Implementation for xcorr_kernel callback set

.gitignore

@@ -6,4 +6,7 @@ src/rnnoise_data.h
 src/rnnoise_data.c
 CMakeUserPresets.json
 .vscode/
-sample_audios/
+.DS_Store
+sample_audios/
+evaluation_output/
+rnnoise_env/

CMakeLists.txt

@@ -3,6 +3,15 @@ project(rnnoise)
 
 option(RNNOISE_COMPILE_OPUS OFF)
 option(RNNOISE_COMPILE_DEMO ON)
+option(RNNOISE_BAREMETAL OFF)
+
+if(NOT RNNOISE_BAREMETAL)
+        find_package(mimalloc REQUIRED)
+endif()
+
+if ("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fprofile-instr-generate -fcoverage-mapping")
+endif()
 
 if(RNNOISE_COMPILE_OPUS)
 	add_definitions(-DCOMPILE_OPUS)
@@ -68,7 +77,7 @@ set(RN_NOISE_SRC
         src/opus_types.h
         src/pitch.h
         src/rnn.h
-        src/rnnoise_data.h
+        src/rnnoise_data_little.h
         include/rnnoise.h
         src/vec.h
 
@@ -79,7 +88,7 @@ set(RN_NOISE_SRC
             src/denoise.c
             src/kiss_fft.c
             src/parse_lpcnet_weights.c
-            src/rnnoise_data.c
+            src/rnnoise_data_little.c
             #write_weights.c
             #dump_features.c
             src/nnet.c
@@ -96,8 +105,16 @@ target_include_directories(rnnoise PUBLIC
         $<INSTALL_INTERFACE:include>
         PRIVATE src)
 
+
+if(NOT RNNOISE_BAREMETAL)
+        target_link_libraries(rnnoise PUBLIC mimalloc-static)
+        target_compile_definitions(rnnoise PUBLIC USE_MIMALLOC_ALLOCATOR)
+endif()
+
 if (UNIX)
         target_link_libraries(rnnoise PUBLIC m)
 endif(UNIX)
 
-add_subdirectory(examples)
+if(RNNOISE_COMPILE_DEMO)
+        add_subdirectory(examples)
+endif()

README

@@ -118,3 +118,8 @@ model, rename rnnoise_data_little.c to rnnoise_data.c. It is possible
 to build both the regular and little binary weights and load any of them
 at run time since the little model has the same size as the regular one
 (except for the increased sparsity).
+
+## Build with Vitis Xilinx toolchain:
+```shell
+cmake -G"Unix Makefiles" -DCMAKE_BUILD_TYPE="Release" DCMAKE_TOOLCHAIN_FILE="${PWD}/cmake/vitis_toolchain.cmake" -S="$PWD}" -B="${PWD}/build" -DRNNOISE_COMPILE_DEMO=OFF DRNNOISE_BAREMETAL=ON
+```

conanfile.txt

@@ -2,7 +2,7 @@
 libsndfile/1.2.2
 cxxopts/3.2.0
 fmt/10.2.1
-spdlog/1.14.0
+mimalloc/2.1.2
 
 [generators]
 CMakeDeps

examples/CMakeLists.txt

@@ -7,10 +7,13 @@ target_link_libraries(rnnoise_demo PRIVATE rnnoise)
 find_package(SndFile REQUIRED)
 find_package(cxxopts REQUIRED)
 find_package(fmt REQUIRED)
-find_package(spdlog REQUIRED)
+find_package(mimalloc REQUIRED)
 
-
-add_executable(rnnoise_libsoundfile rnnoise_libsndfile.cpp)
+add_executable(rnnoise_libsoundfile
+    rnnoise_libsndfile.cpp
+    profiling/xcorr_impl.cpp
+    profiling/xcorr_offload_kernel.cpp
+)
 
 if(CMAKE_SYSTEM_NAME STREQUAL "Windows")
     message(WARNING "Compiling RNNOISE for Windows...")
@@ -24,6 +27,6 @@ target_link_libraries(
     SndFile::sndfile
     cxxopts::cxxopts
     fmt::fmt
-    spdlog::spdlog
+    mimalloc-static
 )
 target_include_directories(rnnoise_libsoundfile PRIVATE ${CMAKE_CURRENT_LIST_DIR})

examples/README.md

@@ -5,14 +5,11 @@ mkdir build
 
 ## For debug version of the app
 conan install conanfile.txt --build=missing --settings=build_type=Debug
-cmake -G"Unix Makefiles" -B=build --preset=conan-debug .
-
+cmake -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=./Debug/generators/conan_toolchain.cmake -DCMAKE_BUILD_TYPE=Debug -B build
+cmake --build build
 
