diff options
| author | SiCong Li <sicong.li@arm.com> | 2020-08-03 15:39:45 +0100 |
|---|---|---|
| committer | SiCong Li <sicong.li@arm.com> | 2020-08-06 15:38:22 +0000 |
| commit | b972ae62dd877eb53e6ad56ee124cfbc89441e2d (patch) | |
| tree | faaf38a072dd4c1c277567772f3106ad39dd9f3a /src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp | |
| parent | 187558fdacc457887d21b61f50425cd55187722a (diff) | |
| download | ComputeLibrary-b972ae62dd877eb53e6ad56ee124cfbc89441e2d.tar.gz | |
COMPMID-3652 Fix CLFullyConnectedLayer failure on S10
* Fix out-of-bound mem reads in cases where M < M0 in
CLGEMMMatrixMultiplyNativeKernel and
CLGEMMMatrixMultiplyReshapedOnlyRHSKernel, as a result of the new
boundary-aware reading logics.
* Add fixture tests (alongside the padding configuration tests) in
these 2 kernels to catch all 5 possible scenarios with block dimension
configurations, which includes this particular bug as the
"...BoundaryHandlingFullInXSinglePartialInY" test case
Change-Id: I8a10ab67594171e3ea4fb6e35c84ddc4ab964fba
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3650
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp | 6 |
1 files changed, 5 insertions, 1 deletions
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp index 27520c6072..e65726b234 100644 --- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp @@ -263,6 +263,10 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext const unsigned int partial_store_m0 = internal_m % lhs_info.m0; const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0; + // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads. + // NOTE: This might have implications on heuristics and performance + const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0); + // Create build options CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); @@ -282,7 +286,7 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m)); build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n)); build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k)); - build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); + build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0)); build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0)); |