/******************************************************************************* * Copyright 2019 Intel Corporation * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. *******************************************************************************/ #ifndef CPU_JIT_AVX512_CORE_BF16_CONVOLUTION_HPP #define CPU_JIT_AVX512_CORE_BF16_CONVOLUTION_HPP #include "c_types_map.hpp" #include "memory_tracking.hpp" #include "mkldnn_thread.hpp" #include "utils.hpp" #include "cpu_barrier.hpp" #include "cpu_convolution_pd.hpp" #include "cpu_reducer.hpp" #include "jit_transpose_src_utils.hpp" #include "jit_avx512_core_bf16_conv_kernel.hpp" #include "bfloat16_utils.hpp" namespace mkldnn { namespace impl { namespace cpu { template struct _jit_avx512_core_bf16_convolution_fwd_t : public cpu_primitive_t { struct pd_t : public cpu_convolution_fwd_pd_t { pd_t(engine_t *engine, const convolution_desc_t *adesc, const primitive_attr_t *attr, const typename pd_t::base_class *hint_fwd_pd) : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) , jcp_() { } DECLARE_COMMON_PD_T( JIT_IMPL_NAME_HELPER("jit_bf16:", avx512_core, ""), _jit_avx512_core_bf16_convolution_fwd_t); virtual status_t init() override { using namespace prop_kind; assert(this->engine()->kind() == engine_kind::cpu); bool ok = true && mayiuse(avx512_core) && utils::one_of(this->desc()->prop_kind, forward_training, forward_inference) && utils::one_of(this->desc()->alg_kind, alg_kind::convolution_auto, alg_kind::convolution_direct) && !this->has_zero_dim_memory() && this->desc()->src_desc.data_type == data_type::bf16 && this->desc()->weights_desc.data_type == data_type::bf16 && this->desc()->dst_desc.data_type == dst_type && IMPLICATION(this->with_bias(), utils::one_of( this->desc()->bias_desc.data_type, data_type::f32, data_type::bf16)); if (!ok) return status::unimplemented; status_t status = jit_avx512_core_bf16_fwd_kernel::init_conf( jcp_, *this->desc(), this->src_pd_, this->weights_pd_, this->dst_pd_, this->bias_pd_, *this->attr(), mkldnn_get_max_threads()); if (status != status::success) return status; if (status == status::success && this->desc()->alg_kind == alg_kind::convolution_auto) CHECK(this->set_alg_kind(alg_kind::convolution_direct)); init_scratchpad(); return status::success; } inline int ndims() const { return this->desc()->src_desc.ndims; } jit_conv_conf_t jcp_; private: void init_scratchpad() { using namespace memory_tracking::names; auto scratchpad = scratchpad_registry().registrar(); if (jcp_.with_bias && jcp_.oc != jcp_.oc_without_padding) scratchpad.book(key_conv_padded_bias, jcp_.typesize_bia * jcp_.oc); } }; _jit_avx512_core_bf16_convolution_fwd_t(const pd_t *apd, const input_vector &inputs, const output_vector &outputs) : cpu_primitive_t(apd, inputs, outputs) { kernel_ = new jit_avx512_core_bf16_fwd_kernel(pd()->jcp_, *pd()->attr()); } ~_jit_avx512_core_bf16_convolution_fwd_t() { delete kernel_;} typedef typename prec_traits::type src_data_t; typedef typename prec_traits::type wei_data_t; typedef typename prec_traits::type dst_data_t; virtual void execute(event_t *e) const { if (pd()->ndims() == 3) execute_forward_1d(); else if (pd()->ndims() == 4) execute_forward_2d(); else if (pd()->ndims() == 5) execute_forward_3d(); else assert(false); /*TODO: zero pad dst */ e->set_state(event_t::ready); } private: void prepare_padded_bias(const char *&bias) const; void execute_forward_1d() const; void execute_forward_2d() const; void