diff --git a/hpvm/lib/Transforms/HPVM2NVDLA/HPVM2NVDLAPass.cpp b/hpvm/lib/Transforms/HPVM2NVDLA/HPVM2NVDLAPass.cpp
index e8f47ff5a0b357c76edb3a1d6022e84ac3841197..977e4b6075805d6913fcf7738a27f4f2e17e9c24 100644
--- a/hpvm/lib/Transforms/HPVM2NVDLA/HPVM2NVDLAPass.cpp
+++ b/hpvm/lib/Transforms/HPVM2NVDLA/HPVM2NVDLAPass.cpp
@@ -1,3 +1,4 @@
+
#define ENABLE_ASSERTS
#define DEBUG_TYPE "DFG2NVDLA"
@@ -329,6 +330,7 @@ Weights CGT_NVDLA::readTrainedWeights(User *TensorPtr,
DEBUG(errs() << "Data file is not found. Aborting.\n");
abort();
}
+
fseek(file, file_header_size, SEEK_CUR);
size_t bytes_read = fread(tensor_data, 1, size_in_bytes, file);
DEBUG(errs() << "BYTES READ: " << bytes_read << "\n");
@@ -1380,27 +1382,31 @@ fail:
}
bool HPVM2NVDLA::runOnModule(Module &M) {
+
DEBUG(errs() << "**************HPVM2NVDLA PASS****************\n");
NvDlaError e = NvDlaError_TestApplicationFailed;
TestAppArgs testAppArgs = defaultTestAppArgs;
// Get the HPVM IR graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
- std::vector<DFInternalNode *> Roots = DFG.getRoots();
+ BuildDFG &DFG = getAnalysis<BuildDFG>();
+ std::vector<DFInternalNode *> Roots = DFG.getRoots();
- // Visitor for Code Generation Graph Traversal
- CGT_NVDLA *CGTVisitor = new CGT_NVDLA(M, DFG);
+ // Visitor for Code Generation Graph Traversal
+ CGT_NVDLA *CGTVisitor = new CGT_NVDLA(M, DFG);
if(ComputePrecision == "INT8" || ComputePrecision == "int8") {
testAppArgs.computePrecision = nvdla::DataType::INT8;
testAppArgs.quantizationMode = nvdla::QuantizationMode::PER_KERNEL;
testAppArgs.configtarget = std::string("nv_small");
+ errs() << " INT8 mode \n";
} else {
testAppArgs.computePrecision = nvdla::DataType::HALF;
testAppArgs.quantizationMode = nvdla::QuantizationMode::NONE;
testAppArgs.configtarget = std::string("nv_full");
+ errs() << "FP16 mode \n";
}
+
testAppArgs.profileName = std::string("hpvm-mod");
testAppArgs.calibTable = CalibTablePath;//std::string("output_scales.txt");
testAppArgs.outputPath = std::string(".");
@@ -1419,7 +1425,9 @@ bool HPVM2NVDLA::runOnModule(Module &M) {
return false;
}
+
NvDlaError CGT_NVDLA::compileProfile(const TestAppArgs* appArgs, TestInfo* i) {
+
NvDlaError e = NvDlaSuccess;
std::string profileName = "";
std::string targetConfigName = "";
@@ -1438,26 +1446,38 @@ NvDlaError CGT_NVDLA::compileProfile(const TestAppArgs* appArgs, TestInfo* i) {
targetConfigName = appArgs->configtarget;
+ errs()<<" Compiling NVDLA code *** \n";
+
// Determine profile
PROPAGATE_ERROR_FAIL(generateProfile(appArgs, &profileName, i));
+ errs() << "compiling profile \"%s\"... config \"%s\"...\n" << profileName.c_str() << targetConfigName.c_str() <<"\n";
+
// Compile
DEBUG(NvDlaDebugPrintf("compiling profile \"%s\"... config \"%s\"...\n", profileName.c_str(), targetConfigName.c_str()));
- PROPAGATE_ERROR_FAIL(compiler->compile(profileName.c_str(), targetConfigName.c_str(), &i->compiledLoadable));
+ //PROPAGATE_ERROR_FAIL(compiler->compile(profileName.c_str(), targetConfigName.c_str(), &i->compiledLoadable));
+
+ compiler->compile(profileName.c_str(), targetConfigName.c_str(), &i->compiledLoadable);
+
// Get loadable buffer and dump it into a file
PROPAGATE_ERROR_FAIL(compiler->getLoadableImageSize(profileName.c_str(),
- &size));
+ &size));
if (size == 0) {
ORIGINATE_ERROR_FAIL(NvDlaError_BadParameter,
"Invalid size for a loadable");
