feat: Upgrade TRT to 8.4 #1152
Merged
feat: Upgrade TRT to 8.4 #1152
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
BREAKING CHANGE: Removing deprecated settings like min timing iterations
Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
github-actions
bot
added
component: build system
narendasan
reviewed
Jun 30, 2022
| // Lower threshold because FP16 | ||
| ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results, trt_results, 2e-1)); | ||
| } | ||
| // TEST(Partitioning, ComputeResNet50FallbackGraphCorrectly) { |
There was a problem hiding this comment.
@bowang007 Can you work with @peri044 here to determine the best testing strategy here?
|
@peri044 can you open a second PR into |
Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
This reverts commit d736499.
…A memory options Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
github-actions
bot
added
component: api [Python]
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
There are some changes that do not conform to C++ style guidelines:
diff --git a/workspace/py/torch_tensorrt/csrc/tensorrt_classes.cpp b/tmp/changes.txt
index 5aeac3b..775c71d 100644
--- a/workspace/py/torch_tensorrt/csrc/tensorrt_classes.cpp
+++ b/tmp/changes.txt
@@ -225,11 +225,17 @@ core::CompileSpec CompileSpec::toInternalCompileSpec() {
info.convert_info.engine_settings.num_avg_timing_iters = num_avg_timing_iters;
TORCHTRT_CHECK(workspace_size >= 0, "workspace_size must be 0 or greater");
info.convert_info.engine_settings.workspace_size = workspace_size;
- TORCHTRT_CHECK(dla_sram_size >= 4096, "DLA managed SRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 MiB");
+ TORCHTRT_CHECK(
+ dla_sram_size >= 4096,
+ "DLA managed SRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 MiB");
info.convert_info.engine_settings.dla_sram_size = dla_sram_size;
- TORCHTRT_CHECK(dla_local_dram_size >= 4096, "DLA Local DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 GiB");
+ TORCHTRT_CHECK(
+ dla_local_dram_size >= 4096,
+ "DLA Local DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 GiB");
info.convert_info.engine_settings.dla_local_dram_size = dla_local_dram_size;
- TORCHTRT_CHECK(dla_global_dram_size >= 4096, "DLA Global DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 512 MiB");
+ TORCHTRT_CHECK(
+ dla_global_dram_size >= 4096,
+ "DLA Global DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 512 MiB");
info.convert_info.engine_settings.dla_global_dram_size = dla_global_dram_size;
return info;
}
diff --git a/workspace/py/torch_tensorrt/csrc/register_tensorrt_classes.cpp b/tmp/changes.txt
index 310f23d..a171190 100644
--- a/workspace/py/torch_tensorrt/csrc/register_tensorrt_classes.cpp
+++ b/tmp/changes.txt
@@ -67,7 +67,8 @@ void RegisterTRTCompileSpec() {
ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, workspace_size);
ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_sram_size);
ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_local_dram_size);
- ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_global_dram_size);
+ ADD_FIELD_GET_SET_REGISTRATION(
+ TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_global_dram_size);
ADD_FIELD_GET_SET_REGISTRATION(
TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, truncate_long_and_double);
}
diff --git a/workspace/core/conversion/conversionctx/ConversionCtx.cpp b/tmp/changes.txt
index 688aaa7..d123ee4 100644
--- a/workspace/core/conversion/conversionctx/ConversionCtx.cpp
+++ b/tmp/changes.txt
@@ -107,7 +107,7 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
}
cfg->setAvgTimingIterations(settings.num_avg_timing_iters);
- if (settings.workspace_size != 0){
+ if (settings.workspace_size != 0) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kWORKSPACE, settings.workspace_size);
}
@@ -124,13 +124,13 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
settings.enabled_precisions.find(nvinfer1::DataType::kFLOAT) == settings.enabled_precisions.end(),
"DLA supports only fp16 or int8 precision");
