in the build instructions. Pull Request resolved: pytorch#1701 Reviewed By: sryap Differential Revision: D45011147 Pulled By: q10 fbshipit-source-id: 704ce5bd3cfbd62c31f434c830a7300e5d645024 * Fix a build error from -Wno-unused-but-set-variable (pytorch#1702) Summary: This project is compiled with -Wall and -Werror (see pytorch#868) and is throwing an error for the unused variable here. This code appears to be debugging code that was used to verify that the function it's contained in was originally implemented properly so the most straightforward solution is to just remove it. Pull Request resolved: pytorch#1702 Reviewed By: sryap Differential Revision: D45011174 Pulled By: q10 fbshipit-source-id: 2c252cfa6063789371f5fba5f642c2f4fb72455f * Fix exception in QuantUtilsTest (pytorch#1703) Summary: This test mistakenly calls reserve() to set a vector's length instead of resize(). reserve() allocates memory for the specified number of elements, but does not actually increase the number of elements that can legally be stored in the vector. This test runs with ASAN enabled which is catching this illegal access and causing the test to fail. This change fixes the code to instead call resize(); the test now passes. Pull Request resolved: pytorch#1703 Reviewed By: sryap Differential Revision: D45011317 Pulled By: q10 fbshipit-source-id: 2840d7bfcfb46ca1523f55e77a3834a1d561c045 * Support EXACT_ADAGRAD in `get_optimizer_state` (pytorch#1700) Summary: Pull Request resolved: pytorch#1700 This diff support `get_optimizer_state` for exact_adagrad. Exact_adagrad is not supported in `get_optimizer_state`. However, this is needed for creating fused optimizer in torchrec. Reviewed By: r-barnes Differential Revision: D44963975 fbshipit-source-id: e2f523dfc1e1d17a4925e7ce4a9e65829f1cf1b0 * Split the Rendering of `embedding_forward_quantized_split_template.cu` into Smaller Files (pytorch#1694) Summary: `embedding_forward_quantized_split_template.cu` is a very large jinja template that renders 30+ C++ templates, which are then instantiated to over 600+ kernel functions. There are three sets of jinja templates in `embedding_forward_quantized_split_template.cu`: those related to `int_nbit_split_embedding_*`, `pruned_hashmap_lookup_*` and `pruned_array_lookup_*`.. Currently, the rendering produces a single file, which takes a large amount of time to compile. This PR does two things at a high level. First, it breaks up the jinja template into multiple jinja templates. Then, it forces each of these smaller jinja templates to render multiple source files instead of a single source file. This change will enable build parallelization and overall build time savings. Details: - Port improvements to `embedding_forward_quantized_split_template.cu` from D44707812 - Move the non-jinja-template code inside `embedding_forward_quantized_split_template.cu` over to `embedding_forward_template_helpers.cuh` - Move `pruned_hashmap_lookup_*` and `pruned_array_lookup_*` sets of jinja templates out to non-jinja-template `embedding_forward_quantized_split_lookup.cu`, since the template generated functions are redundant. - Break the `int_nbit_split_embedding_*` set of jinja templates into two files, one for rendering kernel-side code (`embedding_forward_quantized_split_nbit_kernel_template.cu`) and the other for rendering host-side code (`embedding_forward_quantized_split_nbit_host_template.cu`) - For the `int_nbit_split_embedding_*` host-side jinja template, make it render `weighted`, `unweighted`, and `unweighted nobag` variants into separate source files - For the `int_nbit_split_embedding_*` kernel-side jinja template, make it render into N = [`weighted`, `unweighted`, and `unweighted nobag` variants ] x [ 6 embedding types ] separate source files, each containing a single C++ template kernel function. Also generate the code to explicitly instantiate the kernel templates. For each of the C++ templates being generated, there will be 2 {device-only bool} x [3-4] (output types) x [3-5] (cases) = 18-40 actual template instantiations - To help with debugging missing template instantiations, print out all undefined symbols in the output library after build to inspect whether or not templates have been un-instantiated - Update build scripts to print out `cc`, `c++`, and `nvcc` preprocessor defines - Handle the case where `TORCH_CUDA_ARCH_LIST` is pre-defined in the environment Pull Request resolved: pytorch#1694 Reviewed By: sryap, r-barnes Differential Revision: D44842524 Pulled By: q10 fbshipit-source-id: 96f92e40ab2fec598aeb8c483e94997ac050aae7 * Back out "Prune CPU/GPU TBE optimizer codegen" (pytorch#1706) Summary: Pull Request resolved: pytorch#1706 Original commit changeset: f04710e66498 Original Phabricator Diff: D44484764 Reviewed By: q10, brad-mengchi, jianyuh, shintaro-iwasaki Differential Revision: D45054051 fbshipit-source-id: 9d14504c76eb93b2f1b14f4c2ec4c5b807c7fc4a * Use CUB kernel for 2D asynchronous_complete_cumsum (pytorch#1707) Summary: Pull Request resolved: pytorch#1707 Temporarily use the CUB kernel instead of the custom kernel for 2D `asynchronous_complete_cumsum` Reviewed By: q10, brad-mengchi, jianyuh Differential Revision: D45062784 fbshipit-source-id: cebe3992ff8ebec9c0f554e729b8d79a1eced1de * Split the Code Generation for `embedding_backward_split_template.cu` into Smaller Files (pytorch#1705) Summary: `embedding_backward_split_template.cu` contains both jinja-template and non-jinja-template code, and some of the templating is unneccessary. Furthermore, the template generates both the vanilla and `nobag` variants of unweighted into the same source file. This PR moves the non-jinja-template code out of the template, de-duplicates code are unneccessarily templated, and splits the generation of the code to three files per optimizer, one for `weighted`, `unweighted nobag`, and `unweighted`. Details: - Migrate non-jinja-templated code out of `embedding_backward_split_template.cu` and into `embedding_backward_template_helpers.cuh` - De-templatize `split_embedding_backward_codegen_{{ optimizer }}_{{ wdesc }}_find_long_segments` into `split_embedding_backward_codegen_find_long_segments` since there is no implementation difference between the optimizers and weighted vs unweighted - Migrate `grad_mean_kernel` and `split_embedding_backward_codegen_find_long_segments` into a separate non-template source file to de-duplicate code generation and compilation - Split the code generation of `embedding_backward_split_template.cu` into 3 files per optimizer, according to weighted, unweighted_nobag, and unweighted Pull Request resolved: pytorch#1705 Reviewed By: sryap Differential Revision: D45073273 Pulled By: q10 fbshipit-source-id: e82ea643f8e67ad5aa0b3de03562532c5735453d * Add jagged slice op for cpu (pytorch#1690) Summary: Pull Request resolved: pytorch#1690 The context why this is needed is as follows 1) For really long sparse features we want to split them into multiple chunks that can be fed into the model 2) Slicing requires users to require per row start point & a maximum L. Based on these requirements, a custom op mimicing the slice semantics of a normal tensor works best. An example usage using pseudo code ``` input_jagged_tensor = [[1, 2, 3, 4], [1, 2, 3], [1, 2, 3, 4, 5, 6], [1], [1, 2]] start = [0, 0, 0, 0, 0] slice_length = 3 >> jagged_slice(input_jagged_tensor, start, slice_length) output_jagged_tensor = [[1, 2, 3], [1, 2, 3], [1, 2, 3], [1], [1, 2]] ``` A corresponding operation for dense tensor would look like ``` dense_tensor = torch.randn((8, 10)) slice_dense_tensor = dense_tensor[:, 1:3] ``` Reviewed By: sryap Differential Revision: D44299744 fbshipit-source-id: 44996f2f2ec5fc5f31dda4cb3bd8f0241497df66 * Move radix sort to common utilities and add the possibility to handle negative integers (pytorch#1672) Summary: Move the `radix_sort` implementation to common utilities, so it can be used in PyTorch in case it was not built with FBGEMM GPU. Add the possibility to handle negative integers, which is crucial for reusing `radix_sort` in PyTorch's `sort` operation. Details: This PR addresses two issues: 1. `radix_sort` is currently used in [scatter_reduce](https://github.com/dszwicht/pytorch/blob/master/aten/src/ATen/native/cpu/ScatterGatherKernel.cpp#L630) (please view this [comment](https://github.com/pytorch/pytorch/pull/82703/files#r1045360609) for more information). Till now `radix_sort` was under `fbgemm_gpu` subproject. It means that implementation was not available in PyTorch in case it was built for CPU - that's why `radix_sort` was copy pasted under aten directory in PyTorch. This PR moves `radix_sort` implementation to common utilities. 2. In GNN workloads we often sort 1D integer data with non-negative values, for example, when converting CSR to CSC format. Unfortunately, `torch.sort` for 1D data works sequentially. `radix_sort` seems to be a perfect match to accelerate described case. However, suppose we want to do that on the PyTorch site. In that case, we have to either fallback to a regular path after detecting negative numbers in the tensor or perform post-processing, by swapping positive and negative blocks of data (data like `[2, -1, -2, 1]` after sorting will be in the following form `[1, 2, -2, -1]`, due to the fact of how numbers are stored). Both these solutions are not elegant. As an alternative, I propose the extension of `radix_sort` algorithm, by giving it capability to work with negative numbers. This can be enabled by passing an optional parameter, `maybe_with_neg_vals`. If set to `true`, we will perform all passes (up to the most significant sign bit) and apply a special prefix sum combination in the last pass. An example of how we can reuse fbgemm in PyTorch can be found in my private fork, [here](DamianSzwichtenberg/pytorch#2) (I also provide speedup data). The above changes have several consequences: 1. `TORCH_CHECK` was replaced with `assert` as fbgemm CPU does not have PyTorch in its dependencies. 2. `__builtin_clz` was replaced with manual implementation as `__builtin_clz` is not portable. Additional information for reviewers: I did perform benchmarks of `radix_sort` before and after my code modification. I didn't observe any performance drop. Pull Request resolved: pytorch#1672 Reviewed By: sryap Differential Revision: D44616959 Pulled By: q10 fbshipit-source-id: f34594478c94ec6610c05545feb2044b58d79d66 * Daily `arc lint --take CLANGFORMAT` Reviewed By: bigfootjon Differential Revision: D45141964 fbshipit-source-id: 58308a31522a3b1446835e358a93483b611c4b15 --------- Co-authored-by: Banit Agrawal <[email protected]> Co-authored-by: Sabin Devkota <[email protected]> Co-authored-by: Junjie Yang <[email protected]> Co-authored-by: Benson Ma <[email protected]> Co-authored-by: Alfredo Tupone <[email protected]> Co-authored-by: Sarunya Pumma <[email protected]> Co-authored-by: Doe Hyun Yoon <[email protected]> Co-authored-by: Matt Galloway <[email protected]> Co-authored-by: Richard Barnes <[email protected]> Co-authored-by: Xiao Sun <[email protected]> Co-authored-by: Rengan Xu <[email protected]> Co-authored-by: siwasaki <[email protected]> Co-authored-by: Jianyu Huang <[email protected]> Co-authored-by: Yue Dong <[email protected]> Co-authored-by: Geet Sethi <[email protected]> Co-authored-by: Janet Yang <[email protected]> Co-authored-by: Wang Zhou <[email protected]> Co-authored-by: Jongsoo Park <[email protected]> Co-authored-by: Tran Le <[email protected]> Co-authored-by: Ryan Landay <[email protected]> Co-authored-by: Devashish Tyagi <[email protected]> Co-authored-by: Szwichtenberg, Damian <[email protected]> Co-authored-by: generatedunixname89002005325676 <[email protected]>

