optimization for tiny_yolov4 (#591)

Type: code improvment

Signed-off-by: Chen <jack.chen@verisilicon.com>
Co-authored-by: Chen <jack.chen@verisilicon.com>

cmake/openssl.patch

@@ -61,9 +61,18 @@ index 22ab3cc..00bf90d 100644
 +        DESTINATION ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR})
  endif()
 diff --git a/cmake/BuildOpenSSL.cmake b/cmake/BuildOpenSSL.cmake
-index e31f4bb..a5c5f72 100644
+index e31f4bb..e5e99a6 100644
 --- a/cmake/BuildOpenSSL.cmake
 +++ b/cmake/BuildOpenSSL.cmake
+@@ -29,7 +29,7 @@ include(ExternalProject)
+ 
+ # find packages
+ find_package(Git REQUIRED)
+-find_package(PythonInterp 3 REQUIRED)
++find_package(PythonInterp 2 REQUIRED)
+ 
+ # # used to apply various patches to OpenSSL
+ find_program(PATCH_PROGRAM patch)
 @@ -50,29 +50,25 @@ endif()
  if ((EXISTS ${OPENSSL_LIBSSL_PATH}) AND (EXISTS ${OPENSSL_LIBCRYPTO_PATH}))
      message(WARNING "Not building OpenSSL again. Remove ${OPENSSL_LIBSSL_PATH} and ${OPENSSL_LIBCRYPTO_PATH} for rebuild")

include/tim/vx/graph.h

@@ -27,13 +27,17 @@
 #ifdef BUILD_WITH_BAZEL
 #include "vsi_feat_ops_def.h"
 #endif
-
+#ifdef ENABLE_TENSOR_CACHE
+#include <openssl/evp.h>
+#include <string>
+#endif
 #include <memory>
 #include <vector>
-
 namespace tim {
 namespace vx {
-
+#ifdef ENABLE_TENSOR_CACHE
+const std::string calculateMd5Secret32(const std::string& src);
+#endif
 class Tensor;
 struct TensorSpec;
 struct DmaBufferDesc;

include/tim/vx/ops.h

@@ -99,6 +99,7 @@
 #include "tim/vx/ops/conv3d.h"
 #include "tim/vx/ops/custom_base.h"
 #include "tim/vx/ops/topk.h"
+#include "tim/vx/ops/tiny_yolov4_postprocess.h"
 #include "tim/vx/ops/bidirectional_sequence_lstm.h"
 #include "tim/vx/ops/hashtable_lookup.h"
 #include "tim/vx/ops/embedding_lookup.h"

include/tim/vx/ops/tiny_yolov4_postprocess.h

@@ -0,0 +1,44 @@
+/****************************************************************************
+*
+*    Copyright (c) 2020-2023 Vivante Corporation
+*
+*    Permission is hereby granted, free of charge, to any person obtaining a
+*    copy of this software and associated documentation files (the "Software"),
+*    to deal in the Software without restriction, including without limitation
+*    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+*    and/or sell copies of the Software, and to permit persons to whom the
+*    Software is furnished to do so, subject to the following conditions:
+*
+*    The above copyright notice and this permission notice shall be included in
+*    all copies or substantial portions of the Software.
+*
+*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+*    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+*    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+*    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+*    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+*    DEALINGS IN THE SOFTWARE.
+*
+*****************************************************************************/
+#ifndef TIM_VX_OPS_TINY_YOLOV4_H_
+#define TIM_VX_OPS_TINY_YOLOV4_H_
+#include "tim/vx/builtin_op.h"
+#ifdef VSI_FEAT_OP_CUSTOM_TINY_YOLOV4_POSTPROCESS
+namespace tim {
+namespace vx {
+namespace ops {
+
+class TinyYolov4Postprocess : public BuiltinOp {
+ public:
+  TinyYolov4Postprocess(Graph* graph);
+
+  std::shared_ptr<Operation> Clone(std::shared_ptr<Graph>& graph) const override;
+};
+
+}  // namespace ops
+}  // namespace vx
+}  // namespace tim
+
+#endif //(FEAT_OP_CUSTOM_TINY_YOLOV4_POSTPROCESS)
+#endif /* TIM_VX_OPS_TINY_YOLOV4_H_ */

