diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp | 230 |
1 files changed, 199 insertions, 31 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp index 73d0c272a6..321c97262f 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 Arm Limited. + * Copyright (c) 2019-2020, 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,61 +25,228 @@ #include "arm_gemm.hpp" -#include "kernels/a64_hybrid_s8s32_dot_16x4.hpp" -#include "kernels/a64_smallK_hybrid_s8s32_dot_4x6.hpp" -#include "kernels/a64_smallK_hybrid_s8s32_dot_4x8.hpp" -#include "kernels/sve_hybrid_s8s32_dot_4VLx4.hpp" -#include "kernels/sve_smallK_hybrid_s8s32_dot_1VLx8.hpp" +#include "kernels/a64_gemm_s16_8x12.hpp" +#include "kernels/a64_gemm_s8_4x4.hpp" +#include "kernels/a64_gemm_s8_8x12.hpp" +#include "kernels/a64_hybrid_s8qa_dot_4x16.hpp" +#include "kernels/a64_hybrid_s8qa_mmla_4x16.hpp" +#include "kernels/a64_hybrid_s8qs_dot_6x16.hpp" +#include "kernels/a64_hybrid_s8qs_mmla_6x16.hpp" +#include "kernels/a64_hybrid_s8s32_dot_6x16.hpp" +#include "kernels/a64_hybrid_s8s32_mmla_6x16.hpp" +#include "kernels/a64_interleaved_s8s32_mmla_8x12.hpp" +#include "kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp" +#include "kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp" +#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_SME2 +#include "kernels/sme2_gemv_s8qa_dot_16VL.hpp" +#include "kernels/sme2_interleaved_nomerge_s8q_mopa_1VLx4VL.hpp" +#include "kernels/sme2_interleaved_nomerge_s8q_mopa_2VLx2VL.hpp" +#include "kernels/sme2_interleaved_nomerge_s8q_mopa_4VLx1VL.hpp" +#endif // ARM_COMPUTE_ENABLE_SME2 + +#include "kernels/sve_hybrid_s8qa_dot_4x4VL.hpp" +#include "kernels/sve_hybrid_s8qa_mmla_4x4VL.hpp" +#include "kernels/sve_hybrid_s8qs_dot_6x4VL.hpp" +#include "kernels/sve_hybrid_s8qs_mmla_6x4VL.hpp" +#include "kernels/sve_hybrid_s8s32_dot_6x4VL.hpp" +#include "kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp" +#include "kernels/sve_interleaved_s8s32_dot_8x3VL.hpp" +#include "kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp" +#endif // ARM_COMPUTE_ENABLE_SVE + +#include "gemm_hybrid_indirect.hpp" #include "gemm_hybrid_quantized.hpp" +#include "gemm_hybrid_quantized_inline.hpp" +#include "gemm_interleaved.hpp" +#include "gemv_pretransposed.hpp" #include "quantize_wrapper.hpp" +#include "utils.hpp" namespace arm_gemm { static const GemmImplementation<int8_t, int8_t, Requantize32> gemm_qint8_methods[] = { -#ifdef __ARM_FEATURE_SVE +#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_SME2 { - GemmMethod::GEMM_HYBRID_QUANTIZED, - "smallK_hybrid_s8s32_dot_1VLx8", - [](const GemmArgs &args, const Requantize32 &) { return args._Ksize<=64 && !args._trA && args._pretransposed_hint; }, + GemmMethod::GEMV_PRETRANSPOSED, + "sme2_gemv_s8qa_dot_16VL", + [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sme2() && quant_hybrid_asymmetric(qp) && args._Msize == 1 && !args._indirect_input && args._nbatches == 1; }, nullptr, - [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_s8s32_dot_1VLx8, int8_t, int8_t>(args, qp); } + [](const GemmArgs &args, const Requantize32 &qp) { return new GemvPretransposed<cls_sme2_gemv_s8qa_dot_16VL, int8_t, int8_t, Requantize32>(args, qp); } }, { - GemmMethod::GEMM_HYBRID_QUANTIZED, - "hybrid_s8s32_dot_4VLx4", - [](const GemmArgs &args, const Requantize32 &) { return args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; }, - [](const GemmArgs &args, const Requantize32 &) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, - [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<hybrid_s8s32_dot_4VLx4, int8_t, int8_t>(args, qp); } + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_s8q_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<int32_t>(); + 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_s8q_mopa_1VLx4VL, int8_t, int8_t>(args, qp); } }, -#endif { - GemmMethod::GEMM_HYBRID_QUANTIZED, - "smallK_hybrid_s8s32_dot_4x8", - [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._trA && args._pretransposed_hint; }, + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_s8q_mopa_4VLx1VL", + [](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<int32_t>(); + return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_4VLx1VL, int8_t, int8_t>(args, qp); } +}, +{ + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_s8q_mopa_2VLx2VL", + [](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)));}, nullptr, - [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_s8s32_dot_4x8, int8_t, int8_t>(args, qp); } + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_2VLx2VL, int8_t, int8_t>(args, qp); } }, +#endif // ARM_COMPUTE_ENABLE_SME2 +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "sve_hybrid_s8qa_mmla_4x4VL", + [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_asymmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "sve_hybrid_s8qs_mmla_6x4VL", + [](const GemmArgs &args, const