+ errs()<<" ERROR: Invalid size for loadable \n";
}
+ errs()<<" size of loadable module = " << size;
+
buffer = (NvU8 *) NvDlaAlloc(size);
if (buffer == NULL) {
ORIGINATE_ERROR_FAIL(NvDlaError_InsufficientMemory,
"Failed to allocate buffer for loadable");
+ errs() << "ERROR: Failed to allocate buffer or loadable \n";
}
+
PROPAGATE_ERROR_FAIL(compiler->getLoadableImage(profileName.c_str(),
buffer));
fileName = profileName + ".nvdla";
@@ -1588,6 +1608,8 @@ fail:
NvDlaError CGT_NVDLA::generateTensorScales(const TestAppArgs* appArgs, TestInfo* i, nvdla::INetwork* network) {
NvDlaError e = NvDlaSuccess;
+ errs() << " generateTenorScales \n";
+
std::vector<nvdla::ILayer*> networkLayers = network->getLayers();
std::vector<nvdla::ITensor*> networkInputs = network->getInputs();
@@ -1626,13 +1648,16 @@ fail:
return e;
}
+
+
NvDlaError CGT_NVDLA::readTensorScales(const TestAppArgs* appArgs, TestInfo *i, nvdla::INetwork* network) {
- NvDlaError e = NvDlaSuccess;
+
+ NvDlaError e = NvDlaSuccess;
NvDlaStatType stat;
- std::string calibTableFile = /*i->calibTablesPath + "/" + */appArgs->calibTable;
+ std::string calibTableFile = appArgs->calibTable;
- //PROPAGATE_ERROR_FAIL(NvDlaStat(calibTableFile.c_str(), &stat));
- DEBUG(errs() << "***********READING TENSOR SCALESi*************\n");
+ errs() << "***********READING TENSOR SCALESi*************\n";
+
std::ifstream infile(calibTableFile.c_str());
std::string line;
std::map<std::string, float> LayerNameToScaleMap;
@@ -1643,13 +1668,16 @@ NvDlaError CGT_NVDLA::readTensorScales(const TestAppArgs* appArgs, TestInfo *i,
std::string delimiter = ":";
std::string layer_name = line.substr(0, line.find(delimiter));
std::string Scale = line.substr(line.find(delimiter) + 1);
- DEBUG(errs() << "LAYER NAME: " << layer_name << "\n");
- DEBUG(errs() << "SCALE: " << Scale << "\n");
+
+ errs() << "LAYER NAME: " << layer_name << "\n";
+ errs() << "SCALE: " << Scale << "\n";
+
size_t size;
LayerNameToScaleMap[layer_name] = std::stof(Scale, &size);
}
+
infile.close();
- DEBUG(errs() << "GOT TENSOR SCALES FROM CALIB TABLE\n");
+ errs() << "GOT TENSOR SCALES FROM CALIB TABLE \n";
std::vector<nvdla::ILayer*> networkLayers = network->getLayers();
std::vector<nvdla::ITensor*> networkInputs = network->getInputs();
@@ -1665,36 +1693,42 @@ NvDlaError CGT_NVDLA::readTensorScales(const TestAppArgs* appArgs, TestInfo *i,
PROPAGATE_ERROR_FAIL(Input->setChannelDynamicRange(-1, min, max) );
const_cast<TestAppArgs*>(appArgs)->tensorScales.insert(std::pair<std::string, NvF32>("data", scale));
}
- DEBUG(errs() << "PER LAYER CALIB\n");
+
+ errs() << "PER LAYER CALIB \n";
+
for (auto *Layer : networkLayers) {
NvF32 scale = 0.0f;
- NvF32 min = 0.0f;
- NvF32 max = 0.0f;
- std::string tName = Layer->getName();
- DEBUG(errs() << "SETTING SCALE FOR LAYER NAME: " << tName << "\n");
- nvdla::ITensor* outTensor = Layer->getOutput(0);
- auto it = LayerNameToScaleMap.find(tName);
- if (it != LayerNameToScaleMap.end()) {
- DEBUG(errs() << "SET SCALE FOR NAME: " << tName << "\n");
- DEBUG(errs() << "SCALE: " << it->second << "\n");
- scale = it->second;
- min = scale * -127.0f;
- max = scale * 127.0f;
- } else {
- DEBUG(errs() << "SET DEFAULT SCALE FOR NAME: " << tName << "\n");
- DEBUG(errs() << "SCALE: 1\n");
- scale = 1;
- min = scale * -127.0f;
- max = scale * 127.0f;
- }
- //else {