src/tim/CMakeLists.txt

@@ -126,15 +126,21 @@ include_directories(${INC_DIRS})
 # convert op list as compile flags so that we can implement compile compatable easier
 if(${TIM_VX_USE_EXTERNAL_OVXLIB})
     file(STRINGS "${OVXLIB_INC}/interface/ops.def" ops_file_content)
+    file(STRINGS "${OVXLIB_INC}/custom/custom_ops.def" custom_ops_file_content)
 else()
     file(STRINGS "./vx/internal/include/interface/ops.def" ops_file_content)
+    file(STRINGS "./vx/internal/include/custom/custom_ops.def" custom_ops_file_content)
 endif()
 string(LENGTH "/*;   Add new ops to the end.;*_/" comment_len)
 string(SUBSTRING "${ops_file_content}" ${comment_len} -1 op_list_only)
 string(REGEX REPLACE "DEF_OP\\(" "-DVSI_FEAT_OP_" op_list_tmp ${op_list_only})
 string(REGEX REPLACE "\\)" " " op_as_flags ${op_list_tmp})
 
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${op_as_flags}")
+string(LENGTH "/*;   Add custom ops to the end.;*_/" custom_comment_len)
+string(SUBSTRING "${custom_ops_file_content}" ${custom_comment_len} -1 custom_op_list_only)
+string(REGEX REPLACE "DEF_OP\\(" "-DVSI_FEAT_OP_" custom_op_list_tmp ${custom_op_list_only})
+string(REGEX REPLACE "\\)" " " custom_op_as_flags ${custom_op_list_tmp})
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${op_as_flags} ${custom_op_as_flags}")
 
 add_library(${TARGET_NAME} ${${TARGET_NAME}_SRCS})
 target_include_directories(${TARGET_NAME} PRIVATE ${INC_DIRS})
@@ -170,7 +176,13 @@ endif()
 # Install
 install(TARGETS ${TARGET_NAME} ${TARGET_NAME}
 	DESTINATION ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR})
-
+    
+if(TIM_VX_ENABLE_TENSOR_CACHE)
+install(
+    FILES
+        ${openssl-cmake_BINARY_DIR}/usr/local/include/openssl/evp.h
+    DESTINATION ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_INCLUDEDIR}/openssl)
+endif()
 install(
     FILES
         ${CMAKE_SOURCE_DIR}/include/tim/vx/builtin_op.h

src/tim/transform/layout_inference.cc

@@ -63,6 +63,7 @@
 #include "ops/conv3d_layout_inference.h"
 #include "ops/default_layout_inference.h"
 #include "ops/transpose_layout_inference.h"
+#include "ops/yolov4_layout_inference.h"
 #include "ops/unidirectional_lstm_layout_inference.h"
 #include "ops/broadcast_layout_inference.h"
 #include "ops/unidirectional_rnn_layout_inference.h"
@@ -274,6 +275,9 @@ std::vector<std::shared_ptr<vx::Tensor>> HandleLayoutInfer(
     REGIST_LAYOUT_INFERENCE(VSI_NN_OP_EXPAND_BROADCAST, Broadcast);
     REGIST_LAYOUT_INFERENCE(VSI_NN_OP_UNIDIRECTIONAL_SEQUENCE_RNN, UnidirectionalRnn);
     REGIST_LAYOUT_INFERENCE(VSI_NN_OP_BIDIRECTIONAL_SEQUENCE_RNN, BidirectionalRnn);
+#ifdef ENABLE_TENSOR_CACHE
+    REGIST_LAYOUT_INFERENCE(VSI_NN_OP_CUSTOM_TINY_YOLOV4_POSTPROCESS, Yolov4);
+#endif
     REGIST_LOGICAL_LAYOUT_INFERENCE(VSI_NN_OP_LOGICAL_OPS);
     REGIST_REDUCE_LAYOUT_INFERENCE(VSI_NN_OP_REDUCE);
     // use default layout inference