* using different mechanism for host mapped pinned memory (#1638) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1638 This diff adds another mechanism for allocating the host mapped pinned memory to reduce adverse affect on other processes running on the same host when one process is doing some large allocations. Reviewed By: zyan0, jianyuh Differential Revision: D43950253 fbshipit-source-id: 41a434cb63354509d32e00c851c5f3a2d68be686 * disable use_cpu test (#1635) Summary: This PR addresses the issue https://github.com/pytorch/FBGEMM/issues/1636 akin to https://github.com/pytorch/FBGEMM/blob/8616ed701015f8b9e4c2825ce592b204b4cfaf28/fbgemm_gpu/test/split_table_batched_embeddings_test.py#L1009 Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1635 Reviewed By: shintaro-iwasaki Differential Revision: D44033725 Pulled By: q10 fbshipit-source-id: 49f28fc2f1c20948a42728eebf3defc5195baa5d * Update API interface and reroute backend for exact_rowwise_adagrad FE when using freq based methods (#1352) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1352 1. Update interface to accomadate rowwise_adagrad_with_counter. 2. Route backend for rowwise_adagrad to the new rowwise_adagrad_with_counter when freq based methods (e.g. freq sgd, counter adjusted regularization) are used. Reviewed By: csmiler Differential Revision: D36788395 fbshipit-source-id: 8eb5da8a5c8b52bc1e237af1054aac9f7245c443 * Remove sync point in jagged_dense_elementwise_add_jagged_output backward (#1642) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1642 Remove sync point in jagged_dense_elementwise_add_jagged_output backward Reviewed By: brad-mengchi Differential Revision: D44039901 fbshipit-source-id: 8e7e23e4d9e01359e67e5b166adc57f894a1224d * Add Comprehensive Build Instructions and Isolate CPU and ROCm Builds (#1639) Summary: - Remove `.post0` suffix from the autogenerated package version - Document the full FBGEMM_GPU OSS build process in a separate Markdown file - Remove installation of packages not needed for ROCm builds - Migrate CPU and ROCm jobs to run on top of Docker containers instead of bare metal instances - Update GitHub workflow configuration to cancel previous jobs for a PR if a new commit is pushed to the PR Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1639 Reviewed By: shintaro-iwasaki Differential Revision: D44076312 Pulled By: q10 fbshipit-source-id: 6b2d083022feb7421b26da2d998678e00c11f283 * include cstdint (#1640) Summary: fix build with gcc-13 Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1640 Reviewed By: shintaro-iwasaki Differential Revision: D44044422 Pulled By: q10 fbshipit-source-id: 692ec9c34f4aaf726294a2b643fbceabf8159033 * Add support for group size > 54 in group_index_select (#1611) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1611 If group size is larger than 54, internally breaks the group down into smaller groups (each subgroup size is less than or equal to 54). Reviewed By: jianyuh Differential Revision: D43585937 fbshipit-source-id: bf14eeb79881a5737dcf7660e3e0f56d21f7b326 * Implement cache miss emulation in UVM_CACHING (#1637) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1637 Enforce cache misses (even if trace-driven testing doesn't experience cache miss due to limited trace size) so that we can evaluate performance under cache misses. Note that it's not exactly cache misses; enforce access to UVM by overriding lxu_cache_locations -- N / 256 requests. Reviewed By: YuzeDaiMeta Differential Revision: D42194019 fbshipit-source-id: ab04c1cc7a749e84d605cfe4f1687489ceab5725 * Add TensorAccessor with memcheck (#1602) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1602 Illegal memory access is a common problem during GPU kernel execution. The FBGEMM GPU relies on PyTorch's `C10_CUDA_KERNEL_LAUNCH_CHECK()` and the CUDA runtime to detect such problems and throw an error. However, there are a few known issues with this approach. (1) `C10_CUDA_KERNEL_LAUNCH_CHECK()` detects errors on the host. However, due to the non-blocking, asynchronous nature of GPU kernel execution, the error is caught on the host at a later point than where the problematic kernel was launched. This can cause the stack trace to be inaccurate and make debugging more difficult. Although the issue can be fixed by running the code with `CUDA_LAUNCH_BLOCKING=1`, this can change the state of the execution and cause Heisenbugs. (2) Not all illegal memory accesses are caught by the runtime. This means that the system may not always throw an error when illegal memory access occurs. (3) Although the runtime throws an error for illegal memory access, it is difficult to pinpoint the specific kernel and memory buffer/address that is causing the problem. For all the aforementioned reasons, we attempt to catch and throw an error as soon as possible in the kernel when illegal memory accesses occur in FBGEMM GPU. We introduce the `FBGEMM_GPU_MEMCHECK` flag to enable memory checking during compile time. We copy PyTorch's `TensorAccessor.h` into the FBGEMM GPU and extend it to check every memory access through the `PackedTensorAccessor`. If an invalid memory access occurs, we throw an error using `CUDA_KERNEL_ASSERT`. The error message includes the name of the tensor and the kernel that caused the problem. If `FBGEMM_GPU_MEMCHECK` is enabled, FBGEMM operators will use `fbgemm::PackedTensorAccessor`. Otherwise, they will use `at::PackedTensorAccessor` `FBGEMM_GPU_MEMCHECK` integration in FBGEMM ops will be done in subsequent diffs Reviewed By: r-barnes Differential Revision: D43421838 fbshipit-source-id: c8ef04970d94bb097cb5f09b42f994db72845167 * Fix compiling with Xcode 14.3 (#1648) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1648 This hack is not needed in Xcode 14.3 anymore, where the clang version is 14.0.3. So change the workaround to only include up to 14.0.2. Reviewed By: MatzeB Differential Revision: D44130421 fbshipit-source-id: 1fb2948567941bdf6ee9487ccfaa9dfb2caf92dd * Add support for building FBGEMM_GPU against Python 3.11 in OSS (#1646) Summary: - Parallelize the FBGEMM CI builds to build and test static and shared libraries independently instead of in serial - Move the FBGEMM CI builds to run inside Docker containers - Add support for building FBGEMM_GPU against Python 3.11 in OSS - Move all FBGEMM_GPU nightly and release build jobs to run inside `amazonlinux:2023` Docker container - Assuming no build errors or resource starvation, the full OSS build process now runs under 30 minutes. Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1646 Reviewed By: shintaro-iwasaki Differential Revision: D44157228 Pulled By: q10 fbshipit-source-id: 6403ea9955856157785c50837b0b8e4c0cd26d53 * Remove magic numbers from fbgemm/Types.h (#1629) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1629 Replaces magic numbers with constexpr variables Reviewed By: sryap Differential Revision: D43776442 fbshipit-source-id: 5cef7566816f8730f5daa08948ee3260367787aa * added check to avoid div 0 errors in cache report (#1645) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1645 as in title Reviewed By: jianyuh Differential Revision: D44096435 fbshipit-source-id: a7a87a14ffecc2fb6e0be74d199d385357946672 * jagged_dense_bmm operator optimization (#1643) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1643 This diff optimizes the jagged_dense_bmm operator with the following optimizations: * tiling across thread blocks, and use GPU shared memory for thread block * tiling across threads within a thread block, and use registers for each thread Reviewed By: brad-mengchi Differential Revision: D43674845 fbshipit-source-id: 85f0abf89fa958f79636ef59c3070a1c569b73c2 * jagged_dense_bmm: fix ROCm test failures (#1655) Summary: This patch fixes test failures on AMD GPUs. 1. Remove `__restrict__ `. I don't think it is needed even for CUDA, but it confuses HIPCC. 