- // ORIGINATE_ERROR_FAIL(NvDlaError_BadParameter, "Atleast 1 of scale or min-max should be specified for %s\n", tName.c_str());
- //}
+ NvF32 min = 0.0f;
+ NvF32 max = 0.0f;
+
+ std::string tName = Layer->getName();
+ errs() << "SETTING SCALE FOR LAYER NAME: " << tName << "\n";
+
+ nvdla::ITensor* outTensor = Layer->getOutput(0);
+ auto it = LayerNameToScaleMap.find(tName);
+
+ if (it != LayerNameToScaleMap.end()) {
+ errs() << "SET SCALE FOR NAME: " << tName << "\n";
+ errs() << "SCALE: " << it->second << "\n";
+ scale = it->second;
+ min = scale * -127.0f;
+ max = scale * 127.0f;
+ } else {
+
+ DEBUG(errs() << "SET DEFAULT SCALE FOR NAME: " << tName << "\n");
+ DEBUG(errs() << "SCALE: 1\n");
+
+ scale = 1;
+ min = scale * -127.0f;
+ max = scale * 127.0f;
+ }
+
PROPAGATE_ERROR_FAIL( outTensor->setChannelDynamicRange(-1, min, max) );
const_cast<TestAppArgs*>(appArgs)->tensorScales.insert(std::pair<std::string, NvF32>(tName, scale));
}
- DEBUG(errs() << "DONE PARSING CALIBRATION TABLE\n");
+ errs() << "DONE PARSING CALIBRATION TABLE ----- \n";
+
fail:
return e;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/miniera-hpvm/src/miniera-hpvm.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/miniera-hpvm/src/miniera-hpvm.cpp
index 3ab28ea74edd80e9850b3ee95370b7b99d5f30c1..9c9099cda1afaa4b4ba8812ab5ef84729b12c667 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/miniera-hpvm/src/miniera-hpvm.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/miniera-hpvm/src/miniera-hpvm.cpp
@@ -369,29 +369,29 @@ int main(){
//std::string input_path = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/input_fp16.bin";
std::string labels_path = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/labels_fp16.bin";
//char conv2d_1_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_1_w_fp16.bin";
- void* conv2d_1_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_1_w.bin", 0,32,3,3,3);
- //char conv2d_1_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_1_b_fp16.bin";
- void* conv2d_1_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_1_b.bin", 0,1,32, 1, 1);//30,30);
- //char conv2d_2_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_w_fp16.bin";
- void* conv2d_2_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_w.bin", 0,32,32,3,3);
- //char conv2d_2_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_b_fp16.bin";
- void* conv2d_2_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_b.bin", 0,1,32, 1, 1);//28,28);
- //char conv2d_3_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_w_fp16.bin";
- void* conv2d_3_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_w.bin", 0,64,32,3,3);
- //char conv2d_3_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_b_fp16.bin";
- void* conv2d_3_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_b.bin", 0,1,64, 1, 1);//12,12);
- //char conv2d_4_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_w_fp16.bin";
- void* conv2d_4_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_w.bin", 0,64,64,3,3);
- //char conv2d_4_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_b_fp16.bin";
- void* conv2d_4_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_b.bin", 0,1,64, 1, 1);//10,10);
- //char dense_1_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_w_fp16.bin";
- void* dense_1_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_w.bin", 0,1,1,1600,256);
- //char dense_1_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_b_fp16.bin";
- void* dense_1_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_b.bin", 0,1,256,1,1);
- //char dense_2_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_w_fp16.bin";
- void* dense_2_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_w.bin", 0,1,1,256,5);
- //char dense_2_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_b_fp16.bin";
- void* dense_2_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_b.bin", 0,1,5,1,1);
+ void* conv2d_1_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_1_w_fp16.bin", 0,32,3,3,3);
+ //char conv2d_1_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_1_b_fp16_fp16.bin";
+ void* conv2d_1_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_1_b_fp16.bin", 0,1,32, 1, 1);//30,30);
+ //char conv2d_2_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_w_fp16_fp16.bin";
+ void* conv2d_2_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_w_fp16.bin", 0,32,32,3,3);
+ //char conv2d_2_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_b_fp16_fp16.bin";
+ void* conv2d_2_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_2_b_fp16.bin", 0,1,32, 1, 1);//28,28);
+ //char conv2d_3_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_w_fp16_fp16.bin";
+ void* conv2d_3_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_w_fp16.bin", 0,64,32,3,3);
+ //char conv2d_3_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_b_fp16_fp16.bin";
+ void* conv2d_3_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_3_b_fp16.bin", 0,1,64, 1, 1);//12,12);
+ //char conv2d_4_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_w_fp16_fp16.bin";
+ void* conv2d_4_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_w_fp16.bin", 0,64,64,3,3);
+ //char conv2d_4_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_b_fp16_fp16.bin";
+ void* conv2d_4_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/conv2d_4_b_fp16.bin", 0,1,64, 1, 1);//10,10);
+ //char dense_1_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_w_fp16_fp16.bin";
+ void* dense_1_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_w_fp16.bin", 0,1,1,1600,256);
+ //char dense_1_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_b_fp16_fp16.bin";
+ void* dense_1_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_1_b_fp16.bin", 0,1,256,1,1);
+ //char dense_2_w_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_w_fp16_fp16.bin";
+ void* dense_2_w = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_w_fp16.bin", 0,1,1,256,5);
+ //char dense_2_b_path[] = "../../../../../projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_b_fp16_fp16.bin";
+ void* dense_2_b = readTrainedWeights("/home/hsharif3/Gitlab/old_hpvm_nvdla/hpvm/llvm/projects/hpvm-tensor-rt/model_params/legacy/hpvm_mio/dense_2_b_fp16.bin", 0,1,5,1,1);
//void* input = readTrainedWeights(input_path, 0,1,3,32,32);
//uint32_t* labels = readLabels3(labels_path, 500);
diff --git a/hpvm/tools/hpvm-clang/main.py.in b/hpvm/tools/hpvm-clang/main.py.in
index 0c047de0437285b84227a0f4ef7614761a2daed2..7d4d8644e687cf4a810e49a1b7bb0b4b294b7768 100644
--- a/hpvm/tools/hpvm-clang/main.py.in
+++ b/hpvm/tools/hpvm-clang/main.py.in
@@ -60,7 +60,8 @@ def compile_hpvm_c(
passes += ["LLVMHPVM2NVDLAPass"]
pass_flags += [
"hpvm-nvdla",
- "cprecision=fp16",
+ #"cprecision=fp16",
+ "cprecision=int8",
"calib-table=calib.txt"
]