Merge branch 'grid' (including also upstream/master merge with hel ma…

…dgraph5#960, mac madgraph5#974, nvcc madgraph5#966) into cmsdy Fix conflict in tlau/fromgridpacks/parseGridpackLogs.sh (use the currenmt cmsdy version: git checkout b125b65 tlau/fromgridpacks/parseGridpackLogs.sh)
valassi · Aug 22, 2024 · 5223f79 · 5223f79
2 parents ea87504 + 289a973
commit 5223f79
Show file tree

Hide file tree

Showing 37 changed files with 1,436 additions and 282 deletions.
diff --git a/.github/workflows/c-cpp.yml b/.github/workflows/c-cpp.yml
@@ -38,7 +38,7 @@ jobs:
   CPU_MAC:
     runs-on: macos-latest
     env:
-      FC: gfortran-11
+      FC: gfortran-14 # see #971
     strategy:
       matrix:
         folder: [ epochX/cudacpp/ee_mumu.mad/SubProcesses/P1_epem_mupmum, epochX/cudacpp/gg_ttgg.mad/SubProcesses/P1_gg_ttxgg ]

diff --git a/epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/launch_plugin.py b/epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/launch_plugin.py
@@ -37,7 +37,7 @@ def compile(self, *args, **opts):
             misc.sprint('FPTYPE checked')
         cudacpp_supported_backends = [ 'fortran', 'cuda', 'hip', 'cpp', 'cppnone', 'cppsse4', 'cppavx2', 'cpp512y', 'cpp512z', 'cppauto' ]
         if args and args[0][0] == 'madevent' and hasattr(self, 'run_card'):            
-            if self.run_card['cudacpp_bldall'] == 'True': # pre-build all backends #945
+            if self.run_card['cudacpp_bldall'] == True: # pre-build all backends #945
                 logger.info("Pre-building madevent in madevent_interface.py with ALL matrix elements")
                 args[0][0] = 'bldall'
                 misc.compile(nb_core=self.options['nb_core'], *args, **opts)
@@ -104,11 +104,10 @@ def default_setup(self):
                        fct_mod=(self.reset_makeopts,(),{}), # AV: I assume this forces a 'make cleanavx' if FPTYPE changes?
                        allowed=['m','d','f']
                        )
-        self.add_param('cudacpp_bldall', 'False',
+        self.add_param('cudacpp_bldall', False,
                        include=False, # AV: no need to add this parameter to run_card.inc
                        hidden=False, # AV: add cudacpp_bldall to runcard template and keep 'hidden='False'
                        fct_mod=(self.reset_makeopts,(),{}), # AV: I assume this will raise an exception if cudacpp_bldall changes?
-                       allowed=['False','True']
                        )
         self['vector_size'] = 16 # already setup in default class (just change value)
         self['aloha_flag'] = '--fast-math'

diff --git a/...hX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/madgraph/iolibs/template_files/gpu/cudacpp.mk b/...hX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/madgraph/iolibs/template_files/gpu/cudacpp.mk
@@ -116,15 +116,31 @@ override CUDA_HOME = $(patsubst %%/bin/nvcc,%%,$(shell which nvcc 2>/dev/null))
 # Set HIP_HOME from the path to hipcc, if it exists
 override HIP_HOME = $(patsubst %%/bin/hipcc,%%,$(shell which hipcc 2>/dev/null))
 
-# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists
-# (FIXME? Is there any equivalent of NVTX FOR HIP? What should be configured if both CUDA and HIP are installed?)
-ifneq ($(CUDA_HOME),)
-  USE_NVTX ?=-DUSE_NVTX
-  CUDA_INC = -I$(CUDA_HOME)/include/
+# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists (see #965)
+ifeq ($(CUDA_HOME),)
+  # CUDA_HOME is empty (nvcc not found)
+  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/),)
+  # CUDA_HOME is defined (nvcc was found) but $(CUDA_HOME)/include/ does not exist?
+  override CUDA_INC=
 else
+  CUDA_INC = -I$(CUDA_HOME)/include/
+endif
+###$(info CUDA_INC=$(CUDA_INC))
+
+# Configure NVTX if a CUDA include directory exists and NVTX headers exist (see #965)
+ifeq ($(CUDA_INC),)
+  # $(CUDA_HOME)/include/ does not exist
   override USE_NVTX=
-  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/nvtx3/nvToolsExt.h),)
+  # $(CUDA_HOME)/include/ exists but NVTX headers do not exist?
+  override USE_NVTX=
+else
+  # $(CUDA_HOME)/include/nvtx.h exists: use NVTX
+  # (NB: the option to disable NVTX if 'USE_NVTX=' is defined has been removed)
+  override USE_NVTX=-DUSE_NVTX
 endif
+###$(info USE_NVTX=$(USE_NVTX))
 