cfg->setDLACore(settings.device.dla_core);
- if (settings.dla_sram_size != 1048576){
+ if (settings.dla_sram_size != 1048576) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kDLA_MANAGED_SRAM, settings.dla_sram_size);
}
- if (settings.dla_local_dram_size != 1073741824){
+ if (settings.dla_local_dram_size != 1073741824) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kDLA_LOCAL_DRAM, settings.dla_local_dram_size);
}
- if (settings.dla_global_dram_size != 536870912){
+ if (settings.dla_global_dram_size != 536870912) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kDLA_GLOBAL_DRAM, settings.dla_global_dram_size);
}
}
diff --git a/workspace/cpp/bin/torchtrtc/main.cpp b/tmp/changes.txt
index 075af18..a0a91be 100644
--- a/workspace/cpp/bin/torchtrtc/main.cpp
+++ b/tmp/changes.txt
@@ -117,8 +117,7 @@ int main(int argc, char** argv) {
parser, "num_iters", "Number of averaging timing iterations used to select kernels", {"num-avg-timing-iters"});
args::ValueFlag<uint64_t> workspace_size(
parser, "workspace_size", "Maximum size of workspace given to TensorRT", {"workspace-size"});
- args::ValueFlag<uint64_t> dla_sram_size(
- parser, "dla_sram_size", "DLA managed SRAM size", {"dla-sram-size"});
+ args::ValueFlag<uint64_t> dla_sram_size(parser, "dla_sram_size", "DLA managed SRAM size", {"dla-sram-size"});
args::ValueFlag<uint64_t> dla_local_dram_size(
parser, "dla_local_dram_size", "DLA Local DRAM size", {"dla-local-dram-size"});
args::ValueFlag<uint64_t> dla_global_dram_size(
ERROR: Some files do not conform to style guidelines…nd added DLA memory size configurations Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
There was a problem hiding this comment.
There are some changes that do not conform to C++ style guidelines:
diff --git a/workspace/py/torch_tensorrt/csrc/tensorrt_classes.cpp b/tmp/changes.txt
index 5aeac3b..775c71d 100644
--- a/workspace/py/torch_tensorrt/csrc/tensorrt_classes.cpp
+++ b/tmp/changes.txt
@@ -225,11 +225,17 @@ core::CompileSpec CompileSpec::toInternalCompileSpec() {
info.convert_info.engine_settings.num_avg_timing_iters = num_avg_timing_iters;
TORCHTRT_CHECK(workspace_size >= 0, "workspace_size must be 0 or greater");
info.convert_info.engine_settings.workspace_size = workspace_size;
- TORCHTRT_CHECK(dla_sram_size >= 4096, "DLA managed SRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 MiB");
+ TORCHTRT_CHECK(
+ dla_sram_size >= 4096,
+ "DLA managed SRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 MiB");
info.convert_info.engine_settings.dla_sram_size = dla_sram_size;
- TORCHTRT_CHECK(dla_local_dram_size >= 4096, "DLA Local DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 GiB");
+ TORCHTRT_CHECK(
+ dla_local_dram_size >= 4096,
+ "DLA Local DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 1 GiB");
info.convert_info.engine_settings.dla_local_dram_size = dla_local_dram_size;
- TORCHTRT_CHECK(dla_global_dram_size >= 4096, "DLA Global DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 512 MiB");
+ TORCHTRT_CHECK(
+ dla_global_dram_size >= 4096,
+ "DLA Global DRAM size must be at least 4 KiB and must be a power of 2. This defaults to 512 MiB");
info.convert_info.engine_settings.dla_global_dram_size = dla_global_dram_size;
return info;
}
diff --git a/workspace/py/torch_tensorrt/csrc/register_tensorrt_classes.cpp b/tmp/changes.txt
index 9165b21..ba2e168 100644
--- a/workspace/py/torch_tensorrt/csrc/register_tensorrt_classes.cpp
+++ b/tmp/changes.txt
@@ -65,7 +65,8 @@ void RegisterTRTCompileSpec() {
ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, workspace_size);
ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_sram_size);
ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_local_dram_size);
- ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_global_dram_size);
+ ADD_FIELD_GET_SET_REGISTRATION(
+ TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, dla_global_dram_size);
ADD_FIELD_GET_SET_REGISTRATION(
