Update nccl and gpu targets, add restore GPU build workflows

browsermt · Jul 30, 2023 · 31bb989 · 31bb989
1 parent 6a6bbb6
commit 31bb989
Show file tree

Hide file tree

Showing 3 changed files with 95 additions and 46 deletions.
diff --git a/.github/workflows/ubuntu.yml b/.github/workflows/ubuntu.yml
@@ -13,7 +13,15 @@ jobs:
         include:
           # Ubuntu 20.04 supports CUDA 11+
           # Unit tests and examples are not compiled to save disk space
-          - name: "Ubuntu 20.04 gcc-9"
+          - name: "Ubuntu 20.04 gcc-9 full"
+            os: ubuntu-20.04
+            cuda: "11.2"
+            gcc: 9
+            cpu: true
+            gpu: true
+            unit_tests: false
+            examples: false
+          - name: "Ubuntu 20.04 gcc-9 CPU"
             os: ubuntu-20.04
             cuda: "11.2"
             gcc: 9
@@ -22,7 +30,16 @@ jobs:
             unit_tests: false
             examples: false
           # Unit tests and examples are not compiled to save disk space
-          - name: "Ubuntu 22.04 gcc-11"
+          - name: "Ubuntu 22.04 gcc-11 full"
+            os: ubuntu-22.04
+            cuda: "11.7"
+            gcc: 11
+            clang: ""
+            cpu: true
+            gpu: true
+            unit_tests: false
+            examples: false
+          - name: "Ubuntu 22.04 gcc-11 CPU"
             os: ubuntu-22.04
             cuda: "11.7"
             gcc: 11

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -436,48 +436,71 @@ if(CUDA_FOUND)
   # We want to compile as many targets as possible but different CUDA versions support different targets.
   # Let's instead enable options based on what cuda version we have.
   if((CUDA_VERSION VERSION_EQUAL "9.0" OR CUDA_VERSION VERSION_GREATER "9.0") AND CUDA_VERSION VERSION_LESS "11.0")
-    option(COMPILE_CUDA_SM35 "Compile GPU version with SM35 support" ON)
-    option(COMPILE_CUDA_SM50 "Compile GPU version with SM50 support" ON)
-    option(COMPILE_CUDA_SM60 "Compile GPU version with SM60 support" ON)
-    option(COMPILE_CUDA_SM70 "Compile GPU version with SM70 support" ON)
+    option(COMPILE_KEPLER  "Compile GPU version with SM35 support" OFF)
+    option(COMPILE_MAXWELL "Compile GPU version with SM50 support" OFF)
+    option(COMPILE_PASCAL  "Compile GPU version with SM60 support" ON)
+    option(COMPILE_VOLTA   "Compile GPU version with SM70 support" ON)
   endif()
   if((CUDA_VERSION VERSION_EQUAL "10.0" OR CUDA_VERSION VERSION_GREATER "10.0") AND CUDA_VERSION VERSION_LESS "11.0")
-    option(COMPILE_CUDA_SM35 "Compile GPU version with SM35 support" ON)
-    option(COMPILE_CUDA_SM50 "Compile GPU version with SM50 support" ON)
-    option(COMPILE_CUDA_SM60 "Compile GPU version with SM60 support" ON)
-    option(COMPILE_CUDA_SM70 "Compile GPU version with SM70 support" ON)
-    option(COMPILE_CUDA_SM75 "Compile GPU version with SM75 support" ON)
+    option(COMPILE_KEPLER  "Compile GPU version with SM35 support" OFF)
+    option(COMPILE_MAXWELL "Compile GPU version with SM50 support" OFF)
+    option(COMPILE_PASCAL  "Compile GPU version with SM60 support" ON)
+    option(COMPILE_VOLTA   "Compile GPU version with SM70 support" ON)
+    option(COMPILE_TURING  "Compile GPU version with SM75 support" ON)
   endif()
   if(CUDA_VERSION VERSION_EQUAL "11.0" OR CUDA_VERSION VERSION_GREATER "11.0")
-    option(COMPILE_CUDA_SM35 "Compile GPU version with SM35 support" ON)
-    option(COMPILE_CUDA_SM50 "Compile GPU version with SM50 support" ON)
-    option(COMPILE_CUDA_SM60 "Compile GPU version with SM60 support" ON)
-    option(COMPILE_CUDA_SM70 "Compile GPU version with SM70 support" ON)
-    option(COMPILE_CUDA_SM75 "Compile GPU version with SM75 support" ON)
-    option(COMPILE_CUDA_SM80 "Compile GPU version with SM80 support" ON)
+    option(COMPILE_KEPLER  "Compile GPU version with SM35 support" OFF) # deprecated for CUDA 11
+    option(COMPILE_MAXWELL "Compile GPU version with SM50 support" OFF) # deprecated for CUDA 11
+    option(COMPILE_PASCAL  "Compile GPU version with SM60 support" ON)
+    option(COMPILE_VOLTA   "Compile GPU version with SM70 support" ON)
+    option(COMPILE_TURING  "Compile GPU version with SM75 support" ON)
+    option(COMPILE_AMPERE  "Compile GPU version with SM80 support" ON)
+    LIST(APPEND COMPUTE -Wno-deprecated-gpu-targets)
+  endif()
+  if(CUDA_VERSION VERSION_EQUAL "11.1" OR CUDA_VERSION VERSION_GREATER "11.1")
+    option(COMPILE_KEPLER      "Compile GPU version with SM35 support" OFF) # deprecated for CUDA 11
+    option(COMPILE_MAXWELL     "Compile GPU version with SM50 support" OFF) # deprecated for CUDA 11
+    option(COMPILE_PASCAL      "Compile GPU version with SM60 support" ON)
+    option(COMPILE_VOLTA       "Compile GPU version with SM70 support" ON)
+    option(COMPILE_TURING      "Compile GPU version with SM75 support" ON)
+    option(COMPILE_AMPERE      "Compile GPU version with SM80 support" ON)
+    option(COMPILE_AMPERE_RTX  "Compile GPU version with SM86 support" ON)
+    LIST(APPEND COMPUTE -Wno-deprecated-gpu-targets)
   endif()
 
