41 #ifdef ARM_COMPUTE_ENABLE_SVE
42 #ifdef ARM_COMPUTE_ENABLE_SME2
47 #endif // ARM_COMPUTE_ENABLE_SME2
57 #endif // ARM_COMPUTE_ENABLE_SVE
69 static const GemmImplementation<int8_t, int8_t, Requantize32> gemm_qint8_methods[] =
71 #ifdef ARM_COMPUTE_ENABLE_SVE
72 #ifdef ARM_COMPUTE_ENABLE_SME2
75 "sme2_gemv_s8qa_dot_16VL",
78 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemvPretransposed<cls_sme2_gemv_s8qa_dot_16VL, int8_t, int8_t, Requantize32>(
args, qp); }
82 "sme2_interleaved_nomerge_s8q_mopa_1VLx4VL",
83 [](
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)));},
84 [](
const GemmArgs &
args,
const Requantize32 &) {
const auto VL = sme::get_vector_length<int32_t>();
85 return args._Msize <= VL || (2*VL <
args._Msize &&
args._Msize <= 3*VL); },
86 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_1VLx4VL, int8_t, int8_t>(
args, qp); }
90 "sme2_interleaved_nomerge_s8q_mopa_4VLx1VL",
91 [](
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)));},
92 [](
const GemmArgs &
args,
const Requantize32 &) {
const auto VL = sme::get_vector_length<int32_t>();
93 return args._Nsize <= VL || (2*VL <
args._Nsize &&
args._Nsize <= 3*VL); },
94 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_4VLx1VL, int8_t, int8_t>(
args, qp); }
98 "sme2_interleaved_nomerge_s8q_mopa_2VLx2VL",
99 [](
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)));},
101 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedPretransposedNoMergeQuantizedInline<cls_sme2_interleaved_nomerge_s8q_mopa_2VLx2VL, int8_t, int8_t>(
args, qp); }
103 #endif // ARM_COMPUTE_ENABLE_SME2
106 "sve_hybrid_s8qa_mmla_4x4VL",
108 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
109 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_sve_hybrid_s8qa_mmla_4x4VL, int8_t, int8_t, Requantize32>(
args, qp); }
113 "sve_hybrid_s8qs_mmla_6x4VL",
115 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
116 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_sve_hybrid_s8qs_mmla_6x4VL, int8_t, int8_t, Requantize32>(
args, qp); }
120 "sve_interleaved_s8s32_mmla_8x3VL",
121 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_svei8mm() && (
args._Ksize>8); },
122 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(
args); },
123 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int8_t>(
args, qp); }
127 "sve_hybrid_s8s32_mmla_6x4VL",
128 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_svei8mm(); },
129 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(
args); },
130 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int8_t, Requantize32, true>(
args, qp); }
134 "sve_hybrid_s8qs_dot_6x4VL",
136 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
137 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_sve_hybrid_s8qs_dot_6x4VL, int8_t, int8_t, Requantize32>(
args, qp); }
141 "sve_hybrid_s8qa_dot_4x4VL",
143 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
144 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_sve_hybrid_s8qa_dot_4x4VL, int8_t, int8_t, Requantize32>(
args, qp); }
148 "sve_hybrid_s8s32_dot_6x4VL",
149 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_sve(); },
150 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(
args); },
151 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int8_t, Requantize32, true>(
args, qp); }
155 "sve_interleaved_s8s32_dot_8x3VL",
156 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_sve() && (
args._Ksize>4); },
157 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>::estimate_cycles<int8_t>(
args); },
158 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedQuantized<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int8_t>(
args, qp); }
163 "a64_hybrid_s8qa_mmla_4x16",
165 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
166 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_a64_hybrid_s8qa_mmla_4x16, int8_t, int8_t, Requantize32>(
args, qp); }
170 "a64_hybrid_s8qs_mmla_6x16",
172 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
173 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_a64_hybrid_s8qs_mmla_6x16, int8_t, int8_t, Requantize32>(
args, qp); }
177 "a64_interleaved_s8s32_mmla_8x12",
178 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_i8mm() && (
args._Ksize>8); },
179 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(
args); },
180 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedQuantized<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int8_t>(
args, qp); }
184 "a64_hybrid_s8s32_mmla_6x16",
185 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_i8mm(); },
186 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(
args); },
187 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int8_t, Requantize32, true>(
args, qp); }
191 "a64_smallK_hybrid_s8s32_dot_8x4",
192 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_dotprod() && (
args._Nsize % 4 == 0) && (
args._Ksize<=32) && !
args._indirect_input; },
193 [](
const GemmArgs &
args,
const Requantize32 &) {
return !(
args._ci->has_svei8mm() ||
args._ci->has_i8mm()); },
194 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int8_t>(
args, qp); }
198 "a64_smallK_hybrid_s8s32_dot_6x4",
199 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_dotprod() && (
args._Nsize % 4 == 0) && (
args._Ksize>32) && (
args._Ksize<=64) && !
args._indirect_input; },
200 [](
const GemmArgs &
args,
const Requantize32 &) {
return !(
args._ci->has_svei8mm() ||
args._ci->has_i8mm()); },
201 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int8_t>(
args, qp); }
207 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->get_cpu_model() == CPUModel::A53 && ((
args._Msize > 28) || ((
args._Msize % 8) > 4)); },
208 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedQuantized<cls_a64_gemm_s16_8x12, int8_t, int8_t>(
args, qp); }
212 "a64_hybrid_s8qs_dot_6x16",
214 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
215 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_a64_hybrid_s8qs_dot_6x16, int8_t, int8_t, Requantize32>(
args, qp); }
219 "a64_hybrid_s8qa_dot_4x16",
221 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>::estimate_cycles<int8_t>(
args); },
222 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_a64_hybrid_s8qa_dot_4x16, int8_t, int8_t, Requantize32>(
args, qp); }
226 "a64_hybrid_s8s32_dot_6x16",
227 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_dotprod(); },
228 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>::estimate_cycles<int8_t>(
args); },
229 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int8_t, Requantize32, true>(
args, qp); }
234 [](
const GemmArgs &
args,
const Requantize32 &) {
return args._ci->has_dotprod(); },
235 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>::estimate_cycles<int8_t>(
args); },
236 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedQuantized<cls_a64_gemm_s8_8x12, int8_t, int8_t>(
args, qp); }
242 [](
const GemmArgs &
args,
const Requantize32 &) {
return GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>::estimate_cycles<int8_t>(
args); },
243 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new GemmInterleavedQuantized<cls_a64_gemm_s8_4x4, int8_t, int8_t>(
args, qp); }
248 [](
const GemmArgs &
args,
const Requantize32 &) {
return !
args._indirect_input; },
249 [](
const GemmArgs &,
const Requantize32 &) {
return false; },
250 [](
const GemmArgs &
args,
const Requantize32 &qp) {
return new QuantizeWrapper<int8_t, int8_t, int32_t>(
args, qp); }
262 const GemmImplementation<int8_t, int8_t, Requantize32> *gemm_implementation_list<int8_t, int8_t, Requantize32>() {
263 return gemm_qint8_methods;
266 template UniqueGemmCommon<int8_t, int8_t> gemm<int8_t, int8_t, Requantize32>(
const GemmArgs &
args,
const Requantize32 &os);
267 template bool has_opt_gemm<int8_t, int8_t, Requantize32>(
WeightFormat &weight_format,
const GemmArgs &
args,
const Requantize32 &os);
268 template KernelDescription get_gemm_method<int8_t, int8_t, Requantize32>(
const GemmArgs &
args,
const Requantize32 &os);
269 template std::vector<KernelDescription> get_compatible_kernels<int8_t, int8_t, Requantize32>(
const GemmArgs &
args,
const Requantize32 &os);
273 #endif // __aarch64__