inference-engine/src/gna_plugin/lstm.hpp

   1 // Copyright (C) 2018-2019 Intel Corporation
   2 // SPDX-License-Identifier: Apache-2.0
   3 //
   4
   5 #pragma once
   6
   7 #define LSTM_GIFO_X_C (component_index)
   8 #define LSTM_GIFO_R_C (component_index+1)
   9 #define LSTM_INPUT_GATE_C (component_index+2)
  10 #define LSTM_INPUT_SIGMOID_C (component_index+3)
  11 #define LSTM_FORGET_GATE_C (component_index+4)
  12 #define LSTM_FORGET_SIGMOID_C (component_index+5)
  13 #define LSTM_CELL_INPUT1_C (component_index+6)
  14 #define LSTM_CELL_INPUT1_TANH_C (component_index+7)
  15 #define LSTM_CELL_INPUT2_C (component_index+8)
  16 #define LSTM_CELL_OUTPUT1_C (component_index+9)
  17 #define LSTM_CELL_TANH_C (component_index+10)
  18 #define LSTM_CELL_OUTPUT2_C (component_index+11)
  19 #define LSTM_CELL_CLIPPING_C (component_index+12)
  20 #define LSTM_OUTPUT_GATE_C (component_index+13)
  21 #define LSTM_OUTPUT_SIGMOID_C (component_index+14)
  22 #define LSTM_HIDDEN_C (component_index+15)
  23 #define LSTM_HIDDEN_IDENTITY_C (component_index+16)
  24 #define LSTM_PROJECTED_C (component_index+17)
  25 #define LSTM_PROJECTED_IDENTITY_C (component_index+18)
  26 #define NUM_LSTM_COMPONENTS 19
  27
  28 #define BILSTM_GIFO_X_FW_C (component_index)
  29 #define BILSTM_GIFO_R_FW_C (component_index+1)
  30 #define BILSTM_INPUT_GATE_FW_C (component_index+2)
  31 #define BILSTM_INPUT_SIGMOID_FW_C (component_index+3)
  32 #define BILSTM_FORGET_GATE_FW_C (component_index+4)
  33 #define BILSTM_FORGET_SIGMOID_FW_C (component_index+5)
  34 #define BILSTM_CELL_INPUT1_FW_C (component_index+6)
  35 #define BILSTM_CELL_INPUT1_TANH_FW_C (component_index+7)
  36 #define BILSTM_CELL_INPUT2_FW_C (component_index+8)
  37 #define BILSTM_CELL_GATE_FW_C (component_index+9)
  38 #define BILSTM_CELL_OUTPUT1_FW_C (component_index+10)
  39 #define BILSTM_CELL_TANH_FW_C (component_index+11)
  40 #define BILSTM_CELL_COPY_FW_C (component_index+12)
  41 #define BILSTM_OUTPUT_GATE_FW_C (component_index+13)
  42 #define BILSTM_OUTPUT_SIGMOID_FW_C (component_index+14)
  43 #define BILSTM_HIDDEN_FW_C (component_index+15)
  44 #define BILSTM_HIDDEN_IDENTITY_FW_C (component_index+16)
  45 #define BILSTM_GIFO_X_BW_C (component_index+17)
  46 #define BILSTM_GIFO_R_BW_C (component_index+18)
  47 #define BILSTM_INPUT_GATE_BW_C (component_index+19)
  48 #define BILSTM_INPUT_SIGMOID_BW_C (component_index+20)
  49 #define BILSTM_FORGET_GATE_BW_C (component_index+21)
  50 #define BILSTM_FORGET_SIGMOID_BW_C (component_index+22)
  51 #define BILSTM_CELL_INPUT1_BW_C (component_index+23)
  52 #define BILSTM_CELL_INPUT1_TANH_BW_C (component_index+24)
  53 #define BILSTM_CELL_INPUT2_BW_C (component_index+25)
  54 #define BILSTM_CELL_GATE_BW_C (component_index+26)
  55 #define BILSTM_CELL_OUTPUT1_BW_C (component_index+27)
  56 #define BILSTM_CELL_TANH_BW_C (component_index+28)
  57 #define BILSTM_CELL_COPY_BW_C (component_index+29)
  58 #define BILSTM_OUTPUT_GATE_BW_C (component_index+30)
  59 #define BILSTM_OUTPUT_SIGMOID_BW_C (component_index+31)
  60 #define BILSTM_HIDDEN_BW_C (component_index+32)
  61 #define BILSTM_HIDDEN_IDENTITY_BW_C (component_index+33)
  62 #define NUM_BILSTM_COMPONENTS 34
  63
  64 #include "gna-api.h"
  65
  66 #define ACTIVATION_SCALE_IG  1024.0f
  67 #define ACTIVATION_SCALE_CI1 1024.0f
  68 #define ACTIVATION_SCALE_CO1 2048.0f
  69 #define ACTIVATION_SCALE_OG  2048.0f
  70 #define ACTIVATION_SCALE_HID 2048.0f
  71 #define MAX_WEIGHT_IFO_GATE  1024.0f
  72 #define NUM_WEIGHT_BYTES_IN        2
  73 #define NUM_WEIGHT_BYTES_PROJ    2
  74
  75 typedef struct {
  76     float min;
  77     float max;
  78     float sum;
  79     float sum_squared;
  80     uint32_t num_saturations;
  81     uint32_t num_elements;
  82 } intel_buffer_stats_t;
  83
  84 typedef struct {
  85     intel_nnet_layer_t in;        // combined input transform
  86     intel_nnet_layer_t rec;        // combined recurrent transform
  87     intel_nnet_layer_t ig;        // input gate
  88     intel_nnet_layer_t fg;        // forget gate