src/tim/transform/ops/yolov4_layout_inference.h

@@ -0,0 +1,77 @@
+/****************************************************************************
+ *
+ *    Copyright (c) 2020-2023 Vivante Corporation
+ *
+ *    Permission is hereby granted, free of charge, to any person obtaining a
+ *    copy of this software and associated documentation files (the "Software"),
+ *    to deal in the Software without restriction, including without limitation
+ *    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ *    and/or sell copies of the Software, and to permit persons to whom the
+ *    Software is furnished to do so, subject to the following conditions:
+ *
+ *    The above copyright notice and this permission notice shall be included in
+ *    all copies or substantial portions of the Software.
+ *
+ *    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ *    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ *    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ *    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ *    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ *    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ *    DEALINGS IN THE SOFTWARE.
+ *
+ *****************************************************************************/
+#ifndef TIM_LAYOUT_INFER_YOLOV4_LAYOUT_INFERENCE_H_
+#define TIM_LAYOUT_INFER_YOLOV4_LAYOUT_INFERENCE_H_
+
+#include "tim/vx/ops/tiny_yolov4_postprocess.h"
+
+#include "ops/op_layout_inference.h"
+#include "permute_vector.h"
+#include "builtin_op_impl.h"
+
+namespace tim {
+namespace transform {
+
+class Yolov4LayoutInfer : public OpLayoutInfer {
+ public:
+  Yolov4LayoutInfer(
+      const std::shared_ptr<vx::Operation> op,
+      std::shared_ptr<layout_inference_impl::LayoutInferContext>& context)
+      : OpLayoutInfer(op, context) {}
+
+  void OnInputs(
+      std::vector<std::shared_ptr<vx::Tensor>>& next_tensors) override {
+    auto cloned_op = op_->Clone(context_->infer_graph_);
+
+    for (const auto& i_src : op_->impl()->InputsTensor()) {
+      if (i_src->IsConstTensor()) {
+        std::vector<uint8_t> dataRef(i_src->GetSpec().GetByteSize());
+        i_src->CopyDataFromTensor(dataRef.data());
+        auto i_infer = context_->infer_graph_->CreateTensor(
+            i_src->GetSpec(), (const void*)dataRef.data());
+        context_->SetPermuteVector(i_src, MakeShared(4));
+        context_->UpdateTensorMap(i_src, i_infer);
+      } 
+      (*cloned_op).BindInput(context_->GetMapedTensor(i_src));
+    }
+
+    std::vector<std::shared_ptr<IPermuteVector>> required_pv_lst;
+    for (auto out_tensor: op_->impl()->OutputsTensor()) {
+      required_pv_lst.push_back(MakeShared(out_tensor->GetShape().size()));
+    }
+    auto out_infer = CreateOutputsTensor(required_pv_lst);
+
+    (*cloned_op).BindOutputs(out_infer);
+    uint32_t i = 0;
+    for (auto out_tensor : op_->impl()->OutputsTensor()) {
+      context_->SetPermuteVector(out_tensor, required_pv_lst[i++]);
+      next_tensors.push_back(out_tensor);
+    }
+  }
+};
+
+}  // namespace transform
+}  // namespace tim
+
+#endif

