diff options
| author | David Mansell <David.Mansell@arm.com> | 2023-10-17 13:33:24 +0100 |
|---|---|---|
| committer | David Mansell <David.Mansell@arm.com> | 2023-10-17 13:57:48 +0000 |
| commit | 0fa92b849fd4892a341a3cda5e2ff9092093f841 (patch) | |
| tree | efb30c7e4cd79a611156a8f3b91f8fd8bfa4aa0e /src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp | |
| parent | 4ea9bac9d7d198b1bb06b42ded6aff9c9a653343 (diff) | |
| download | ComputeLibrary-0fa92b849fd4892a341a3cda5e2ff9092093f841.tar.gz | |
arm_gemm: Add SME2 FP16 GEMV using FP16->FP32 dot product.
Signed-off-by: David Mansell <David.Mansell@arm.com>
Change-Id: If02f7809f9b6e84979121698c5e7a62cbb41e2c3
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10487
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp | 9 |
1 files changed, 9 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp index efdaeeb170..3b444ae333 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp @@ -32,6 +32,7 @@ #include "gemm_hybrid_indirect.hpp" #include "gemm_implementation.hpp" #include "gemm_interleaved.hpp" +#include "gemv_pretransposed.hpp" #include "kernels/a32_sgemm_8x6.hpp" #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS @@ -42,6 +43,7 @@ #include "kernels/a64_hybrid_fp16_mla_6x32.hpp" #include "kernels/a64_sgemm_8x12.hpp" #ifdef ARM_COMPUTE_ENABLE_SME2 +#include "kernels/sme2_gemv_fp16fp32fp16_dot_16VL.hpp" #include "kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_1VLx4VL.hpp" #include "kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_2VLx2VL.hpp" #include "kernels/sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL.hpp" @@ -59,6 +61,13 @@ static const GemmImplementation<__fp16, __fp16> gemm_fp16_methods[] = { #ifdef ARM_COMPUTE_ENABLE_SVE #ifdef ARM_COMPUTE_ENABLE_SME2 { + GemmMethod::GEMM_HYBRID, + "sme2_gemv_fp16fp32fp16_dot_16VL", + [](const GemmArgs &args) { return args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input; }, + nullptr, + [](const GemmArgs &args) { return new GemvPretransposed<cls_sme2_gemv_fp16fp32fp16_dot_16VL, __fp16, __fp16>(args); } +}, +{ GemmMethod::GEMM_INTERLEAVED, "sme2_interleaved_nomerge_fp16fp32fp16_mopa_4VLx1VL", [](const GemmArgs &args) { return args._ci->has_sme2(); }, |