| 228 | } |
| 229 | |
| 230 | SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoIm2col::dispatch_kerns( |
| 231 | const NCBKernSizeParam& param) const { |
| 232 | MIDOUT_BEGIN(megdnn_fallback_im2col, 0, 1) { |
| 233 | size_t OH = param.osz[0]; |
| 234 | size_t OW = param.osz[1]; |
| 235 | size_t OC = param.filter_meta.ocpg; |
| 236 | size_t ohw = OH * OW; |
| 237 | size_t oc_tile_size = 0, ohw_tile_size = 0; |
| 238 | |
| 239 | auto matmul_desc = m_matmul_algo->matmul_description(); |
| 240 | |
| 241 | bool default_pack = matmul_desc.packmode == Pack_Mode::DEFAULT; |
| 242 | bool no_pack = matmul_desc.packmode == Pack_Mode::NO_PACK; |
| 243 | bool only_packA = matmul_desc.packmode == Pack_Mode::ONLY_PACKA; |
| 244 | bool enable_filter_preprocess = is_enable_filter_preprocess(param); |
| 245 | choice_ohw_oc_block( |
| 246 | param, oc_tile_size, ohw_tile_size, matmul_desc.innerblocksize.m, |
| 247 | matmul_desc.innerblocksize.n, m_ohw_tile_size, matmul_desc.packmode); |
| 248 | |
| 249 | size_t packa_parallel_times = 0; |
| 250 | size_t pack_oc_size = pack_size(param.filter_meta.format); |
| 251 | if (only_packA) { |
| 252 | packa_parallel_times = div_ceil<size_t>(OC, oc_tile_size); |
| 253 | } else if (default_pack) { |
| 254 | packa_parallel_times = div_ceil<size_t>(OC, matmul_desc.innerblocksize.m); |
| 255 | } |
| 256 | |
| 257 | auto matmul_param = get_matmul_kern_param( |
| 258 | param, ohw_tile_size, default_pack ? OC : oc_tile_size); |
| 259 | |
| 260 | WorkspaceBundle bundle = |
| 261 | get_bundle(param, m_matmul_algo, oc_tile_size, ohw_tile_size); |
| 262 | WorkspaceBundle bundle_thread = get_thread_bundle( |
| 263 | param, m_matmul_algo, matmul_param, matmul_desc, oc_tile_size, |
| 264 | ohw_tile_size); |
| 265 | |
| 266 | StrategyParam strategyparam; |
| 267 | strategyparam.ohw = ohw; |
| 268 | strategyparam.is_dst_8bit = |
| 269 | (param.src_type.enumv() == DTypeEnum::QuantizedS8 && |
| 270 | param.dst_type.enumv() == DTypeEnum::QuantizedS8) || |
| 271 | (param.src_type.enumv() == DTypeEnum::Quantized8Asymm && |
| 272 | param.dst_type.enumv() == DTypeEnum::Quantized8Asymm); |
| 273 | strategyparam.is_ohw_size_bigger = (ohw_tile_size >= ohw); |
| 274 | strategyparam.skip_copy_dst = |
| 275 | strategyparam.is_ohw_size_bigger && !strategyparam.is_dst_8bit; |
| 276 | strategyparam.oc_tile_size = oc_tile_size; |
| 277 | strategyparam.pack_oc_size = pack_oc_size; |
| 278 | strategyparam.enable_filter_preprocess = enable_filter_preprocess; |
| 279 | strategyparam.packA_group_size = packA_group_size( |
| 280 | m_matmul_algo, matmul_param, matmul_desc, packa_parallel_times); |
| 281 | |
| 282 | SmallVector<ConvBiasImpl::NCBKern> ret_kern; |
| 283 | StrategyBase* im2colstrategy = |
| 284 | Factory::get_im2col_strategy(param, m_matmul_algo); |
| 285 | if (default_pack) { |
| 286 | MIDOUT_BEGIN( |
| 287 | megdnn_fallback_im2col, |
nothing calls this directly
no test coverage detected