diff options
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_int8.cpp | 2 | ||||
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp | 2 | ||||
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp | 2 | ||||
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp | 2 |
4 files changed, 4 insertions, 4 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp index 31f225002e..f081558c40 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp @@ -105,7 +105,7 @@ static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = { GemmMethod::GEMM_INTERLEAVED, "a64_gemm_s16_8x12", nullptr, - [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Ksize>4; }, + [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s16_8x12, int8_t, int32_t>(args); }, }, { diff --git a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp index 05c5116bf3..0ffe8080f4 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp @@ -128,7 +128,7 @@ static const GemmImplementation<int8_t, int8_t, Requantize32> gemm_qint8_methods GemmMethod::GEMM_INTERLEAVED, "a64_gemm_s16_8x12", nullptr, - [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53; }, + [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); }, [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s16_8x12, int8_t, int8_t>(args, qp); } }, { diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp index 7342fda5d1..84628c0c48 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp @@ -118,7 +118,7 @@ static const GemmImplementation<uint8_t, uint8_t, Requantize32> gemm_quint8_meth GemmMethod::GEMM_INTERLEAVED, "a64_gemm_u16_8x12", nullptr, - [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53; }, + [](const GemmArgs &args, const Requantize32 &) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; }, [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_u16_8x12, uint8_t, uint8_t>(args, qp); }, }, { diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp index c300b8cdf9..7d24ea68d6 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp @@ -106,7 +106,7 @@ static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = { GemmMethod::GEMM_INTERLEAVED, "a64_gemm_u16_8x12", nullptr, - [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53; }, + [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u16_8x12, uint8_t, uint32_t>(args); }, }, { |