Requantize32 &qp) { return quant_hybrid_symmetric(qp) && args._ci->has_sve2() && args._ci->has_svei8mm(); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_interleaved_s8s32_mmla_8x3VL", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm() && (args._Ksize>8); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_hybrid_s8s32_mmla_6x4VL", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_svei8mm(); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "sve_hybrid_s8qs_dot_6x4VL", + [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_symmetric(qp); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "sve_hybrid_s8qa_dot_4x4VL", + [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_sve2() && quant_hybrid_asymmetric(qp); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "sve_hybrid_s8s32_dot_6x4VL", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve(); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_interleaved_s8s32_dot_8x3VL", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve() && (args._Ksize>4); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>(args, qp); } +), +#endif // ARM_COMPUTE_ENABLE_SVE +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "a64_hybrid_s8qa_mmla_4x16", + [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_asymmetric(qp); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "a64_hybrid_s8qs_mmla_6x16", + [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_i8mm() && quant_hybrid_symmetric(qp); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "a64_interleaved_s8s32_mmla_8x12", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm() && (args._Ksize>8); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "a64_hybrid_s8s32_mmla_6x16", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_i8mm(); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>(args, qp); } +), { GemmMethod::GEMM_HYBRID_QUANTIZED, - "smallK_hybrid_s8s32_dot_4x6", - [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._trA && args._pretransposed_hint; }, - nullptr, - [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_s8s32_dot_4x6, int8_t, int8_t>(args, qp); } + "a64_smallK_hybrid_s8s32_dot_8x4", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; }, + [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int8_t>(args, qp); } }, { GemmMethod::GEMM_HYBRID_QUANTIZED, - "hybrid_s8s32_dot_16x4", - [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; }, - [](const GemmArgs &args, const Requantize32 &) { return args._Nsize<=256 && args._Ksize>128; }, - [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<hybrid_s8s32_dot_16x4, int8_t, int8_t>(args, qp); } + "a64_smallK_hybrid_s8s32_dot_6x4", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; }, + [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int8_t>(args, qp); } }, { - GemmMethod::QUANTIZE_WRAPPER, - "quantized_wrapper", + GemmMethod::GEMM_INTERLEAVED, + "a64_gemm_s16_8x12", nullptr, + [](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); } +}, +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "a64_hybrid_s8qs_dot_6x16", + [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_symmetric(qp); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "a64_hybrid_s8qa_dot_4x16", + [](const GemmArgs &args, const Requantize32 &qp) { return args._ci->has_dotprod() && quant_hybrid_asymmetric(qp); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_HYBRID, + "a64_hybrid_s8s32_dot_6x16", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "a64_gemm_s8_8x12", + [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod(); }, + [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>(args, qp); } +), +GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "a64_gemm_s8_4x4", nullptr, + [](const GemmArgs &args, const Requantize32 &) { return GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>::estimate_cycles<int8_t>(args); }, + [](const GemmArgs &args, const Requantize32 &qp) { return new GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>(args, qp); } +), +{ + GemmMethod::QUANTIZE_WRAPPER, + "quantized_wrapper", + [](const GemmArgs &args, const Requantize32 &) { return !args._indirect_input; }, + [](const GemmArgs &, const Requantize32 &) { return false; }, [](const GemmArgs &args, const Requantize32 &qp) { return new QuantizeWrapper<int8_t, int8_t, int32_t>(args, qp); } }, { @@ -97,6 +264,7 @@ const GemmImplementation<int8_t, int8_t, Requantize32> *gemm_implementation_list } template UniqueGemmCommon<int8_t, int8_t> gemm<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os); +template bool has_opt_gemm<int8_t, int8_t, Requantize32>(WeightFormat &weight_format, const GemmArgs &args, const Requantize32 &os); template KernelDescription get_gemm_method<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os); template std::vector<KernelDescription> get_compatible_kernels<int8_t, int8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os); |