chore(gpu): add nvtx tool for profiling

backends/tfhe-cuda-backend/build.rs

@@ -39,6 +39,8 @@ fn main() {
         println!("cargo:rustc-link-lib=cudart");
         println!("cargo:rustc-link-search=native=/usr/lib/x86_64-linux-gnu/");
         println!("cargo:rustc-link-lib=stdc++");
+        println!("cargo:rustc-link-lib=nvToolsExt");
+
     } else {
         panic!(
             "Error: platform not supported, tfhe-cuda-backend not built (only Linux is supported)"

backends/tfhe-cuda-backend/cuda/CMakeLists.txt

@@ -68,6 +68,7 @@ endif()
 add_compile_definitions(CUDA_ARCH=${CUDA_ARCH})
 
 # in production, should use -arch=sm_70 --ptxas-options=-v to see register spills -lineinfo for better debugging
+# to use nvtx when profiling -lnvToolsExt 
 set(CMAKE_CUDA_FLAGS
     "${CMAKE_CUDA_FLAGS} -ccbin ${CMAKE_CXX_COMPILER} -O3 \
   -std=c++17 --no-exceptions  --expt-relaxed-constexpr -rdc=true \

backends/tfhe-cuda-backend/cuda/src/utils/helper_profile.cu

@@ -0,0 +1,39 @@
+#include "helper_profile.cuh"
+
+uint32_t adler32(const unsigned char *data)
+{
+   const uint32_t MOD_ADLER = 65521;
+   uint32_t a = 1, b = 0;
+   size_t index;
+   for (index = 0; data[index] != 0; ++index)
+   {
+      a = (a + data[index]*2) % MOD_ADLER;
+      b = (b + a) % MOD_ADLER;
+   }
+   return (b << 16) | a;
+}
+
+void cuda_nvtx_label_with_color(const char *name) {
+   int color_id = adler32((const unsigned char*)name);
+   int r,g,b;
+   r=color_id & 0x000000ff;
+   g=(color_id & 0x000ff000) >> 12;
+   b=(color_id & 0x0ff00000) >> 20;
+   if (r<64 & g<64 & b<64) {
+      r=r*3;
+      g=g*3+64;
+      b=b*4;
+   }
+
+   color_id = 0xff000000 | (r << 16) | (g << 8) | (b);
+   nvtxEventAttributes_t eventAttrib = {0};
+   eventAttrib.version = NVTX_VERSION;
+   eventAttrib.size = NVTX_EVENT_ATTRIB_STRUCT_SIZE;
+   eventAttrib.colorType = NVTX_COLOR_ARGB;
+   eventAttrib.color = color_id;
+   eventAttrib.messageType = NVTX_MESSAGE_TYPE_ASCII;
+   eventAttrib.message.ascii = name;
+   nvtxRangePushEx(&eventAttrib);
+}
+
+

backends/tfhe-cuda-backend/cuda/src/utils/helper_profile.cuh

@@ -0,0 +1,12 @@
+#ifndef HELPER_PROFILE
+#define HELPER_PROFILE
+#include <nvToolsExt.h>
+
+
+void cuda_nvtx_label_with_color(const char *name); 
+
+#define PUSH_RANGE(name) { cuda_nvtx_label_with_color(name);}
+#define POP_RANGE() nvtxRangePop();
+
+
+#endif