2. Use `uint32_t` instead of `auto`: old ROCm (including ROCm <= 5.3) does not have `+=` operator for the type of `blockIdx.z`, causing a compilation error. We observed that this issue is fixed in ROCm 5.4.3, but let's use `uint32_t` for now. We should revisit and use `auto` later. See this for details: https://github.com/ROCm-Developer-Tools/hipamd/commit/86a1634c642daeda1e984d4124bcc2aeba5c4e19 Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1655 Test Plan: GitHub Actions' AMD CI Reviewed By: q10, brad-mengchi Differential Revision: D44242622 Pulled By: shintaro-iwasaki fbshipit-source-id: c9b88155ebf1ed881b2d03e3be0e8991b4b30174 * Support embedding dim 1024 ~ 2048 (#1656) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1656 wushirong reported the failure on https://fburl.com/code/hae91ra7 . - The embedding config is from f418615450 . - `max_int8_128b_rows` is 10 --> D = 1280 Our embedding dim has grown to 1024 + ? Note that the static shared memory can only go up to 48 KB: > Kernels relying on shared memory allocations over 48 KB per block are architecture-specific, as such they must use dynamic shared memory (rather than statically sized arrays) in https://docs.nvidia.com/cuda/cuda-c-programming-guide/ for ptx shared mem error: ``` [2023-03-21T22:04:33.899-07:00] ptxas error : Entry function '_ZN4nbit60INT8_split_embedding_codegen_forward_weighted_kernel_small_LIiN3c104HalfELm2ELm4ELm4E Lm8ELm16ELb1EEEvN2at27GenericPackedTensorAccessorIhLm1ENS3_17RestrictPtrTraitsElEES6_NS4_IiLm1ES5_iEENS4_IlLm1ES5_iEENS4_IhLm1ES5_iEES7_N10fbgemm_gpu12FixedDiv isorENS4_IT_Lm1ES5_iEESD_llNS4_IfLm1ES5_iEENS4_IT0_Lm2ES5_iEENS4_IhLm2ES5_lEES7_' uses too much shared data (0x10080 bytes, 0xc000 max) ``` Currently we reduce `InputRowsInFlight` to bypass the issue (the static shared memory used in the kernel is ``` typedef uint4 AllBuffers[WarpsPerBlock][OutputRowsPerThread][InputRowsInFlight][NumUint4LoadsPerRow]; __shared__ AllBuffers buffers; ``` Long term, we can change the static shared memory to dynamic shared memory, and increase the shared memory size to be 64 KB +. Reviewed By: wushirong Differential Revision: D44270081 fbshipit-source-id: 367ae838ea073dfe58d859ea3c0e6c7190beca6a * Containerize the remaining FBGEMM_GPU CI jobs (#1658) Summary: - Containerize the remaining FBGEMM_GPU CI jobs - Add Conda cleanups to make PyTorch and CUDA installs more reliable - Update post-install checks for PyTorch to work with ROCm - Update the CI to continue running on jobs that fail on just a few variants - Use PIP to install PyTorch GPU nightly as the nightly packages show up in PIP more reliably than in Conda Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1658 Reviewed By: shintaro-iwasaki Differential Revision: D44306708 Pulled By: q10 fbshipit-source-id: 5f0862f18eca7151759d9983aa97849222539d7d * Add tbe_input_combine_with_length for GPU (#1647) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1647 Implement `tbe_input_combine_with_length` for GPU. The operator takes 3 lists of tensors (`indices`, `lengths`, and `per_sample_weights`) and concatenates each one into a single tensor. Implicit type casting is also performed if the input types are different from the output types. `indices` and `lengths` tensors can be of type `int32_t` or `int64_t`. The outputs for `indices` concatenation and `lengths` concatenation are fixed to `int32_t`. `per_sample_weights` must be `float`. Reviewed By: bangshengtang Differential Revision: D44076452 fbshipit-source-id: f6ce8628e7345093bb55835f9523870c2914516f * jagged_jagged_bmm operator optimization (#1644) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1644 This diff optimizes the jagged_jagged_bmm operator using tiling across thread blocks and GPU shared memory. Reviewed By: brad-mengchi Differential Revision: D44029528 fbshipit-source-id: fa5cd5a26893f935427bce5efb7dfcc731c3f47d * Specify device to emulate_cache_miss kernel (#1660) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1660 When enabled emulate cache miss, it caused illegal memory access, if we're using more than one GPU. It turns out that previous diff didn't specify device within emulate_cache_miss kernel. This diff fixes it. In addition, cleaned up a bit (e.g., no need to used index_t based kernel launch for emulate_cache_miss kernel, as lxu_cache_locations is always with int32_t. Reviewed By: sryap, YuzeDaiMeta Differential Revision: D44340131 fbshipit-source-id: d99ba2364e9030cbca6c1166e578d24d99646bb1 * Add C++17 Support to FBGEMM and FBGEMM_GPU OSS builds (#1652) Summary: - Add C++17 support for the entire FBGEMM_GPU build - Add C++17 support for the entire FBGEMM build - Update FBGEMM tests and benchmarks to be C++17-compatible - Make FBGEMM builds output more logging - Cherry-pick code changes from D43776442 v4 now that C++17 is fully supported Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1652 Reviewed By: shintaro-iwasaki Differential Revision: D44287321 Pulled By: q10 fbshipit-source-id: 4bf2bcf66d528939865d42b6deafc470bee55d17 * Prune CPU/GPU TBE optimizer codegen (#1659) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1659 This diff aims to reduce the build time and libary size of `//deeplearning/fbgemm/fbgemm_gpu/codegen:embedding_ops`. The diff modifies the build target to generate and compile only the necessary files. This is based on the fact that CPU and GPU do not support all optimizers in `SplitTBE`. (Before this diff, all optimizers were generated and compiled for both CPU and GPU.) The following is the list of supported optimizers |OptimType|Generated optimizer|Supported on CPU|Supported on GPU| |EXACT_ADAGRAD|adagrad|x|x| |EXACT_ROWWISE_ADAGRAD|rowwise_adagrad_with_counter|x|x| ||rowwise_adagrad|x|x| |EXACT_ROWWISE_WEIGHTED_ADAGRAD|rowwise_weighted_adagrad|x|x| |EXACT_SGD|sgd|x|x| |SGD|approx_sgd|x|x| |ROWWISE_ADAGRAD|approx_rowwise_adagrad_with_counter|x|| ||approx_rowwise_adagrad|x|| |ADAM|adam||x| |LAMB|lamb||x| |LARS_SGD|lars_sgd||x| |PARTIAL_ROWWISE_ADAM|partial_rowwise_adam||x| |PARTIAL_ROWWISE_LAMB|partial_rowwise_lamb||x| |-|rowwise_adagrad_with_weight_decay||| |-|approx_rowwise_adagrad_with_weight_decay||| Note: x = supported Reviewed By: jianyuh Differential Revision: D44326540 fbshipit-source-id: 02413256b4a675f13ada8e8820820cb5112cb405 * Fix the Documentation Build Job (#1673) Summary: - Rewrite the documentation builds job to use the build infrastructure tooling - Rename workflow files for consistency Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1673 Reviewed By: shintaro-iwasaki Differential Revision: D44472660 Pulled By: q10 fbshipit-source-id: 60434c1f7098b7efa8c750133bb22f14fc98d5dc * Back out "Prune CPU/GPU TBE optimizer codegen" (#1675) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1675 Original commit changeset: 02413256b4a6 Original Phabricator Diff: D44326540 Reviewed By: q10, jianyuh Differential Revision: D44475251 fbshipit-source-id: 5be66944a833e03a2737fc6d1baaa5c351455b2c * Prepare bounds_check_indices for VBE (#1633) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1633 Prepare `bounds_check_indices` for variable batch size TBE (VBE). - Update the frontend API to accept VBE args - Update the backend logic to process VBE data Reviewed By: jianyuh Differential Revision: D43253703 fbshipit-source-id: 2870f0c41a96265650281a9b6362d4e6dc48009b * Move pruning/index_remapping support to embedding inplace update files (#1667) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1667 As title. This diff moves pruning/index_remapping support to embedding inplace update files. Reviewed By: jianyuh Differential Revision: D44409419 fbshipit-source-id: 93fc91d83502eb95cb0feca2a8a03b003c336078 * jagged_softmax forward optimization (#1661) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1661 This diff optimizes jagged_softmax forward with more efficient reduction from cub library. Reviewed By: brad-mengchi Differential Revision: D44161021 fbshipit-source-id: bf2e059d14ef4d7ad311edac65155a463ba653ff * jagged_softmax backward optimization (#1662) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1662 This diff optimizes jagged_softmax backward with more efficient reduction from cub library Reviewed By: brad-mengchi Differential Revision: D44205819 fbshipit-source-id: cd1d7a886d6ba68201dc1ad782c2e8cde7ff706b * multi-gpu all_to_one improvements (#1674) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1674 improved multi-gpu all_to_one with: 1. new intermediate hop selection taking advantage of distinct NVLinks 2. overlapping of intermediate hop transfers with each-other and with direct-peer transfers Reviewed By: doehyun Differential Revision: D44285941 fbshipit-source-id: 0202083f04388b5ba60b8155809433f334993ef4 * Extract and export weights offsets/placements initialization functions (#1669) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1669 Extract portions initializing the weights_placements/offsets tensors into separate functions and jit.export them. SplitState is