 ## For Release version:
 conan install conanfile.txt --build=missing
-cd build
-cmake -G"Unix Makefiles" -B=build --preset=conan-release .
-
-
-cmake --build .
+cmake -G"Unix Makefiles" -DCMAKE_TOOLCHAIN_FILE=./Release/generators/conan_toolchain.cmake -DCMAKE_BUILD_TYPE=Release -B build
+cmake --build buil
 ```

examples/profiling/xcorr_impl.cpp

@@ -0,0 +1,83 @@
+#include "xcorr_impl.h"
+#include <cassert>
+#include <fmt/core.h>
+#include <chrono>
+
+inline float MAC16_16(float c, float a, float b){
+    return c+a*b;
+}
+
+void xcorr_native_impl(const float * x, const float * y, float sum[4], int len)
+{    
+   int j;
+   float y_0, y_1, y_2, y_3;
+   assert(len>=3);
+   fmt::print("Xcorr native impl called with len {} \n",len);
+   auto start = std::chrono::high_resolution_clock::now();
+
+   y_3=0; /* gcc doesn't realize that y_3 can't be used uninitialized */
+   y_0=*y++;
+   y_1=*y++;
+   y_2=*y++;
+   for (j=0;j<len-3;j+=4)
+   {
+      float tmp;
+      tmp = *x++;
+      y_3=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_0);
+      sum[1] = MAC16_16(sum[1],tmp,y_1);
+      sum[2] = MAC16_16(sum[2],tmp,y_2);
+      sum[3] = MAC16_16(sum[3],tmp,y_3);
+      tmp=*x++;
+      y_0=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_1);
+      sum[1] = MAC16_16(sum[1],tmp,y_2);
+      sum[2] = MAC16_16(sum[2],tmp,y_3);
+      sum[3] = MAC16_16(sum[3],tmp,y_0);
+      tmp=*x++;
+      y_1=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_2);
+      sum[1] = MAC16_16(sum[1],tmp,y_3);
+      sum[2] = MAC16_16(sum[2],tmp,y_0);
+      sum[3] = MAC16_16(sum[3],tmp,y_1);
+      tmp=*x++;
+      y_2=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_3);
+      sum[1] = MAC16_16(sum[1],tmp,y_0);
+      sum[2] = MAC16_16(sum[2],tmp,y_1);
+      sum[3] = MAC16_16(sum[3],tmp,y_2);
+   }
+   if (j++<len)
+   {
+      float tmp = *x++;
+      y_3=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_0);
+      sum[1] = MAC16_16(sum[1],tmp,y_1);
+      sum[2] = MAC16_16(sum[2],tmp,y_2);
+      sum[3] = MAC16_16(sum[3],tmp,y_3);
+   }
+   if (j++<len)
+   {
+      float tmp=*x++;
+      y_0=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_1);
+      sum[1] = MAC16_16(sum[1],tmp,y_2);
+      sum[2] = MAC16_16(sum[2],tmp,y_3);
+      sum[3] = MAC16_16(sum[3],tmp,y_0);
+   }
+   if (j<len)
+   {
+      float tmp=*x++;
+      y_1=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_2);
+      sum[1] = MAC16_16(sum[1],tmp,y_3);
+      sum[2] = MAC16_16(sum[2],tmp,y_0);
+      sum[3] = MAC16_16(sum[3],tmp,y_1);
+   }
+
+   auto end = std::chrono::high_resolution_clock::now();
+
+   auto duration = std::chrono::duration_cast<std::chrono::microseconds>(end - start);
+
+   fmt::print("Execution time is:{} us\n",duration.count());
+}

examples/profiling/xcorr_impl.h

@@ -0,0 +1,10 @@
+#ifndef XCORR_IMPL_H
+#define XCORR_IMPL_H
+
+extern "C"
+{
+
+void xcorr_native_impl(const float * x, const float * y, float sum[4], int len);
+
+}
+#endif

examples/profiling/xcorr_offload_kernel.cpp

@@ -0,0 +1,103 @@
+#include <cassert>
+#include <cstring>
+#include "xcorr_offload_kernel.hpp"
+
+#include <stdio.h>
+
+inline float MAC16_16(float c, float a, float b){ 
+    return c+a*b; 
+} 
+
+#define MAX_PROCESSING_BLOCK_SIZE 1024
+void xcorr_kernel(const float * x, const float * y, float* sum, int len) 
+{ 
+