TRTCompileSpecTSRegistration, torch_tensorrt::pyapi::CompileSpec, truncate_long_and_double);
}
diff --git a/workspace/core/conversion/conversionctx/ConversionCtx.cpp b/tmp/changes.txt
index 688aaa7..d123ee4 100644
--- a/workspace/core/conversion/conversionctx/ConversionCtx.cpp
+++ b/tmp/changes.txt
@@ -107,7 +107,7 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
}
cfg->setAvgTimingIterations(settings.num_avg_timing_iters);
- if (settings.workspace_size != 0){
+ if (settings.workspace_size != 0) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kWORKSPACE, settings.workspace_size);
}
@@ -124,13 +124,13 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
settings.enabled_precisions.find(nvinfer1::DataType::kFLOAT) == settings.enabled_precisions.end(),
"DLA supports only fp16 or int8 precision");
cfg->setDLACore(settings.device.dla_core);
- if (settings.dla_sram_size != 1048576){
+ if (settings.dla_sram_size != 1048576) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kDLA_MANAGED_SRAM, settings.dla_sram_size);
}
- if (settings.dla_local_dram_size != 1073741824){
+ if (settings.dla_local_dram_size != 1073741824) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kDLA_LOCAL_DRAM, settings.dla_local_dram_size);
}
- if (settings.dla_global_dram_size != 536870912){
+ if (settings.dla_global_dram_size != 536870912) {
cfg->setMemoryPoolLimit(nvinfer1::MemoryPoolType::kDLA_GLOBAL_DRAM, settings.dla_global_dram_size);
}
}
diff --git a/workspace/core/conversion/converters/converter_util.cpp b/tmp/changes.txt
index a6a2bbd..7452615 100644
--- a/workspace/core/conversion/converters/converter_util.cpp
+++ b/tmp/changes.txt
@@ -207,13 +207,13 @@ nvinfer1::ITensor* clamp(
nvinfer1::ITensor* lower_bound,
nvinfer1::ITensor* upper_bound,
std::string const& name) {
-
auto max_layer = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kMAX, x, lower_bound, "max layer for " + name);
TORCHTRT_CHECK(max_layer, "Unable to create max layer for clamp");
LOG_DEBUG(ctx->logger, "Create " << max_layer->getName() << " for clamp");
auto max_itensor = max_layer->getOutput(0);
- auto min_layer = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kMIN, max_itensor, upper_bound, "min layer for " + name);
+ auto min_layer =
+ add_elementwise(ctx, nvinfer1::ElementWiseOperation::kMIN, max_itensor, upper_bound, "min layer for " + name);
TORCHTRT_CHECK(min_layer, "Unable to create min layer for clamp");
LOG_DEBUG(ctx->logger, "Create " << min_layer->getName() << " for clamp");
auto min_itensor = min_layer->getOutput(0);
@@ -227,13 +227,13 @@ nvinfer1::ITensor* clamp_to_input_dim(
nvinfer1::ITensor* input_dim,
int nbdims,
std::string const& name) {
-
auto zero = torch::zeros({nbdims}).to(torch::kI32);
auto zero_itensor = tensor_to_const(ctx, zero);
auto one = torch::ones({nbdims}).to(torch::kI32);
auto one_itensor = tensor_to_const(ctx, one);
- auto upper_bound_layer = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUB, input_dim, one_itensor, "sub layer for " + name);
+ auto upper_bound_layer =
+ add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUB, input_dim, one_itensor, "sub layer for " + name);
TORCHTRT_CHECK(upper_bound_layer, "Unable to create sub layer for clamp to inputDim");
LOG_DEBUG(ctx->logger, "Create " << upper_bound_layer->getName() << " for clamp to inputDim");
auto upper_bound = upper_bound_layer->getOutput(0);
@@ -243,7 +243,8 @@ nvinfer1::ITensor* clamp_to_input_dim(
LOG_DEBUG(ctx->logger, "Create " << max_layer->getName() << " for clamp to inputDim");
auto max_itensor = max_layer->getOutput(0);
- auto min_layer = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kMIN, max_itensor, upper_bound, "min layer for " + name);
+ auto min_layer =
+ add_elementwise(ctx, nvinfer1::ElementWiseOperation::kMIN, max_itensor, upper_bound, "min layer for " + name);
TORCHTRT_CHECK(min_layer, "Unable to create min_layer for clamp to inputDim");
LOG_DEBUG(ctx->logger, "Create " << min_layer->getName() << " for clamp to inputDim");
auto min_itensor = min_layer->getOutput(0);