src/tim/vx/graph.cc

@@ -26,6 +26,7 @@
 
 #ifdef ENABLE_TENSOR_CACHE
 #include <openssl/evp.h>
+#include <cstring>
 #endif
 
 #include "context_private.h"
@@ -39,6 +40,44 @@
 
 namespace tim {
 namespace vx {
+#ifdef ENABLE_TENSOR_CACHE
+#define MD5_SECRET_LEN_16 (16)
+#define MD5_BYTE_STRING_LEN (4)
+const std::string calculateMd5Secret32(const std::string& src) {
+    std::string md5String;
+    EVP_MD_CTX *mdctx;
+    const EVP_MD *md;
+    uint32_t md_len;
+    unsigned char md_value[MD5_SECRET_LEN_16] = {0};
+    char tmp[MD5_BYTE_STRING_LEN] = {0};
+
+    md = EVP_md5();
+    if (md == NULL) {
+      printf("Unknown EVP_md5 message.");
+    }
+    mdctx = EVP_MD_CTX_new();
+    if (!EVP_DigestInit_ex(mdctx, md, NULL)) {
+      printf("EVP_MD_CTX initialization failed.");
+      EVP_MD_CTX_free(mdctx);
+    }
+    if (!EVP_DigestUpdate(mdctx, src.c_str(), src.size())) {
+      printf("EVP_MD_CTX update failed.");
+      EVP_MD_CTX_free(mdctx);
+    }
+    if (!EVP_DigestFinal_ex(mdctx, md_value, &md_len)) {
+      printf("EVP_MD_CTX finalization failed.");
+      EVP_MD_CTX_free(mdctx);
+    }
+    EVP_MD_CTX_free(mdctx);
+
+    for (int i = 0; i < MD5_SECRET_LEN_16; ++i) {
+      memset(tmp, 0x00, sizeof(tmp));
+      snprintf(tmp, sizeof(tmp), "%02X", md_value[i]);
+      md5String += tmp;
+    }
+    return md5String;
+  }
+#endif
 
 const std::vector<std::shared_ptr<Tensor>> Graph::GetConstantInputs() const {
     std::vector<std::shared_ptr<Tensor>> const_inputs;
@@ -64,44 +103,7 @@ std::map<std::string, std::shared_ptr<tim::vx::Tensor>>& GraphImpl::GetTensorCac
   return cached_tensor_;
 }
 
-#define MD5_SECRET_LEN_16 (16)
-#define MD5_BYTE_STRING_LEN (4)
-const std::string GraphImpl::caclulateMd5Secret32(const std::string& src) {
-  std::string md5String;
-  EVP_MD_CTX *mdctx;
-  const EVP_MD *md;
-  uint32_t md_len;
-  unsigned char md_value[MD5_SECRET_LEN_16] = {0};
-  char tmp[MD5_BYTE_STRING_LEN] = {0};
-
-  md = EVP_md5();
-  if (md == NULL) {
-    VSILOGE("Unknown EVP_md5 message.");
-  }
-  mdctx = EVP_MD_CTX_new();
-  if (!EVP_DigestInit_ex(mdctx, md, NULL)) {
-    VSILOGE("EVP_MD_CTX initialization failed.");
-    EVP_MD_CTX_free(mdctx);
-  }
-  if (!EVP_DigestUpdate(mdctx, src.c_str(), src.size())) {
-    VSILOGE("EVP_MD_CTX update failed.");
-    EVP_MD_CTX_free(mdctx);
-  }
-  if (!EVP_DigestFinal_ex(mdctx, md_value, &md_len)) {
-    VSILOGE("EVP_MD_CTX finalization failed.");
-    EVP_MD_CTX_free(mdctx);
-  }
-  EVP_MD_CTX_free(mdctx);
-
-  for (int i = 0; i < MD5_SECRET_LEN_16; ++i) {
-    memset(tmp, 0x00, sizeof(tmp));
-    snprintf(tmp, sizeof(tmp), "%02X", md_value[i]);
-    md5String += tmp;
-  }
-  return md5String;
-}
-
-const std::string GraphImpl::CaclulateCacheKey(const TensorSpec& spec, const void* data) {
+const std::string GraphImpl::CalculateCacheKey(const TensorSpec& spec, const void* data) {
   std::string md5_key;
   uint32_t data_size = 1;
   for (auto it = spec.shape_.begin(); it != spec.shape_.end(); ++it) {
@@ -125,9 +127,9 @@ const std::string GraphImpl::CaclulateCacheKey(const TensorSpec& spec, const voi
       break;
   }
   if (data_size < 512) {
-    md5_key = caclulateMd5Secret32(std::string((const char*)data, data_size));
+    md5_key = calculateMd5Secret32(std::string((const char*)data, data_size));
   } else {
-    md5_key = caclulateMd5Secret32(
+    md5_key = calculateMd5Secret32(
         std::string((const char*)data, 512));  //Take first 512 bytes
   }
   return md5_key;
@@ -135,7 +137,7 @@ const std::string GraphImpl::CaclulateCacheKey(const TensorSpec& spec, const voi
 
