Merge b1c424d1a6 into 593de54e52

CMakeLists.txt

@@ -0,0 +1,24 @@
+cmake_minimum_required(VERSION 4.0)
+
+project(nccl LANGUAGES CUDA CXX VERSION 2.27.7)
+
+option(VERBOSE "VERBOSE" OFF)
+option(KEEP "KEEP" OFF)
+option(TRACE "TRACE" OFF)
+option(PROFAPI "PROFAPI" OFF)
+option(NVTX "NVTX" ON)
+option(NET_PROFILER "NET_PROFILER" OFF)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+add_subdirectory(src)
+
+install(
+  TARGETS nccl nccl_static
+  EXPORT NCCLConfig
+  FILE_SET public_headers
+  DESTINATION include)
+
+install(
+  EXPORT NCCLConfig
+  DESTINATION lib/cmake/nccl
+  NAMESPACE NCCL::)

cmake/common.cmake

@@ -0,0 +1,39 @@
+function(nccl_add_target_options target)
+  target_compile_options(${target} PRIVATE $<$<CONFIG:Debug>:-ggdb3>)
+  target_compile_options(${target} PRIVATE $<$<NOT:$<CONFIG:Debug>>:-O3>)
+  target_compile_options(
+    ${target} PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--expt-extended-lambda -Xptxas
+                      -maxrregcount=96 -Xfatbin -compress-all -fPIC>)
+  target_compile_options(${target} PRIVATE -fPIC -Wall -Wno-unused-function
+                                           -Wno-sign-compare -Wvla)
+  set_property(TARGET ${target} PROPERTY CXX_STANDARD 17)
+  set_property(TARGET ${target} PROPERTY CUDA_STANDARD 17)
+  set_property(TARGET ${target} PROPERTY CXX_VISIBILITY_PRESET hidden)
+  set_property(TARGET ${target} PROPERTY VISIBILITY_INLINES_HIDDEN 1)
+  set_property(TARGET ${target} PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
+  if(VERBOSE)
+    target_compile_options(${target} PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xptxas
+                                             -v -Xcompiler -Wall,-Wextra>)
+    target_compile_options(${target} PRIVATE -Wall -Wextra)
+  endif()
+
+  if(TRACE)
+    target_compile_options(${target} PRIVATE ENABLE_TRACE)
+  endif()
+
+  if(NOT NVTX)
+    target_compile_options(${target} PRIVATE NVTX_DISABLE)
+  endif()
+
+  if(KEEP)
+    target_compile_options(${target} PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-keep>)
+  endif()
+
+  if(PROFAPI)
+    target_compile_options(${target} PRIVATE PROFAPI)
+  endif()
+
+  if(NET_PROFILER)
+    target_compile_options(${target} PRIVATE NET_PROFILER)
+  endif()
+endfunction()

src/CMakeLists.txt

@@ -0,0 +1,51 @@
+include(../cmake/common.cmake)
+
+find_package(CUDAToolkit REQUIRED)
+
+set(nccl_Major ${nccl_VERSION_MAJOR})
+set(nccl_Minor ${nccl_VERSION_MINOR})
+set(nccl_Patch ${nccl_VERSION_PATCH})
+# NCCL_VERSION(X,Y,Z) ((X) * 10000 + (Y) * 100 + (Z))
+math(
+  EXPR
+  nccl_Version
+  "${nccl_VERSION_MAJOR} * 10000 + ${nccl_VERSION_MINOR} * 100 + ${nccl_VERSION_PATCH}"
+)
+set(nccl_Suffix)
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/nccl.h.in
+               ${CMAKE_CURRENT_SOURCE_DIR}/include/nccl.h)
+
+file(
+  GLOB
+  SRC_FILES
+  "${CMAKE_CURRENT_SOURCE_DIR}/*.cc"
+  "${CMAKE_CURRENT_SOURCE_DIR}/misc/*.cc"
+  "${CMAKE_CURRENT_SOURCE_DIR}/transport/*.cc"
+  "${CMAKE_CURRENT_SOURCE_DIR}/collectives/*.cc"
+  "${CMAKE_CURRENT_SOURCE_DIR}/graph/*.cc")
+
+set(HEADER_FILES "${CMAKE_CURRENT_SOURCE_DIR}/include/nccl.h")
+
+set(NCCL_LIBS nccl;nccl_static)
+
+add_library(nccl SHARED ${SRC_FILES})
+add_library(nccl_static STATIC ${SRC_FILES})
+
+foreach(lib_name IN LISTS NCCL_LIBS)
+  nccl_add_target_options(${lib_name})
+  target_include_directories(
+    ${lib_name}
+    PRIVATE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+            $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include/plugin>)
+  target_include_directories(${lib_name} PRIVATE "${CUDAToolkit_INCLUDE_DIRS}")
+  target_sources(
+    ${lib_name}
+    PUBLIC FILE_SET
+           public_headers
+           TYPE
+           HEADERS
+           BASE_DIRS
+           "${CMAKE_CURRENT_SOURCE_DIR}"
+           FILES
+           ${HEADER_FILES})
+endforeach()

