3D forward convolution solver with non-packed input tensors #2418
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
replace constant with appropriate block dimension WIP: add strides as a kernel parameter for non-packed tensors WIP: created non-packed variants for remaining 3D conv kernels fix: treat input and output as 6D NDHWGC tensor use std::call_once for initializing random seed revamp naive kernels to use strides in the non-packed tensors case (controlled by a bool flag) WIP: fixed kernel compilation issues but unable to load kernel code object WIP: fixed issue with hip rtc split channel strides into group, channels_per_group in solver fix indexing to left-to-right order fix bug with too much padding between kernel args num channels should be a multiple of num groups re-enable naive ref kernels with strides array 2D forward tests are all working now WIP: debugging bwd tests WIP: tests up till 3D bwd conv passing fix bug in bwd ndhwc kernel fix formatting disable prints fix readability-inconsistent-declaration-parameter-name fix clang-format fix hip tidy issue reverting the change to static init of random seed address comments and tidy issues. Remove extra print removed blank line change remove unneeded include env var for choosing packed vs non-packed reference kernel fix warnings from hip-tidy address comment about array initialization clear a tiny hip tidy issue
replace constant with appropriate block dimension WIP: add strides as a kernel parameter for non-packed tensors WIP: created non-packed variants for remaining 3D conv kernels fix: treat input and output as 6D NDHWGC tensor use std::call_once for initializing random seed revamp naive kernels to use strides in the non-packed tensors case (controlled by a bool flag) WIP: fixed kernel compilation issues but unable to load kernel code object WIP: fixed issue with hip rtc split channel strides into group, channels_per_group in solver fix indexing to left-to-right order fix bug with too much padding between kernel args num channels should be a multiple of num groups re-enable naive ref kernels with strides array 2D forward tests are all working now WIP: debugging bwd tests WIP: tests up till 3D bwd conv passing fix bug in bwd ndhwc kernel fix formatting disable prints fix readability-inconsistent-declaration-parameter-name fix clang-format fix hip tidy issue reverting the change to static init of random seed address comments and tidy issues. Remove extra print removed blank line change remove unneeded include env var for choosing packed vs non-packed reference kernel fix warnings from hip-tidy address comment about array initialization clear a tiny hip tidy issue
replace constant with appropriate block dimension WIP: add strides as a kernel parameter for non-packed tensors WIP: created non-packed variants for remaining 3D conv kernels fix: treat input and output as 6D NDHWGC tensor use std::call_once for initializing random seed revamp naive kernels to use strides in the non-packed tensors case (controlled by a bool flag) WIP: fixed kernel compilation issues but unable to load kernel code object WIP: fixed issue with hip rtc split channel strides into group, channels_per_group in solver fix indexing to left-to-right order fix bug with too much padding between kernel args num channels should be a multiple of num groups re-enable naive ref kernels with strides array 2D forward tests are all working now WIP: debugging bwd tests WIP: tests up till 3D bwd conv passing fix bug in bwd ndhwc kernel fix formatting disable prints fix readability-inconsistent-declaration-parameter-name fix clang-format fix hip tidy issue reverting the change to static init of random seed address comments and tidy issues. Remove extra print removed blank line change remove unneeded include env var for choosing packed vs non-packed reference kernel fix warnings from hip-tidy address comment about array initialization clear a tiny hip tidy issue
Co-authored-by: Alex Eremin <CAHEK7@yandex.ru>
|
@iq136boy please resolve the conflict by merging the latest |
|
@CAHEK7 @averinevg @DrizztDoUrden @atamazov @JehandadKhan @amberhassaan : |
I do not understand why. According to the PR description this should be merged AFTER #2394. Moreover, the fix for the critical problems of #2334 (listed at #2440) must be merged before #2394. Proposed order of actions: |
@amberhassaan Yes. If we have enough time, the order should be #2394 first and then this one. But this PR is a p1 gating issue for rocm6.0 and today is the last day of branching date if I am correct. And the content of this PR should go in mainline. What I suggest and am doing now is that I create a PR that merge latest #2394 and let that PR run through MIOpen CI separately. If that PR run through the CI without any issue before the branching cutting date, then we can merge 2394 first. Otherwise, I still suggest that this PR should be merged first to unblock the mainline gating status. |
|
@atamazov Yes. The procedure that you mentioned is correct if we do not have a urgent hard deadline today. However, the MIOpen CI issue blocks the PRs for a couple days and caused some delays. So I proposed a solution here. Generally, I would like to merge 2394 into this PR and let it run through MIOpen CI separately. If something happened, then this PR should be a backup plan to unblock the rocm6.0 p1 gating. |
|
@junliume @amberhassaan @CAHEK7 Well, it seems there is NO conflict between this PR and #2394. I merged #2394 with this PR and there is NO merge conflict. Also, I tested it on MI200 machine development machine locally and all tests passed after the merge. So we are good to merge the two PRs no matter the orders. |
|
@iq136boy I understand your motives. However, I advise to never cut corners unless there is an explicit and specific request from management. From personal experience, I see that cutting corners to save time often leads to problems that then require more time, efforts, and stress to resolve than doing things the right way. 🤝 |
| @@ -314,7 +320,7 @@ bool ConvHipImplicitGemm3DGroupFwdXdlops::IsApplicable( | |||
| if(!problem.IsLayoutNHWC()) | |||
| return false; | |||
| const std::string& arch = ctx.GetStream().GetDeviceName(); | |||
| if(!(arch == "gfx908" || arch == "gfx90a")) | |||
| if(miopen::StartsWith(arch, "gfx11") || miopen::StartsWith(arch, "gfx10")) | |||
There was a problem hiding this comment.
