DRAFT CFe fuse Mul Add to Fullyconnected

on-going draft to fuse Mul Add to Fullyconnected.

Signed-off-by: SaeHie Park <[email protected]>

compiler/circle2circle-dredd-recipe-test/test.lst

@@ -21,6 +21,9 @@ Add(MaxPoolWithArgmax_000 PASS resolve_customop_max_pool_with_argmax)
 Add(MaxPoolWithArgmax_001 PASS resolve_customop_max_pool_with_argmax)
 Add(MaxPoolWithArgmax_002 PASS resolve_customop_max_pool_with_argmax)
 Add(Net_Add_FloorMod_Gather_000 PASS remove_gather_guard)
+Add(Net_Add_FullyConnected_000 PASS fuse_add_to_fullyconnected_bias)
+Add(Net_Add_FullyConnected_001 PASS fuse_add_to_fullyconnected_bias)
+Add(Net_Add_FullyConnected_002 PASS fuse_add_to_fullyconnected_bias)
 Add(Net_BroadcastTo_AddV2_000 PASS resolve_customop_add)
 Add(Net_BroadcastTo_AddV2_001 PASS resolve_customop_add)
 Add(Net_BroadcastTo_AddV2_002 PASS resolve_customop_add)
@@ -61,6 +64,9 @@ Add(Net_Mul_Add_002 PASS remove_unnecessary_add)
 Add(Net_Mul_Add_003 PASS remove_unnecessary_add)
 Add(Net_Mul_Div_000 PASS fuse_mul_with_div)
 Add(Net_Mul_Div_001 PASS fuse_mul_with_div)
+Add(Net_Mul_FullyConnected_000 PASS fuse_mul_to_fullyconnected_weights fold_mul)
+Add(Net_Mul_FullyConnected_001 PASS fuse_mul_to_fullyconnected_weights fold_mul)
+Add(Net_Mul_FullyConnected_002 PASS fuse_mul_to_fullyconnected_weights fold_mul)
 Add(Net_Preactivation_BN_000 PASS fuse_preactivation_batchnorm)
 Add(Net_Reshape_Reshape_000 PASS remove_redundant_reshape)
 Add(Net_Shape_Add_000 PASS fold_shape)

compiler/circle2circle/src/Circle2Circle.cpp

@@ -81,6 +81,7 @@ int entry(int argc, char **argv)
   add_switch(arser, "--fold_fully_connected",
              "This will fold FullyConnected operator with constant inputs");
   add_switch(arser, "--fold_gather", "This will fold Gather operator");
+  add_switch(arser, "--fold_mul", "This will fold Mul operator");
   add_switch(arser, "--fold_reshape", "This will fold Reshape operator");
   add_switch(arser, "--fold_shape", "This will fold Shape operator");
   add_switch(arser, "--fold_sparse_to_dense", "This will fold SparseToDense operator");
@@ -105,6 +106,10 @@ int entry(int argc, char **argv)
   add_switch(arser, "--fuse_batchnorm_with_tconv",
              "This will fuse BatchNorm operators to Transposed Convolution operator");
   add_switch(arser, "--fuse_bcq", "This will fuse operators and apply Binary Coded Quantization");
+  add_switch(arser, "--fuse_add_to_fullyconnected_bias",
+             "This will fuse Add to following FullyConnected bias");
+  add_switch(arser, "--fuse_mul_to_fullyconnected_weights",
+             "This will fuse Mul to following FullyConnected weights");
   add_switch(arser, "--fuse_instnorm", "This will fuse operators to InstanceNorm operator");
   add_switch(arser, "--fuse_mean_with_mean",
              "This will fuse two Mean operations when they follow one by one. This will fold them "
@@ -275,6 +280,8 @@ int entry(int argc, char **argv)
     options->enable(Algorithms::FoldFullyConnected);
   if (arser.get<bool>("--fold_gather"))
     options->enable(Algorithms::FoldGather);
+  if (arser.get<bool>("--fold_mul"))
+    options->enable(Algorithms::FoldMul);
   if (arser.get<bool>("--fold_reshape"))
     options->enable(Algorithms::FoldReshape);
   if (arser.get<bool>("--fold_shape"))
@@ -293,6 +300,8 @@ int entry(int argc, char **argv)
     options->enable(Algorithms::FuseHorizontalFullyConnected);
   if (arser.get<bool>("--fuse_batchnorm_with_conv"))
     options->enable(Algorithms::FuseBatchNormWithConv);
+  if (arser.get<bool>("--fuse_add_to_fullyconnected_bias"))
+    options->enable(Algorithms::FuseAddToFullyConnectedBias);
   if (arser.get<bool>("--fuse_add_with_conv"))
     options->enable(Algorithms::FuseAddWithConv);
   if (arser.get<bool>("--fuse_add_with_fully_connected"))
@@ -303,6 +312,8 @@ int entry(int argc, char **argv)
     options->enable(Algorithms::FuseBatchNormWithDwConv);
   if (arser.get<bool>("--fuse_batchnorm_with_tconv"))
     options->enable(Algorithms::FuseBatchNormWithTConv);
+  if (arser.get<bool>("--fuse_mul_to_fullyconnected_weights"))
+    options->enable(Algorithms::FuseMulToFullyConnectedWeights);
   if (arser.get<bool>("--fuse_slice_with_tconv"))
     options->enable(Algorithms::FuseSliceWithTConv);
   if (arser.get<bool>("--fuse_bcq"))

compiler/luci-pass-value-py-test/test.lst

@@ -2,14 +2,17 @@
 # Format:
 #   eval(MODEL PASS)
 # MODEL: tflite model file name in build/compiler/common-artifacts folder.
-# PASS: Optimization Pass to test. Supports only one Pass for now.
+# PASS: Optimization Pass to test. Supports one or Passes.
 #
 