@@ -257,7 +258,6 @@ nvinfer1::ITensor* normalize_indices(
nvinfer1::ITensor* indices,
int nbdims,
std::string const& name) {
-
auto zero = torch::zeros({nbdims}).to(torch::kI32);
auto neg = -torch::ones({nbdims}).to(torch::kI32);
auto zero_itensor = tensor_to_const(ctx, zero);
@@ -307,17 +307,20 @@ nvinfer1::ITensor* get_slice_size(
at::Tensor one_tensor = torch::ones({nbdims}).to(torch::kI32);
auto one_itensor = tensor_to_const(ctx, one_tensor);
- auto sub_layer = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUB, end, start, "get_slice_size sub layer for " + name);
+ auto sub_layer =
+ add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUB, end, start, "get_slice_size sub layer for " + name);
TORCHTRT_CHECK(sub_layer, "Unable to create sub layer in calculate_output_size");
LOG_DEBUG(ctx->logger, "Create " << sub_layer->getName() << " for calculate_output_size");
auto sub_itensor = sub_layer->getOutput(0);
- auto div_layer = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kDIV, sub_itensor, stride, "get_slice_size div layer for " + name);
+ auto div_layer = add_elementwise(
+ ctx, nvinfer1::ElementWiseOperation::kDIV, sub_itensor, stride, "get_slice_size div layer for " + name);
TORCHTRT_CHECK(div_layer, "Unable to create div layer in calculate_output_size");
LOG_DEBUG(ctx->logger, "Create " << div_layer->getName() << " for calculate_output_size");
auto div_itensor = div_layer->getOutput(0);
- auto add_layer = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUM, div_itensor, one_itensor, "get_slice_size sum layer for " + name);
+ auto add_layer = add_elementwise(
+ ctx, nvinfer1::ElementWiseOperation::kSUM, div_itensor, one_itensor, "get_slice_size sum layer for " + name);
TORCHTRT_CHECK(add_layer, "Unable to create add layer in calculate_output_size");
LOG_DEBUG(ctx->logger, "Create " << add_layer->getName() << " for calculate_output_size");
auto size_itensor = add_layer->getOutput(0);
diff --git a/workspace/core/conversion/converters/impl/select.cpp b/tmp/changes.txt
index 3599ab9..d33f09a 100644
--- a/workspace/core/conversion/converters/impl/select.cpp
+++ b/tmp/changes.txt
@@ -103,121 +103,118 @@ nvinfer1::ITensor* roll(
auto select_registrations TORCHTRT_UNUSED =
RegisterNodeConversionPatterns()
- .pattern(
- {"aten::select.int(Tensor(a) self, int dim, int index) -> (Tensor(a))",
- [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
- auto in = args[0].ITensorOrFreeze(ctx);
- auto maxDim = static_cast<int64_t>(in->getDimensions().nbDims);
- auto dim = args[1].unwrapToInt();
- // Handle negative axis by refering to nbDims of input Tensor
- dim = dim < 0 ? dim + maxDim : dim;
- auto ind = (int32_t)args[2].unwrapToInt();
- // Along the specified dimension, handle negative index by subtracting along length of dimension.
- ind = ind < 0 ? ind + in->getDimensions().d[dim] : ind;
- LOG_DEBUG("Gather input dimensions: " << in->getDimensions());
- LOG_DEBUG("Dimension to select: " << dim);
- LOG_DEBUG("Index: " << ind);
-
- // index to access needs to be an at::Tensor
- at::Tensor indices = torch::tensor({ind}).to(torch::kI32);
- auto const_out = tensor_to_const(ctx, indices);
-
- // IGatherLayer takes in input tensor, the indices, and the axis
- // of input tensor to take indices from
- auto gather_layer = ctx->net->addGather(*in, *const_out, dim);
- TORCHTRT_CHECK(gather_layer, "Unable to create gather layer from node: " << *n);
- auto out = gather_layer->getOutput(0);
+ .pattern({"aten::select.int(Tensor(a) self, int dim, int index) -> (Tensor(a))",
+ [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+ auto in = args[0].ITensorOrFreeze(ctx);
+ auto maxDim = static_cast<int64_t>(in->getDimensions().nbDims);
+ auto dim = args[1].unwrapToInt();
+ // Handle negative axis by refering to nbDims of input Tensor
+ dim = dim < 0 ? dim + maxDim : dim;
+ auto ind = (int32_t)args[2].unwrapToInt();
+ // Along the specified dimension, handle negative index by subtracting along length of dimension.
