[luci] Introduce Compress weights pass

This commit introduces CopressWeightsPass for Conv2D ONE-DCO-1.0-Signed-off-by: Vyacheslav Bazhenov <[email protected]>
Samsung · Aug 21, 2024 · ef1caef · ef1caef
1 parent 2b9d81f
commit ef1caef
Show file tree

Hide file tree

Showing 60 changed files with 4,684 additions and 52 deletions.
diff --git a/Makefile.template b/Makefile.template
@@ -159,7 +159,7 @@ prepare_nncc_internal: $(WORKSPACE)
 ifeq (,$(findstring android,$(TARGET_OS)))
 	EXTERNAL_FLATC=$(BUILDTOOL_PATH)/bin/flatc ./nncc configure -DBUILD_GTEST=OFF -DENABLE_TEST=OFF -DEXTERNALS_BUILD_THREADS=$(NPROCS) -DCMAKE_BUILD_TYPE=$(BUILD_TYPE) \
 		-DCMAKE_INSTALL_PREFIX=$(OVERLAY_FOLDER) \
-		-DBUILD_WHITELIST="luci;foder;pepper-csv2vec;loco;locop;logo;logo-core;mio-circle08;luci-compute;oops;hermes;hermes-std;angkor;pp;pepper-strcast;pepper-str" \
+		-DBUILD_WHITELIST="luci;foder;pepper-csv2vec;loco;locop;logo;logo-core;mio-circle09;luci-compute;oops;hermes;hermes-std;angkor;pp;pepper-strcast;pepper-str" \
 		$(OPTIONS_NNCC)
 	./nncc build -j$(NPROCS)
 	cmake --install $(NNCC_FOLDER) $(INSTALL_OPTIONS)

diff --git a/compiler/circle-inspect/CMakeLists.txt b/compiler/circle-inspect/CMakeLists.txt
@@ -1,6 +1,6 @@
-if(NOT TARGET mio_circle08)
+if(NOT TARGET mio_circle09)
   return()
-endif(NOT TARGET mio_circle08)
+endif(NOT TARGET mio_circle09)
 
 set(DRIVER "driver/Driver.cpp")
 
@@ -10,6 +10,6 @@ add_executable(circle-inspect ${DRIVER} ${SOURCES})
 target_include_directories(circle-inspect PRIVATE src)
 target_link_libraries(circle-inspect arser)
 target_link_libraries(circle-inspect foder)
-target_link_libraries(circle-inspect mio_circle08)
-target_link_libraries(circle-inspect mio_circle08_helper)
+target_link_libraries(circle-inspect mio_circle09)
+target_link_libraries(circle-inspect mio_circle09_helper)
 target_link_libraries(circle-inspect safemain)
diff --git a/compiler/circle-inspect/requires.cmake b/compiler/circle-inspect/requires.cmake
@@ -1,4 +1,4 @@
 require("arser")
 require("foder")
-require("mio-circle08")
+require("mio-circle09")
 require("safemain")
diff --git a/compiler/circle-operator/CMakeLists.txt b/compiler/circle-operator/CMakeLists.txt
@@ -1,6 +1,6 @@
-if(NOT TARGET mio_circle08)
+if(NOT TARGET mio_circle09)
   return()
-endif(NOT TARGET mio_circle08)
+endif(NOT TARGET mio_circle09)
 
 set(DRIVER "driver/Driver.cpp")
 
@@ -10,8 +10,8 @@ add_executable(circle-operator ${DRIVER} ${SOURCES})
 target_include_directories(circle-operator PRIVATE src)
 target_link_libraries(circle-operator arser)
 target_link_libraries(circle-operator foder)
-target_link_libraries(circle-operator mio_circle08)
-target_link_libraries(circle-operator mio_circle08_helper)
+target_link_libraries(circle-operator mio_circle09)
+target_link_libraries(circle-operator mio_circle09_helper)
 target_link_libraries(circle-operator safemain)
 
 install(TARGETS circle-operator DESTINATION bin)
diff --git a/compiler/circle-operator/requires.cmake b/compiler/circle-operator/requires.cmake
@@ -1,4 +1,4 @@
 require("arser")
 require("foder")
-require("mio-circle08")
+require("mio-circle09")
 require("safemain")
diff --git a/compiler/circle-tensordump/CMakeLists.txt b/compiler/circle-tensordump/CMakeLists.txt
@@ -1,6 +1,6 @@
-if(NOT TARGET mio_circle08)
+if(NOT TARGET mio_circle09)
   return()
-endif(NOT TARGET mio_circle08)
+endif(NOT TARGET mio_circle09)
 
 nnas_find_package(HDF5 COMPONENTS STATIC QUIET)
 