I would suggest we convert this to a white list, so, list the arches that CK supports.
There was a problem hiding this comment.
There was a problem hiding this comment.
Good suggestion and should be used in all CK-based solvers.
There was a problem hiding this comment.
I would suggest we convert this to a white list, so, list the arches that CK supports.
Great idea. I have applied the change to all ck based solvers and currently doing the local test before push the changes.
bghimireamd
reviewed
Oct 12, 2023
| StartsWith(device_name, "gfx942"); | ||
| return is_wrw ? res_wrw | ||
| : (res_wrw || StartsWith(device_name, "gfx900") || | ||
| StartsWith(device_name, "gfx906")); |
There was a problem hiding this comment.
This is just a recommendation : place StartsWith(device_name, "gfx900") || StartsWith(device_name, "gfx906") in sperate bool function. That way we don't need to pass is_wrw.
junliume
approved these changes
Oct 12, 2023
junliume
force-pushed
the
dfeng/ck_nonpack_conv3d_fwd
branch
from
fe98e67
to
8091d1c
Compare
cderb
added a commit
that referenced
this pull request
Jun 28, 2024
* [Windows] roctracer: disable on Windows (not supported) (#2404) Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> * [MI200] Refresh kdb using db_sync (#2411) * Removal of convolution context (#2402) * [Jenkins][CI] clean workspace after each stage (#2412) * [tests] convert test_conv_igemm_mlir_fwd to gTest (#2291) * Revert "cmake: enable finding installed ZStd library (#2362)" This reverts commit e608b43. * Revert "Revert "cmake: enable finding installed ZStd library (#2362)"" This reverts commit 1e325a7. * Bump cryptography from 41.0.3 to 41.0.4 in /docs/.sphinx (#2408) Bumps [cryptography](https://github.com/pyca/cryptography) from 41.0.3 to 41.0.4. - [Changelog](https://github.com/pyca/cryptography/blob/main/CHANGELOG.rst) - [Commits](pyca/cryptography@41.0.3...41.0.4) --- updated-dependencies: - dependency-name: cryptography dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [SWDEV-416089][Doc] convolution API in MIOpen is restricted to alpha = 1.0 and beta = 0.0 (#2419) * [HotFix] zstd dependency on multi Linux distributes (#2417) * [CI][Jenkins] Enable rebooting in CI stages for CI stages with GPU use (#2420) * conf_reboot * configs_chg * [Bug Fixes] miopen_rocblas_gemm_ex3 call - invoker cache extra elements - conv direct naive input cast (#2414) * bugfixes miopen_rocblas_gemm_ex3 call would always throw error invoker cache adding extra elements conv direct naive yielding incorrect input cast for kernel arg * clear clang format issue --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [CI][Jenkins] Disabling smoke stages for CI branch runs (#2422) * [Tests] disable solver ConvHipImplicitGemm3DGroupWrwXdlops on Vega10 (#2432) * [Dockerfile] Upgrade cmake so that MIOpen docker can compile Composable Kernel (#2424) * upgrade cmake so that MIOpen docker can compile Composable Kernel * pin the cmake version to 3.27.5 * [Bug Fix] Compilation fix for -DMIOPEN_USE_ROCBLAS=Off (#2435) * bg/lwpmiopen 193 : Integrate CK's batch norm backward training into non-tunable MIOpen solver (#2385) * Reference kernel for 3D convolution for non-packed tensors (#2334) * [Doc] Bump rocm-docs-core from 0.24.2 to 0.25.0 in /docs/.sphinx (#2434) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.24.2 to 0.25.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.24.2...v0.25.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Fix weight tensor intialization to replace old PR1950 (#2436) * Add typecast to config key (#2413) * add typecast value to config key, as optional arg to fdb_key * fix clang-format issue * Save space in db key and optimize code. Do not print casting value when casting is not actually necessary. Co-authored-by: Artem Tamazov <artem.tamazov@gmail.com> * do not print casting to confkey when unnecessary, code cleanup, datatype rename * move GetDataTypeName to problem_descrption_base.hpp, organize includes * fix missing header --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> Co-authored-by: Artem Tamazov <artem.tamazov@gmail.com> * [Bugfix] Add cast swapping for swapped gemm inputs. (#2443) * add swapping for cast types when swapping A+B for gemm * [Bugfix] Kernel name fix, compilation err fix (#2446) * Bump gitpython from 3.1.35 to 3.1.37 in /docs/.sphinx (#2445) Bumps [gitpython](https://github.com/gitpython-developers/GitPython) from 3.1.35 to 3.1.37. - [Release notes](https://github.com/gitpython-developers/GitPython/releases) - [Changelog](https://github.com/gitpython-developers/GitPython/blob/main/CHANGES) - [Commits](gitpython-developers/GitPython@3.1.35...3.1.37) --- updated-dependencies: - dependency-name: gitpython dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Add MIOPEN_BETA_API defines around f8 (#2430) --------- Co-authored-by: JD <jahandad@gmail.com> * Remove INT8x4 support (#2441) * Test non-packed inputs with naive reference convolution kernels (#2394) * 3D forward convolution solver with non-packed input tensors (#2418) * Bump CK comit for ROCm 6.0 (#2439) * [Jenkins][CI] Enabling Nightly Runs for Nightly Branch w/ build_smoke_(fp32 + aux1 + fp16_bf16_int8) (#2437) * Remove ck solver's strides restriction (#2438) * remove ck solver's strides restriction * bn_cleanup: rename to in_strides * [tests] remove std::rand usage (#2400) * remove std::rand usage * remove deprecated code * Bump rocm-docs-core from 0.25.0 to 0.26.0 in /docs/.sphinx (#2451) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.25.0 to 0.26.