Update NNPI Backend to v0.4.2.5 (#3968)

Summary:
Updating NNPI Backend to version 0.4.2.5.
Pull Request resolved: #3968

Reviewed By: hyuen, jfix71

Differential Revision: D19300112

Pulled By: arunm-git

fbshipit-source-id: 63f3644d7b4bcf7fb579d24603020678bfacb668

Author: omromano

lib/Backends/NNPI/CMakeLists.txt

@@ -97,6 +97,14 @@ message(STATUS "[NNPI] NNPI_MG_LIB_DIR             = ${NNPI_MG_LIB}")
 message(STATUS "[NNPI] GLOW_BINARY_DIR             = ${GLOW_BINARY_DIR}")
 message(STATUS "[NNPI] NNPI_COLLECT_MEM_USAGE      = ${NNPI_MEM_PROFILING}")
 
+if (UNIX)
+  EXEC_PROGRAM(cat ARGS "/proc/cpuinfo" OUTPUT_VARIABLE CPUINFO)
+  STRING(REGEX REPLACE "^.*(avx512f).*$" "\\1" AVX512F_FLAG ${CPUINFO})
+  if("avx512f" STREQUAL "${AVX512F_FLAG}")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -mavx512vl -mavx512bw")
+    add_definitions(-DUSE_AVX)
+  endif()
+endif()
 add_subdirectory(ClassGen)
 
 include_directories(

lib/Backends/NNPI/DebugMacros.h

@@ -19,7 +19,9 @@
 #include "glow/Support/Error.h"
 #include "nnpi_inference.h"
 #include "nnpi_transformer.h"
+#include <chrono>
 #include <glog/logging.h>
+#include <string>
 
 // Macro for memory instrumentation.
 #if NNPI_COLLECT_MEM_USAGE
@@ -255,4 +257,20 @@ GetNNPIInferenceErrorDesc(NNPIInferenceErrorCode err) {
     if (exp_res != NNPI_INF_NO_ERROR)                                          \
       callback(MAKE_ERR(ErrorValue::ErrorCode::RUNTIME_ERROR, msg));           \
   }
+
+// Used in debugging.
+#define NNPI_TIMER_START(timer_name)                                           \
+  auto timer_name = std::chrono::high_resolution_clock::now();
+#define NNPI_TIMER_STOP(timer_name, msg_prefix_)                               \
+  {                                                                            \
+    auto timer_end_ = std::chrono::high_resolution_clock::now();               \
+    std::cout                                                                  \
+        << std::string(msg_prefix_) +                                          \
+               std::to_string(                                                 \
+                   std::chrono::duration_cast<std::chrono::microseconds>(      \
+                       timer_end - timer_name)                                 \
+                       .count()) +                                             \
+               "\n";                                                           \
+  }
+
 #endif // GLOW_NNPI_DEBUG_MACROS_H

lib/Backends/NNPI/Importer.cpp

@@ -941,7 +941,7 @@ class ConcatNodeImporter : public INNPINodeImporter {
     auto numInputs = glowConcat->getNumInputs();
     NNPIObjectName *inputs = new NNPIObjectName[numInputs];
     LOG_AND_RETURN_IF_NOT(ERROR, inputs, "No inputs", NNPI_INVALID_PARAM);
-    std::set<std::string> inputTensors;
+    std::unordered_set<std::string> inputTensors;
 
     for (unsigned i = 0; i < numInputs; i++) {
       auto nvName = nodeValueName(glowConcat->getNthInput(i));
@@ -969,11 +969,29 @@ class TileNodeImporter : public INNPINodeImporter {
     importer.setUsedTensors({nodeValueName(glowTile->getInput())},
                             {nodeValueName(glowTile->getResult())});
 