+#pragma HLS INTERFACE mode=s_axilite port=return bundle=CONTROL_BUS 
+#pragma HLS INTERFACE mode=s_axilite port=len bundle=CONTROL_BUS 
+
+#pragma HLS INTERFACE m_axi port=x offset=slave bundle=INPUT 
+#pragma HLS INTERFACE m_axi port=y offset=slave bundle=INPUT 
+#pragma HLS INTERFACE m_axi port=sum offset=slave bundle=OUTPUT 
+
+   int j; 
+   float y_0, y_1, y_2, y_3; 
+   assert(len>=3); 
+
+   float x_copy[MAX_PROCESSING_BLOCK_SIZE]; 
+   float y_copy[MAX_PROCESSING_BLOCK_SIZE];
+
+   float sum_copy[4]; 
+
+   float* x_arr_ptr = x_copy; 
+   float* y_arr_ptr = y_copy; 
+
+   memcpy(x_copy, x, len); 
+   memcpy(y_copy, y, len); 
+   memcpy(sum_copy, sum, 4); 
+
+   y_3=0; /* gcc doesn't realize that y_3 can't be used uninitialized */ 
+   y_0=*y_arr_ptr++; 
+   y_1=*y_arr_ptr++; 
+   y_2=*y_arr_ptr++; 
+
+xcoor_1st_loop: 
+   for (j=0;j<len-3;j+=4) 
+   { 
+      float tmp; 
+      tmp = *x_arr_ptr++; 
+      y_3=*y_arr_ptr++; 
+      sum_copy[0] = MAC16_16(sum_copy[0],tmp,y_0); 
+      sum_copy[1] = MAC16_16(sum_copy[1],tmp,y_1); 
+      sum_copy[2] = MAC16_16(sum_copy[2],tmp,y_2); 
+      sum_copy[3] = MAC16_16(sum_copy[3],tmp,y_3); 
+      tmp=*x_arr_ptr++; 
+      y_0=*y_arr_ptr++; 
+      sum_copy[0] = MAC16_16(sum_copy[0],tmp,y_1); 
+      sum_copy[1] = MAC16_16(sum_copy[1],tmp,y_2); 
+      sum_copy[2] = MAC16_16(sum_copy[2],tmp,y_3); 
+      sum_copy[3] = MAC16_16(sum_copy[3],tmp,y_0); 
+      tmp=*x_arr_ptr++; 
+      y_1=*y_arr_ptr++; 
+      sum_copy[0] = MAC16_16(sum_copy[0],tmp,y_2); 
+      sum_copy[1] = MAC16_16(sum_copy[1],tmp,y_3); 
+      sum_copy[2] = MAC16_16(sum_copy[2],tmp,y_0); 
+      sum_copy[3] = MAC16_16(sum_copy[3],tmp,y_1); 
+      tmp=*x_arr_ptr++; 
+      y_2=*y_arr_ptr++; 
+      sum_copy[0] = MAC16_16(sum_copy[0],tmp,y_3); 
+      sum_copy[1] = MAC16_16(sum_copy[1],tmp,y_0); 
+      sum_copy[2] = MAC16_16(sum_copy[2],tmp,y_1); 
+      sum_copy[3] = MAC16_16(sum_copy[3],tmp,y_2); 
+   } 
+xcoor_2nd_loop: 
+   if (j++<len) 
+   { 
+      float tmp = *x_arr_ptr++; 
+      y_3=*y_arr_ptr++; 
+      sum_copy[0] = MAC16_16(sum_copy[0],tmp,y_0); 
+      sum_copy[1] = MAC16_16(sum_copy[1],tmp,y_1); 
+      sum_copy[2] = MAC16_16(sum_copy[2],tmp,y_2); 
+      sum_copy[3] = MAC16_16(sum_copy[3],tmp,y_3); 
+   } 
+xcoor_3rd_loop: 
+   if (j++<len) 
+   { 
+      float tmp=*x_arr_ptr++; 
+      y_0=*y_arr_ptr++; 
+      sum_copy[0] = MAC16_16(sum_copy[0],tmp,y_1); 
+      sum_copy[1] = MAC16_16(sum_copy[1],tmp,y_2); 
+      sum_copy[2] = MAC16_16(sum_copy[2],tmp,y_3); 
+      sum_copy[3] = MAC16_16(sum_copy[3],tmp,y_0); 
+   } 
+xcoor_4th_loop: 
+   if (j<len) 
+   { 
+      float tmp=*x_arr_ptr++; 
+      y_1=*y_arr_ptr++; 
+      sum_copy[0] = MAC16_16(sum_copy[0],tmp,y_2); 
+      sum_copy[1] = MAC16_16(sum_copy[1],tmp,y_3); 
+      sum_copy[2] = MAC16_16(sum_copy[2],tmp,y_0); 
+      sum_copy[3] = MAC16_16(sum_copy[3],tmp,y_1); 
+   } 
+   memcpy(sum,sum_copy,4); 
+}

examples/profiling/xcorr_offload_kernel.hpp

@@ -0,0 +1,6 @@
+#ifndef XCORR_OFFLOAD_KERNEL_HPP
+#define XCORR_OFFLOAD_KERNEL_HPP
+
+void xcorr_kernel(const float * x, const float * y, float* sum, int len);
+
+#endif

examples/rnnoise_demo.c

@@ -59,7 +59,6 @@ int main(int argc, char **argv) {
     if (!first) fwrite(tmp, sizeof(short), FRAME_SIZE, fout);
     first = 0;
   }
-  rnnoise_destroy(st);
   fclose(f1);
   fclose(fout);
 #ifdef USE_WEIGHTS_FILE