 # eval(Net_Preactivation_BN_000 fuse_preactivation_batchnorm) : value diff exist
 # --> https://github.com/Samsung/ONE/issues/5782
 eval(FullyConnected_007 replace_non_const_fc_with_batch_matmul)
 eval(HardSwish_001 decompose_hardswish)
 eval(Net_Add_FloorMod_Gather_000 remove_gather_guard)
+eval(Net_Add_FullyConnected_000 fuse_add_to_fullyconnected_bias)
+eval(Net_Add_FullyConnected_001 fuse_add_to_fullyconnected_bias)
+eval(Net_Add_FullyConnected_002 fuse_add_to_fullyconnected_bias)
 eval(Net_Conv_Add_000 fuse_add_with_conv)
 eval(Net_Conv_Add_001 fuse_add_with_conv)
 # eval(Net_Conv_Add_002 fuse_add_with_conv) --> Conv2D w/o bias fails in tflite interpreter
@@ -40,6 +43,9 @@ eval(Net_Mul_Add_002 remove_unnecessary_add)
 eval(Net_Mul_Add_003 remove_unnecessary_add)
 eval(Net_Mul_Div_000 fuse_mul_with_div)
 eval(Net_Mul_Div_001 fuse_mul_with_div)
+eval(Net_Mul_FullyConnected_000 fuse_mul_to_fullyconnected_weights)
+eval(Net_Mul_FullyConnected_001 fuse_mul_to_fullyconnected_weights)
+eval(Net_Mul_FullyConnected_002 fuse_mul_to_fullyconnected_weights)
 eval(Net_Reshape_Mean_000 forward_reshape_to_unaryop)
 eval(Net_Reshape_Neg_000 forward_reshape_to_unaryop)
 eval(Net_Reshape_Reshape_000 remove_redundant_reshape)

compiler/luci-pass-value-py-test/test_luci_eval.py

@@ -95,8 +95,13 @@ def luci_eval_verify(test_name,
             assert np.allclose(
                 luci_output_data, intp_output_data, rtol=rtolint, atol=atolint), err_msg
         elif output_details["dtype"] == np.float32:
-            assert np.allclose(
-                luci_output_data, intp_output_data, rtol=rtolf32, atol=atolf32), err_msg
+            diff_comp = np.allclose(
+                luci_output_data, intp_output_data, rtol=rtolf32, atol=atolf32)
+            if not diff_comp:
+                print("\r\ntflite:\r\n", intp_output_data, flush=True)
+                print("\r\ncircle:\r\n", luci_output_data, flush=True)
+                print("\r\nDiff:\r\n", intp_output_data - luci_output_data, flush=True)
+                assert diff_comp, err_msg
         elif output_details["dtype"] == np.int64:
             assert np.allclose(
                 luci_output_data, intp_output_data, rtol=rtolint, atol=atolint), err_msg

compiler/luci/pass/include/luci/CircleOptimizer.h

@@ -34,12 +34,14 @@ class CircleOptimizer final
   {
     enum Algorithm
     {
+      FuseAddToFullyConnectedBias,
       FuseAddWithConv,
       FuseAddWithFullyConnected,
       FuseAddWithTConv,
       FuseBatchNormWithConv,
       FuseBatchNormWithDwConv,
       FuseBatchNormWithTConv,
+      FuseMulToFullyConnectedWeights,
       FuseSliceWithTConv,
       FuseBCQ,
       FuseHorizontalFullyConnected,
@@ -61,6 +63,7 @@ class CircleOptimizer final
       FoldFullyConnected,
       FoldDequantize,
       FoldGather,
+      FoldMul,
       FoldReshape,
       FoldShape,
       FoldSparseToDense,

compiler/luci/pass/include/luci/Pass/FoldMulPass.h

@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __LUCI_FOLD_MUL_PASS_H__
+#define __LUCI_FOLD_MUL_PASS_H__
+
+#include <logo/Pass.h>
+
+namespace luci
+{
+
+/**
+ * @brief  Class to fold Mul to a constant tensor
+ *
+ */
+struct FoldMulPass final : public logo::Pass
+{
+  const char *name(void) const final { return "luci::FoldMulPass"; }
+
+  bool run(loco::Graph *g) final;
+};
+
+} // namespace luci
+
+#endif // __LUCI_FOLD_MUL_PASS_H__

