acl_operator.cc 6.3 KB
Newer Older
S
sharper 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212
/*!
 * Copyright (c) 2016 by Contributors
 * \file acl_operator.cc
 * \brief
 * \author Joey
 */
#if USE_ACL == 1
#include "acl_operator.h"
unsigned int bypass_acl_class_layer =
    (0 | FLAGS_ENABLE_ACL_CONCAT |
     /*0xffffffff |*/ /*FLAGS_ENABLE_ACL_FC |*/ /*FLAGS_ENABLE_ACL_LRN
                                                   |*/
     0);

int enable_schedule = 0;

#ifdef USE_PROFILING

#include "arm_neon.h"

unsigned int acl_log_flags =
    (0 | MASK_LOG_APP_TIME | /*MASK_LOG_ALLOCATE | */  /*MASK_LOG_ALLOCATE | */
     /*MASK_LOG_RUN      | */ /*MASK_LOG_CONFIG   | */ /*MASK_LOG_COPY     | */
     MASK_LOG_ABSVAL | MASK_LOG_BNLL | MASK_LOG_CONV | MASK_LOG_FC |
     MASK_LOG_LRN | MASK_LOG_POOLING | MASK_LOG_RELU | MASK_LOG_SIGMOID |
     MASK_LOG_SOFTMAX | MASK_LOG_TANH | MASK_LOG_LC | MASK_LOG_BN |
     MASK_LOG_CONCAT | 0);
#include <stdio.h>  /* printf */
#include <stdlib.h> /* getenv */
#endif              // USE_PROFILING

static bool force_enable_gpu = false;
bool AclEnableSchedule(int enable) {
  enable_schedule = enable;
  if (enable) {
    force_enable_gpu = true;
  }
  return true;
}
int isScheduleEnable() { return enable_schedule; }

namespace paddle_mobile {
namespace operators {
namespace acl {

bool ACLOperator::init_gpu_env = true;
#ifdef USE_OPENCL
bool ACLOperator::support_opencl_ = false;
bool opencl_is_available() { return arm_compute::opencl_is_available(); }
#elif defined(USE_OPENGLES)
bool ACLOperator::support_opengles_ = false;
#endif
ACLOperator::ACLOperator(bool is_gpu)
    : operator_state_(operator_not_init),
      force_bypass_acl_path_(false),
      target_hint_(TargetHint::DONT_CARE),
      convolution_method_hint_(ConvolutionMethodHint::GEMM),
      _group(1),
      name_(""),
      input_idx_(0),
      output_idx_(0),
      is_gpu_(is_gpu) {
  const char* pBypassACL;
  if (init_gpu_env) {
#ifdef USE_OPENCL
    try {
      if (opencl_is_available()) {
        arm_compute::CLScheduler::get().default_init();
        support_opencl_ = true;
      }
    } catch (std::exception& e) {
      support_opencl_ = false;
    }
#elif defined(USE_OPENGLES)
    try {
      arm_compute::GCScheduler::get().default_init();
      support_opengles_ = true;
    } catch (std::exception& e) {
      support_opengles_ = false;
    }
#endif
    init_gpu_env = false;
  }
  if (force_enable_gpu) is_gpu_ = true;
  pBypassACL = getenv("BYPASSACL");
  if (pBypassACL) {
    unsigned int bacl;
    sscanf(pBypassACL, "%i", &bacl);
    if (bacl != bypass_acl_class_layer) {
      bypass_acl_class_layer = bacl;
      printf("BYPASSACL<%s>\n", pBypassACL);
      printf("BYPASSACL: %x\n", bypass_acl_class_layer);
    }
  }