execute_forward_3d() const; jit_avx512_core_bf16_fwd_kernel *kernel_; const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); } }; template using jit_avx512_core_bf16_convolution_fwd_t = _jit_avx512_core_bf16_convolution_fwd_t; template struct _jit_avx512_core_bf16_convolution_bwd_data_t: public cpu_primitive_t { struct pd_t: public cpu_convolution_bwd_data_pd_t { pd_t(engine_t *engine, const convolution_desc_t *adesc, const primitive_attr_t *attr, const convolution_fwd_pd_t *hint_fwd_pd) : cpu_convolution_bwd_data_pd_t(engine, adesc, attr, hint_fwd_pd) , jcp_() {} DECLARE_COMMON_PD_T( JIT_IMPL_NAME_HELPER("jit_bf16:", avx512_core, ""), _jit_avx512_core_bf16_convolution_bwd_data_t); virtual status_t init() override { using namespace prop_kind; assert(this->engine()->kind() == engine_kind::cpu); bool ok = true && mayiuse(avx512_core) && utils::one_of(this->desc()->prop_kind, backward_data) && utils::one_of(this->desc()->alg_kind, alg_kind::convolution_auto, alg_kind::convolution_direct) && this->desc()->alg_kind == alg_kind::convolution_direct && this->desc()->diff_dst_desc.data_type == data_type::bf16 && this->desc()->weights_desc.data_type == data_type::bf16 && this->desc()->diff_src_desc.data_type == diff_src_type && this->set_default_params() == status::success && !this->has_zero_dim_memory(); if (!ok) return status::unimplemented; status_t status = jit_avx512_core_bf16_bwd_data_kernel::init_conf( jcp_, *this->desc(), *this->diff_src_pd_.desc(), *this->weights_pd_.desc(), *this->diff_dst_pd_.desc()); if (status != status::success) return status; if (status == status::success && this->desc()->alg_kind == alg_kind::convolution_auto) CHECK(this->set_alg_kind(alg_kind::convolution_direct)); return status::success; } jit_conv_conf_t jcp_; protected: memory_format_t src_format() { using namespace memory_format; return utils::pick(ndims() - 3, nCw16c, nChw16c, nCdhw16c); } memory_format_t wei_format() { using namespace memory_format; return this->with_groups() ? utils::pick(ndims() - 3, gOIw8o16i2o, gOIhw8o16i2o, gOIdhw8o16i2o) : utils::pick(ndims() - 3, OIw8o16i2o, OIhw8o16i2o, OIdhw8o16i2o); } virtual status_t set_default_params() override { using namespace memory_format; if (this->diff_src_pd_.desc()->format == any) CHECK(this->diff_src_pd_.set_format(src_format())); if (this->diff_dst_pd_.desc()->format == any) CHECK(this->diff_dst_pd_.set_format(src_format())); if (this->weights_pd_.desc()->format == any) CHECK(this->weights_pd_.set_format(wei_format())); return status::success; } }; _jit_avx512_core_bf16_convolution_bwd_data_t(const pd_t *apd, const input_vector &inputs, const output_vector &outputs) : cpu_primitive_t(apd, inputs, outputs) { kernel_ = new jit_avx512_core_bf16_bwd_data_kernel(pd()->jcp_); } ~_jit_avx512_core_bf16_convolution_bwd_data_t() { delete kernel_; }; typedef typename prec_traits::type diff_dst_data_t; typedef typename prec_traits::type wei_data_t; typedef typename prec_traits::type diff_src_data_t; virtual void execute(event_t *e) const { if (pd()->ndims() < 5) execute_backward_data(); else if (pd()->ndims() == 5) execute_backward_data_3d(); else assert(!"invalid dimension"); e->set_state(event_t::ready); } private: void execute_backward_data() const; void execute_backward_data_3d() const; const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); } jit_avx512_core_bf16_bwd_data_kernel *kernel_; }; template using jit_avx512_core_bf16_convolution_bwd_data_t = _jit_avx512_core_bf16_convolution_bwd_data_t; template struct _jit_avx512_core_bf16_convolution_bwd_weights_t: public cpu_primitive_t { struct pd_t: public cpu_convolution_bwd_weights_pd_t { pd_t(engine_t *engine, const convolution_desc_t *adesc, const primitive_attr_t *attr, const convolution_fwd_pd_t *hint_fwd_pd) : cpu_convolution_bwd_weights_pd_t(engine, adesc, attr, hint_fwd_pd) , jcp_() {} DECLARE_COMMON_PD_T( JIT_IMPL_NAME_HELPER("jit_bf16:", avx512_core, ""), _jit_avx512_core_bf16_convolution_bwd_weights_t); virtual status_t init() override { assert(this->engine()->kind() == engine_kind::cpu); bool ok = true && mayiuse(avx512_core) && this->desc()->prop_kind == prop_kind::backward_weights && this->desc()->alg_kind == alg_kind::convolution_direct && !this->has_zero_dim_memory() && this->desc()->src_desc.data_type == data_type::bf16 && this->desc()->diff_dst_desc.data_type == data_type::bf16 && this->desc()->diff_weights_desc.data_type == diff_weights_type && IMPLICATION(this->with_bias(), utils::one_of( this->desc()->diff_bias_desc.data_type, data_type::f32, data_type::bf16)); if (!ok) return status::unimplemented; status_t status = jit_avx512_core_bf16_conv_bwd_weights_kernel_f32::init_conf(jcp_, *this->desc(), this->src_pd_, this->diff_weights_pd_, this->diff_bias_pd_, this->diff_dst_pd_); if (status != status::success) return status; init_balancers(); auto scratchpad = scratchpad_registry().registrar(); jit_avx512_core_bf16_conv_bwd_weights_kernel_f32::init_scratchpad( scratchpad, jcp_); auto reducer_bia_scratchpad = memory_tracking::registrar_t( scratchpad, memory_tracking::names::prefix_reducer_bia); reducer_bia_conf_.init_scratchpad(reducer_bia_scratchpad); return status; } jit_conv_conf_t jcp_; typename cpu_reducer_t::conf_t reducer_bia_conf_; private: void init_balancers() { const size_t max_buffer_size = jcp_.nthr * 3 * 5 * 5 * 16 * 16; if (with_bias()) { reducer_bia_conf_.init(reduce_balancer_t(jcp_.nthr, jcp_.oc_block, jcp_.ngroups * jcp_.nb_oc, jcp_.mb, max_buffer_size)); } } }; _jit_avx512_core_bf16_convolution_bwd_weights_t(const pd_t *pd, const input_vector &inputs, const output_vector &outputs); ~_jit_avx512_core_bf16_convolution_bwd_weights_t() { delete kernel_; #ifndef BF16_CONV_BWD_W_JIT_KER_USES_PERMW_TRANSPOSITION delete trans_kernel_; delete trans_dst_kernel_; #endif delete acc_ker_; delete reducer_bias_; } typedef typename prec_traits::type src_data_t; typedef typename prec_traits::type diff_dst_data_t; typedef typename prec_traits::type diff_weights_data_t; virtual void execute(event_t *e) const { execute_backward_weights(); e->set_state(event_t::ready); } private: struct thread_info_t; void execute_backward_weights() const; void prepare_scratchpad_data() const; void compute_diff_weights(const thread_info_t *) const; void reduce_and_convert_diff_weights(const thread_info_t *) const; void compute_diff_bias(const thread_info_t *) const; const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); } int nthr_, nthr_mb_, nthr_g_, nthr_oc_b_, nthr_ic_b_; jit_avx512_core_bf16_conv_bwd_weights_kernel_f32 *kernel_; cpu_accumulator_1d_t *acc_ker_; cpu_reducer_t *reducer_bias_; #ifndef BF16_CONV_BWD_W_JIT_KER_USES_PERMW_TRANSPOSITION jit_trans_src_t *trans_kernel_; jit_trans_dst_t *trans_dst_kernel_; #endif }; template using jit_avx512_core_bf16_convolution_bwd_weights_t = _jit_avx512_core_bf16_convolution_bwd_weights_t; } } } #endif // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s