compiler/luci/pass/include/luci/Pass/FuseAddToFullyConnectedBiasPass.h

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __LUCI_FUSE_ADD_TO_FULLY_CONNECTED_BIAS_PASS_H__
+#define __LUCI_FUSE_ADD_TO_FULLY_CONNECTED_BIAS_PASS_H__
+
+#include <logo/Pass.h>
+
+namespace luci
+{
+
+/**
+ * @brief  Class to fuse Add to following FC bias
+ */
+struct FuseAddToFullyConnectedBiasPass final : public logo::Pass
+{
+  const char *name(void) const final { return "luci::FuseAddToFullyConnectedBiasPass"; }
+
+  bool run(loco::Graph *g) final;
+};
+
+} // namespace luci
+
+#endif // __LUCI_FUSE_ADD_TO_FULLY_CONNECTED_BIAS_PASS_H__

compiler/luci/pass/include/luci/Pass/FuseMulToFullyConnectedWeightsPass.h

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2024 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef __LUCI_FUSE_MUL_TO_FULLY_CONNECTED_WEIGHTS_PASS_H__
+#define __LUCI_FUSE_MUL_TO_FULLY_CONNECTED_WEIGHTS_PASS_H__
+
+#include <logo/Pass.h>
+
+namespace luci
+{
+
+/**
+ * @brief  Class to fuse Mul into following FullyConnected
+ */
+struct FuseMulToFullyConnectedWeightsPass final : public logo::Pass
+{
+  const char *name(void) const final { return "luci::FuseMulToFullyConnectedWeightsPass"; }
+
+  bool run(loco::Graph *g) final;
+};
+
+} // namespace luci
+
+#endif // __LUCI_FUSE_MUL_TO_FULLY_CONNECTED_WEIGHTS_PASS_H__

compiler/luci/pass/src/CircleOptimizer.cpp

@@ -26,20 +26,23 @@
 #include "luci/Pass/FoldDequantizePass.h"
 #include "luci/Pass/FoldFullyConnectedPass.h"
 #include "luci/Pass/FoldGatherPass.h"
+#include "luci/Pass/FoldMulPass.h"
 #include "luci/Pass/FoldReshapePass.h"
 #include "luci/Pass/FoldShapePass.h"
 #include "luci/Pass/FoldSparseToDensePass.h"
 #include "luci/Pass/FoldSqueezePass.h"
 #include "luci/Pass/ForwardReshapeToUnaryOpPass.h"
 #include "luci/Pass/ForwardTransposeOpPass.h"
 #include "luci/Pass/FuseActivationFunctionPass.h"
+#include "luci/Pass/FuseAddToFullyConnectedBiasPass.h"
 #include "luci/Pass/FuseAddWithConvPass.h"
 #include "luci/Pass/FuseAddWithFullyConnectedPass.h"
 #include "luci/Pass/FuseAddWithTConvPass.h"
 #include "luci/Pass/FuseBatchNormWithConvPass.h"
 #include "luci/Pass/FuseBatchNormWithDwConvPass.h"
 #include "luci/Pass/FuseBatchNormWithTConvPass.h"
 #include "luci/Pass/FuseBCQPass.h"
+#include "luci/Pass/FuseMulToFullyConnectedWeightsPass.h"
 #include "luci/Pass/FuseInstanceNormPass.h"
 #include "luci/Pass/FuseMeanWithMeanPass.h"
 #include "luci/Pass/FuseMulWithConvPass.h"
@@ -317,6 +320,10 @@ void CircleOptimizer::optimize(loco::Graph *g) const
   {
     phase.emplace_back(std::make_unique<FuseSliceWithTConvPass>());
   }
+  if (_options->query(Options::Algorithm::FuseAddToFullyConnectedBias))
+  {
+    phase.emplace_back(std::make_unique<FuseAddToFullyConnectedBiasPass>());
+  }
   if (_options->query(Options::Algorithm::FuseAddWithConv))
   {
     phase.emplace_back(std::make_unique<FuseAddWithConvPass>());
@@ -333,6 +340,10 @@ void CircleOptimizer::optimize(loco::Graph *g) const
   {
     phase.emplace_back(std::make_unique<FuseActivationFunctionPass>());
   }
+  if (_options->query(Options::Algorithm::FuseMulToFullyConnectedWeights))
+  {
+    phase.emplace_back(std::make_unique<FuseMulToFullyConnectedWeightsPass>());
+  }
   if (_options->query(Options::Algorithm::FusePRelu))
   {
     phase.emplace_back(std::make_unique<FusePReluPass>());
@@ -381,6 +392,10 @@ void CircleOptimizer::optimize(loco::Graph *g) const
   {
     phase.emplace_back(std::make_unique<luci::FoldGatherPass>());
   }
+  if (_options->query(Options::Algorithm::FoldMul))
+  {
+    phase.emplace_back(std::make_unique<luci::FoldMulPass>());
+  }
   if (_options->query(Options::Algorithm::FoldReshape))
   {
     phase.emplace_back(std::make_unique<luci::FoldReshapePass>());