Skip to content

Commit 918e8a1

Browse files
brockhargreaves-amdassistant-librarian[bot]
brockhargreaves-amd
authored and
assistant-librarian[bot]
committed
[rocm-libraries] ROCm/rocm-libraries#6400 (commit c0b3c95)
[MIOPEN] [CK] Revert "[CK] Disable test cases affected by compiler codegen bugs on gfx90a" (#6400) Reverts ROCm/rocm-libraries#6343 This is causing failures in miopen, namely Dbsync gfx942 even though it shouldn't be affected so this needs to be investigated. Please add miopen as a label to the new PR for addressing the compiler codegen bug so that this can be addressed simultaneously.
1 parent a92fd0d commit 918e8a1

4 files changed

Lines changed: 0 additions & 42 deletions

File tree

‎CMakeLists.txt‎

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -274,11 +274,6 @@ if (SUPPORTED_GPU_TARGETS MATCHES "gfx950" AND NOT FORCE_DISABLE_XDL)
274274
add_definitions(-DCK_USE_GFX950)
275275
set(CK_USE_GFX950 "ON")
276276
endif()
277-
if (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" AND NOT FORCE_DISABLE_XDL)
278-
add_definitions(-DCK_USE_GFX90A)
279-
set(CK_USE_GFX90A "ON")
280-
endif()
281-
282277

283278
# new macro CK_TILE_USE_WMMA in order to separately compile examples for MFMA/WMMA
284279
set(CK_TILE_USE_WMMA 0)

‎include/ck_tile/core/config.hpp‎

Lines changed: 0 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -209,17 +209,6 @@
209209
#endif
210210
#endif
211211

212-
// workaround for AMDGPU compiler VGPR aliasing bug in dropout codegen (ROCm >= 7.12)
213-
// Philox RNG VGPR parameters get aliased under high register pressure (d256 tile).
214-
// fp16 is affected; bf16 is not (different type conversion codegen path).
215-
#ifndef CK_TILE_WORKAROUND_ROCM_7_12_FP16_DROPOUT_MISCOMPILE
216-
#if(HIP_VERSION_MAJOR == 7 && HIP_VERSION_MINOR >= 12) || (HIP_VERSION_MAJOR > 7)
217-
#define CK_TILE_WORKAROUND_ROCM_7_12_FP16_DROPOUT_MISCOMPILE 1
218-
#else
219-
#define CK_TILE_WORKAROUND_ROCM_7_12_FP16_DROPOUT_MISCOMPILE 0
220-
#endif
221-
#endif
222-
223212
#ifndef CK_TILE_DEBUG_LOG
224213
#define CK_TILE_DEBUG_LOG 0
225214
#endif

‎library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp‎

Lines changed: 0 additions & 18 deletions
Original file line numberDiff line numberDiff line change
@@ -95,16 +95,7 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f16_instances = std::tuple
9595
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
9696
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
9797
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
98-
99-
// Problematic instance on gfx90a - accuracy tests fail for 3D bwd weight conv as the instance produces incorrect results.
100-
// The problem occurs at least for compiler version
101-
// 22.0.0git (https://github.com/ROCm/llvm-project.git 2de9eb6063dd56b109cf139a75550b7b06808273+PATCHED:9a6ac45c97a1e511db838c5b46257324d2de1780)
102-
// Older compilers from the 20.0 family produce correct results.
103-
#if defined(CK_USE_GFX90A)
104-
#else
10598
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
106-
#endif
107-
10899
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 5, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 5, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
109100
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 7, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 7, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>
110101
// clang-format on
@@ -177,16 +168,7 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_bf16_instances = std::tupl
177168
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
178169
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
179170
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
180-
181-
// Problematic instance on gfx90a - accuracy tests fail for 3D bwd weight conv as the instance produces incorrect results.
182-
// The problem occurs at least for compiler version
183-
// 22.0.0git (https://github.com/ROCm/llvm-project.git 2de9eb6063dd56b109cf139a75550b7b06808273+PATCHED:9a6ac45c97a1e511db838c5b46257324d2de1780)
184-
// Older compilers from the 20.0 family produce correct results.
185-
#if defined(CK_USE_GFX90A)
186-
#else
187171
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
188-
#endif
189-
190172
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 5, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 5, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
191173
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 7, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 7, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>
192174
//clang-format on

‎test/ck_tile/fmha/test_fmha_fwd.cpp‎

Lines changed: 0 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -601,14 +601,6 @@ TEST_P(Dropout, DataTypeConfig)
601601
auto [drop_seed, drop_offset, drop_prefs] = drop_seed_offset_prefs;
602602
auto [batch, nhead, nhead_k, seqlen_q, seqlen_k, mask_str] = dims_mask;
603603

604-
#if CK_TILE_WORKAROUND_ROCM_7_12_FP16_DROPOUT_MISCOMPILE
605-
if constexpr(std::is_same_v<DataTypeConfig, FmhaFwdFp16>)
606-
{
607-
if(hdim_q > 128 && mode == mode_enum::batch)
608-
GTEST_SKIP() << "Skipped: fp16 dropout d256 batch — compiler bug (ROCm >= 7.12)";
609-
}
610-
#endif
611-
612604
auto result = fmha_fwd_run<DataTypeConfig>(mode,
613605
batch,
614606
nhead,

0 commit comments

Comments
 (0)