src/device/CMakeLists.txt

@@ -0,0 +1,35 @@
+set(CU_FILES onerank_reduce.cu functions.cu)
+
+add_library(colldevice OBJECT ${CU_FILES})
+
+set(datatypes "i8;u8;i32;u32;i64;u64;f16;f32;f64")
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL "11")
+  list(APPEND datatypes bf16)
+endif()
+set(ops "sum;prod;min;max;premulsum;sumpostdiv")
+list(LENGTH ops op_num)
+math(EXPR op_num "${op_num} - 1")
+list(LENGTH datatypes datatype_num)
+math(EXPR datatype_num "${datatype_num} - 1")
+set(base_files "sendrecv;all_reduce;all_gather;broadcast;reduce;reduce_scatter")
+foreach(base IN LISTS base_files)
+  foreach(opn RANGE ${op_num})
+    list(GET ops ${opn} op)
+    foreach(dtn RANGE ${datatype_num})
+      list(GET datatypes ${dtn} dt)
+      set(new_file ${CMAKE_CURRENT_BINARY_DIR}/${base}_${op}_${dt}.cu)
+      configure_file(${CMAKE_CURRENT_SOURCE_DIR}/${base}.cu ${new_file}
+                     COPYONLY)
+      set_property(SOURCE ${new_file} PROPERTY COMPILE_DEFINITIONS
+                                               NCCL_OP=${opn} NCCL_TYPE=${dtn})
+      target_sources(colldevice PRIVATE ${new_file})
+    endforeach()
+  endforeach()
+endforeach()
+
+target_include_directories(
+  colldevice PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../../include
+                     ${CMAKE_CURRENT_SOURCE_DIR})
+
+# Compiled kernels and collectives with relocatable device code ...
+set_property(TARGET colldevice PROPERTY CUDA_SEPARABLE_COMPILATION ON)

src/nccl.h.in

@@ -16,12 +16,12 @@
 #include <cuda_fp8.h>
 #endif
 
-#define NCCL_MAJOR ${nccl:Major}
-#define NCCL_MINOR ${nccl:Minor}
-#define NCCL_PATCH ${nccl:Patch}
-#define NCCL_SUFFIX "${nccl:Suffix}"
+#define NCCL_MAJOR ${nccl_Major}
+#define NCCL_MINOR ${nccl_Minor}
+#define NCCL_PATCH ${nccl_Patch}
+#define NCCL_SUFFIX "${nccl_Suffix}"
 
-#define NCCL_VERSION_CODE ${nccl:Version}
+#define NCCL_VERSION_CODE ${nccl_Version}
 #define NCCL_VERSION(X,Y,Z) (((X) <= 2 && (Y) <= 8) ? (X) * 1000 + (Y) * 100 + (Z) : (X) * 10000 + (Y) * 100 + (Z))
 
 #ifdef __cplusplus