#ifdef USE_PROFILING
  const char* pLogACL;
  pLogACL = getenv("LOGACL");
  if (pLogACL) {
    unsigned int alf;
    sscanf(pLogACL, "%i", &alf);
    if (alf != acl_log_flags) {
      acl_log_flags = alf;
      printf("LOGACL<%s>\n", pLogACL);
      printf("LOGACL: %x\n", acl_log_flags);
    }
  }
#endif  // USE_PROFILING
  const char* pEnableSchedule;
  pEnableSchedule = getenv("ENABLESCHEDULE");
  if (pEnableSchedule) {
    int bshedule;
    sscanf(pEnableSchedule, "%i", &bshedule);
    if (bshedule != enable_schedule) {
      enable_schedule = bshedule;
      printf("ENABLESCHEDULE<%s>\n", pEnableSchedule);
      printf("ENABLESCHEDULE: %x\n", enable_schedule);
    }
    if (enable_schedule) {
      AclEnableSchedule(1);
    }
  }
}
ACLOperator::~ACLOperator() {}

bool ACLOperator::new_tensor(std::unique_ptr<ACLTensor>& tensor,
                             arm_compute::TensorShape& shape, void* mem,
                             bool commit) {
  auto acl_tensor =
      new ACLTensor(arm_compute::TensorInfo(shape, arm_compute::Format::F32));
  acl_tensor->set_target(getTargetHint());
  acl_tensor->bindmem(mem);
  if (commit) acl_tensor->commit();
  tensor = (std::unique_ptr<ACLTensor>)std::move(acl_tensor);
  return true;
}
bool ACLOperator::new_tensor(std::unique_ptr<ACLSubTensor>& tensor,
                             std::unique_ptr<ACLTensor>& parent,
                             arm_compute::TensorShape& shape,
                             arm_compute::Coordinates& coord) {
  auto acl_tensor = new ACLSubTensor(parent, shape, coord);
  acl_tensor->set_target(getTargetHint());
  tensor = (std::unique_ptr<ACLSubTensor>)std::move(acl_tensor);
  return true;
}

void ACLTensor::commit(TensorType type) {
  settensortype(type);
  if (mem_) {
    if (!allocate_) {
#ifdef USE_PROFILING
      logtime_util log_time(ACL_ALLOCATE_INFO);
#endif  // USE_PROFILING
      allocate();
      allocate_ = true;
    }
    if (type_ != tensor_output) {
      tensor_copy(mem_);
    }
    mem_ = nullptr;
  }
}

int BaseACLTensor::tensor_copy(arm_compute::ITensor* tensor, void* mem,
                               bool toTensor) {
#ifdef USE_PROFILING
  logtime_util log_time(ACL_COPY_INFO);
#endif  // USE_PROFILING
  arm_compute::Window window;
  // Iterate through the rows (not each element)
  window.use_tensor_dimensions(tensor->info()->tensor_shape(),
                               /* first_dimension =*/arm_compute::Window::DimY);

  int width = tensor->info()->tensor_shape()[0];
  int height = tensor->info()->tensor_shape()[1];
  int deepth = tensor->info()->tensor_shape()[2];
  map();
  // Create an iterator:
  arm_compute::Iterator it(tensor, window);
  // Except it works for an arbitrary number of dimensions
  if (toTensor) {  // mem->tensor
    arm_compute::execute_window_loop(
        window,
        [&](const arm_compute::Coordinates& id) {
          memcpy(it.ptr(),
                 ((char*)mem) +
                     ((id[3] * (width * height * deepth) +
                       id.z() * (width * height) + id.y() * width + id.x()) *
                      tensor->info()->element_size()),
                 width * tensor->info()->element_size());
        },
        it);
  } else {  // tensor-->mem
    arm_compute::execute_window_loop(
        window,
        [&](const arm_compute::Coordinates& id) {
          memcpy(((char*)mem) + ((id[3] * (width * height * deepth) +
                                  id.z() * (width * height) + id.y() * width) *
                                 tensor->info()->element_size()),
                 it.ptr(), width * tensor->info()->element_size());
        },
        it);
  }
  unmap();

  return 0;
}

}  // namespace acl
}  // namespace operators
}  // namespace paddle_mobile
#endif