[Tuning] Remove Solver ConvHipImplicitGemmBwdDataV4R1Xdlops FP32 #1209
Conversation
|
@cderb Thanks! Let me try to target 'Fp16 Hip All Install gfx908" stage and see if it passed there. |
Platform-specific tests all passed |
| @@ -30,10 +30,6 @@ | |||
| #include <miopen/solver/implicitgemm_util.hpp> | |||
| #include <cstddef> | |||
|
|
|||
| /// Disable ConvHipImplicitGemmBwdDataV4R1Xdlops by default. | |||
There was a problem hiding this comment.
This W/A must be not removed but narrowed to FP32 in this PR at least. Otherwise the next tuning round will bring the problem back. And I am not sure that ConvHipImplicitGemmBwdDataV4R1Xdlops is good for FP16.
test_regression_half_mi100 has failed because the failing test case is intended for ConvHipImplicitGemmBwdDataV4R1Xdlops specifically. This test should have been disabled in #1207 (and in #1208) under the WORKAROUND_ISSUE_1206
There was a problem hiding this comment.
Another reason for keeping the W/A is that the end user can activate the Normal Find mode.
There was a problem hiding this comment.
The fdb update is good as it recovers some performance.
|
This PR wasn't squashed, and git history does not show |
Removes FP32 ConvHipImplicitGemmBwdDataV4R1Xdlops from find_db replacing with second best Backward Data Implicit GEMM solver.