converted to a NamedTuple since we can't jit.script a dataclass that also holds an enum. Reviewed By: houseroad Differential Revision: D44338256 fbshipit-source-id: e1c12e5956f7217d51cd190958c3764d220e521d * Fix the ROCm Test Job (#1668) Summary: - Clean up the ROCm test job and re-enable ROCm testing on the rocm instances. - Update the build scripts framework to build FBGEMM_GPU against the correct hardware target that it is intended to be tested on. One thing that was discovered was that if FBGEMM_GPU was built with `PYTORCH_ROCM_ARCH=gfx90a` but run on `gfx908` target, the tests will fail with a segfault. While the failure is expected, the segfault can be unfriendly and confusing for users. - Enable correct compilation of `merge_pooled_embeddings` operator under ROCm - Fix existing code in `jagged_tensor_ops` from PR https://github.com/pytorch/FBGEMM/issues/1661 and https://github.com/pytorch/FBGEMM/issues/1662 that break its compilation under ROCm 5.3 Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1668 Reviewed By: shintaro-iwasaki Differential Revision: D44453594 Pulled By: q10 fbshipit-source-id: 2030cd0e00c6ff9694c2783dfd62c31cf5543da2 * Use exported functions instead of calling initialize_weights in weights loading (#1676) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1676 Export a function to reset the embedding specs by target location Reviewed By: RoshanPAN, houseroad Differential Revision: D44338258 fbshipit-source-id: 502733e9f3a164450a02656d2822492fbf69f994 * Extract index remappings array initialization and jit.export it (#1670) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1670 ATT Reviewed By: RoshanPAN, houseroad Differential Revision: D44338257 fbshipit-source-id: c091666c7a4d294c283f5e3774d0494089fc3478 * Disable COUNTER in FBGEMM test (#1683) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1683 Disable FBGEMM test on COUNTER mode temporarily. Reviewed By: sryap Differential Revision: D44589052 fbshipit-source-id: f2af6f9e3cce75d4c599c4708055e5f52ac705e2 * update hipify_torch and remove manual mapping of C10 macros (#1682) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1682 Reviewed By: shintaro-iwasaki Differential Revision: D44599348 Pulled By: q10 fbshipit-source-id: 8f968a7c21b09358eac070a35ee15d5b767ea94c * Install NVIDIA Drivers on Instances Missing the Drivers (#1684) Summary: - Use the pytorch/test-infra action ot install NVIDIA drivers properly if the instance is missing the drivers Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1684 Reviewed By: shintaro-iwasaki Differential Revision: D44603925 Pulled By: q10 fbshipit-source-id: 712bdf5c2af67c5a6f540567abcc47ed892912c1 * Clean up the linting job (#1686) Summary: Sumary: - Clean up the linting job to use the build scripts infrastructure - Delete the Conda prefix directory before creating a new environment, if it exists Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1686 Reviewed By: shintaro-iwasaki Differential Revision: D44646234 Pulled By: q10 fbshipit-source-id: d754efeadffb265c9e55bc302606fc1e60ef8b51 * reduce_to_one (#1571) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1571 reduce_to_one for row-wise sharding in inference Similar approach to all_to_one but without having the source waiting for target to be ready for potential WAR and WAW dependency violation because in this reduce_to_one implementation we create a new destination tensor. Reviewed By: xing-liu, jianyuh Differential Revision: D34263436 fbshipit-source-id: 7b1630b395311cfd6fef124113436f87f51a6fba * Reorganize the build scripts (#1685) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1685 Reviewed By: r-barnes, shintaro-iwasaki Differential Revision: D44654808 Pulled By: q10 fbshipit-source-id: a58987b4a3970139bba72db8cecc89c0256fba76 * Prune CPU/GPU TBE optimizer codegen (#1678) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1678 This diff aims to reduce the build time and libary size of `//deeplearning/fbgemm/fbgemm_gpu/codegen:embedding_ops`. [1/2] Update `lookup_invoker` to enable the function invoker based on `has_cpu_support` and `has_gpu_support` [2/2] Update the code generation part The diff modifies the build target to generate and compile only the necessary files. This is based on the fact that CPU and GPU do not support all optimizers in `SplitTBE`. (Before this diff, all optimizers were generated and compiled for both CPU and GPU.) The following is the list of supported optimizers |OptimType|Generated optimizer|Supported on CPU|Supported on GPU| |EXACT_ADAGRAD|adagrad|x|x| |EXACT_ROWWISE_ADAGRAD|rowwise_adagrad_with_counter|x|x| ||rowwise_adagrad|x|x| |EXACT_ROWWISE_WEIGHTED_ADAGRAD|rowwise_weighted_adagrad|x|x| |EXACT_SGD|sgd|x|x| |SGD|approx_sgd|x|x| |ROWWISE_ADAGRAD|approx_rowwise_adagrad_with_counter|x|| ||approx_rowwise_adagrad|x|| |ADAM|adam||x| |LAMB|lamb||x| |LARS_SGD|lars_sgd||x| |PARTIAL_ROWWISE_ADAM|partial_rowwise_adam||x| |PARTIAL_ROWWISE_LAMB|partial_rowwise_lamb||x| |-|rowwise_adagrad_with_weight_decay||| |-|approx_rowwise_adagrad_with_weight_decay||| Reviewed By: q10 Differential Revision: D44484764 fbshipit-source-id: f04710e66498bdcbdad619d48411c2403316901c * thread tiling for jagged_jagged_bmm (#1691) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1691 This diff adds thread tiling optimization in jagged_jagged_bmm operator, where each thread will process a tile of elements instead of one. The implementation is similar to the one applied to jagged_dense_bmm: D43674845. Reviewed By: brad-mengchi Differential Revision: D44764339 fbshipit-source-id: ca4cf257bac755ab97754fdc6605072cfbfb1c4d * tune the tile sizes for jagged_dense_bmm (#1692) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1692 Tune the tile sizes based on the input tensor size. If M > N, then use larger tile size in M dimension, otherwise use larger tile size in N dimension. Reviewed By: brad-mengchi Differential Revision: D44791699 fbshipit-source-id: 348a66089d781e9fef141b63d7a56e6dfa5da905 * Populate supported optims to match OSS Pytorch state dict (#1632) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1632 ATT. Reviewed By: jianyuh Differential Revision: D43887969 fbshipit-source-id: 048ff61a925113b29c547abf20d7acdc4a50b8d7 * Build Scripts and README Improvements (#1695) Summary: - Update build scripts to print out cc, c++, and nvcc preprocessor defines - Print out all undefined symbols in the output library after build to inspect whether or not templates have been un-instantiated - Handle the case where `TORCH_CUDA_ARCH_LIST` is pre-defined in the environment - Clean up the FBGEMM_GPU READMEs to consolidate all FBGEMM_GPU build instructions into `docs/BuildInstructions.md` - Fix the build badges for FBGEMM and FBGEMM_GPU - Add Slack contact information to the READMEs - Remove deprecated GitHub workflows and build scripts in favor of the new scripts, which cover all the functionality of the old scripts Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1695 Reviewed By: shintaro-iwasaki Differential Revision: D44901368 Pulled By: q10 fbshipit-source-id: bef6045347c905a051970e4e5f8630175e0f5ef6 * Add Documentation to Work Around GCC 12 Regressions (#1697) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1697 Reviewed By: shintaro-iwasaki Differential Revision: D44935915 Pulled By: q10 fbshipit-source-id: e1bdd4ebff18bd9708208a5b659ef9a93ebc866a * Fix build instructions (#1701) Summary: This change fixes a missing step (cd) in the build instructions. Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1701 Reviewed By: sryap Differential Revision: D45011147 Pulled By: q10 fbshipit-source-id: 704ce5bd3cfbd62c31f434c830a7300e5d645024 * Fix a build error from -Wno-unused-but-set-variable (#1702) Summary: This project is compiled with -Wall and -Werror (see https://github.com/pytorch/FBGEMM/pull/868) and is throwing an error for the unused variable here. This code appears to be debugging code that was used to verify that the function it's contained in was originally implemented properly so the most straightforward solution is to just remove it. Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1702 Reviewed By: sryap Differential Revision: D45011174 Pulled By: q10 fbshipit-source-id: 2c252cfa6063789371f5fba5f642c2f4fb72455f * Fix exception in QuantUtilsTest (#1703) Summary: This test mistakenly calls reserve() to set a vector's length instead of resize(). reserve() allocates memory for the specified number of elements, but does not actually increase the number of elements that can legally be stored in the vector. This test runs with ASAN enabled which is catching this illegal access and causing the test to fail. This change fixes the code to instead call resize(); the test now passes. Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1703 Reviewed By: sryap Differential Revision: D45011317 Pulled By: q10 fbshipit-source-id: 2840d7bfcfb46ca1523f55e77a3834a1d561c045 * Support EXACT_ADAGRAD in `get_optimizer_state` (#1700) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1700 This diff support `get_optimizer_state` for exact_adagrad. Exact_adagrad is not supported in `get_optimizer_state`. However, this is needed for creating fused optimizer in torchrec. Reviewed By: r-barnes Differential Revision: D44963975 fbshipit-source-id: e2f523dfc1e1d17a4925e7ce4a9e65829f1cf1b0 * Split the Rendering of `embedding_forward_quantized_split_template.cu` into Smaller Files (#1694) Summary: `embedding_forward_quantized_split_template.cu` is a very large jinja template that renders 30+ C++ templates, which are then instantiated to over 600+ kernel functions. There are three sets of jinja templates in `embedding_forward_quantized_split_template.cu`: those related to `int_nbit_split_embedding_*`, `pruned_hashmap_lookup_*` and `pruned_array_lookup_*`.. Currently, the rendering produces a single file, which takes a large amount of time to compile. This PR does two things at a high level. First, it breaks up the jinja template into multiple jinja templates. Then, it forces each of these smaller jinja templates to render multiple source files instead of a single source file. This change will enable build parallelization and overall build time savings. Details: - Port improvements to `embedding_forward_quantized_split_template.cu` from D44707812 - Move the non-jinja-template code inside `embedding_forward_quantized_split_template.cu` over to `embedding_forward_template_helpers.cuh` - Move `pruned_hashmap_lookup_*` and `pruned_array_lookup_*` sets of jinja templates out to non-jinja-template `embedding_forward_quantized_split_lookup.cu`, since the template generated functions are redundant. - Break the `int_nbit_split_embedding_*` set of jinja templates into two files, one for rendering kernel-side code (`embedding_forward_quantized_split_nbit_kernel_template.cu`) and the other for rendering host-side code (`embedding_forward_quantized_split_nbit_host_template.cu`) - For the `int_nbit_split_embedding_*` host-side jinja template, make it render `weighted`, `unweighted`, and `unweighted nobag` variants into separate source files - For the `int_nbit_split_embedding_*` kernel-side jinja template, make it render into N = [`weighted`, `unweighted`, and `unweighted nobag` variants ] x [ 6 embedding types ] separate source files, each containing a single C++ template kernel function. Also generate the code to explicitly instantiate the kernel templates. For each of the C++ templates being generated, there will be 2 {device-only bool} x [3-4] (output types) x [3-5] (cases) = 18-40 actual template instantiations - To help with debugging missing template instantiations, print out all undefined symbols in the output library after build to inspect whether or not templates have been un-instantiated - Update build scripts to print out `cc`, `c++`, and `nvcc` preprocessor defines - Handle the case where `TORCH_CUDA_ARCH_LIST` is pre-defined in the environment Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1694 Reviewed By: sryap, r-barnes Differential Revision: D44842524 Pulled By: q10 fbshipit-source-id: 96f92e40ab2fec598aeb8c483e94997ac050aae7 * Back out "Prune CPU/GPU TBE optimizer codegen" (#1706) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1706 Original commit changeset: f04710e66498 Original Phabricator Diff: D44484764 Reviewed By: q10, brad-mengchi, jianyuh, shintaro-iwasaki Differential Revision: D45054051 fbshipit-source-id: 9d14504c76eb93b2f1b14f4c2ec4c5b807c7fc4a * Use CUB kernel for 2D asynchronous_complete_cumsum (#1707) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1707 Temporarily use the CUB kernel instead of the custom kernel for 2D `asynchronous_complete_cumsum` Reviewed By: q10, brad-mengchi, jianyuh Differential Revision: D45062784 fbshipit-source-id: cebe3992ff8ebec9c0f554e729b8d79a1eced1de * Split the Code Generation for `embedding_backward_split_template.cu` into Smaller Files (#1705) Summary: `embedding_backward_split_template.cu` contains both jinja-template and non-jinja-template code, and some of the templating is unneccessary. Furthermore, the template generates both the vanilla and `nobag` variants of unweighted into the same source file. This PR moves the non-jinja-template code out of the template, de-duplicates code are unneccessarily templated, and splits the generation of the code to three files per optimizer, one for `weighted`, `unweighted nobag`, and `unweighted`. Details: - Migrate non-jinja-templated code out of `embedding_backward_split_template.cu` and into `embedding_backward_template_helpers.cuh` - De-templatize `split_embedding_backward_codegen_{{ optimizer }}_{{ wdesc }}_find_long_segments` into `split_embedding_backward_codegen_find_long_segments` since there is no implementation difference between the optimizers and weighted vs unweighted - Migrate `grad_mean_kernel` and `split_embedding_backward_codegen_find_long_segments` into a separate non-template source file to de-duplicate code generation and compilation - Split the code generation of `embedding_backward_split_template.cu` into 3 files per optimizer, according to weighted, unweighted_nobag, and unweighted Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1705 Reviewed By: sryap Differential Revision: D45073273 Pulled By: q10 fbshipit-source-id: e82ea643f8e67ad5aa0b3de03562532c5735453d * Add jagged slice op for cpu (#1690) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1690 The context why this is needed is as follows 1) For really long sparse features we want to split them into multiple chunks that can be fed into the model 2) Slicing requires users to require per row start point & a maximum L. Based on these requirements, a custom op mimicing the slice semantics of a normal tensor works best. An example usage using pseudo code ``` input_jagged_tensor = [[1, 2, 3, 4], [1, 2, 3], [1, 2, 3, 4, 5, 6], [1], [1, 2]] start = [0, 0, 0, 0, 0] slice_length = 3 >> jagged_slice(input_jagged_tensor, start, slice_length) output_jagged_tensor = [[1, 2, 3], [1, 2, 3], [1, 2, 3], [1], [1, 2]] ``` A corresponding operation for dense tensor would look like ``` dense_tensor = torch.randn((8, 10)) slice_dense_tensor = dense_tensor[:, 1:3] ``` Reviewed By: sryap Differential Revision: D44299744 fbshipit-source-id: 44996f2f2ec5fc5f31dda4cb3bd8f0241497df66 * Move radix sort to common utilities and add the possibility to handle negative integers (#1672) Summary: Move the `radix_sort` implementation to common utilities, so it can be used in PyTorch in case it was not built with FBGEMM GPU. Add the possibility to handle negative integers, which is crucial for reusing `radix_sort` in PyTorch's `sort` operation. Details: This PR addresses two issues: 1. `radix_sort` is currently used in [scatter_reduce](https://github.com/dszwicht/pytorch/blob/master/aten/src/ATen/native/cpu/ScatterGatherKernel.cpp#L630) (please view this [comment](https://github.com/pytorch/pytorch/pull/82703/files#r1045360609) for more information). Till now `radix_sort` was under `fbgemm_gpu` subproject. It means that implementation was not available in PyTorch in case it was built for CPU - that's why `radix_sort` was copy pasted under aten directory in PyTorch. This PR moves `radix_sort` implementation to common utilities. 