Merge pull request #1252 from inocsin/scatter

feat: support scatter.value and scatter.src
pytorch · Aug 15, 2022 · 298c3a3 · 298c3a3
2 parents a64a3ac + 01e6541
commit 298c3a3
Show file tree

Hide file tree

Showing 6 changed files with 158 additions and 24 deletions.
diff --git a/core/conversion/converters/converter_util.cpp b/core/conversion/converters/converter_util.cpp
@@ -363,6 +363,27 @@ nvinfer1::ITensor* get_slice_size(
   return size_itensor;
 }
 
+nvinfer1::ITensor* scalar_to_tensor(ConversionCtx* ctx, at::Scalar s) {
+  nvinfer1::ITensor* out;
+  if (s.isIntegral(false)) {
+    auto s_int = s.to<int64_t>();
+    auto s_t = torch::tensor({s_int}).to(at::kInt);
+    out = tensor_to_const(ctx, s_t);
+  } else if (s.isBoolean()) {
+    auto s_bool = s.to<bool>();
+    auto s_t = torch::tensor({s_bool}).to(at::kBool);
+    out = tensor_to_const(ctx, s_t);
+  } else if (s.isFloatingPoint()) {
+    auto other_float = s.to<float>();
+    auto s_t = torch::tensor({other_float});
+    out = tensor_to_const(ctx, s_t);
+  } else {
+    out = nullptr;
+    TORCHTRT_THROW_ERROR("Unsupported data type for scalar. Found: (" << s.type() << ")");
+  }
+  return out;
+}
+
 } // namespace converters
 } // namespace conversion
 } // namespace core

diff --git a/core/conversion/converters/converter_util.h b/core/conversion/converters/converter_util.h
@@ -80,6 +80,8 @@ nvinfer1::ITensor* get_slice_size(
     int nbdims,
     std::string const& name);
 
+nvinfer1::ITensor* scalar_to_tensor(ConversionCtx* ctx, at::Scalar s);
+
 } // namespace converters
 } // namespace conversion
 } // namespace core

diff --git a/core/conversion/converters/impl/element_wise.cpp b/core/conversion/converters/impl/element_wise.cpp
@@ -25,26 +25,7 @@ nvinfer1::ITensor* clamp_util(
   return clamp_layer_out;
 }
 
-nvinfer1::ITensor* scalar_to_tensor(ConversionCtx* ctx, at::Scalar s) {
-  nvinfer1::ITensor* out;
-  if (s.isIntegral(false)) {
-    auto s_int = s.to<int64_t>();
-    auto s_t = torch::tensor({s_int}).to(at::kInt);
-    out = tensor_to_const(ctx, s_t);
-  } else if (s.isBoolean()) {
-    auto s_bool = s.to<bool>();
-    auto s_t = torch::tensor({s_bool}).to(at::kBool);
-    out = tensor_to_const(ctx, s_t);
-  } else if (s.isFloatingPoint()) {
-    auto other_float = s.to<float>();
-    auto s_t = torch::tensor({other_float});
-    out = tensor_to_const(ctx, s_t);
-  } else {
-    out = nullptr;
-    TORCHTRT_THROW_ERROR("Unsupported data type for scalar. Found: (" << s.type() << ")");
-  }
-  return out;
-}
+
 
 auto element_wise_registrations TORCHTRT_UNUSED =
     RegisterNodeConversionPatterns()

diff --git a/core/conversion/converters/impl/select.cpp b/core/conversion/converters/impl/select.cpp
@@ -464,6 +464,53 @@ auto select_registrations TORCHTRT_UNUSED =
                auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], new_layer->getOutput(0));
                LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
                return true;
+             }})
+        .pattern(
+            {"aten::scatter.value(Tensor self, int dim, Tensor index, Scalar value) -> (Tensor)",
+             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+               auto self = args[0].ITensorOrFreeze(ctx);
+               int dim = args[1].unwrapToInt();
+               auto index = args[2].ITensorOrFreeze(ctx);
+               auto index_dim = index->getDimensions();
+               std::vector<int64_t> dim_vec;
+               for (int i = 0; i < index_dim.nbDims; i++) {
+                 dim_vec.push_back(index_dim.d[i]);
+               }
+               auto value = args[3].unwrapToScalar() * torch::ones(dim_vec);
+               auto value_tensor = tensor_to_const(ctx, value, "");
+               if (self->getType() != value_tensor->getType()) {
+                 value_tensor = castITensor(ctx, value_tensor, self->getType());
+               }
+
+               auto layer = ctx->net->addScatter(*self, *index, *value_tensor, nvinfer1::ScatterMode::kELEMENT);
+               layer->setAxis(dim);
+
+               TORCHTRT_CHECK(layer, "Unable to create layer for aten::scatter.value");
+
+               layer->setName(util::node_info(n).c_str());
+
+               auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], layer->getOutput(0));
+               LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
+               return true;
+             }})
+        .pattern(
+            {"aten::scatter.src(Tensor self, int dim, Tensor index, Tensor src) -> (Tensor)",
+             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+               auto self = args[0].ITensorOrFreeze(ctx);
+               int dim = args[1].unwrapToInt();
+               auto index = args[2].ITensorOrFreeze(ctx);
+               auto src = args[3].ITensorOrFreeze(ctx);
+
+               auto layer = ctx->net->addScatter(*self, *index, *src, nvinfer1::ScatterMode::kELEMENT);
+               layer->setAxis(dim);
+
+               TORCHTRT_CHECK(layer, "Unable to create layer for aten::scatter.src");
+
+               layer->setName(util::node_info(n).c_str());
+
+               auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], layer->getOutput(0));
+               LOG_DEBUG("Output shape: " << out_tensor->getDimensions());
+               return true;
              }});
 