 # NB: NEW LOGIC FOR ENABLING AND DISABLING CUDA OR HIP BUILDS (AV Feb-Mar 2024)
 # - In the old implementation, by default the C++ targets for one specific AVX were always built together with either CUDA or HIP.
@@ -424,13 +440,18 @@ endif
 # (NB: allow HASCURAND=hasCurand even if $(GPUCC) does not point to nvcc: assume CUDA_HOME was defined correctly...)
 ifeq ($(HASCURAND),)
   ifeq ($(GPUCC),) # CPU-only build
-    ifneq ($(CUDA_HOME),)
+    ifeq ($(CUDA_INC),)
+      # $(CUDA_HOME)/include/ does not exist (see #965)
+      override HASCURAND = hasNoCurand
+    else ifeq ($(wildcard $(CUDA_HOME)/include/curand.h),)
+      # $(CUDA_HOME)/include/ exists but CURAND headers do not exist? (see #965)
+      override HASCURAND = hasNoCurand
+    else
       # By default, assume that curand is installed if a CUDA installation exists
       override HASCURAND = hasCurand
-    else
-      override HASCURAND = hasNoCurand
     endif
   else ifeq ($(findstring nvcc,$(GPUCC)),nvcc) # Nvidia GPU build
+    # By default, assume that curand is installed if a CUDA build is requested
     override HASCURAND = hasCurand
   else # non-Nvidia GPU build
     override HASCURAND = hasNoCurand

diff --git a/epochX/cudacpp/ee_mumu.mad/SubProcesses/cudacpp.mk b/epochX/cudacpp/ee_mumu.mad/SubProcesses/cudacpp.mk
@@ -116,15 +116,31 @@ override CUDA_HOME = $(patsubst %/bin/nvcc,%,$(shell which nvcc 2>/dev/null))
 # Set HIP_HOME from the path to hipcc, if it exists
 override HIP_HOME = $(patsubst %/bin/hipcc,%,$(shell which hipcc 2>/dev/null))
 
-# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists
-# (FIXME? Is there any equivalent of NVTX FOR HIP? What should be configured if both CUDA and HIP are installed?)
-ifneq ($(CUDA_HOME),)
-  USE_NVTX ?=-DUSE_NVTX
-  CUDA_INC = -I$(CUDA_HOME)/include/
+# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists (see #965)
+ifeq ($(CUDA_HOME),)
+  # CUDA_HOME is empty (nvcc not found)
+  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/),)
+  # CUDA_HOME is defined (nvcc was found) but $(CUDA_HOME)/include/ does not exist?
+  override CUDA_INC=
 else
+  CUDA_INC = -I$(CUDA_HOME)/include/
+endif
+###$(info CUDA_INC=$(CUDA_INC))
+
+# Configure NVTX if a CUDA include directory exists and NVTX headers exist (see #965)
+ifeq ($(CUDA_INC),)
+  # $(CUDA_HOME)/include/ does not exist
   override USE_NVTX=
-  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/nvtx3/nvToolsExt.h),)
+  # $(CUDA_HOME)/include/ exists but NVTX headers do not exist?
+  override USE_NVTX=
+else
+  # $(CUDA_HOME)/include/nvtx.h exists: use NVTX
+  # (NB: the option to disable NVTX if 'USE_NVTX=' is defined has been removed)
+  override USE_NVTX=-DUSE_NVTX
 endif
+###$(info USE_NVTX=$(USE_NVTX))
 