-  if(COMPILE_CUDA_SM35)
-    LIST(APPEND COMPUTE -arch=sm_35; -gencode=arch=compute_35,code=sm_35;)                             # Tesla K40 and above
-  endif(COMPILE_CUDA_SM35)
-  if(COMPILE_CUDA_SM50)
+  if(COMPILE_KEPLER)
+    message(STATUS "Compiling code for Kepler GPUs")
+    LIST(APPEND COMPUTE -gencode=arch=compute_35,code=sm_35;) # Tesla K40 and above
+  endif(COMPILE_KEPLER)
+  if(COMPILE_MAXWELL)
+    message(STATUS "Compiling code for Maxwell GPUs")
     LIST(APPEND COMPUTE -gencode=arch=compute_50,code=sm_50; -gencode=arch=compute_52,code=sm_52;)     # Maxwell GPUs
-  endif(COMPILE_CUDA_SM50)
-  if(COMPILE_CUDA_SM60)
+  endif(COMPILE_MAXWELL)
+  if(COMPILE_PASCAL)
+    message(STATUS "Compiling code for Pascal GPUs")
     LIST(APPEND COMPUTE -gencode=arch=compute_60,code=sm_60; -gencode=arch=compute_61,code=sm_61;)     # Pascal GPUs
-  endif(COMPILE_CUDA_SM60)
-  if(COMPILE_CUDA_SM70)
-    LIST(APPEND COMPUTE -gencode=arch=compute_70,code=sm_70; -gencode=arch=compute_70,code=compute_70) # Volta GPUs
-  endif(COMPILE_CUDA_SM70)
+  endif(COMPILE_PASCAL)
+  if(COMPILE_VOLTA)
+    message(STATUS "Compiling code for Volta GPUs")
+    LIST(APPEND COMPUTE -arch=sm_70; -gencode=arch=compute_70,code=sm_70; -gencode=arch=compute_70,code=compute_70) # Volta GPUs
+  endif(COMPILE_VOLTA)
   if(CUDA_VERSION VERSION_EQUAL "10.0" OR CUDA_VERSION VERSION_GREATER "10.0")
-    if(COMPILE_CUDA_SM75)
+    if(COMPILE_TURING)
+        message(STATUS "Compiling code for Turing GPUs")
         LIST(APPEND COMPUTE -gencode=arch=compute_75,code=sm_75; -gencode=arch=compute_75,code=compute_75) # Turing GPUs
-    endif(COMPILE_CUDA_SM75)
+    endif(COMPILE_TURING)
   endif()
   if(CUDA_VERSION VERSION_EQUAL "11.0" OR CUDA_VERSION VERSION_GREATER "11.0")
-    if(COMPILE_CUDA_SM80)
+    if(COMPILE_AMPERE)
+        message(STATUS "Compiling code for Ampere GPUs")
         LIST(APPEND COMPUTE -gencode=arch=compute_80,code=sm_80; -gencode=arch=compute_80,code=compute_80) # Ampere GPUs
-    endif(COMPILE_CUDA_SM80)
+    endif(COMPILE_AMPERE)
+  endif()
+  if(CUDA_VERSION VERSION_EQUAL "11.1" OR CUDA_VERSION VERSION_GREATER "11.1")
+    if(COMPILE_AMPERE_RTX)
+        message(STATUS "Compiling code for Ampere RTX GPUs")
+        LIST(APPEND COMPUTE -gencode=arch=compute_86,code=sm_86; -gencode=arch=compute_86,code=compute_86) # Ampere RTX GPUs
+    endif(COMPILE_AMPERE_RTX)
   endif()
 