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.25.0...v0.26.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [tests] test_tensor_api enhancement (#2450) * [MI100][MI200] Kernel db updates (#2454) * [tests] bg/ck_gfx_white_list: start using ck_utility::is_ck_whitelist to restrict tests to applicable platforms (#2458) * bg/ck_gfx_white_list : start using ck_utility::is_ck_whitelist function for all CK solvers * bg/ck_gfx_white_list: fix review comments * Rename function and return invalid solution instead of throwing an error (#2457) * Fix the return code on the workspace API (#2460) * regression: do not use file system symbolic/hard links (#2425) Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> Co-authored-by: JD <jahandad@gmail.com> Co-authored-by: Jun Liu <Liu.Jun@amd.com> * bg/fix_ck_guard_in_bn : fix CK guard around bn (#2464) Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [Bugfix] Layernorm Test: add missing hip_runtime.h include (#2465) * Remove redundancy. Replace test_layernorm_test with test_layernorm (#2467) * add missing hip_runtime.h include * rename the test to remove redundancy * [Doc] fix: update guides locations (#2456) * [MI300] add CI test stages (#2396) * [Windows] comgr: fix compiling with HIP SDK 5.5+ on Windows (#2364) Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [Windows] cmake: add option for building shared libraries (#2361) * [Windows] remove unused sys/time.h header file (#2360) Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> * Finally remove INT8x4 support. (#2452) * [Windows] fix sequences for Windows (#2359) * Fusion Find (#2388) * [tests] tensor_holder enhancement (#2449) * [Windows] cmake: bump up the minimum required version to 3.15 (#2356) Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> * [HotFix] Missing MIO_BN_GFX110X when building kernels (#2473) * [Tuning][MI100][MI200] Gold19 (#2470) * TunaNet Integration: MI250x (#2421) * [tests] remove direct std::random_device usage. (#2397) * Add a check for packed tensors for convolution solvers (#2471) * Bump urllib3 from 1.26.15 to 1.26.18 in /docs/.sphinx (#2462) Bumps [urllib3](https://github.com/urllib3/urllib3) from 1.26.15 to 1.26.18. - [Release notes](https://github.com/urllib3/urllib3/releases) - [Changelog](https://github.com/urllib3/urllib3/blob/main/CHANGES.rst) - [Commits](urllib3/urllib3@1.26.15...1.26.18) --- updated-dependencies: - dependency-name: urllib3 dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [Dependency][CK] Bump CK Commit Hash (#2479) Regular promotion to newer CK commit hash * [Windows] half: fix compiling with HIP SDK 5.7+ on Windows (#2363) * Added an API function for find 2.0 activation problem creation. (#2448) --------- Co-authored-by: Artem Tamazov <artem.tamazov@gmail.com> Co-authored-by: Evgenii Averin <86725875+averinevg@users.noreply.github.com> * [Windows] cmake: generate export header function (#2348) * cmake: generate export header function * incorporate review feedback --------- Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> * [tests] Refactor cache test to gTest (#1652) * Fix transposed convolutions (#2487) * [Windows] cmake: enable testing on Windows (#2380) * [tests] write 3d test that uses 2d gpu kernel (#2401) * Update dependencies and Dockerfile ROCm versions (#2463) * update sqlite3 and boost * fix MIOPEN_USE_COMPOSABLEKERNEL issue; more updates * sync googletest version with FIN * fix merge conflict * suppress float-equal warning in gtest * update FIn to the latest of its develop * update docker rocm to 5.7.1 * [Tests] Removed support for OCL backend. Do not print rocminfo output unless GPU detection failed. Some cleanup. (#2490) * [Windows] cmake: fix cmake/googletest.cmake on Windows (#2350) * cmake: fix cmake/googletest.cmake on Windows * incorporate review feedback --------- Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> * [Windows] cmake: use imported target for threads library instead of variable (#2355) * Remove a check that was missed for packed tensors (#2495) * Fix builds with rocBLAS that does not support F8 (#2480) * fix-build-old-rocblas-no-ck(01) Fix builds with rocBLAS that does not support F8 * fix-build-old-rocblas-no-ck(02) CK BN bugfixes. Fixes for builds without CK. * fix-build-old-rocblas-no-ck(03) Update fin up to the most recent commit in develop * fix-build-old-rocblas-no-ck(06) Resolve review comment * [Workaround] Issue 2496 - disabling the unit test case in wrw solver (#2497) * [NFC] Replace miopen::ProblemDescription with conv::ProblemDescription, part 4 (#2410) * Bump rocm-docs-core from 0.26.0 to 0.27.0 in /docs/.sphinx (#2501) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.26.0 to 0.27.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.26.0...v0.27.