diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp index b85b1c4fcf..93eecf991e 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp @@ -78,7 +78,7 @@ static const GemmImplementation<uint8_t, uint8_t, Requantize32> gemm_quint8_meth "sme2_interleaved_nomerge_u8q_mopa_1VLx4VL", [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && ((qp.per_channel_requant && (qp.per_channel_left_shifts == nullptr)) || (!qp.per_channel_requant && (qp.per_layer_left_shift == 0)));}, [](const GemmArgs &args, const Requantize32 &) { const auto VL = sme::get_vector_length<uint32_t>(); - return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); }, + return args._Nsize >= 8*VL || args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); }, [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_u8q_mopa_1VLx4VL, uint8_t, uint8_t>(args, qp); } }, { |