another attempt for compiling main dll via gcc, still not working

getkeops · Aug 11, 2021 · d8f2d87 · d8f2d87
1 parent 31f7280
commit d8f2d87
Show file tree

Hide file tree

Showing 3 changed files with 25 additions and 8 deletions.
diff --git a/keops/python_engine/compilation/Gpu_link_compile.py b/keops/python_engine/compilation/Gpu_link_compile.py
@@ -1,7 +1,8 @@
 import os
 from keops.python_engine.compilation import link_compile
 from keops.python_engine.config import build_path
-from ctypes import create_string_buffer, CDLL, c_int
+from ctypes import create_string_buffer, CDLL, c_int, RTLD_GLOBAL
+from os import RTLD_LAZY
 from keops.python_engine import jit_binary, use_cuda
 
 
@@ -32,7 +33,10 @@ def __init__(self):
             + self.low_level_code_extension
         ).encode("utf-8")
         # we load the main dll that must be run in order to compile the code
-        self.my_c_dll = CDLL(jit_binary)
+        CDLL("libnvrtc.so", mode=RTLD_GLOBAL)
+        CDLL("libcuda.so", mode=RTLD_GLOBAL)
+        CDLL("libcudart.so", mode=RTLD_GLOBAL)
+        self.my_c_dll = CDLL(jit_binary, mode=RTLD_LAZY)
         # actual dll to be called is the jit binary
         self.true_dllname = jit_binary
         # file to check for existence to detect compilation is needed

diff --git a/.../python_engine/compilation/keops_nvrtc.cu → ...python_engine/compilation/keops_nvrtc.cpp b/.../python_engine/compilation/keops_nvrtc.cu → ...python_engine/compilation/keops_nvrtc.cpp
@@ -3,6 +3,7 @@
 
 #include <nvrtc.h>
 #include <cuda.h>
+#include <cuda_runtime.h>
 #include <stdio.h>
 #include <iostream>
 #include <sstream>
@@ -23,7 +24,7 @@
 #include "utils.cpp"
 #include "ranges_utils.cpp"
 
-extern "C" __host__ int Compile(const char* ptx_file_name, const char* cu_code, int use_half, int device_id) {
+extern "C"  int Compile(const char* ptx_file_name, const char* cu_code, int use_half, int device_id) {
 
     char *ptx;
 
@@ -100,7 +101,7 @@ extern "C" __host__ int Compile(const char* ptx_file_name, const char* cu_code,
 
 
 template < typename TYPE >
-__host__ int launch_keops(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
+ int launch_keops(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
                                         int device_id, int tagI, int tagZero, int use_half, 
 										int tag1D2D, int dimred,
 										int cuda_block_size, int use_chunk_mode,
@@ -337,7 +338,7 @@ __host__ int launch_keops(const char* ptx_file_name, int tagHostDevice, int dimY
 
 
 
-extern "C" __host__ int launch_keops_float(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
+extern "C"  int launch_keops_float(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
                                         int device_id, int tagI, int tagZero, int use_half, 
 										int tag1D2D, int dimred,
 										int cuda_block_size, int use_chunk_mode,
@@ -370,7 +371,7 @@ extern "C" __host__ int launch_keops_float(const char* ptx_file_name, int tagHos
 
 
 
-extern "C" __host__ int launch_keops_double(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
+extern "C"  int launch_keops_double(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
                                         int device_id, int tagI, int tagZero, int use_half, 
 										int tag1D2D, int dimred,
 										int cuda_block_size, int use_chunk_mode,
@@ -402,7 +403,7 @@ extern "C" __host__ int launch_keops_double(const char* ptx_file_name, int tagHo
 
 
 
-extern "C" __host__ int launch_keops_half(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
+extern "C"  int launch_keops_half(const char* ptx_file_name, int tagHostDevice, int dimY, int nx, int ny, 
                                         int device_id, int tagI, int tagZero, int use_half, 
 										int tag1D2D, int dimred,
 										int cuda_block_size, int use_chunk_mode,

diff --git a/keops/python_engine/config.py b/keops/python_engine/config.py
@@ -16,16 +16,28 @@ def get_jit_binary(gpu_props_compile_flags, check_compile=True):
         + os.path.sep
         + "compilation"
         + os.path.sep
-        + "keops_nvrtc.cu"
+        + "keops_nvrtc.cpp"
     )
     jit_binary = build_path + "keops_nvrtc.so"
     if check_compile and not os.path.exists(jit_binary):
         print("[KeOps] Compiling main dll...", flush=True, end="")
         bindings_source_dir = base_dir_path + "binders"
+
+        # nvcc
         flags = "-shared -Xcompiler -fPIC -lnvrtc -lcuda "
         flags += gpu_props_compile_flags
         # jit_compile_command = f"nvcc -I {bindings_source_dir} {flags} {jit_source_file} -o {jit_binary}"
         jit_compile_command = f"nvcc {flags} {jit_source_file} -o {jit_binary}"
+
+        # g++
+        flags = "-L/usr/lib/x86_64-linux-gnu -L/opt/cuda/lib64 -L/opt/cuda/targets/x86_64-linux/lib/ "
+        flags += "-I/usr/local/cuda-11.0/targets/x86_64-linux/include/ -I/opt/cuda/targets/x86_64-linux/include/ -I/opt/cuda/targets/x86_64-linux/include/ "
+        flags += "-Wl,-rpath,/usr/lib/x86_64-linux-gnu "
+        flags += "-shared -fPIC -fpermissive -lcudart -lcuda -lnvrtc "
+        flags += gpu_props_compile_flags
+        # jit_compile_command = f"nvcc -I {bindings_source_dir} {flags} {jit_source_file} -o {jit_binary}"
+        jit_compile_command = f"g++ --verbose {flags} {jit_source_file} -o {jit_binary}"
+
         os.system(jit_compile_command)
         print("Done.", flush=True)
     return jit_binary