   if(USE_STATIC_LIBS)

diff --git a/src/3rd_party/CMakeLists.txt b/src/3rd_party/CMakeLists.txt
@@ -173,34 +173,43 @@ if(CUDA_FOUND)
     # disables compilation for sm_30 to avoid ptxas warning... that is general Kepler support. But K80s are supported for instance by sm_35
 
     set(GENCODE "")
-    if(COMPILE_CUDA_SM35)
+    if(CUDA_VERSION VERSION_EQUAL "11.0" OR CUDA_VERSION VERSION_GREATER "11.0")
+      set(GENCODE "${GENCODE} -Wno-deprecated-gpu-targets")
+    endif()
+    if(COMPILE_KEPLER)
       set(GENCODE "${GENCODE} -gencode=arch=compute_35,code=sm_35")
-    endif(COMPILE_CUDA_SM35)
-    if(COMPILE_CUDA_SM50)
+    endif(COMPILE_KEPLER)
+    if(COMPILE_MAXWELL)
       set(GENCODE "${GENCODE} -gencode=arch=compute_50,code=sm_50")
-    endif(COMPILE_CUDA_SM50)
-    if(COMPILE_CUDA_SM60)
+    endif(COMPILE_MAXWELL)
+    if(COMPILE_PASCAL)
       set(GENCODE "${GENCODE} -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_61,code=sm_61")
-    endif(COMPILE_CUDA_SM60)
-    if(COMPILE_CUDA_SM70)
-      set(GENCODE "${GENCODE} -gencode=arch=compute_70,code=sm_70")
-    endif(COMPILE_CUDA_SM70)
-    if(COMPILE_CUDA_SM75)
+    endif(COMPILE_PASCAL)
+    if(COMPILE_VOLTA)
+      set(GENCODE "${GENCODE} -arch=sm_70 -gencode=arch=compute_70,code=sm_70")
+    endif(COMPILE_VOLTA)
+    if(COMPILE_TURING)
       set(GENCODE "${GENCODE} -gencode=arch=compute_75,code=sm_75; -gencode=arch=compute_75,code=compute_75")
-    endif(COMPILE_CUDA_SM75)
-    if(COMPILE_CUDA_SM80)
+    endif(COMPILE_TURING)
+    if(COMPILE_AMPERE)
       set(GENCODE "${GENCODE} -gencode=arch=compute_80,code=sm_80; -gencode=arch=compute_80,code=compute_80")
-    endif(COMPILE_CUDA_SM80)
+    endif(COMPILE_AMPERE)
 
     # install nccl in ${CMAKE_BINARY_DIR}/local similar to /usr/local linux installation
+    # Using $(MAKE) instead of $CMAKE_MAKE_PROGRAM in order to make parallelization in NCCL compilation work with make -j16.
+    # Apparently this does not get properly propagated otherwise and builts with only a single thread/process.
     ExternalProject_Add(nccl_install
       SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/nccl
       BINARY_DIR ${CMAKE_CURRENT_SOURCE_DIR}/nccl
       CONFIGURE_COMMAND ""
       BUILD_COMMAND
-        ${CMAKE_MAKE_PROGRAM} -f ${CMAKE_CURRENT_SOURCE_DIR}/nccl/Makefile src.build
+        # Note: $(MAKE) here causes CMake cache generation via Ninja failing (Windows build with
+        # CMake only) because Ninja complains about unescaped $ signs. This happens only if
+        # COMPILE_CUDA=on and because of that, USE_NCCL=off is used in CMakeSettings.json for now.
+        # @TODO: find proper escaping that works for MSVC builds.
+        "\$(MAKE)" -f ${CMAKE_CURRENT_SOURCE_DIR}/nccl/Makefile src.build
         BUILDDIR=${CMAKE_BINARY_DIR}/local CUDA_HOME=${CUDA_TOOLKIT_ROOT_DIR}
-        CUDA8_GENCODE=${GENCODE} CXX=${CMAKE_CXX_COMPILER}
+        CUDA8_GENCODE=${GENCODE} CXX=${CMAKE_CXX_COMPILER} CXX_FLAGS=${NCCL_FLAGS}
       INSTALL_COMMAND "")
 
     set_target_properties(nccl PROPERTIES IMPORTED_LOCATION ${CMAKE_BINARY_DIR}/local/lib/libnccl_static.a)