 # NB: NEW LOGIC FOR ENABLING AND DISABLING CUDA OR HIP BUILDS (AV Feb-Mar 2024)
 # - In the old implementation, by default the C++ targets for one specific AVX were always built together with either CUDA or HIP.
@@ -424,13 +440,18 @@ endif
 # (NB: allow HASCURAND=hasCurand even if $(GPUCC) does not point to nvcc: assume CUDA_HOME was defined correctly...)
 ifeq ($(HASCURAND),)
   ifeq ($(GPUCC),) # CPU-only build
-    ifneq ($(CUDA_HOME),)
+    ifeq ($(CUDA_INC),)
+      # $(CUDA_HOME)/include/ does not exist (see #965)
+      override HASCURAND = hasNoCurand
+    else ifeq ($(wildcard $(CUDA_HOME)/include/curand.h),)
+      # $(CUDA_HOME)/include/ exists but CURAND headers do not exist? (see #965)
+      override HASCURAND = hasNoCurand
+    else
       # By default, assume that curand is installed if a CUDA installation exists
       override HASCURAND = hasCurand
-    else
-      override HASCURAND = hasNoCurand
     endif
   else ifeq ($(findstring nvcc,$(GPUCC)),nvcc) # Nvidia GPU build
+    # By default, assume that curand is installed if a CUDA build is requested
     override HASCURAND = hasCurand
   else # non-Nvidia GPU build
     override HASCURAND = hasNoCurand

diff --git a/epochX/cudacpp/ee_mumu.sa/SubProcesses/cudacpp.mk b/epochX/cudacpp/ee_mumu.sa/SubProcesses/cudacpp.mk
@@ -116,15 +116,31 @@ override CUDA_HOME = $(patsubst %/bin/nvcc,%,$(shell which nvcc 2>/dev/null))
 # Set HIP_HOME from the path to hipcc, if it exists
 override HIP_HOME = $(patsubst %/bin/hipcc,%,$(shell which hipcc 2>/dev/null))
 
-# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists
-# (FIXME? Is there any equivalent of NVTX FOR HIP? What should be configured if both CUDA and HIP are installed?)
-ifneq ($(CUDA_HOME),)
-  USE_NVTX ?=-DUSE_NVTX
-  CUDA_INC = -I$(CUDA_HOME)/include/
+# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists (see #965)
+ifeq ($(CUDA_HOME),)
+  # CUDA_HOME is empty (nvcc not found)
+  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/),)
+  # CUDA_HOME is defined (nvcc was found) but $(CUDA_HOME)/include/ does not exist?
+  override CUDA_INC=
 else
+  CUDA_INC = -I$(CUDA_HOME)/include/
+endif
+###$(info CUDA_INC=$(CUDA_INC))
+
+# Configure NVTX if a CUDA include directory exists and NVTX headers exist (see #965)
+ifeq ($(CUDA_INC),)
+  # $(CUDA_HOME)/include/ does not exist
   override USE_NVTX=
-  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/nvtx3/nvToolsExt.h),)
+  # $(CUDA_HOME)/include/ exists but NVTX headers do not exist?
+  override USE_NVTX=
+else
+  # $(CUDA_HOME)/include/nvtx.h exists: use NVTX
+  # (NB: the option to disable NVTX if 'USE_NVTX=' is defined has been removed)
+  override USE_NVTX=-DUSE_NVTX
 endif
+###$(info USE_NVTX=$(USE_NVTX))
 
 # NB: NEW LOGIC FOR ENABLING AND DISABLING CUDA OR HIP BUILDS (AV Feb-Mar 2024)
 # - In the old implementation, by default the C++ targets for one specific AVX were always built together with either CUDA or HIP.
@@ -424,13 +440,18 @@ endif
 # (NB: allow HASCURAND=hasCurand even if $(GPUCC) does not point to nvcc: assume CUDA_HOME was defined correctly...)
 ifeq ($(HASCURAND),)
   ifeq ($(GPUCC),) # CPU-only build
-    ifneq ($(CUDA_HOME),)
+    ifeq ($(CUDA_INC),)
+      # $(CUDA_HOME)/include/ does not exist (see #965)
+      override HASCURAND = hasNoCurand
+    else ifeq ($(wildcard $(CUDA_HOME)/include/curand.h),)
+      # $(CUDA_HOME)/include/ exists but CURAND headers do not exist? (see #965)
+      override HASCURAND = hasNoCurand
+    else
       # By default, assume that curand is installed if a CUDA installation exists
       override HASCURAND = hasCurand
-    else
-      override HASCURAND = hasNoCurand
     endif
   else ifeq ($(findstring nvcc,$(GPUCC)),nvcc) # Nvidia GPU build
+    # By default, assume that curand is installed if a CUDA build is requested
     override HASCURAND = hasCurand
   else # non-Nvidia GPU build
     override HASCURAND = hasNoCurand

diff --git a/epochX/cudacpp/gg_tt.mad/SubProcesses/cudacpp.mk b/epochX/cudacpp/gg_tt.mad/SubProcesses/cudacpp.mk
@@ -116,15 +116,31 @@ override CUDA_HOME = $(patsubst %/bin/nvcc,%,$(shell which nvcc 2>/dev/null))
 # Set HIP_HOME from the path to hipcc, if it exists
 override HIP_HOME = $(patsubst %/bin/hipcc,%,$(shell which hipcc 2>/dev/null))
 
-# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists
-# (FIXME? Is there any equivalent of NVTX FOR HIP? What should be configured if both CUDA and HIP are installed?)
-ifneq ($(CUDA_HOME),)
-  USE_NVTX ?=-DUSE_NVTX
-  CUDA_INC = -I$(CUDA_HOME)/include/
+# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists (see #965)
+ifeq ($(CUDA_HOME),)
+  # CUDA_HOME is empty (nvcc not found)
+  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/),)
+  # CUDA_HOME is defined (nvcc was found) but $(CUDA_HOME)/include/ does not exist?
+  override CUDA_INC=
 else
+  CUDA_INC = -I$(CUDA_HOME)/include/
+endif
+###$(info CUDA_INC=$(CUDA_INC))
+
+# Configure NVTX if a CUDA include directory exists and NVTX headers exist (see #965)
+ifeq ($(CUDA_INC),)
+  # $(CUDA_HOME)/include/ does not exist
   override USE_NVTX=
-  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/nvtx3/nvToolsExt.h),)
+  # $(CUDA_HOME)/include/ exists but NVTX headers do not exist?
+  override USE_NVTX=
+else
+  # $(CUDA_HOME)/include/nvtx.h exists: use NVTX
+  # (NB: the option to disable NVTX if 'USE_NVTX=' is defined has been removed)
+  override USE_NVTX=-DUSE_NVTX
 endif
+###$(info USE_NVTX=$(USE_NVTX))
 
 # NB: NEW LOGIC FOR ENABLING AND DISABLING CUDA OR HIP BUILDS (AV Feb-Mar 2024)
 # - In the old implementation, by default the C++ targets for one specific AVX were always built together with either CUDA or HIP.
@@ -424,13 +440,18 @@ endif
 # (NB: allow HASCURAND=hasCurand even if $(GPUCC) does not point to nvcc: assume CUDA_HOME was defined correctly...)
 ifeq ($(HASCURAND),)
   ifeq ($(GPUCC),) # CPU-only build
-    ifneq ($(CUDA_HOME),)
+    ifeq ($(CUDA_INC),)
+      # $(CUDA_HOME)/include/ does not exist (see #965)
+      override HASCURAND = hasNoCurand
+    else ifeq ($(wildcard $(CUDA_HOME)/include/curand.h),)
+      # $(CUDA_HOME)/include/ exists but CURAND headers do not exist? (see #965)
+      override HASCURAND = hasNoCurand
+    else
       # By default, assume that curand is installed if a CUDA installation exists
       override HASCURAND = hasCurand
-    else
-      override HASCURAND = hasNoCurand
     endif
   else ifeq ($(findstring nvcc,$(GPUCC)),nvcc) # Nvidia GPU build
+    # By default, assume that curand is installed if a CUDA build is requested
     override HASCURAND = hasCurand
   else # non-Nvidia GPU build
     override HASCURAND = hasNoCurand

diff --git a/epochX/cudacpp/gg_tt.mad/bin/internal/launch_plugin.py b/epochX/cudacpp/gg_tt.mad/bin/internal/launch_plugin.py
@@ -37,7 +37,7 @@ def compile(self, *args, **opts):
             misc.sprint('FPTYPE checked')
         cudacpp_supported_backends = [ 'fortran', 'cuda', 'hip', 'cpp', 'cppnone', 'cppsse4', 'cppavx2', 'cpp512y', 'cpp512z', 'cppauto' ]
         if args and args[0][0] == 'madevent' and hasattr(self, 'run_card'):            
-            if self.run_card['cudacpp_bldall'] == 'True': # pre-build all backends #945
+            if self.run_card['cudacpp_bldall'] == True: # pre-build all backends #945
                 logger.info("Pre-building madevent in madevent_interface.py with ALL matrix elements")
                 args[0][0] = 'bldall'
                 misc.compile(nb_core=self.options['nb_core'], *args, **opts)
@@ -104,11 +104,10 @@ def default_setup(self):
                        fct_mod=(self.reset_makeopts,(),{}), # AV: I assume this forces a 'make cleanavx' if FPTYPE changes?
                        allowed=['m','d','f']
                        )
-        self.add_param('cudacpp_bldall', 'False',
+        self.add_param('cudacpp_bldall', False,
                        include=False, # AV: no need to add this parameter to run_card.inc
                        hidden=False, # AV: add cudacpp_bldall to runcard template and keep 'hidden='False'
                        fct_mod=(self.reset_makeopts,(),{}), # AV: I assume this will raise an exception if cudacpp_bldall changes?
-                       allowed=['False','True']
                        )
         self['vector_size'] = 16 # already setup in default class (just change value)
         self['aloha_flag'] = '--fast-math'