+ ind = ind < 0 ? ind + in->getDimensions().d[dim] : ind;
+ LOG_DEBUG("Gather input dimensions: " << in->getDimensions());
+ LOG_DEBUG("Dimension to select: " << dim);
+ LOG_DEBUG("Index: " << ind);
+
+ // index to access needs to be an at::Tensor
+ at::Tensor indices = torch::tensor({ind}).to(torch::kI32);
+ auto const_out = tensor_to_const(ctx, indices);
+
+ // IGatherLayer takes in input tensor, the indices, and the axis
+ // of input tensor to take indices from
+ auto gather_layer = ctx->net->addGather(*in, *const_out, dim);
+ TORCHTRT_CHECK(gather_layer, "Unable to create gather layer from node: " << *n);
+ auto out = gather_layer->getOutput(0);
+
+ LOG_DEBUG("Gather tensor shape: " << out->getDimensions());
+
+ if (out->getDimensions().nbDims != 1) {
+ // IShuffleLayer removes redundant dimensions
+ auto shuffle_layer = ctx->net->addShuffle(*out);
+ TORCHTRT_CHECK(shuffle_layer, "Unable to create shuffle layer from node: " << *n);
+ shuffle_layer->setReshapeDimensions(util::squeezeDims(out->getDimensions(), dim));
+ shuffle_layer->setName(util::node_info(n).c_str());
+ out = shuffle_layer->getOutput(0);
+ }
+
+ out = ctx->AssociateValueAndTensor(n->outputs()[0], out);
+
+ LOG_DEBUG("Output tensor shape: " << out->getDimensions());
- LOG_DEBUG("Gather tensor shape: " << out->getDimensions());
+ return true;
+ }})
+ .pattern({"aten::narrow(Tensor(a) self, int dim, int start, int length) -> Tensor(a)",
+ [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+ auto in = args[0].ITensor();
+ auto axis = args[1].unwrapToInt();
+ auto start = (int32_t)args[2].unwrapToInt();
+ auto length = (int32_t)args[3].unwrapToInt();
- if (out->getDimensions().nbDims != 1) {
- // IShuffleLayer removes redundant dimensions
- auto shuffle_layer = ctx->net->addShuffle(*out);
- TORCHTRT_CHECK(shuffle_layer, "Unable to create shuffle layer from node: " << *n);
- shuffle_layer->setReshapeDimensions(util::squeezeDims(out->getDimensions(), dim));
- shuffle_layer->setName(util::node_info(n).c_str());
- out = shuffle_layer->getOutput(0);
- }
+ // index to access needs to be an at::Tensor
+ at::Tensor indices = torch::arange(start, start + length, 1).to(torch::kI32);
+ auto weights = Weights(ctx, indices);
- out = ctx->AssociateValueAndTensor(n->outputs()[0], out);
+ // IConstantLayer to convert indices from Weights to ITensor
+ auto const_layer = ctx->net->addConstant(weights.shape, weights.data);
+ TORCHTRT_CHECK(const_layer, "Unable to create constant layer from node: " << *n);
+ auto const_out = const_layer->getOutput(0);
- LOG_DEBUG("Output tensor shape: " << out->getDimensions());
+ // IGatherLayer takes in input tensor, the indices, and the axis
+ // of input tensor to take indices from
+ auto gather_layer = ctx->net->addGather(*in, *const_out, axis);
+ TORCHTRT_CHECK(gather_layer, "Unable to create gather layer from node: " << *n);
+ auto gather_out = gather_layer->getOutput(0);
- return true;
- }})
- .pattern(
- {"aten::narrow(Tensor(a) self, int dim, int start, int length) -> Tensor(a)",
- [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
- auto in = args[0].ITensor();
- auto axis = args[1].unwrapToInt();
- auto start = (int32_t)args[2].unwrapToInt();
- auto length = (int32_t)args[3].unwrapToInt();
-
- // index to access needs to be an at::Tensor
- at::Tensor indices = torch::arange(start, start + length, 1).to(torch::kI32);
- auto weights = Weights(ctx, indices);
-
- // IConstantLayer to convert indices from Weights to ITensor
- auto const_layer = ctx->net->addConstant(weights.shape, weights.data);
- TORCHTRT_CHECK(const_layer, "Unable to create constant layer from node: " << *n);