  89     intel_nnet_layer_t ci1;        // cell gate input part 1
  90     intel_nnet_layer_t ci2;        // cell gate input part 2
  91     intel_nnet_layer_t co1;        // cell gate output part 1
  92     intel_nnet_layer_t co2;        // cell gate output part 2
  93     intel_nnet_layer_t og;        // output gate
  94     intel_nnet_layer_t hid;        // hidden gated output
  95     intel_nnet_layer_t proj;    // projected output
  96 } intel_lstm_projected_layer_t;
  97
  98 typedef struct {
  99     intel_affine_layer_t *in;        // combined input transform
 100     intel_affine_layer_t *rec;        // combined recurrent transform
 101     intel_affine_layer_t *ig;        // input gate
 102     intel_affine_layer_t *fg;        // forget gate
 103     intel_affine_layer_t *ci1;        // cell gate input part 1
 104     intel_affine_layer_t *ci2;        // cell gate input part 2
 105     intel_affine_layer_t *co1;        // cell gate output part 1
 106     intel_affine_layer_t *co2;        // cell gate output part 2
 107     intel_affine_layer_t *og;        // output gate
 108     intel_affine_layer_t *hid;        // hidden gated output
 109     intel_affine_layer_t *proj;        // projected output
 110 } intel_lstm_projected_transform_t;
 111
 112 typedef struct {
 113     intel_buffer_stats_t in;        // combined input transform
 114     intel_buffer_stats_t rec;        // combined recurrent transform
 115     intel_buffer_stats_t ig;        // input gate
 116     intel_buffer_stats_t fg;        // forget gate
 117     intel_buffer_stats_t ci1;        // cell gate input part 1
 118     intel_buffer_stats_t ci2;        // cell gate input part 2
 119     intel_buffer_stats_t co1;        // cell gate output part 1
 120     intel_buffer_stats_t co2;        // cell gate output part 2
 121     intel_buffer_stats_t og;        // output gate
 122     intel_buffer_stats_t hid;        // hidden gated output
 123     intel_buffer_stats_t proj;    // projected output
 124 } intel_lstm_projected_stats_t;
 125
 126 typedef struct {
 127     intel_nnet_layer_t rec;        // combined recurrent transform
 128     intel_nnet_layer_t ig;        // input gate
 129     intel_nnet_layer_t fg;        // forget gate
 130     intel_nnet_layer_t ci1;        // cell gate input part 1
 131     intel_nnet_layer_t ci2;        // cell gate input part 2
 132     intel_nnet_layer_t co1;        // cell gate output part 1
 133     intel_nnet_layer_t co2;        // cell gate output part 2
 134     intel_nnet_layer_t og;        // output gate
 135     intel_nnet_layer_t hid;        // hidden gated output
 136     intel_nnet_layer_t proj;    // projected output
 137 } intel_lstm_partial_layer_t;
 138
 139 typedef struct {
 140     intel_affine_layer_t *rec;        // combined recurrent transform
 141     intel_affine_layer_t *ig;        // input gate
 142     intel_affine_layer_t *fg;        // forget gate
 143     intel_affine_layer_t *ci1;        // cell gate input part 1
 144     intel_affine_layer_t *ci2;        // cell gate input part 2
 145     intel_affine_layer_t *co1;        // cell gate output part 1
 146     intel_affine_layer_t *co2;        // cell gate output part 2
 147     intel_affine_layer_t *og;        // output gate
 148     intel_affine_layer_t *hid;        // hidden gated output
 149     intel_affine_layer_t *proj;        // projected output
 150 } intel_lstm_partial_transform_t;
 151
 152 typedef struct {
 153     intel_buffer_stats_t rec;        // combined recurrent transform
 154     intel_buffer_stats_t ig;        // input gate
 155     intel_buffer_stats_t fg;        // forget gate
 156     intel_buffer_stats_t ci1;        // cell gate input part 1
 157     intel_buffer_stats_t ci2;        // cell gate input part 2
 158     intel_buffer_stats_t co1;        // cell gate output part 1
 159     intel_buffer_stats_t co2;        // cell gate output part 2
 160     intel_buffer_stats_t og;        // output gate
 161     intel_buffer_stats_t hid;        // hidden gated output
 162     intel_buffer_stats_t proj;    // projected output
 163 } intel_lstm_partial_stats_t;