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Update the SyncDB tests to use multi-threading (#2407) * Workaround for issue #2492 - disable ConvBinWinoRxS when granularity loss is huge (#2507) * workaround_issue_2492(01) Disable ConvBinWinoRxS when granularity loss is > 0.995 (performance drops 200 times) * workaround_issue_2492(02) Allow disabing the W/A by setting MIOPEN_DEBUG_WORKAROUND_ISSUE_2493=0 in the env. * Workaround for issue #2492 part 2 (improvement) (#2510) * workaround_issue_2492(01) Disable ConvBinWinoRxS when granularity loss is > 0.995 (performance drops 200 times) * workaround_issue_2492(02) Allow disabing the W/A by setting MIOPEN_DEBUG_WORKAROUND_ISSUE_2493=0 in the env. * workaround_issue_2492(03) [debug] Disable MIOPEN_DEBUG_WORKAROUND_ISSUE_2493 during driver warm-up. * workaround_issue_2492(04) [quality] Make the compuation of max granularity loss more clear. * workaround_issue_2492_01(02) [debug] Log granularity loss when ConvBinWinogradRxSf2x3* solver is skipped. * workaround_issue_2492_01(03) [tests] test_db_sync: Disable WORKAROUND_ISSUE_2493 via environment. Support reading legacy fdb (WORKAROUND_ISSUE_1987). Allow FDB testing on gfx1030 (SKIP_KDB_PDB_TESTING). Add W/A for ConvOclDirectFwdFused on gfx1030. Print number of failures per testing thread. * workaround_issue_2492_01(04) Remove leftovers from gfx1030 testing * workaround_issue_2492_01(05) More gfx1030 leftovers removed * Find 2.0 problem fusing (#2466) * [Windows] fix for end of line issue on Windows (#2515) Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [Windows] cmake: strip mingw32 support and cross-compilation out (#2352) * [Windows] cmake: cleanup outdated code (#2349) * [Windows] cmake: make building tests optional (#2351) * [Tests] Relocate Googletest to dev requirements (#2512) * Step 0: build pass but tests fail * Make test build passes * Adopt reviewer comments * Integrate CK's layer norm into MIOpen solver (#2481) * [Tests] Limit layernorm CK test applicability (#2528) * Bump rocm-docs-core from 0.27.0 to 0.28.0 in /docs/.sphinx (#2534) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.27.0 to 0.28.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.27.0...v0.28.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [Windows] cmake: a few fixes for multi-config generators (#2357) * [clang-tidy] Use config file for clang-tidy configuration (#2489) * [Tests] split teardown to runtest and verify in layernorm gtest (#2535) split teardown to runtest and verify * [Enhancement] Add checks on workspace params (#2498) * added checks on workspace params * addressed review comments * fix release build warning * Revert "[Enhancement] Add checks on workspace params (#2498)" This reverts commit aa878a8. * [HOTFIX] Fix offline HIP builds after PR #2357. (#2544) * [conv] Remove clamping to MAX from CastTensor used in Bwd and WrW (#2538) * conv-bwd-wrw-disable-clamping(01) [wip] Add clamping parameter to CastTensor() and set it to proper value. * conv-bwd-wrw-disable-clamping(02) Clamping in SubTensorOpWithCastTensor1d. * conv-bwd-wrw-disable-clamping(03) Clamping in SubTensorOpWithCastTensor2d/3d/4d/5d * conv-bwd-wrw-disable-clamping(04) Removed WORKAROUND_ISSUE_2496 * [CK] Bump CK commit hash (#2540) * [Windows] cmake: use built in operator to component-wise version comparison (#2353) Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [Hotfix] when MLIR is not used in MIOpen (#2549) * [Doc] Standardize documentation for ReadtheDocs (#2548) Relates to ROCm/rocm-docs-core#330 * [quality] Fix: always define MIOPEN_LIBMLIR_SUPPORTS_GFX103X_DEFAULT (#2552) * ConvOclDirectFwdGen: Fixed out-of-bounds memory access (#2546) * Find 2.0 fusion (#2486) * [Doc] Remove dated comments in test CmakeLists.txt (#2551) * Bump cryptography from 41.0.4 to 41.0.6 in /docs/sphinx (#2561) Bumps [cryptography](https://github.com/pyca/cryptography) from 41.0.4 to 41.0.6. - [Changelog](https://github.com/pyca/cryptography/blob/main/CHANGELOG.rst) - [Commits](pyca/cryptography@41.0.4...41.0.6) --- updated-dependencies: - dependency-name: cryptography dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [Test] fix gemm driver dataType initialization (#2558) fix gemm dataType initialization and make gemm driver more dataType friendly * Environment variables update (#2514) * Use split CK libraries. (#2526) * Sum enhancement in case of inner dim reduce (#2543) * Initialize sum, modify layernorm * FLOAT to FLOAT_ACCUM in kernel, fix kernel index and host test and, split teardown to runtest and verify * remove unused var, int64_t to size_t, add two kernel profile, fix kernel index error, change reqd_work_item_cnt * Use GetMaxComputeUnits, fix GetSumWorkspaceSize flow * Add doxygen, add test case * remove MIOPEN_BETA_API * modify tolerance, add solver list * alignment * add IsImprovementOverROCm, reduce to sqrt(reduce), modify test case * throw to return false in performance check, duplicate code to function, fix wrong allocate memmory size * add experimental caution in doc, add memory copy check in driver, add detail in verify result of driver * modify tolerance * modify get input in driver * [Windows] cmake: replace UNIX with NATIVE command for separate_arguments() (#2555) * [Find 2.0] Bias for Find 2.0 fusion (#2525) * [HotFix] Env Var set conflicts between #2543 and #2514 (#2571) * [Doc] Bump rocm-docs-core from 0.29.0 to 0.30.0 in /docs/sphinx (#2572) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.29.0 to 0.30.