/******************************************************************************* * Copyright 2017-2018 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_COMMON_CONVOLUTION_WINOGRAD_HPP #define CPU_JIT_AVX512_COMMON_CONVOLUTION_WINOGRAD_HPP #include "c_types_map.hpp" #include "cpu_convolution_pd.hpp" #include "cpu_engine.hpp" #include "scratchpad.hpp" #include "mkldnn_thread.hpp" #include "jit_avx512_common_conv_winograd_kernel_f32.hpp" namespace mkldnn { namespace impl { namespace cpu { namespace winograd { struct winograd_scratchpad_t { public: winograd_scratchpad_t(const jit_conv_winograd_conf_t &jcp) { get_scratchpad_size_(jcp); allocate_scratchpad_(jcp); } ~winograd_scratchpad_t() { if (scratchpad_ != nullptr) delete scratchpad_; } char *U_ptr() { /* buffer for wei transform U*/ return scratchpad_->get() + U_offset_; } char *V_ptr() { /* buffer for src transform V*/ return scratchpad_->get() + V_offset_; } char *M_ptr() { /* buffer for dst transform M*/ return scratchpad_->get() + M_offset_; } char *bias_ptr() { /* buffer for bias update in bwdw*/ return scratchpad_->get() + bias_offset_; } char *src_transpose_ptr() { /* buffer for src transpose in bwdw using qfma*/ return scratchpad_->get() + src_transpose_offset_; } int num_threads(){ return nthreads_; } private: inline void get_scratchpad_size_(const jit_conv_winograd_conf_t &jcp) { nthreads_ = mkldnn_get_max_threads(); U_sz_ = (size_t)alpha * alpha * jcp.ic * jcp.oc * sizeof(float); V_sz_ = (size_t)alpha * alpha * jcp.mb * jcp.ic * (jcp.itiles * jcp.jtiles + jcp.tile_4fma_padding) * sizeof(float); M_sz_ = (size_t)alpha * alpha * jcp.mb * jcp.oc * (jcp.itiles * jcp.jtiles + jcp.tile_4fma_padding) * sizeof(float); switch (jcp.sched_policy) { case WSCHED_DATA_W_SGD: V_sz_ = (size_t)nthreads_ * alpha * alpha * jcp.nb_tile_block_ur * jcp.tile_block_ur * jcp.ic * sizeof(float); M_sz_ = (size_t)nthreads_* alpha * alpha * jcp.nb_tile_block_ur * jcp.tile_block_ur * jcp.oc * sizeof(float); break; case WSCHED_WEI_SDGt_W: U_sz_ = (size_t)nthreads_ * U_sz_; V_sz_ = (size_t)nthreads_ * alpha * alpha * (jcp.nb_tile_block_ur * jcp.tile_block_ur + jcp.tile_4fma_padding) * jcp.ic * sizeof(float); M_sz_ = (size_t)nthreads_ * alpha * alpha * (jcp.nb_tile_block_ur * jcp.tile_block_ur + jcp.tile_4fma_padding) * jcp.oc * sizeof(float); bias_sz_ = nthreads_ * jcp.oc * sizeof(float); break; case WSCHED_WEI_SDGtWo: U_sz_ = (size_t)nthreads_ * alpha * alpha * jcp.oc_block * jcp.oc_simd_block * jcp.ic * sizeof(float); M_sz_ = (size_t)nthreads_ * alpha * alpha * (jcp.nb_tile_block_ur * jcp.tile_block_ur + jcp.tile_4fma_padding) * jcp.oc_simd_block * jcp.oc_block * sizeof(float); bias_sz_ = nthreads_ * jcp.oc * sizeof(float); break; case WSCHED_WEI_S_D_Giot_W: U_sz_ = (size_t)(nthreads_ + 1) * alpha * alpha * jcp.ic * jcp.oc * sizeof(float); V_sz_ = (size_t)alpha * alpha * (jcp.itiles * jcp.jtiles + jcp.tile_4fma_padding) * jcp.ic * jcp.mb * sizeof(float); M_sz_ = (size_t)alpha * alpha * (jcp.itiles * jcp.jtiles + jcp.tile_4fma_padding) * jcp.oc * jcp.mb * sizeof(float); bias_sz_ = nthreads_ * jcp.oc * sizeof(float); src_transpose_sz_ = jcp.ver == ver_4fma ? ((size_t)nthreads_ * alpha * alpha * jcp.tile_4fma * jcp.ic_simd_block * sizeof(float)) : 0; break; case WSCHED_WEI_S_D_G_W: src_transpose_sz_ = jcp.ver == ver_4fma ? ((size_t)nthreads_ * alpha * alpha * jcp.tile_4fma * jcp.ic_simd_block * sizeof(float)) : 0; bias_sz_ = jcp.with_bias ? nthreads_ * jcp.oc * sizeof(float) : 0; break; default: break; } } inline void allocate_scratchpad_(const jit_conv_winograd_conf_t &jcp) { const size_t page_size = PAGE_2M; U_offset_ = 0; V_offset_ = utils::rnd_up(U_sz_, page_size); M_offset_ = V_offset_ + utils::rnd_up(V_sz_, page_size); scratchpad_sz_ = M_offset_ + M_sz_; if (src_transpose_sz_) { src_transpose_offset_ = M_offset_ + utils::rnd_up(M_sz_, page_size); scratchpad_sz_ = src_transpose_offset_ + src_transpose_sz_; } if (bias_sz_) { bias_offset_ = src_transpose_sz_ ? src_transpose_offset_ + utils::rnd_up(src_transpose_sz_, page_size) : M_offset_ + utils::rnd_up(M_sz_, page_size); scratchpad_sz_ = bias_offset_ + bias_sz_; } scratchpad_ = create_scratchpad(scratchpad_sz_); } scratchpad_t *scratchpad_; int nthreads_; size_t scratchpad_sz_ = 0, U_sz_ = 0, V_sz_ = 0, M_sz_ = 0, bias_sz_ = 0, src_transpose_sz_ = 0; size_t U_offset_ = 0; size_t V_offset_ = 0; size_t M_offset_ = 0; size_t bias_offset_ = 0; size_t src_transpose_offset_ = 0; // only relevant for bwdw using qfma }; } template <bool is_fwd> struct _jit_avx512_common_convolution_winograd_t { _jit_avx512_common_convolution_winograd_t( const jit_conv_winograd_conf_t &jcp, const primitive_attr_t *attr) : kernel_(nullptr), scratchpad_(nullptr), attr_(attr) { kernel_ = new _jit_avx512_common_conv_winograd_data_kernel_f32(jcp); scratchpad_ = new winograd::winograd_scratchpad_t(jcp); } ~_jit_avx512_common_convolution_winograd_t() { delete kernel_; delete scratchpad_; }; protected: void _execute_data_W_S_G_D(const int MB, float *inp_ptr, float *out_ptr, float *wei_ptr, float *bias_ptr = NULL); void _execute_data_W_SGD(const int MB, float *inp_ptr, float *out_ptr, float *wei_ptr, float *bias_ptr = NULL); _jit_avx512_common_conv_winograd_data_kernel_f32 *kernel_; // Buffer required to store transforms in the frequency domain winograd::winograd_scratchpad_t *scratchpad_; const primitive_attr_t *attr_; }; template <bool with_relu> struct _jit_avx512_common_convolution_winograd_fwd_t : _jit_avx512_common_convolution_winograd_t<true> , public cpu_primitive_t { struct pd_t : public _cpu_convolution_fwd_pd_t<with_relu> { pd_t(engine_t *engine, const typename pd_t::base_desc_t *adesc, const primitive_attr_t *attr, const typename pd_t::base_class *hint_fwd_pd) : _cpu_convolution_fwd_pd_t<with_relu>(engine, adesc, attr, hint_fwd_pd) , jcp_() {} DECLARE_COMMON_PD_T( JIT_IMPL_NAME_HELPER("jit_wino:", avx512_common, ""), _jit_avx512_common_convolution_winograd_fwd_t<with_relu>); virtual status_t init() override { using namespace prop_kind; assert(this->engine()->kind() == engine_kind::cpu); bool ok = true && this->set_default_params() == status::success && utils::one_of(this->cdesc_().prop_kind, forward_training, forward_inference) && this->cdesc_().alg_kind == alg_kind::convolution_winograd && !this->has_zero_dim_memory() && utils::everyone_is(data_type::f32, this->cdesc_().src_desc.data_type, this->cdesc_().weights_desc.data_type, this->cdesc_().dst_desc.data_type) && utils::implication(this->with_bias(), data_type::f32 == this->cdesc_().bias_desc.data_type) && mkldnn_thr_syncable(); ok = ok && this->dst_pd_.desc()->format == memory_format::nChw16c && this->src_pd_.desc()->format == memory_format::nChw16c; if (!ok) return status::unimplemented; return jit_avx512_common_conv_winograd_fwd_kernel_f32::init_conf( jcp_, this->cdesc_(), *this->src_pd_.desc(), *this->weights_pd_.desc(), *this->dst_pd_.desc(), *this->attr(), with_relu, this->negative_slope()); } jit_conv_winograd_conf_t jcp_; protected: virtual status_t set_default_params() override { using namespace memory_format; if (this->src_pd_.desc()->format == any) CHECK(this->src_pd_.set_format(nChw16c)); if (this->dst_pd_.desc()->format == any) CHECK(this->dst_pd_.set_format(nChw16c)); if (this->weights_pd_.desc()->format == any) CHECK(this->weights_pd_.set_format( this->with_groups() ? gOIhw16i16o : OIhw16i16o)); if (this->bias_pd_.desc()->format == any) CHECK(this->bias_pd_.set_format(x)); return status::success; } }; _jit_avx512_common_convolution_winograd_fwd_t(const pd_t *pd, const input_vector &inputs, const output_vector &outputs) : _jit_avx512_common_convolution_winograd_t<true>(pd->jcp_, pd->attr()) , cpu_primitive_t(&conf_, inputs, outputs) , conf_(*pd) {} ~_jit_avx512_common_convolution_winograd_fwd_t(){}; typedef typename prec_traits<data_type::f32>::type data_t; virtual void execute(event_t *e) { float *src = (float *)this->input_memory(0); float *dst = (float *)this->memory(); float *weights = (float *)this->input_memory(1); float *bias = (float *)this->input_memory(2); switch ((conf_.jcp_).sched_policy) { case WSCHED_DATA_W_S_G_D: this->_execute_data_W_S_G_D(conf_.MB(), src, dst, weights, bias); break; case WSCHED_DATA_W_SGD: this->_execute_data_W_SGD(conf_.MB(), src, dst, weights, bias); break; default: break; } e->set_state(event_t::ready); } private: pd_t conf_; }; using jit_avx512_common_convolution_winograd_fwd_t = _jit_avx512_common_convolution_winograd_fwd_t<false>; using jit_avx512_common_convolution_winograd_relu_t = _jit_avx512_common_convolution_winograd_fwd_t<true>; struct jit_avx512_common_convolution_winograd_bwd_data_t : _jit_avx512_common_convolution_winograd_t<false>, 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_wino:", avx512_common, ""), jit_avx512_common_convolution_winograd_bwd_data_t); virtual status_t init() override { using namespace prop_kind; assert(this->engine()->kind() == engine_kind::cpu); bool ok = true && this->set_default_params() == status::success && utils::one_of(this->desc()->prop_kind, backward_data) && this->desc()->alg_kind == alg_kind::convolution_winograd && !this->has_zero_dim_memory() && utils::everyone_is(data_type::f32, this->desc()->diff_src_desc.data_type, this->desc()->weights_desc.data_type, this->desc()->diff_dst_desc.data_type) && mkldnn_thr_syncable(); if (!ok) return status::unimplemented; return jit_avx512_common_conv_winograd_bwd_data_kernel_f32:: init_conf(jcp_, *this->desc(), *this->diff_src_pd_.desc(), *this->weights_pd_.desc(), *this->diff_dst_pd_.desc()); } jit_conv_winograd_conf_t jcp_; protected: 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(nChw16c)); if (this->diff_dst_pd_.desc()->format == any) CHECK(this->diff_dst_pd_.set_format(nChw16c)); if (this->weights_pd_.desc()->format == any) CHECK(this->weights_pd_.set_format( this->with_groups() ? gOIhw16i16o : OIhw16i16o)); return status::success; } }; jit_avx512_common_convolution_winograd_bwd_data_t(const pd_t *pd, const input_vector &inputs, const output_vector &outputs) : _jit_avx512_common_convolution_winograd_t<false>(pd->jcp_, pd->attr()) , cpu_primitive_t(&conf_, inputs, outputs) , conf_(*pd) {} ~jit_avx512_common_convolution_winograd_bwd_data_t(){}; typedef typename prec_traits<data_type::f32>::type data_t; virtual void execute(event_t *e) { float *diff_dst = (float *)this->input_memory(0); float *diff_src = (float *)this->memory(); float *weights = (float *)this->input_memory(1); if (conf_.desc()->prop_kind == prop_kind::backward_data) { switch ((conf_.jcp_).sched_policy) { case WSCHED_DATA_W_S_G_D: this->_execute_data_W_S_G_D(conf_.MB(), diff_dst, diff_src, weights, NULL); break; case WSCHED_DATA_W_SGD: this->_execute_data_W_SGD(conf_.MB(), diff_dst, diff_src, weights, NULL); break; default: break; } } else { assert(!"invalid prop_kind"); } e->set_state(event_t::ready); } private: pd_t conf_; }; struct