2. In GNN workloads we often sort 1D integer data with non-negative values, for example, when converting CSR to CSC format. Unfortunately, `torch.sort` for 1D data works sequentially. `radix_sort` seems to be a perfect match to accelerate described case. However, suppose we want to do that on the PyTorch site. In that case, we have to either fallback to a regular path after detecting negative numbers in the tensor or perform post-processing, by swapping positive and negative blocks of data (data like `[2, -1, -2, 1]` after sorting will be in the following form `[1, 2, -2, -1]`, due to the fact of how numbers are stored). Both these solutions are not elegant. As an alternative, I propose the extension of `radix_sort` algorithm, by giving it capability to work with negative numbers. This can be enabled by passing an optional parameter, `maybe_with_neg_vals`. If set to `true`, we will perform all passes (up to the most significant sign bit) and apply a special prefix sum combination in the last pass. An example of how we can reuse fbgemm in PyTorch can be found in my private fork, [here](https://github.com/dszwicht/pytorch/pull/2) (I also provide speedup data). The above changes have several consequences: 1. `TORCH_CHECK` was replaced with `assert` as fbgemm CPU does not have PyTorch in its dependencies. 2. `__builtin_clz` was replaced with manual implementation as `__builtin_clz` is not portable. Additional information for reviewers: I did perform benchmarks of `radix_sort` before and after my code modification. I didn't observe any performance drop. Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1672 Reviewed By: sryap Differential Revision: D44616959 Pulled By: q10 fbshipit-source-id: f34594478c94ec6610c05545feb2044b58d79d66 * Daily `arc lint --take CLANGFORMAT` Reviewed By: bigfootjon Differential Revision: D45141964 fbshipit-source-id: 58308a31522a3b1446835e358a93483b611c4b15 * `CMakeLists.txt` Cleanups (#1712) Summary: - Re-organize and comment the `CMakeLists.txt` for FBGEMM_GPU for better clarity - Disable verbose HIPCC warnings that are non-actionable when building the ROCm variant of FBGEMM_GPU Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1712 Reviewed By: shintaro-iwasaki Differential Revision: D45189904 Pulled By: q10 fbshipit-source-id: 3df6ff3b957886c64bc13fc6bc7a0147b74ee783 * support indices broadcast for reorder_batched_ad_indices (#1711) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1711 this is to support the case for request-only combined input sparse feature broadcast when `broadcast_indices` is enabled, the assumption for the inputs: - `cat_ad_offsets` and `cat_ad_indices` only contain the offsets and indices for the combined batches, where each batch only contain one instance (potentially multiple tables) - `reordered_cat_ad_offsets` needs to be after broadcasted, and contains `num_ads_in_batch * num_tables + 1` elements - `batch_offsets` is also after broadcasted - `num_indices_after_broadcast` is required to allocate the output buffer added coverage for the newly added branch Reviewed By: r-barnes Differential Revision: D45155887 fbshipit-source-id: 67f96d60168aa83cf24fef459addee89f06e1c6b * Add a check that get_filelist python exec process worked (#1715) Summary: Add a check that get_filelist python exec worked. If bad params (python, args, ...), get_filelist() was continuing without noticing/warning/erroring out, making cmake failing later for weird reasons ("no sources"). Adds a safety check on the RESULT_VARIABLE of cmake execute_process(). Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1715 Reviewed By: shintaro-iwasaki Differential Revision: D45235231 Pulled By: q10 fbshipit-source-id: 049eae1fc5d7f42d73048e81c02c2f282d8859b0 * Fix compilation error under ROCm 5.3 (#1719) Summary: - Fix bug introduced by PR 1711 (D45155887), which broke compilation of FBGEMM_GPU under ROCm 5.3 Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1719 Reviewed By: sryap Differential Revision: D45238536 Pulled By: q10 fbshipit-source-id: de9d2aa01ced0a37be1ea7903a361e3a24beed8d * Backward Split, pt. 2: Migrate `*_warp_per_row` and `*_cta_per_row` kernel templates out of `embedding_backward_split_template.cu` (#1710) Summary: - Migrate the definition of `split_embedding_*_backward_codegen_*_*_kernel_warp_per_row_1` from `embedding_backward_split_template.cu` over to `embedding_backward_split_kernel_warp_template.cu` and explicitly instantiate the templates separately - Migrate the definition of `split_embedding_*_backward_codegen_*_*_kernel_cta_per_row_1` from `embedding_backward_split_template.cu` over to `embedding_backward_split_kernel_cta_template.cu` and explicitly instantiate the templates separately Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1710 Reviewed By: sryap Differential Revision: D45205217 Pulled By: q10 fbshipit-source-id: 96b34e9389e70b64d8391f2c9d39f4009f3d65ce * Add CLI support (M,N,K) to GEMMsBenchmark (#1721) Summary: Add CLI support (M,N,K) to GEMMsBenchmark Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1721 Reviewed By: sryap Differential Revision: D45281533 Pulled By: q10 fbshipit-source-id: 0ce5b38f54877acb26421dead1d2dc63cd11a2a1 * Fix data conversion in radix sort that can cause data loss (#1718) Summary: Fix data conversion in `radix_sort` that can cause data loss. Details: When `elements_count` is passed to the internal kernel implementation it is implicitly converted from `int64_t` to `int`. It can cause data loss, resulting in a partially sorted array. This PR fixes this issue. As a result of changing the `elements_count` type in internal functions to `int64_t`, `histogram` and `histogram_ps` types also were updated (to not generate further conversions). This is a follow-up for https://github.com/pytorch/FBGEMM/issues/1672. Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1718 Reviewed By: sryap Differential Revision: D45253811 Pulled By: q10 fbshipit-source-id: a5368a4401f05ebc471cb17107297a48f43a75c0 * support lengths broadcast for reorder_batched_ad_lengths (#1716) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1716 similar to D45155887 when `broadcast_lengths` is enabled, the lengths are copied from the only instance of each batch, this is also to facilitate request-only broadcast Reviewed By: r-barnes Differential Revision: D45208736 fbshipit-source-id: 2c06cd4e9aae0c9c4e0668098de7db6f6da8c06b * remove optional for two ops (#1722) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1722 remove unnecessary optional decorators for the two newly added sparse ops Reviewed By: r-barnes Differential Revision: D45286152 fbshipit-source-id: 26109548db1acbc8fdf1a5183977eb8c64b45d41 * Prepare bounds_check_indices for VBE (#1713) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1713 Prepare bounds_check_indices for variable batch size TBE (VBE). - Update arg names Reviewed By: jspark1105, r-barnes Differential Revision: D45203680 fbshipit-source-id: 396c4122058db8dd1fc9eb5f0d620e8179c3e7a9 * Add check on configs and logging (#1728) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1728 Freq-SGD requires to set both `weight_decay_mode=WeightDecayMode.COUNTER` and `counter_based_regularization` to kick in. Previously we checked when `weight_decay_mode` is set but no config provided. There's another missing case when the config is provided but users forget to set `weight_decay_mode`. We add the check in this diff. In addition, added logging to print out whether **internally** counter is used or not to make debugging easier. Reviewed By: dvksabin Differential Revision: D45329516 fbshipit-source-id: 30389671c34a17d4baf48726f28096a670ede0b6 * Prepare transpose_embedding_input for VBE (#1717) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1717 Prepare `transpose_embedding_input` for variable batch size TBE (VBE). - Update the frontend API with new args Reviewed By: yuguo68 Differential Revision: D45212897 fbshipit-source-id: 5ad11a737130777fbe119aed6c7086e892752f4a * Convert GEMMsBench timebreakdown to a runtime cli option (#1725) Summary: Convert timebreakdown to a runtime cli option. Note: there is no code to measure packing, compute, kernel time ... so these are (atm) reported as 0, only total time is measured. ``` M, N, K, Type, Packing (us), Kernel(us), Postproc (us), Total (us), GOPs 64, 800, 320, FBGEMM_i8_acc32, 0, 0, 0, 218.593, 149.9 64, 800, 320, FBGEMM_i8_acc16, 0.0, 0.0, 0.0, 187.6, 174.7 ``` Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1725 Reviewed By: sryap Differential Revision: D45361847 Pulled By: q10 fbshipit-source-id: 4f2991a6208f0a5ae780729ce19bee611720953b * Fix error with empty row_counter_dev (#1730) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1730 In some cases, `torch.max(row_counter_dev)` causes failure because `row_counter_dev` is an empty tensor, example flow (f431977946). Here we guard the op by first checking if `row_counter_dev` is empty. Reviewed By: sryap Differential Revision: D45342010 fbshipit-source-id: 756a481c1098095f71dbb278ea84a01e89783790 * padding for fp8-rowwise quantization for varying length of 1D Tensor (#1729) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1729 As all gather becomes expensive for tensor/sequential parallel training, we create padded rowwise quantization/dequantization kernels for flattened tensor to convert between fp8 (stored as uint8 for gpu <= A100) and fp32 formats. Since the activations/grads will be concat into 1d tensor for all gather, the scaling to fit into fp8 format's range might be tricky as small elements will be quantized to zero if the scale is chosen to accommodate the largest element in the model. Thus, we continue to use row-wise quantization used in the previous all2all kernel. Every block with the size of "row_dim" will be quantized with the scale choose to accommodate the largest value in the block. Since the total length of the flattened tensor will not always be divisible by row_dim, we'll pad the 1D tensor to multiple of row_dim. As such, the padding/unpadding is handled by quantize/dequantize kernels and will be invisible to API calling them. Reviewed By: rohan-varma Differential Revision: D42721325 Privacy Context Container: L1138451 fbshipit-source-id: 33c712ba2fae709d29babee5ee4a8af6c7637b68 * Improve `TORCH_CHECK` diagnostics in files including deeplearning/fbgemm/fbgemm_gpu/codegen/embedding_forward_split_cpu.cpp (#1732) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1732 `TORCH_CHECK` produces pretty generic error messages. Using, eg, `TORCH_CHECK_GE` produces a message that shows the names of the variables being compared as well as their values at the time of comparison. This makes debugging easier. - If you approve of this diff, please use the "Accept & Ship" button :-) (7 files modified.) Reviewed By: bangshengtang Differential Revision: D45402701 fbshipit-source-id: 42501350543e31455e430b240e53f8e1883eb1ba * Improve `TORCH_CHECK` diagnostics in files including deeplearning/fbgemm/fbgemm_gpu/codegen/embedding_backward_dense_host.cpp (#1733) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1733 `TORCH_CHECK` produces pretty generic error messages. Using, eg, `TORCH_CHECK_GE` produces a message that shows the names of the variables being compared as well as their values at the time of comparison. This makes debugging easier. - If you approve of this diff, please use the "Accept & Ship" button :-) (7 files modified.) Reviewed By: bangshengtang Differential Revision: D45402700 fbshipit-source-id: 275bf837341a00d1cd4642b31bf9168455fa6c77 * Build cleanups (#1731) Summary: - Further break up `setup_env.bash` into separate domain scripts for easier maintenance - Update FBGEMM `CMakeLists.txt` to remove warning (https://github.com/pytorch/FBGEMM/issues/1714) Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1731 Reviewed By: sryap Differential Revision: D45406676 Pulled By: q10 fbshipit-source-id: 3ff5a7e2486b6898cb450d268a092371da5c2717 * Improve `TORCH_CHECK` diagnostics in files including deeplearning/fbgemm/fbgemm_gpu/fb/src/split_embeddings_utils.cu (#1735) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1735 `TORCH_CHECK` produces pretty generic error messages. Using, eg, `TORCH_CHECK_GE` produces a message that shows the names of the variables being compared as well as their values at the time of comparison. This makes debugging easier. - If you approve of this diff, please use the "Accept & Ship" button :-) (7 files modified.) Reviewed By: bangshengtang Differential Revision: D45402704 fbshipit-source-id: 9e9b1c1f526a398bbe50c99055187195ab751fa2 * Improve `TORCH_CHECK` diagnostics in files including deeplearning/fbgemm/fbgemm_gpu/src/split_embeddings_utils.cu (#1737) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1737 `TORCH_CHECK` produces pretty generic error messages. Using, eg, `TORCH_CHECK_GE` produces a message that shows the names of the variables being compared as well as their values at the time of comparison. This makes debugging easier. - If you approve of this diff, please use the "Accept & Ship" button :-) (3 files modified.) Reviewed By: bangshengtang Differential Revision: D45402697 fbshipit-source-id: c490d39bc826eab44ec16cbcc86273f8d7258fd9 * Use volatile pointer in inclusive_sum_scan_kernel (#1739) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1739 In the multi-block cumsum case, the `inclusive_sum_scan_kernel` implements the stream-scan technique in which each thread block has to consume the preceding sum result from the previous block. The sum result is passed via the `block_sums` buffer (global memory). To ensure that the sum results are visible for inter-thread-block consumption, the buffer has to be declared as `volatile` to prevent the compiler from caching the results in registers. This diff adds the `volatile` keyword to `block_sums`. Reviewed By: q10 Differential Revision: D45435897 fbshipit-source-id: f81a25b43eda18ae1eb18bed33f595fc27ef2707 * BF16 support for HBC ops. (#1744) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1744 Adding BF16 support for HBC ops, and updates on tests. Reviewed By: q10, sryap Differential Revision: D45449360 fbshipit-source-id: 8321155b426143d80064f12a910c0626bdfafbba * Use designated initializers & kernel launch checks in deeplearning/fbgemm/include/fbgemm/Utils.h (#1746) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1746 Designated initializers can make the code cleaner - If you approve of this diff, please use the "Accept & Ship" button :-) (1 files modified.) Reviewed By: sryap Differential Revision: D45464948 fbshipit-source-id: 28e38dc60b893fe7c91db0d791e069a6de87b420 * Dynamically determine platform name in FBGEMM scripts (#1742) Summary: Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1742 Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1738 Instead of hardcoding x86_64 when installing dependencies, let's now dynamically determine the platform name Reviewed By: excelle08 Differential Revision: D45246996 fbshipit-source-id: d9031e76a915c2362be62c85a3c1f0786828ca8b * Split the Rendering of `embedding_forward_split_template.cu` into Smaller Files (#1723) Summary: - Migrate `*_embedding_*_codegen_forward_*_kernel` out of `embedding_forward_split_template.cu` and into `embedding_forward_split_kernel_template.cu` - Migrate `*_embedding_nobag_codegen_forward_unweighted_small_kernel` out of `embedding_forward_split_template.cu` and into `embedding_forward_split_kernel_small_template.cu` Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1723 Reviewed By: sryap Differential Revision: D45363388 Pulled By: q10 fbshipit-source-id: 563ca610b15830aca854bc00d6a31fd6e8cb8a53 * Installation instructions for OSS (#1750) Summary: - Add installation instructions for OSS - Migrate Installation, Test, and Documentation information out of the README - Add link to GitHub Discussions in the README - Migrate the Netlify configuration from website to TOML file in the repo so that build jobs are configurable by developers Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/1750 Reviewed By: sryap, shintaro-iwasaki Differential Revision: D45540724 Pulled By: q10 fbshipit-source-id: beaab824cc5d441b96b89daea2a71f541e21f2ec --------- Co-authored-by: Banit Agrawal <[email protected]> Co-authored-by: Sabin Devkota <[email protected]> Co-authored-by: Junjie Yang <[email protected]> Co-authored-by: Benson Ma <[email protected]> Co-authored-by: Alfredo Tupone <[email protected]> Co-authored-by: Sarunya Pumma <[email protected]> Co-authored-by: Doe Hyun Yoon <[email protected]> Co-authored-by: Matt Galloway <[email protected]> Co-authored-by: Richard Barnes <[email protected]> Co-authored-by: Xiao Sun <[email protected]> Co-authored-by: Rengan Xu <[email protected]> Co-authored-by: siwasaki <[email protected]> Co-authored-by: Jianyu Huang <[email protected]> Co-authored-by: Yue Dong <[email protected]> Co-authored-by: Geet Sethi <[email protected]> Co-authored-by: Janet Yang <[email protected]> Co-authored-by: Wang Zhou <[email protected]> Co-authored-by: Jongsoo Park <[email protected]> Co-authored-by: Tran Le <[email protected]> Co-authored-by: Ryan Landay <[email protected]> Co-authored-by: Devashish Tyagi <[email protected]> Co-authored-by: Szwichtenberg, Damian <[email protected]> Co-authored-by: generatedunixname89002005325676 <[email protected]> Co-authored-by: Bangsheng Tang <[email protected]> Co-authored-by: William Tambellini <[email protected]> Co-authored-by: Jason Park <[email protected]>