- auto const_out = const_layer->getOutput(0);
-
- // IGatherLayer takes in input tensor, the indices, and the axis
- // of input tensor to take indices from
- auto gather_layer = ctx->net->addGather(*in, *const_out, axis);
- TORCHTRT_CHECK(gather_layer, "Unable to create gather layer from node: " << *n);
- auto gather_out = gather_layer->getOutput(0);
-
- // IShuffleLayer removes redundant dimensions
- auto shuffle_layer = ctx->net->addShuffle(*gather_out);
- TORCHTRT_CHECK(shuffle_layer, "Unable to create shuffle layer from node: " << *n);
- shuffle_layer->setReshapeDimensions(util::unpadDims(gather_out->getDimensions()));
- shuffle_layer->setName(util::node_info(n).c_str());
- auto shuffle_out = shuffle_layer->getOutput(0);
+ // IShuffleLayer removes redundant dimensions
+ auto shuffle_layer = ctx->net->addShuffle(*gather_out);
+ TORCHTRT_CHECK(shuffle_layer, "Unable to create shuffle layer from node: " << *n);
+ shuffle_layer->setReshapeDimensions(util::unpadDims(gather_out->getDimensions()));
+ shuffle_layer->setName(util::node_info(n).c_str());
+ auto shuffle_out = shuffle_layer->getOutput(0);
- auto out = ctx->AssociateValueAndTensor(n->outputs()[0], shuffle_out);
+ auto out = ctx->AssociateValueAndTensor(n->outputs()[0], shuffle_out);
- LOG_DEBUG("Output tensor shape: " << out->getDimensions());
+ LOG_DEBUG("Output tensor shape: " << out->getDimensions());
- return true;
- }})
- .pattern(
- {"aten::narrow.Tensor(Tensor(a) self, int dim, Tensor start, int length) -> Tensor(a)",
- [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
- auto in = args[0].ITensor();
- auto axis = args[1].unwrapToInt();
- torch::Tensor start = args[2].IValue()->toTensor().to(torch::kI32);
- int32_t startIdx = start.item().to<int32_t>();
- auto length = (int32_t)args[3].unwrapToInt();
-
- // index to access needs to be an at::Tensor
- at::Tensor indices = torch::arange(startIdx, startIdx + length, 1).to(torch::kI32);
- auto weights = Weights(ctx, indices);
-
- // IConstantLayer to convert indices from Weights to ITensor
- auto const_layer = ctx->net->addConstant(weights.shape, weights.data);
- TORCHTRT_CHECK(const_layer, "Unable to create constant layer from node: " << *n);
- auto const_out = const_layer->getOutput(0);
-
- // IGatherLayer takes in input tensor, the indices, and the axis
- // of input tensor to take indices from
- auto gather_layer = ctx->net->addGather(*in, *const_out, axis);
- TORCHTRT_CHECK(gather_layer, "Unable to create gather layer from node: " << *n);
- auto gather_out = gather_layer->getOutput(0);
-
- // IShuffleLayer removes redundant dimensions
- auto shuffle_layer = ctx->net->addShuffle(*gather_out);
- TORCHTRT_CHECK(shuffle_layer, "Unable to create shuffle layer from node: " << *n);
- shuffle_layer->setReshapeDimensions(util::unpadDims(gather_out->getDimensions()));
- shuffle_layer->setName(util::node_info(n).c_str());
- auto shuffle_out = shuffle_layer->getOutput(0);
-
- auto out = ctx->AssociateValueAndTensor(n->outputs()[0], shuffle_out);
-
- LOG_DEBUG("Output tensor shape: " << out->getDimensions());
+ return true;
+ }})
+ .pattern({"aten::narrow.Tensor(Tensor(a) self, int dim, Tensor start, int length) -> Tensor(a)",
+ [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+ auto in = args[0].ITensor();
+ auto axis = args[1].unwrapToInt();
+ torch::Tensor start = args[2].IValue()->toTensor().to(torch::kI32);
+ int32_t startIdx = start.item().to<int32_t>();
+ auto length = (int32_t)args[3].unwrapToInt();
+
+ // index to access needs to be an at::Tensor
+ at::Tensor indices = torch::arange(startIdx, startIdx + length, 1).to(torch::kI32);
+ auto weights = Weights(ctx, indices);
+