 std::shared_ptr<Tensor> GraphImpl::GetTensorFromCache(const TensorSpec& spec, const void* data) {
   std::shared_ptr<tim::vx::Tensor> tensor;
-  std::string md5_key = CaclulateCacheKey(spec, data);
+  std::string md5_key = CalculateCacheKey(spec, data);
   if (GetTensorCacheMap().find(md5_key) != GetTensorCacheMap().end() &&
       GetTensorCacheMap()[md5_key]->GetQuantization().Scales() == spec.quantization_.Scales() &&
       GetTensorCacheMap()[md5_key]->GetQuantization().ZeroPoints() == spec.quantization_.ZeroPoints()) {

src/tim/vx/graph_private.h

@@ -46,8 +46,7 @@ class GraphImpl : public Graph {
   ~GraphImpl();
 #ifdef ENABLE_TENSOR_CACHE
   std::shared_ptr<Tensor> GetTensorFromCache(const TensorSpec& spec, const void* data);
-  const std::string CaclulateCacheKey(const TensorSpec& spec, const void* data);
-  const std::string caclulateMd5Secret32(const std::string& src);
+  const std::string CalculateCacheKey(const TensorSpec& spec, const void* data);
   std::map<std::string, std::shared_ptr<tim::vx::Tensor>>& GetTensorCacheMap();
 #endif
   /// Return the low-level graph object

src/tim/vx/ops/tiny_yolov4_postprocess.cc

@@ -0,0 +1,41 @@
+/****************************************************************************
+*
+*    Copyright (c) 2020-2023 Vivante Corporation
+*
+*    Permission is hereby granted, free of charge, to any person obtaining a
+*    copy of this software and associated documentation files (the "Software"),
+*    to deal in the Software without restriction, including without limitation
+*    the rights to use, copy, modify, merge, publish, distribute, sublicense,
+*    and/or sell copies of the Software, and to permit persons to whom the
+*    Software is furnished to do so, subject to the following conditions:
+*
+*    The above copyright notice and this permission notice shall be included in
+*    all copies or substantial portions of the Software.
+*
+*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+*    IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+*    FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+*    AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+*    LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+*    DEALINGS IN THE SOFTWARE.
+*
+*****************************************************************************/
+#include "tim/vx/ops/tiny_yolov4_postprocess.h"
+
+#include "builtin_op_impl.h"
+#include "vsi_nn_pub.h"
+#ifdef VSI_FEAT_OP_CUSTOM_TINY_YOLOV4_POSTPROCESS
+namespace tim {
+namespace vx {
+namespace ops {
+TinyYolov4Postprocess::TinyYolov4Postprocess(Graph* graph)
+    : BuiltinOp(graph, VSI_NN_OP_CUSTOM_TINY_YOLOV4_POSTPROCESS, 4, 2){}
+
+std::shared_ptr<Operation> TinyYolov4Postprocess::Clone(std::shared_ptr<Graph>& graph) const {
+  return graph->CreateOperation<TinyYolov4Postprocess>();
+}
+}  // namespace ops
+}  // namespace vx
+}  // namespace tim
+#endif //(FEAT_OP_CUSTOM_TINY_YOLOV4_POSTPROCESS)