Skip to content

feat: Support int inputs to aten::max/min and aten::argmax/argmin #1574

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jan 9, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 13 additions & 1 deletion core/conversion/converters/impl/max.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,11 @@ bool min_max_dim(ConversionCtx* ctx, const torch::jit::Node* n, args& args, nvin
if (dim < 0) {
dim = selfDim.size() + dim;
}
bool int_input = self->getType() == nvinfer1::DataType::kINT32;
if (int_input) {
LOG_DEBUG("topk layer does not support int32 inputs, adding cast to float");
self = castITensor(ctx, self, nvinfer1::DataType::kFLOAT, util::node_info(n) + "_input");
}
uint32_t reduce_axes_mask = 1 << dim;
auto topk_layer = ctx->net->addTopK(*self, topKOperation, 1, reduce_axes_mask);
TORCHTRT_CHECK(topk_layer, "Unable to create topk layer from node: " << *n);
Expand All @@ -44,7 +49,10 @@ bool min_max_dim(ConversionCtx* ctx, const torch::jit::Node* n, args& args, nvin
out0 = ctx->AssociateValueAndTensor(n->outputs()[0], topk_layer->getOutput(0));
out1 = ctx->AssociateValueAndTensor(n->outputs()[1], topk_layer->getOutput(1));
}

if (int_input) {
LOG_DEBUG("Adding cast of topK layer output back to int32");
out0 = castITensor(ctx, out0, nvinfer1::DataType::kINT32, util::node_info(n) + "_output");
}
LOG_DEBUG("Output tensor(0) shape: " << out0->getDimensions());
LOG_DEBUG("Output tensor(1) shape: " << out1->getDimensions());

Expand All @@ -59,6 +67,10 @@ bool arg_min_max(ConversionCtx* ctx, const torch::jit::Node* n, args& args, nvin
if (dim < 0) {
dim = selfDim.size() + dim;
}
if (self->getType() == nvinfer1::DataType::kINT32) {
LOG_DEBUG("topk layer does not support int32 inputs, adding cast to float");
self = castITensor(ctx, self, nvinfer1::DataType::kFLOAT, util::node_info(n) + "_input");
}
uint32_t reduce_axes_mask = 1 << dim;
auto topk_layer = ctx->net->addTopK(*self, topKOperation, 1, reduce_axes_mask);
TORCHTRT_CHECK(topk_layer, "Unable to create topk layer from node: " << *n);
Expand Down
45 changes: 45 additions & 0 deletions tests/core/conversion/converters/test_max.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,29 @@ TEST(Converters, ATenMaxDimConvertsCorrectly) {
torch_tensorrt::tests::util::almostEqual(jit_results[1], trt_results[1].reshape_as(jit_results[1]), 2e-6));
}

TEST(Converters, ATenMaxDimIntInputConvertsCorrectly) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
%2 : int = prim::Constant[value=0]()
%3 : bool = prim::Constant[value=0]()
%4 : Tensor, %5 : Tensor = aten::max(%x.1, %2, %3)
return (%4, %5))IR";

auto g = std::make_shared<torch::jit::Graph>();
torch::jit::parseIR(graph, g.get());

auto in = at::randint(-5, 5, {5, 5}, {at::kCUDA});

auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {in});

params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {in});

ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[1], trt_results[1], 2e-6));
}

TEST(Converters, ATenMinDimConvertsCorrectly) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
Expand Down Expand Up @@ -77,6 +100,28 @@ TEST(Converters, ATenArgMaxConvertsCorrectly) {
torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0].reshape_as(jit_results[0]), 2e-6));
}

TEST(Converters, ATenArgMaxIntInputConvertsCorrectly) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
%2 : int = prim::Constant[value=0]()
%3 : bool = prim::Constant[value=0]()
%4 : Tensor = aten::argmax(%x.1, %2, %3)
return (%4))IR";

auto g = std::make_shared<torch::jit::Graph>();
torch::jit::parseIR(graph, g.get());

auto in = at::randint(-5, 5, {5, 5}, {at::kCUDA});

auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {in});

params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {in});

ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
}

TEST(Converters, ATenArgMaxKeepdimConvertsCorrectly) {
const auto graph = R"IR(
graph(%x.1 : Tensor):
Expand Down