+ // IConstantLayer to convert indices from Weights to ITensor
+ auto const_layer = ctx->net->addConstant(weights.shape, weights.data);
+ TORCHTRT_CHECK(const_layer, "Unable to create constant layer from node: " << *n);
+ auto const_out = const_layer->getOutput(0);
+
+ // IGatherLayer takes in input tensor, the indices, and the axis
+ // of input tensor to take indices from
+ auto gather_layer = ctx->net->addGather(*in, *const_out, axis);
+ TORCHTRT_CHECK(gather_layer, "Unable to create gather layer from node: " << *n);
+ auto gather_out = gather_layer->getOutput(0);
+
+ // IShuffleLayer removes redundant dimensions
+ auto shuffle_layer = ctx->net->addShuffle(*gather_out);
+ TORCHTRT_CHECK(shuffle_layer, "Unable to create shuffle layer from node: " << *n);
+ shuffle_layer->setReshapeDimensions(util::unpadDims(gather_out->getDimensions()));
+ shuffle_layer->setName(util::node_info(n).c_str());
+ auto shuffle_out = shuffle_layer->getOutput(0);
+
+ auto out = ctx->AssociateValueAndTensor(n->outputs()[0], shuffle_out);
+
+ LOG_DEBUG("Output tensor shape: " << out->getDimensions());
- return true;
- }})
+ return true;
+ }})
.pattern(
{"aten::embedding(Tensor weight, Tensor indices, int padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> (Tensor)",
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
@@ -239,30 +236,29 @@ auto select_registrations TORCHTRT_UNUSED =
return true;
}})
- .pattern(
- {"aten::roll(Tensor self, int[1] shifts, int[1] dims=[]) -> (Tensor)",
- [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
- auto in = args[0].ITensor();
- auto shifts = args[1].unwrapToIntList().vec();
- auto dims = args[2].unwrapToIntList().vec();
-
- TORCHTRT_CHECK(dims.size() == shifts.size(), "dims.size() should be equal to shifts.size()");
- if (ctx->input_is_dynamic) {
- TORCHTRT_THROW_ERROR("aten::roll is currently not support in dynamic input shape compilation");
- } else {
- auto in_shape = util::toVec(in->getDimensions());
- for (size_t i = 0; i < dims.size(); i++) {
- auto dim = dims[i] < 0 ? (in_shape.size() + dims[i]) : dims[i];
- TORCHTRT_CHECK(dim < in_shape.size(), "Dimension out of range");
- in = roll(ctx, in, shifts[i], dim, in_shape);
- }
- auto out = ctx->AssociateValueAndTensor(n->outputs()[0], in);
-
- LOG_DEBUG("Output tensor shape: " << out->getDimensions());
-
- return true;
- }
- }})
+ .pattern({"aten::roll(Tensor self, int[1] shifts, int[1] dims=[]) -> (Tensor)",
+ [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+ auto in = args[0].ITensor();
+ auto shifts = args[1].unwrapToIntList().vec();
+ auto dims = args[2].unwrapToIntList().vec();
+
+ TORCHTRT_CHECK(dims.size() == shifts.size(), "dims.size() should be equal to shifts.size()");
+ if (ctx->input_is_dynamic) {
+ TORCHTRT_THROW_ERROR("aten::roll is currently not support in dynamic input shape compilation");
+ } else {
+ auto in_shape = util::toVec(in->getDimensions());
+ for (size_t i = 0; i < dims.size(); i++) {
+ auto dim = dims[i] < 0 ? (in_shape.size() + dims[i]) : dims[i];
+ TORCHTRT_CHECK(dim < in_shape.size(), "Dimension out of range");
+ in = roll(ctx, in, shifts[i], dim, in_shape);
+ }
+ auto out = ctx->AssociateValueAndTensor(n->outputs()[0], in);
+
+ LOG_DEBUG("Output tensor shape: " << out->getDimensions());
+
+ return true;
+ }
+ }})
.pattern(
{"aten::index.Tensor(Tensor self, Tensor?[] indices) -> (Tensor)",
[](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
@@ -319,7 +315,8 @@ auto select_registrations TORCHTRT_UNUSED =
int startIdx = 0;
auto startIdxIVal = args[2].IValue();
if (!startIdxIVal->isNone()) {
- startIdx = startIdxIVal->toInt() > std::numeric_limits<int32_t>::max() ? maxDim : startIdxIVal->toInt();