 164
 165 typedef struct {
 166     intel_nnet_layer_t in;                // combined input transform
 167     intel_nnet_layer_t dintl;            // interleave x8
 168     intel_nnet_layer_t intl1;            // deinterleave x2
 169     intel_nnet_layer_t intl2;            // deinterleave x2
 170     intel_nnet_layer_t intl3;            // deinterleave x2
 171     intel_nnet_layer_t intl4;            // deinterleave x2
 172     intel_lstm_partial_layer_t part[4];    // unrolled part
 173     intel_nnet_layer_t intl;            // interleave x4
 174 } intel_lstm_projected_layer_g4_t;
 175
 176 typedef struct {
 177     intel_affine_layer_t *in;                // combined input transform
 178     intel_lstm_partial_transform_t part[4];  // unrolled part
 179 } intel_lstm_projected_transform_g4_t;
 180
 181 typedef struct {
 182     intel_buffer_stats_t in;            // combined input transform
 183     intel_lstm_partial_stats_t part[4];    // unrolled part
 184 } intel_lstm_projected_stats_g4_t;
 185
 186 #define NUM_LSTM_LAYERS 11
 187 #define NUM_LSTM_G4_LAYERS 47
 188
 189 extern const char *intel_lstm_projected_layer_name[NUM_LSTM_LAYERS];
 190 extern const char *intel_lstm_projected_layer_g4_name[NUM_LSTM_G4_LAYERS];
 191 /*
 192 void GetLstmBufferStats(intel_lstm_projected_layer_t *ptr_layer, std::vector<intel_lstm_projected_stats_t> &stats);
 193 void UpdateLstmBufferStats(std::vector<intel_lstm_projected_stats_t> &accum, std::vector<intel_lstm_projected_stats_t> stats);
 194 void ClearLstmBufferStats(std::vector<intel_lstm_projected_stats_t> &stats);
 195 void PrintLstmBufferStats(std::string preamble, std::vector<intel_lstm_projected_stats_t> stats);
 196 uint32_t NumBytesLstmMacroLayer(uint32_t num_inputs, uint32_t num_outputs, uint32_t num_cells, uint32_t num_group_size, uint32_t layer_num, bool is_compact);
 197 void InitLstmMacroLayerG1(intel_lstm_projected_layer_t *ptr_layer, intel_lstm_projected_transform_t *ptr_transform, uint32_t num_inputs, uint32_t num_outputs, uint32_t num_cells);
 198 void InitLstmMacroLayerG4(intel_lstm_projected_layer_g4_t *ptr_layer, intel_lstm_projected_transform_g4_t *ptr_transform, uint32_t num_inputs, uint32_t num_outputs, uint32_t num_cells);
 199 void AllocateLstmMacroLayerG1(intel_lstm_projected_layer_t *ptr_layer, intel_lstm_projected_transform_t *ptr_transform, intel_shared_outputs scratch, uint8_t **ptr_memory, uint32_t *ptr_num_bytes_used, uint32_t num_memory_bytes, bool is_compact);
 200 void AllocateLstmMacroLayerG4(intel_lstm_projected_layer_g4_t *ptr_layer, intel_lstm_projected_transform_g4_t *ptr_transform, intel_shared_outputs scratch, uint8_t **ptr_memory, uint32_t *ptr_num_bytes_used, uint32_t num_memory_bytes, bool is_compact);
 201 void ConnectLstmMacroLayerG1(intel_lstm_projected_layer_t *ptr_layer, intel_lstm_projected_transform_t *ptr_transform);
 202 void ConnectLstmMacroLayerG4(intel_lstm_projected_layer_g4_t *ptr_layer, intel_lstm_projected_transform_g4_t *ptr_transform);
 203 void QuantizeLstmMacroLayerG1(std::vector<intel_dnn_component_t> *ptr_component, uint32_t component_index, intel_lstm_projected_transform_t *ptr_transform, float input_scale, gna_scale_factor_t *scale, uint32_t j);
 204 void QuantizeLstmMacroLayerG4(std::vector<intel_dnn_component_t> *ptr_component, uint32_t component_index, intel_lstm_projected_transform_g4_t *ptr_transform, float input_scale, gna_scale_factor_t *scale, uint32_t j);
 205 void ReQuantizeLstmMacroLayerG1(std::vector<intel_dnn_component_t> *ptr_component, uint32_t component_index, intel_lstm_projected_layer_t *ptr_layer, float input_scale, gna_scale_factor_t *scale, uint32_t j);
 206 void ReQuantizeLstmMacroLayerG4(std::vector<intel_dnn_component_t> *ptr_component, uint32_t component_index, intel_lstm_projected_layer_g4_t *ptr_layer, float input_scale, gna_scale_factor_t *scale, uint32_t j);
 207 void IntegrityCheckLstmMacroLayer(std::vector<intel_dnn_component_t> *ptr_component, uint32_t component_index, intel_lstm_projected_layer_t *ptr_layer, gna_scale_factor_t *scale, uint32_t j);
 208
 209 */