 } // namespace

diff --git a/core/lowering/passes/op_aliasing.cpp b/core/lowering/passes/op_aliasing.cpp
@@ -16,15 +16,25 @@ void AliasOperators(std::shared_ptr<torch::jit::Graph>& graph) {
         graph(%s, %o):
             %1 : Tensor = aten::div(%s, %o)
             return (%1))IR";
-  ;
-
-  // TODO
-  // complete other element wise pass
 
   torch::jit::SubgraphRewriter true_divide_to_div;
   true_divide_to_div.RegisterRewritePattern(true_divide_pattern, div_pattern);
   true_divide_to_div.runOnGraph(graph);
   LOG_GRAPH("Post map true_divide -> div: " << *graph);
+
+  std::string scatter_sub_pattern = R"IR(
+        graph(%data, %dim, %index, %value):
+            %o : Tensor = aten::scatter_(%data, %dim, %index, %value)
+            return (%o))IR";
+  std::string scatter_pattern = R"IR(
+        graph(%data, %dim, %index, %value):
+            %o : Tensor = aten::scatter(%data, %dim, %index, %value)
+            return (%o))IR";
+
+  torch::jit::SubgraphRewriter rewrite_scatter;
+  rewrite_scatter.RegisterRewritePattern(scatter_sub_pattern, scatter_pattern);
+  rewrite_scatter.runOnGraph(graph);
+  LOG_GRAPH("Post map scatter_ -> scatter: " << *graph);
 }
 
 } // namespace passes

diff --git a/tests/core/conversion/converters/test_select.cpp b/tests/core/conversion/converters/test_select.cpp
@@ -855,3 +855,76 @@ TEST(Converters, ATenUnbindNegativeAxisConvertsCorrectly) {
     ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[i], trt, 2e-6));
   }
 }
+
+TEST(Converters, ScatterValueConvertsCorrectly) {
+  const auto graph = R"IR(
+        graph(%data : Tensor,
+              %index.1 : Tensor):
+          %value : int = prim::Constant[value=100]()
+          %dim : int = prim::Constant[value=1]()
+          %5 : NoneType = prim::Constant()
+          %6 : bool = prim::Constant[value=0]()
+          %7 : int = prim::Constant[value=4]()
+          %index : Tensor = aten::to(%index.1, %7, %6, %6, %5) 
+          %10 : Tensor = aten::scatter(%data, %dim, %index, %value)
+          return (%10))IR";
+
+  auto g = std::make_shared<torch::jit::Graph>();
+
+  torch::jit::parseIR(graph, g.get());
+
+  auto index = at::randint(0, 5, {2, 2}, {at::kCUDA});
+  auto data = at::randn({5, 5}, {at::kCUDA});
+
+  auto jit_index = at::clone(index);
+  auto jit_data = at::clone(data);
+  auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_data, jit_index});
+
+  auto trt_index = at::clone(index);
+  auto trt_data = at::clone(data);
+  auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {trt_data, trt_index});
+
+  for (size_t i = 0; i < jit_results.size(); i++) {
+    auto trt = trt_results[i].reshape(jit_results[i].sizes());
+    ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[i], trt, 2e-6));
+  }
+}
+
+TEST(Converters, ScatterSrcConvertsCorrectly) {
+  const auto graph = R"IR(
+        graph(%data : Tensor,
+              %src : Tensor,
+              %index.1 : Tensor):
+          %dim : int = prim::Constant[value=1]()
+          %5 : NoneType = prim::Constant()
+          %6 : bool = prim::Constant[value=0]()
+          %7 : int = prim::Constant[value=4]()
+          %index : Tensor = aten::to(%index.1, %7, %6, %6, %5) 
+          %10 : Tensor = aten::scatter(%data, %dim, %index, %src)
+          return (%10))IR";
+
+  auto g = std::make_shared<torch::jit::Graph>();
+
+  torch::jit::parseIR(graph, g.get());
+
+  auto index = at::randint(0, 4, {2, 2}, {at::kCUDA});
+  auto data = at::randn({5, 5}, {at::kCUDA});
+  auto src = at::randn({2, 2}, {at::kCUDA});
+
+  auto jit_index = at::clone(index);
+  auto jit_data = at::clone(data);
+  auto jit_src = at::clone(src);
+  auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {jit_data, jit_src, jit_index});
+
+  auto trt_index = at::clone(index);
+  auto trt_data = at::clone(data);
+  auto trt_src = at::clone(src);
+  auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {trt_data, trt_src, trt_index});
+
+  for (size_t i = 0; i < jit_results.size(); i++) {
+    auto trt = trt_results[i].reshape(jit_results[i].sizes());
+    ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[i], trt, 2e-6));
+  }
+}