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.29.0...v0.30.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Sam Wu <sam.wu2@amd.com> * [Doc] document NHWC convolution support (#2575) * Forward, backward data and backward weight convolution solver with fp8/bfp8 compute datatype. (#2531) * [HotFix] disable the f8 test cases that failed the f8 reference kernel in gtest (#2576) * [HotFix] Disable f8 gtest cases that might cause CI fails. (#2577) * [CK] Bump CK commit hash for staging (#2581) * Fix the f8 reference kernel issue that failed CI (#2586) * Patch necessary to make FP8 convolution compile with hiprtc (#2584) * [Doc] Bump rocm-docs-core from 0.30.0 to 0.30.1 in /docs/sphinx (#2589) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.30.0 to 0.30.1. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.30.0...v0.30.1) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [DOC] Doxygen change: enable warning as error msg and add missing API comments (#2585) * AI Based Parameter Prediction Model for conv_hip_igemm_group_fwd_xdlops Solver (#2523) * [HotFix] KDB Files should not be in the runtime package (#2591) * [Doc] Adding issue template (#2590) * [Doc] Add documentations for non-packed tensors convolution (#2537) * edit document of convolution * address comments --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [Doc] Fix broken links in README.md (#2595) * Add nightly update workflow (#2579) * Tests for RNN seq API (#2493) * [HotFix] Fix Windows build with disabled CK (after #2523) (#2598) * Properly guard CK usage by MIOPEN_USE_COMPOSABLEKERNEL defines * Update src/solver/conv_hip_implicit_gemm_grouped_fwd_xdlops.cpp Co-authored-by: Artem Tamazov <artem.tamazov@gmail.com> --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> Co-authored-by: Artem Tamazov <artem.tamazov@gmail.com> * [MIOpenDriver] Enabled gemmfp16. [tests] Added smoke test for fp16 and fp32 gemm. (#2592) * fix-gemmfp16(01) [MIOpenDriver] Enable gemmfp16 in the driver * fix-gemmfp16(02) [tests] Add smoke test for fp16 gemm * [Doc] Fix URLs (ROCmSoftwarePlatform -> ROCm) in the doc, comments, and code. + more (#2597) * Update URLs (ROCmSoftwarePlatform -> ROCm) in the documentation and comments in the source code. * (2) Update URLs (ROCmSoftwarePlatform -> ROCm) in the documentation and comments in the source code. * Fix incorrect link * Fix links * [HotFix] Bump CK commit hash for F8 patch (#2603) * [Doc] Fix broken links in CONTRIBUTING.md (#2601) * Fix broken rocmsoftwareplatform.github.io links in CONTRIBUTING.md * Use new organization name for repoistory links * [Windows] use find_package() for Eigen and frugally-deep (#2574) * [Windows] enable compilation on Windows (#2570) * [HotFix] 3D Group Conv Backward data and weight update. Failure noticed when pads and strides are not 1 (#2560) * [CMake] fix find_package(... GLOBAL) for CMake < 3.24 (#2610) * [HotFix][atamazov] multiple undefined behavior discovered with -fsanitize=undefined in DEV builds (#2609) * fix-issue-2602(01) Fix for smoke_miopendriver_gemm * Do not print output parameters in MIOPEN_LOG_FUNCTION calls. --------- Co-authored-by: atamazov <artem.tamazov@gmail.com> * [hipRTC] resolve symbol issues by explicitly link with hipRTC (#2612) * explicitly link with hipRTC * Update formatting * Consider MIOPEN_USE_HIPRTC=Off * Clean up --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> * Standardize workspace abstraction (#2524) * [gtest] conversion for code coverage tests (#2580) * [HotFix] revert #2580 and re-enable smoke tests (#2616) * Revert "[gtest] conversion for code coverage tests (#2580)" This reverts commit c5a2384. * re-enable smoke tests in CI * remove problematic github action * [Windows] use find_package() for SQLite3 (#2564) * [Doc] Bump rocm-docs-core from 0.30.1 to 0.30.2 in /docs/sphinx (#2620) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.30.1 to 0.30.2. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.30.1...v0.30.2) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [Windows] use official ZStd package from Facebook (#2565) * Remove MIOpenGEMM and MIOpenTensile leftovers (#2499) * Remove FIN_OLD_PROBLEM_DESCRIPTION_COMPAT (#2503) * [Jenkins] Add NOMLIR stage. [Workaround] Limit usage of gfx908 nodes in non-nightly builds (#2622) * Get rid of legacy 2GiB offset limits in CallGemm*() and transpose*() internal APIs and kernels. (#2613) * [BugFix] Proper fix for backward passes bwd/wrw for CK group conv 3d (#2619) * [BugFix] asm igemm fwd kernel will have computation error when c <=4 and dilation_y > 1, workaround (#2625) * Fused solver for Fwd Convolution with Residual add, Bias add and then activation function (#2517) * Bump MIOpen version to 3.1.0 and update CI docker (#2519) * [HotFix] resolve unknown type issue after #2517 (#2629) * [Doc] Bump rocm-docs-core from 0.30.2 to 0.30.3 in /docs/sphinx (#2628) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.30.2 to 0.30.3. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.30.2...v0.30.3) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [HOTFIX] Fix build with -DMIOPEN_USE_COMPOSABLEKERNEL=Off after #2517. (#2630) * [Jenkins][Tests] Add stage with -DMIOPEN_USE_COMPOSABLEKERNEL=Off after #2517 #2630. (#2631) * [HOTFIX] Fix build with -DMIOPEN_USE_COMPOSABLEKERNEL=Off after #2517. * add -DMIOPEN_USE_COMPOSABLEKERNEL=Off stage * make NOCK stage anyAPU and build ONLY * Adopt recommended changes * rename config_targets to make_targets * Extend GTest DISCOVERY_TIMEOUT to 5 mins * [Tests] add unit test for #2624 (#2632) * [gtest] Combine gtests into single binary. (#2599) * [Windows] rocblas: disable Beta API on Windows for HIP < 5.7 (#2405) * [tests] Limit applicability of ConvFwdBiasActivAPI/ConvFwdBiasResAddActivTest.ConvFusedAPI (#2635) * [Tests] helper for evn variables update in gtests (#2605) Co-authored-by: xinlipn <xinlipn@gmail.com> * [Windows] fix compilation of math functions on Windows (#2568) * [Windows] fix printf type incompatibility between type specifiers (#2569) * Fix miopen package dependency roctracer etc (#2508) * [Doc][NFC] added rocm v6, mi300, and default component (#2618) * [Windows] add a class to allow os-agnostic process execution (#2567) * [Windows] make BZip2 a required package (#2566) * [Windows] add missing symbol export (#2556) * add missing symbol export * more missing exports * fix format issues --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> Co-authored-by: Alex Eremin <CAHEK7@yandex.ru> * [ROCm 6.1][hipRTC] Fix build failures. [quality] Reorg standard includes in HIP sources. (#2637) * [WORKAROUND] Disable W/A for issue #1359 starting from ROCm 5.4.3. (#2225) Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [Dep] Bump CK commit hash for staging (#2640) * [Windows] default paths to user and system db files on Windows (#2365) * Fix COMgr dependency in MIOpen package (#2645) * [ROCm 6.0.1] Automatically activate the new HIPRTC PCH adaptations starting from the 6.0.24000 version. Fix some build errors. (#2644) * Automatically activate the new HIPRTC PCH adaptations starting from the 6.0.24000 version. Fix some build errors (#2465 + more) (cherry picked from commit 4f695d9) * Remove duplicated includes. * [HOTFIX] Adapt to changes in HIP Mainline 417 (possibly future 6.1 RC) (#2652) * fix-rocm61rc417(01) Disable new kernel build warnings. [NFC] Sort headers properly. * fix-rocm61rc417(02) [ROCm 6.1][HIPRTC] Use custom implementations instead of standard <limits>. This fixes build issues with ROCm 6.1. * fix-rocm61rc417(03) [ROCm 6.1][HIPRTC][Bugfix] Fixed issue in miopen_limits.h that prevented the use of custom implementations. * fix-rocm61rc417(04) [ROCm 6.1 RC][HIPRTC] Disable some of the custom implementations from <type_traits> (like `integral_constant`) for HIP mainline 417. This fixes some build issues. * fix-rocm61rc417(05) [ROCm 6.1 RC][offline compiler] Removed "-mcpu" from build options. This resolves kernel build issues with HIP mainline 417 (offline compiler). Improved diagnostic messages output onto console after offline build failures. * fix-rocm61rc417(06) [tests] Disable some testcase from handle_test as #2600 still persists in Hip Mainline 417. --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> * Correct parameter which prints unused flag in log fusion cmd (#2653) * [MI300][Tuning] Tunings for SWDEV tickets (#2654) * add initial tunings for mi300 * add test to db_sync * [ROCm 6.0.1] Fix merge error in #2652 that affects #2644. (#2658) * [CK] Bump CK commit hash for staging (#2659) * Bump gitpython from 3.1.37 to 3.1.41 in /docs/sphinx (#2662) Bumps [gitpython](https://github.com/gitpython-developers/GitPython) from 3.1.37 to 3.1.41. - [Release notes](https://github.com/gitpython-developers/GitPython/releases) - [Changelog](https://github.com/gitpython-developers/GitPython/blob/main/CHANGES) - [Commits](gitpython-developers/GitPython@3.1.37...3.1.41) --- updated-dependencies: - dependency-name: gitpython dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Bump jinja2 from 3.1.2 to 3.1.3 in /docs/sphinx (#2666) Bumps [jinja2](https://github.com/pallets/jinja) from 3.1.2 to 3.1.3. - [Release notes](https://github.com/pallets/jinja/releases) - [Changelog](https://github.com/pallets/jinja/blob/main/CHANGES.rst) - [Commits](pallets/jinja@3.1.2...3.1.3) --- updated-dependencies: - dependency-name: jinja2 dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [Doc] Updated links to ROCm Repositories (#2667) Changed <old-organization> to "ROCm". * [SWDEV-433582] Search-proofed PrepareInvoker (#2661) * [HotFix] fix clang format issue from #2661 * [FIN] update submodule (#2660) * [Windows] replace [[gnu::noreturn]] with [[noreturn]] (#2656) * [Windows] addkernels: fix operations on path for Windows (#2657) * [Windows] clean up the setting of environment variables cross-platform (#2655) * clean up the setting of environment variables cross-platform * fix clang-tidy * Bump rocm-docs-core from 0.30.3 to 0.31.0 in /docs/sphinx (#2676) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.30.3 to 0.31.