+ startIdx =
+ startIdxIVal->toInt() > std::numeric_limits<int32_t>::max() ? maxDim : startIdxIVal->toInt();
startIdx = maxDim == -1 ? startIdx : std::min(startIdx, maxDim);
}
// Handle case when given tensor index is negative
@@ -331,7 +328,8 @@ auto select_registrations TORCHTRT_UNUSED =
int endIdx = maxDim; // -1 for dynamic shape
auto endIdxIVal = args[3].IValue();
if (!endIdxIVal->isNone()) {
- int truncate_value = endIdxIVal->toInt() > std::numeric_limits<int32_t>::max() ? maxDim : endIdxIVal->toInt();
+ int truncate_value =
+ endIdxIVal->toInt() > std::numeric_limits<int32_t>::max() ? maxDim : endIdxIVal->toInt();
endIdx = maxDim == -1 ? truncate_value : std::min(truncate_value, maxDim);
}
if (maxDim > 0) {
@@ -385,7 +383,8 @@ auto select_registrations TORCHTRT_UNUSED =
// update start and end
nvinfer1::ITensor* out_start;
nvinfer1::ITensor* out_end;
- auto start_end = normalize_start_and_end(ctx, ishape_tensor, start_itensor, end_itensor, nbdims, node_name);
+ auto start_end =
+ normalize_start_and_end(ctx, ishape_tensor, start_itensor, end_itensor, nbdims, node_name);
out_start = start_end[0];
out_end = start_end[1];
@@ -397,7 +396,7 @@ auto select_registrations TORCHTRT_UNUSED =
slice_layer->setInput(2, *size_itensor); // size, must be set if input is dynamic
}
auto slice_out = slice_layer->getOutput(0);
-
+
auto out = ctx->AssociateValueAndTensor(n->outputs()[0], slice_out);
LOG_DEBUG("Slice layer output shape: " << out->getDimensions());
diff --git a/workspace/core/conversion/converters/converter_util.h b/tmp/changes.txt
index cdf2ee5..b155499 100644
--- a/workspace/core/conversion/converters/converter_util.h
+++ b/tmp/changes.txt
@@ -1,8 +1,8 @@
#pragma once
+#include <limits>
#include <map>
#include <string>
-#include <limits>
#include "core/conversion/conversionctx/ConversionCtx.h"
#include "core/conversion/converters/Weights.h"
diff --git a/workspace/tests/core/conversion/converters/test_cast.cpp b/tmp/changes.txt
index 092cdb3..d26c7a0 100644
--- a/workspace/tests/core/conversion/converters/test_cast.cpp
+++ b/tmp/changes.txt
@@ -135,7 +135,6 @@ TEST(Converters, ATenBoolToINT32TensorConvertsCorrectly) {
ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt, 2e-6));
}
-
TEST(Converters, ATenToSingleConvertsCorrectly) {
const auto graph = R"IR(
graph(%y.1 : Tensor):
@@ -164,7 +163,6 @@ TEST(Converters, ATenToSingleConvertsCorrectly) {
ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt, 2e-6));
}
-
TEST(Converters, ATenTypeAsConvertsCorrectly) {
const auto graph = R"IR(
graph(%0 : Tensor,
diff --git a/workspace/cpp/bin/torchtrtc/main.cpp b/tmp/changes.txt
index 6c207d7..51ec2c5 100644
--- a/workspace/cpp/bin/torchtrtc/main.cpp
+++ b/tmp/changes.txt
@@ -117,8 +117,7 @@ int main(int argc, char** argv) {
parser, "num_iters", "Number of averaging timing iterations used to select kernels", {"num-avg-timing-iters"});
args::ValueFlag<uint64_t> workspace_size(
parser, "workspace_size", "Maximum size of workspace given to TensorRT", {"workspace-size"});
- args::ValueFlag<uint64_t> dla_sram_size(
- parser, "dla_sram_size", "DLA managed SRAM size", {"dla-sram-size"});
+ args::ValueFlag<uint64_t> dla_sram_size(parser, "dla_sram_size", "DLA managed SRAM size", {"dla-sram-size"});
args::ValueFlag<uint64_t> dla_local_dram_size(
parser, "dla_local_dram_size", "DLA Local DRAM size", {"dla-local-dram-size"});
args::ValueFlag<uint64_t> dla_global_dram_size(
ERROR: Some files do not conform to style guidelines
# for free
to join this conversation on GitHub.
Already have an account?
# to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Upgrade TRT to 8.4
Type of change
Checklist: