diff options
| author | Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com> | 2023-06-19 14:57:57 +0100 |
|---|---|---|
| committer | Mohmun02 <MohammedSuhail.Munshi@arm.com> | 2023-06-26 11:34:03 +0000 |
| commit | a2bb80ea7111509c24caad8629533089decef430 (patch) | |
| tree | f674572e0cc705af9b66633bfcd9d6ad9e29d970 /src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp | |
| parent | c952596e70f2fe0073029f053e329a4e930ced8c (diff) | |
| download | ComputeLibrary-a2bb80ea7111509c24caad8629533089decef430.tar.gz | |
Use MatMul in fully connected layer with dynamic weights when supported
- Use MatMul kernels in FC layer when using dynamic weights without broadcasting or bias.
- Fix minor typo in IClMatMulNativeKernelConfig.h
Partially Resolves : [COMPMID-6193]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: Id494062b5b4f4e75ff9714c202dde941955afa52
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9797
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp')
| -rw-r--r-- | src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp | 7 |
1 files changed, 4 insertions, 3 deletions
diff --git a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp index 9bbec908a3..38d78c618b 100644 --- a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp +++ b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp @@ -164,9 +164,10 @@ void ClMatMulLowpNativeKernel::configure(const ClCompileContext &compile_context build_opts.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier)); build_opts.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift)); - build_opts.add_option("-DLHS_OFFSET=" + support::cpp11::to_string(-lqinfo.offset)); // Note this is passed as negative to maintain similarity with CLDirectConv2D - build_opts.add_option("-DRHS_OFFSET=" + support::cpp11::to_string(-rqinfo.offset)); // Note this is passed as negative to maintain similarity with CLDirectConv2D - build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(dqinfo.offset)); // Passed as positive (unlike the above two) + // Note : Offset is not negated, unlike gemmlowp kernels + build_opts.add_option("-DLHS_OFFSET=" + support::cpp11::to_string(lqinfo.offset)); + build_opts.add_option("-DRHS_OFFSET=" + support::cpp11::to_string(rqinfo.offset)); + build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(dqinfo.offset)); // Passed as positive (unlike the above two) build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(act_info.a()))); build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(act_info.b()))); |