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.30.3...v0.31.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Argmax enhancement in case of inner dim reduce (#2583) * [Test] Convert conv_igemm_dynamic_dlops etc to gTest (#2553) * [Bugfix] Restore Missing ctests (#2649) * [Windows] fix compilation on Windows (#2677) * [Windows] cmake: unpack kernels into a build directory (#2347) * Remove FIN_OLD_HANDLE_COMPAT and FIN_OLD_BINARY_CACHE_COMPAT (#2627) * Rename transpose* kernels (leftover of #2613) (#2673) * [CK] Bump CK commit hash for staging (#2683) Update CK to the latest staging * [zlib] Update rocm-recipes for more reliable zlib link (#2686) * [OCL] Use OpenCL 2.0 while compiling kernels (#2691) * Fix compilation on SELS/RHEL after #2657 merged (#2690) * [BF16][FP8][BF8] Fixed some specializations from `<limits>` and `<cmath>` (#2669) * conv::ProblemDescription: remove underscores, change return data type (#2685) * Add 2D Group Convolution Backward Data and Weights update solvers. Simplify and unify 3d group conv tests (#2663) * [HOTFIX] Disable "granularity loss" W/A for #2492 and add a new, "tiny tensor" based one. (#2695) * disable 2492 granularity_loss workaround and enable tiny_tensor workaround * workaround_issue_2492_02(01) Macros to uppercase. Add doc for WORKAROUND_ISSUE_2492_TINY_TENSOR. Add conditions N<=4 and C<=4 to the "tiny tensor" W/A. Disable it during warmup, make it controllable by MIOPEN_DEBUG_WORKAROUND_ISSUE_2492. * Update src/solver/conv_winoRxS.cpp --------- Co-authored-by: Jun Liu <Liu.Jun@amd.com> * [Clang-Format] Fix format issue * Bump rocm-docs-core from 0.31.0 to 0.32.0 in /docs/sphinx (#2699) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.31.0 to 0.32.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.31.0...v0.32.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [MI300][Tuning] gold 20 (#2697) * add gfx942 superbench winograd tunings, update gold version to 20 * update with more superbench tunings * Remove support for ROCm < 5.6.0 (#2665) * Remove support for ROCm < 5.7.0 * deprecate-rocm-less-5.7(03) Leftover that fixes build error with "-Werror" * deprecate-rocm-less-5.7(04) Resolve review comment * Bump rocm-docs-core from 0.32.0 to 0.33.0 in /docs/sphinx (#2707) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.32.0 to 0.33.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.32.0...v0.33.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [COMGR] Use OpenCL 2.0. [HIPRTC] Provide min/max limits for int. Fix build errors related to min/max limits for BF16. (#2705) * fix-rocm-mainline-issues-01(01) Removeed `constexpr` from numeric_limits<hip_bfloat16>::min()/max() as BF16 ctor provided by HIP can't be used in const expressions. * fix-rocm-mainline-issues-01(02) [COMGR] Globally engage OpenCL 2.0 * fix-rocm-mainline-issues-01(03) [HIPRTC] Provide min/max limits for int * [DOC] fix broken links in docs (#2696) * lwpmiopen_521_correct_doc_issues: fix broken links in docs * lwpmiopen_521_correct_doc_issues: remove citing * [HotFix] Fix DB install after #2347 (#2702) --------- Co-authored-by: Artur Wojcik <artur.wojcik@outlook.com> Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> * Add GroupNorm forward operation (#2623) * fix not reporting LFS missing files (#2710) * [HotFix][WHL] move the bfloat16 header to the proper guard (#2711) * [HotFix] Update FindDB for finetuning (#2712) * [CK] Update CK commit in requirements.txt for staging (#2713) * [Tests] Fix Gtest single executable build issue (#2715) (#2717) Add the missing build job to Jenkinsfile Fix duplicate class name issue in Gtest * [Windows] Do not use HIP runtime headers on Windows (#2719) * don't use WORKAROUND_DONT_USE_CUSTOM_LIMITS on Windows * don't use workaround SWDEV_413293 on Windows * CI base docker updates to ROCm 6.0.2 (#2714) * Softmax ocl refactoring (#2671) * Add cat forward operation (#2562) * [HotFix] Fix namespace conflict issue in gtest after #2562 (#2725) * Bump cryptography from 41.0.6 to 42.0.0 in /docs/sphinx (#2729) Bumps [cryptography](https://github.com/pyca/cryptography) from 41.0.6 to 42.0.0. - [Changelog](https://github.com/pyca/cryptography/blob/main/CHANGELOG.rst) - [Commits](pyca/cryptography@41.0.6...42.0.0) --- updated-dependencies: - dependency-name: cryptography dependency-type: indirect ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [DB Install] fix installation of *.fdb.txt and *.db files (#2728) * Update CHANGELOG.md (#2720) * bg/update_change_log_lwpmiopen_501: update change long till rocm 6.1.0 (MIOpen-3.1.0) * bg/update_change_log_lwpmiopen_501: remove typo * bg/update_change_log_lwpmiopen_501: fix broken link * bg/update_change_log_lwpmiopen_501: second attempt to fix hyper link * Create placeholder CODEOWNERS (#2718) Add @JehandadKhan and @junliume as CODEOWNERS. * [Solvers] Fix for #2663 ensure tensor dimensions are consumed by solvers correctly (#2716) * [DOC] Add codeowners for documentation (#2692) * Add codeowners for documentation * Update CODEOWNERS --------- Co-authored-by: samjwu <samjwu@users.noreply.github.com> Co-authored-by: Jun Liu <Liu.Jun@amd.com> * Bump rocm-docs-core from 0.33.0 to 0.33.2 in /docs/sphinx (#2733) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.33.0 to 0.33.2. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.33.0...v0.33.2) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Fix build after #2657 and #2690 (boost::filesystem) (#2732) * [Improvements] Replace HasAtLeastOne64BitTensor() with AllTensorsDimsFitIntoInt() (#2731) * Update CK-based 2d/3d convolution solvers to support nchw/ncdhw layout (#2429) * Bump rocm-docs-core from 0.33.2 to 0.34.0 in /docs/sphinx (#2739) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.33.2 to 0.34.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.33.2...v0.34.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * [BugFix] Set System KDB journal_mode to Off (#2724) * [Tests] Converting test_conv3d_extra into GTest (#2554) * [Tests] Convert test_rnn_vanilla , test_gru, test_rnn_extra and test_gru_extra gTests (#2550) * [Doc] Removing unmaintained release notes (#2745) * [CK] Update CK commit in requirements.txt for staging (#2747) * [Tests][gtest] conversion for LSTM (#2545) * Fix for issue #2734: Detect if "-fno-offload-uniform-block" works in HIP compiler. (#2743) * fix-issue-2734 (01) Use "-fno-offload-uniform-block" only if HIP compiler supports it. Resolves #2734. (cherry picked from commit 458c833) Partially changes code from PR #2719 "Do not use HIP runtime headers on Windows" # RESOLVED Conflicts: # CMakeLists.txt * fix-issue-2734(02) Removed W/A from PR #2719 as it is no longer needed. * Enable softmax solver based on attention-softmax implementation (#2737) * [Tests] Replace test_conv_igemm_dynamic_xdlops_bwd with gtest (#2409) * [Tests] Convert ctest to gtest for test_conv_for_implicit_gemm (#2513) * [Tuning][MI300] for m9 tickets (#2754) * [hipRTC] add lowest() for float to MIOpen custom limits (#2753) * [hipRTC] add lowest() to MIOpen custom limits * the earliest trace can be found together with numeric_limits<int> * [Linux] Enhance Compiler flags to avoid Hardcoded ROCm Path (Part 1) (#2694) * Bump rocm-docs-core from 0.34.0 to 0.34.2 in /docs/sphinx (#2755) Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.34.0 to 0.34.2. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/ROCm/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](ROCm/rocm-docs-core@v0.34.0...v0.34.2) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-patch ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: Artur Wojcik <artur.wojcik@outlook.com> Co-authored-by: Artur Wojcik <artur.wojcik@amd.com> Co-authored-by: Vasilii Filippov <DrizztDoUrden@users.noreply.github.com> Co-authored-by: JD <jahandad@gmail.com> Co-authored-by: xinlipn <xinlipn@gmail.com> Co-authored-by: Jun Liu <Liu.Jun@amd.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: mentat <108366729+bghimireamd@users.noreply.github.com> Co-authored-by: Reid Kawaja <74506315+reidkwja@users.noreply.github.com> Co-authored-by: amberhassaan <amber_474@yahoo.com> Co-authored-by: Artem Tamazov <artem.tamazov@gmail.com> Co-authored-by: Daming Feng <dmfeng8898@gmail.com> Co-authored-by: Alex Eremin <CAHEK7@yandex.ru> Co-authored-by: Evgenii Averin <86725875+averinevg@users.noreply.github.com> Co-authored-by: Emmanuel Ferdman <emmanuelferdman@gmail.com> Co-authored-by: M. Saud Ul Hassan <68208941+msaudulhassan@users.noreply.github.com> Co-authored-by: carlushuang <carlus.huang@amd.com> Co-authored-by: saeid-rostami <123997133+saeid-rostami@users.noreply.github.com> Co-authored-by: Seungman Han <120356720+seungmanhan@users.noreply.github.com> Co-authored-by: Sam Wu <sam.wu2@amd.com> Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com> Co-authored-by: Umang Yadav <29876643+umangyadav@users.noreply.github.com> Co-authored-by: Dmantri98 <109552294+Dmantri98@users.noreply.github.com> Co-authored-by: abhimeda <138710508+abhimeda@users.noreply.github.com> Co-authored-by: xu-shawn <50402888+xu-shawn@users.noreply.github.com> Co-authored-by: Kamil Nasyrov <shurale.nkn@gmail.com> Co-authored-by: jasberc <146053952+jasberc@users.noreply.github.com> Co-authored-by: David Galiffi <dgaliffi@amd.com> Co-authored-by: Kyeonghwan Ryu <89056320+kyeonghwanryu@users.noreply.github.com> Co-authored-by: scerzh <102019268+scerzh@users.noreply.github.com> Co-authored-by: Vsevolod Golovko <vsevolod.golovko2@dxc.com> Co-authored-by: Jungkeun Kim <et16kr@gmail.com> Co-authored-by: samjwu <samjwu@users.noreply.github.com> Co-authored-by: Saad Rahim (AMD) <44449863+saadrahim@users.noreply.github.com> Co-authored-by: M.Emin Ozturk <ozturk.27@osu.edu> Co-authored-by: arvindcheru <90783369+arvindcheru@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Note: This PR builds on top of #2394, which itself is based on top of #2334, so the diff set appears bigger than what it is.
Enable 3D forward convolution solver with non-packed input tensors using CK kernels.
The "make check" tests passed on mi200 development machine.
PR current blocked by MIOpen CI.