diff --git a/epochX/cudacpp/gg_tt.sa/SubProcesses/cudacpp.mk b/epochX/cudacpp/gg_tt.sa/SubProcesses/cudacpp.mk
@@ -116,15 +116,31 @@ override CUDA_HOME = $(patsubst %/bin/nvcc,%,$(shell which nvcc 2>/dev/null))
 # Set HIP_HOME from the path to hipcc, if it exists
 override HIP_HOME = $(patsubst %/bin/hipcc,%,$(shell which hipcc 2>/dev/null))
 
-# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists
-# (FIXME? Is there any equivalent of NVTX FOR HIP? What should be configured if both CUDA and HIP are installed?)
-ifneq ($(CUDA_HOME),)
-  USE_NVTX ?=-DUSE_NVTX
-  CUDA_INC = -I$(CUDA_HOME)/include/
+# Configure CUDA_INC (for CURAND and NVTX) and NVTX if a CUDA installation exists (see #965)
+ifeq ($(CUDA_HOME),)
+  # CUDA_HOME is empty (nvcc not found)
+  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/),)
+  # CUDA_HOME is defined (nvcc was found) but $(CUDA_HOME)/include/ does not exist?
+  override CUDA_INC=
 else
+  CUDA_INC = -I$(CUDA_HOME)/include/
+endif
+###$(info CUDA_INC=$(CUDA_INC))
+
+# Configure NVTX if a CUDA include directory exists and NVTX headers exist (see #965)
+ifeq ($(CUDA_INC),)
+  # $(CUDA_HOME)/include/ does not exist
   override USE_NVTX=
-  override CUDA_INC=
+else ifeq ($(wildcard $(CUDA_HOME)/include/nvtx3/nvToolsExt.h),)
+  # $(CUDA_HOME)/include/ exists but NVTX headers do not exist?
+  override USE_NVTX=
+else
+  # $(CUDA_HOME)/include/nvtx.h exists: use NVTX
+  # (NB: the option to disable NVTX if 'USE_NVTX=' is defined has been removed)
+  override USE_NVTX=-DUSE_NVTX
 endif
+###$(info USE_NVTX=$(USE_NVTX))
 
 # NB: NEW LOGIC FOR ENABLING AND DISABLING CUDA OR HIP BUILDS (AV Feb-Mar 2024)
 # - In the old implementation, by default the C++ targets for one specific AVX were always built together with either CUDA or HIP.
@@ -424,13 +440,18 @@ endif
 # (NB: allow HASCURAND=hasCurand even if $(GPUCC) does not point to nvcc: assume CUDA_HOME was defined correctly...)
 ifeq ($(HASCURAND),)
   ifeq ($(GPUCC),) # CPU-only build
-    ifneq ($(CUDA_HOME),)
+    ifeq ($(CUDA_INC),)
+      # $(CUDA_HOME)/include/ does not exist (see #965)
+      override HASCURAND = hasNoCurand
+    else ifeq ($(wildcard $(CUDA_HOME)/include/curand.h),)
+      # $(CUDA_HOME)/include/ exists but CURAND headers do not exist? (see #965)
+      override HASCURAND = hasNoCurand
+    else
       # By default, assume that curand is installed if a CUDA installation exists
       override HASCURAND = hasCurand
-    else
-      override HASCURAND = hasNoCurand
     endif
   else ifeq ($(findstring nvcc,$(GPUCC)),nvcc) # Nvidia GPU build
+    # By default, assume that curand is installed if a CUDA build is requested
     override HASCURAND = hasCurand
   else # non-Nvidia GPU build
     override HASCURAND = hasNoCurand