feat: Rsqrt lowering pass (#1394)

* feat: Add lowering pass for rsqrt operator - Add unpack rsqrt lowering pass - Add test cases for positive inputs, int and float - Add references to new function in headers and BUILD files * Added UnpackRsqrt to lowering passes list
pytorch · Oct 12, 2022 · 85e5e99 · 85e5e99
1 parent e27103a
commit 85e5e99
Show file tree

Hide file tree

Showing 6 changed files with 76 additions and 0 deletions.
diff --git a/core/lowering/lowering.cpp b/core/lowering/lowering.cpp
@@ -60,6 +60,7 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g, LowerInfo lower_info) {
   passes::UnpackAddMM(g);
   // passes::UnpackBatchNorm(g);
   passes::UnpackLogSoftmax(g);
+  passes::UnpackRsqrt(g);
   passes::UnpackStd(g);
   passes::UnpackVar(g);
   passes::RemoveNOPs(g);

diff --git a/core/lowering/passes/BUILD b/core/lowering/passes/BUILD
@@ -33,6 +33,7 @@ cc_library(
         "unpack_hardsigmoid.cpp",
         "unpack_hardswish.cpp",
         "unpack_log_softmax.cpp",
+        "unpack_rsqrt.cpp",
         "unpack_std.cpp",
         "unpack_var.cpp",
         "view_to_reshape.cpp",

diff --git a/core/lowering/passes/CMakeLists.txt b/core/lowering/passes/CMakeLists.txt
@@ -20,6 +20,7 @@ target_sources(${lib_name}
             "${CMAKE_CURRENT_SOURCE_DIR}/unpack_hardsigmoid.cpp"
             "${CMAKE_CURRENT_SOURCE_DIR}/unpack_hardswish.cpp"
             "${CMAKE_CURRENT_SOURCE_DIR}/unpack_log_softmax.cpp"
+            "${CMAKE_CURRENT_SOURCE_DIR}/unpack_rsqrt.cpp"
             "${CMAKE_CURRENT_SOURCE_DIR}/unpack_std.cpp"
             "${CMAKE_CURRENT_SOURCE_DIR}/unpack_var.cpp"
             "${CMAKE_CURRENT_SOURCE_DIR}/view_to_reshape.cpp"

diff --git a/core/lowering/passes/passes.h b/core/lowering/passes/passes.h
@@ -33,6 +33,7 @@ void RemoveUnnecessaryCasts(std::shared_ptr<torch::jit::Graph>& graph);
 void UnpackAddMM(std::shared_ptr<torch::jit::Graph>& graph);
 void UnpackBatchNorm(std::shared_ptr<torch::jit::Graph>& graph);
 void UnpackLogSoftmax(std::shared_ptr<torch::jit::Graph>& graph);
+void UnpackRsqrt(std::shared_ptr<torch::jit::Graph>& graph);
 void UnpackStd(std::shared_ptr<torch::jit::Graph>& graph);
 void UnpackVar(std::shared_ptr<torch::jit::Graph>& graph);
 void AliasOperators(std::shared_ptr<torch::jit::Graph>& graph);

diff --git a/core/lowering/passes/unpack_rsqrt.cpp b/core/lowering/passes/unpack_rsqrt.cpp
@@ -0,0 +1,30 @@
+#include "torch/csrc/jit/passes/subgraph_rewrite.h"
+
+#include "core/util/prelude.h"
+
+namespace torch_tensorrt {
+namespace core {
+namespace lowering {
+namespace passes {
+
+void UnpackRsqrt(std::shared_ptr<torch::jit::Graph>& graph) {
+  std::string rsqrt_pattern = R"IR(
+    graph(%1):
+      %out: Tensor = aten::rsqrt(%1)
+      return (%out))IR";
+  std::string unpacked_pattern = R"IR(
+    graph(%1):
+      %intermediate: Tensor = aten::sqrt(%1)
+      %out: Tensor = aten::reciprocal(%intermediate)
+      return (%out))IR";
+
+  torch::jit::SubgraphRewriter rsqrt_rewriter;
+  rsqrt_rewriter.RegisterRewritePattern(rsqrt_pattern, unpacked_pattern);
+  rsqrt_rewriter.runOnGraph(graph);
+  LOG_GRAPH("Post unpack rsqrt: " << *graph);
+}
+
+} // namespace passes
+} // namespace lowering
+} // namespace core
+} // namespace torch_tensorrt
diff --git a/tests/core/lowering/test_unpack_reduce_ops.cpp b/tests/core/lowering/test_unpack_reduce_ops.cpp
@@ -204,3 +204,45 @@ TEST(LoweringPasses, UnpackStdUnbiasedKeepDimsLowersCorrectly) {
   ASSERT_TRUE(
       torch_tensorrt::tests::util::almostEqual(jit_pre_results[0].toTensor(), jit_post_results[0].toTensor(), 2e-6));
 }
+
+TEST(LoweringPasses, UnpackRsqrtLowersCorrectly) {
+  const auto graph = R"IR(
+      graph(%x.1 : Tensor):
+        %2 : Tensor = aten::rsqrt(%x.1)
+        return (%2))IR";
+
+  // Make range [0.01, 1.01] to ensure positives / avoid NaN with negative sqrt
+  auto in = at::rand({2, 3, 5, 7}, {at::kCUDA}) + 0.01;
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(graph, g.get());
+
+  auto jit_pre_results = torch_tensorrt::tests::util::EvaluateGraphJIT(g, {in});
+  torch_tensorrt::core::lowering::passes::UnpackRsqrt(g);
+  torch::jit::EliminateCommonSubexpression(g);
+  auto jit_post_results = torch_tensorrt::tests::util::EvaluateGraphJIT(g, {in});
+
+  ASSERT_TRUE(
+      torch_tensorrt::tests::util::almostEqual(jit_pre_results[0].toTensor(), jit_post_results[0].toTensor(), 2e-6));
+}
+
+TEST(LoweringPasses, UnpackRsqrtIntLowersCorrectly) {
+  const auto graph = R"IR(
+      graph(%x.1 : Tensor):
+        %2 : Tensor = aten::rsqrt(%x.1)
+        return (%2))IR";
+
+  // Make range of ints [1, 10]
+  auto in = at::randint(1, 11, {2, 3, 5, 7}, {at::kCUDA});
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(graph, g.get());
+
+  auto jit_pre_results = torch_tensorrt::tests::util::EvaluateGraphJIT(g, {in});
+  torch_tensorrt::core::lowering::passes::UnpackRsqrt(g);
+  torch::jit::EliminateCommonSubexpression(g);
+  auto jit_post_results = torch_tensorrt::tests::util::EvaluateGraphJIT(g, {in});
+
+  ASSERT_TRUE(
+      torch_tensorrt::tests::util::almostEqual(jit_pre_results[0].toTensor(), jit_post_results[0].toTensor(), 2e-6));
+}