-    return nnpiNetworkAddTileOp(importer.getNetwork(),
-                                glowTile->getName().begin(),
-                                nodeValueName(glowTile->getInput()).c_str(),
-                                nodeValueName(glowTile->getResult()).c_str(),
-                                glowTile->getCount(), glowTile->getAxis());
+    auto numDims = glowTile->getInput().getType()->dims().size();
+    std::vector<int32_t> repeats(numDims, 1);
+    auto axis = glowTile->getAxis();
+    LOG_AND_RETURN_IF_NOT(ERROR, axis >= 0 && axis < numDims,
+                          "tile axis is invalid", NNPI_INVALID_PARAM);
+    repeats[axis] = glowTile->getCount();
+    NNPITensorDesc desc;
+    desc.attributes.value = 0;
+    desc.attributes.constant = 1;
+    desc.numDims = 1;
+    desc.dims[0] = numDims;
+    desc.quantParams.precision = NNPI_PRECISION_INT32;
+    desc.quantParams.type = NNPI_QUANTIZATION_NONE;
+    desc.layout = NNPI_LAYOUT_ANY;
+
+    auto repeatsTensorName = glowTile->getName().str() + "_repeats";
+
+    importer.addTensor(repeatsTensorName, desc, repeats.data());
+
+    return nnpiNetworkAddTileOp(
+        importer.getNetwork(), glowTile->getName().begin(),
+        nodeValueName(glowTile->getInput()).c_str(), repeatsTensorName.c_str(),
+        nodeValueName(glowTile->getResult()).c_str());
   }
 };
 
@@ -1592,7 +1610,7 @@ class NNPICustomDSPNodeImporter : public INNPINodeImporter {
     auto numInputs = glowDSP->getInputs().size();
     NNPIObjectName *inputs = new NNPIObjectName[numInputs];
     LOG_AND_RETURN_IF_NOT(ERROR, inputs, "No inputs", NNPI_INVALID_PARAM);
-    std::set<std::string> inputTensors;
+    std::unordered_set<std::string> inputTensors;
     uint32_t i = 0;
     for (const auto &nv : glowDSP->getInputs()) {
       auto nvName = nodeValueName(nv);
@@ -1603,7 +1621,7 @@ class NNPICustomDSPNodeImporter : public INNPINodeImporter {
     uint32_t numOutputs = 1;
     NNPIObjectName *outputs = new NNPIObjectName[numOutputs];
     LOG_AND_RETURN_IF_NOT(ERROR, outputs, "No outputs", NNPI_INVALID_PARAM);
-    std::set<std::string> outputTensors;
+    std::unordered_set<std::string> outputTensors;
     auto nvName = nodeValueName(glowDSP->getResult());
     strncpy(outputs[0], nvName.c_str(), sizeof(NNPIObjectName));
     outputTensors.insert(nvName);

lib/Backends/NNPI/Importer.h

@@ -23,9 +23,9 @@
 #include "nnpi_network_builder_EXPERIMENTAL.h"
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/Support/raw_ostream.h"
-#include <map>
-#include <set>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
 
 namespace glow {
 class Function;
@@ -67,8 +67,9 @@ class NNPIImporter {
                          const std::string &offsetTensor = {},
                          bool forceSymlowp = false);
   /// Set given tensor names as inputs/outputs.
-  void setUsedTensors(const std::set<std::string> &readTensors = {},
-                      const std::set<std::string> &writeTensors = {}) {
+  void
+  setUsedTensors(const std::unordered_set<std::string> &readTensors = {},
+                 const std::unordered_set<std::string> &writeTensors = {}) {
     readTensors_.insert(readTensors.begin(), readTensors.end());
     writeTensors_.insert(writeTensors.begin(), writeTensors.end());
   }
@@ -97,13 +98,13 @@ class NNPIImporter {
 
 private:
   /// Map of named external tensors (inputs, outputs, weights, etc...).
-  std::map<std::string, const Tensor *> constants_;
+  std::unordered_map<std::string, const Tensor *> constants_;
   /// Set of tensors written to by the function.
-  std::set<std::string> writeTensors_;
+  std::unordered_set<std::string> writeTensors_;
   /// Set of tensors read from by the function.
-  std::set<std::string> readTensors_;
+  std::unordered_set<std::string> readTensors_;
   /// Set of tensors already defined.
-  std::set<std::string> definedTensors_;
+  std::unordered_set<std::string> definedTensors_;
   /// Number of internal names created for variables.
   size_t internalNameCounter_;