facebook-github-bot added the cla signed label Apr 17, 2023

q10 force-pushed the bm/T150838079/split-embedding-backward branch 3 times, most recently from b99ca15 to bdb6953 Compare April 18, 2023 00:56

q10 changed the title ~~[T150838079][WIP] Split Codegen in embedding_backward_split_template.cu into smaller files~~ [T150838079][WIP] Split the Code Generation for embedding_backward_split_template.cu into Smaller Files Apr 18, 2023

q10 force-pushed the bm/T150838079/split-embedding-backward branch from bdb6953 to 1dba905 Compare April 18, 2023 23:42

facebook-github-bot closed this in 88dc18c Apr 19, 2023

facebook-github-bot added the Merged label Apr 19, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[T150838079][WIP] Split the Code Generation for `embedding_backward_split_template.cu` into Smaller Files #1705

[T150838079][WIP] Split the Code Generation for `embedding_backward_split_template.cu` into Smaller Files #1705

q10 commented Apr 17, 2023 •

edited

Loading

netlify bot commented Apr 17, 2023 •

edited

Loading

facebook-github-bot commented Apr 18, 2023

facebook-github-bot commented Apr 18, 2023

facebook-github-bot commented Apr 19, 2023

[T150838079][WIP] Split the Code Generation for embedding_backward_split_template.cu into Smaller Files #1705

[T150838079][WIP] Split the Code Generation for embedding_backward_split_template.cu into Smaller Files #1705

Conversation

q10 commented Apr 17, 2023 • edited Loading

netlify bot commented Apr 17, 2023 • edited Loading

✅ Deploy Preview for pytorch-fbgemm-docs canceled.

facebook-github-bot commented Apr 18, 2023

facebook-github-bot commented Apr 18, 2023

facebook-github-bot commented Apr 19, 2023

[T150838079][WIP] Split the Code Generation for `embedding_backward_split_template.cu` into Smaller Files #1705

[T150838079][WIP] Split the Code Generation for `embedding_backward_split_template.cu` into Smaller Files #1705

q10 commented Apr 17, 2023 •

edited

Loading

netlify bot commented Apr 17, 2023 •

edited

Loading