@@ -19,8 +19,8 @@ target_include_directories(circle-tensordump PRIVATE ${HDF5_INCLUDE_DIRS})
 target_link_libraries(circle-tensordump PRIVATE ${HDF5_CXX_LIBRARIES})
 target_link_libraries(circle-tensordump PRIVATE arser)
 target_link_libraries(circle-tensordump PRIVATE foder)
-target_link_libraries(circle-tensordump PRIVATE mio_circle08)
-target_link_libraries(circle-tensordump PRIVATE mio_circle08_helper)
+target_link_libraries(circle-tensordump PRIVATE mio_circle09)
+target_link_libraries(circle-tensordump PRIVATE mio_circle09_helper)
 target_link_libraries(circle-tensordump PRIVATE safemain)
 
 install(TARGETS circle-tensordump DESTINATION bin)
diff --git a/compiler/circle-tensordump/requires.cmake b/compiler/circle-tensordump/requires.cmake
@@ -1,4 +1,4 @@
 require("arser")
 require("foder")
-require("mio-circle08")
+require("mio-circle09")
 require("safemain")
diff --git a/compiler/circle2circle/src/Circle2Circle.cpp b/compiler/circle2circle/src/Circle2Circle.cpp
@@ -189,6 +189,8 @@ int entry(int argc, char **argv)
              "This will convert single input Transpose to Reshape");
   add_switch(arser, "--expand_broadcast_const", "This will expand broadcastable constant inputs");
   add_switch(arser, "--unroll_unidirseqlstm", "Unroll UnidirectionalSequenceLSTM operator.");
+  add_switch(arser, "--compress_weights_huffman",
+             "Loseless weights compression with Huffman encoding.");
   add_switch(arser, "--convert_nchw_to_nhwc",
              "Experimental: This will convert NCHW operators to NHWC under the assumption that "
              "input model is NCHW.");
@@ -416,6 +418,8 @@ int entry(int argc, char **argv)
     options->enable(Algorithms::ExpandBroadcastConst);
   if (arser.get<bool>("--unroll_unidirseqlstm"))
     options->enable(Algorithms::UnrollUnidirSeqLSTM);
+  if (arser.get<bool>("--compress_weights_huffman"))
+    options->enable(Algorithms::CompressWeightsHuffman);
 
   // NOTE Experimental options; these will be removed someday
   //      Add experimental options here

diff --git a/compiler/luci-eval-driver/src/EvalDriver.cpp b/compiler/luci-eval-driver/src/EvalDriver.cpp
@@ -49,6 +49,7 @@ void writeDataToFile(const std::string &filename, const char *data, size_t data_
 
 template <typename NodeT> size_t getTensorSize(const NodeT *node)
 {
+  uint32_t tensor_size = loco::size(node->dtype());
   uint32_t tensor_size = luci::size(node->dtype());
   for (uint32_t i = 0; i < node->rank(); ++i)
     tensor_size *= node->dim(i).value();

diff --git a/compiler/luci-interpreter/include/luci_interpreter/core/Tensor.h b/compiler/luci-interpreter/include/luci_interpreter/core/Tensor.h
@@ -18,6 +18,7 @@
 #define LUCI_INTERPRETER_CORE_TENSOR_H
 
 #include "luci_interpreter/core/DataType.h"
+#include <luci/IR/AttrWeightCompression.h>
 
 #include <cassert>
 #include <cstddef>
@@ -144,7 +145,7 @@ class Tensor
 
   void writeData(const void *data_ptr, size_t data_size);
 
-  void resize(const Shape &new_shape);
+  void resize(const Shape &new_shape, size_t raw_size = 0);
 
   void set_data_buffer(uint8_t *buffer)
   {
@@ -173,11 +174,21 @@ class Tensor
 
   void set_offset(int32_t offset) { _offset = offset; }
 
+  luci::CompressionType get_compression() const { return _compression; }
+
+  void set_compression(luci::CompressionType compression) { _compression = compression; }
+
+  size_t get_raw_size(void) const { return _raw_size; }
+  void set_raw_size(size_t size) { _raw_size = size; }
+
 private:
   DataType _element_type;
   Shape _shape;
   AffineQuantization _quantization;
   uint8_t *_data = nullptr;
+  // Used for compressed/sparsed tensors when size != WxHxLxD
+  size_t _raw_size{0};
+
   std::string _name;
   bool _data_allocated = false;
   // Write of tensor is reported to registered Observers only if this tensor is observable
@@ -190,6 +201,8 @@ class Tensor
   // Used by static memory manager.
   // Stores the offset from the beginning of the allocated memory buffer.
   int32_t _offset = -1;
+
+  luci::CompressionType _compression{luci::CompressionType::NONE};
 };
 
 } // namespace luci_interpreter