jit_avx512_common_convolution_winograd_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_wino:", avx512_common, ""), jit_avx512_common_convolution_winograd_bwd_weights_t); virtual status_t init() override { using namespace prop_kind; assert(this->engine()->kind() == engine_kind::cpu); bool ok = true && this->set_default_params() == status::success && utils::one_of(this->desc()->prop_kind, backward_weights) && this->desc()->alg_kind == alg_kind::convolution_winograd && !this->has_zero_dim_memory() && utils::everyone_is(data_type::f32, this->desc()->src_desc.data_type, this->desc()->diff_dst_desc.data_type, this->desc()->diff_weights_desc.data_type) && mkldnn_thr_syncable(); if (!ok) return status::unimplemented; return jit_avx512_common_conv_winograd_bwd_weights_kernel_f32:: init_conf(jcp_, *this->desc(), *this->src_pd_.desc(), *this->diff_dst_pd_.desc(), *this->diff_weights_pd_.desc()); } jit_conv_winograd_conf_t jcp_; protected: virtual status_t set_default_params() override { using namespace memory_format; if (this->src_pd_.desc()->format == any) CHECK(this->src_pd_.set_format(nChw16c)); if (this->diff_dst_pd_.desc()->format == any) CHECK(this->diff_dst_pd_.set_format(nChw16c)); if (this->diff_weights_pd_.desc()->format == any) CHECK(this->diff_weights_pd_.set_format( this->with_groups() ? gOIhw16i16o : OIhw16i16o)); if (diff_bias_pd_.desc()->format == any) CHECK(diff_bias_pd_.set_format(x)); return status::success; } }; jit_avx512_common_convolution_winograd_bwd_weights_t(const pd_t *pd, const input_vector &inputs, const output_vector &outputs) : cpu_primitive_t(&conf_, inputs, outputs) , conf_(*pd) , kernel_(nullptr) , scratchpad_(nullptr) , padded_bias_(nullptr) { auto jcp = conf_.jcp_; kernel_ = new jit_avx512_common_conv_winograd_bwd_weights_kernel_f32( jcp); scratchpad_ = new winograd::winograd_scratchpad_t(jcp); if (conf_.want_padded_bias()) padded_bias_ = (float *)malloc(sizeof(float) * jcp.oc, 64); } ~jit_avx512_common_convolution_winograd_bwd_weights_t() { delete kernel_; delete scratchpad_; free(padded_bias_); }; typedef typename prec_traits<data_type::f32>::type data_t; virtual void execute(event_t *e) { if (conf_.desc()->prop_kind == prop_kind::backward_weights) { const auto &jcp = kernel_->jcp; switch (jcp.sched_policy) { case WSCHED_WEI_S_D_G_W: _execute_backward_weights_S_D_G_W(); break; case WSCHED_WEI_S_D_Giot_W: _execute_backward_weights_S_D_Giot_W(); break; case WSCHED_WEI_SDGtWo: _execute_backward_weights_SDGtWo(); break; case WSCHED_WEI_SDGt_W: _execute_backward_weights_SDGt_W(); break; default: assert(!"Unknown Winograd schedule policy!"); break; } } else assert(!"invalid prop_kind"); e->set_state(event_t::ready); } private: void _execute_backward_weights_S_D_G_W(); void _execute_backward_weights_S_D_Giot_W(); void _execute_backward_weights_SDGtWo(); void _execute_backward_weights_SDGt_W(); void _maybe_execute_diff_bias_copy(); pd_t conf_; jit_avx512_common_conv_winograd_bwd_weights_kernel_f32 *kernel_; // Buffer required to store transforms in the frequency domain winograd::winograd_scratchpad_t *scratchpad_; float *padded_bias_; }; void trans_W_4x4_3x3(float Fw_[6][6][16][16], float F[3][3][16][16]); void trans_O_4x4_3x3(float Mw[6][6][16], float O[4][4][16]); void trans_W_3x3_4x4(float Fw[6][6][16], float F[4][6][16]); void trans_O_3x3_4x4(float Mw[6][6][16][16], float M[3][3][16][16]); void trans_I_4x4_3x3(float Iw[6][6][16], float I[6][6][16]); void trans_W_3x3_4x4_wu(float Fw[6][6][16], float F[4][6][16]); void trans_O_3x3_4x4_wu(float Mw[6][6][16][16], float M[3][3][16][16]); } } } #endif // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s