Run CI also on AMD GPUs from NGT

.github/workflows/c-cpp.yml

@@ -56,21 +56,19 @@ jobs:
       run: make BACKEND=${{ matrix.backend }} OMPFLAGS= FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }}
     - name: make test
       run: make BACKEND=${{ matrix.backend }} OMPFLAGS= FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }} -f cudacpp.mk test
-  GPU:
+
+  GPU_Nvidia:
     runs-on: self-hosted
-    # runs-on: madgraph5-h100
-    # container: registry.cern.ch/ngt/lxplus-like:9
     env:
-      CUDA_HOME: /usr/local/cuda/
       FC: gfortran
     strategy:
       matrix:
-        folder: [ epochX/cudacpp/ee_mumu.mad/SubProcesses/P1_epem_mupmum , epochX/cudacpp/gg_ttgg.mad/SubProcesses/P1_gg_ttxgg ]
-        precision: [ d , f , m ]
+        folder: [ epochX/cudacpp/ee_mumu.mad/SubProcesses/P1_epem_mupmum, epochX/cudacpp/gg_ttgg.mad/SubProcesses/P1_gg_ttxgg ]
+        precision: [ d, f, m ]
         backend: [ cppauto, cuda ]
       fail-fast: false
-    steps:
-    - uses: actions/checkout@v2
+    steps: &gpu_steps # allows to recall the steps
+    - uses: actions/checkout@v4
     - name: path
       run: echo "PATH=$PATH"
     - name: github PR info
@@ -81,3 +79,16 @@ jobs:
       run: make BACKEND=${{ matrix.backend }} FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }}
     - name: make test
       run: make BACKEND=${{ matrix.backend }} FPTYPE=${{ matrix.precision }} -C ${{ matrix.folder }} -f cudacpp.mk test
+
+  GPU_AMD:
+    runs-on: madgraph5-mi300x
+    container: registry.cern.ch/ngt/lxplus-like:9
+    env:
+      FC: gfortran
+    strategy:
+      matrix:
+        folder: [ epochX/cudacpp/ee_mumu.mad/SubProcesses/P1_epem_mupmum, epochX/cudacpp/gg_ttgg.mad/SubProcesses/P1_gg_ttxgg ]
+        precision: [ d, f, m ]
+        backend: [ hip ]
+      fail-fast: false
+    steps: *gpu_steps

epochX/cudacpp/CODEGEN/PLUGIN/CUDACPP_SA_OUTPUT/madgraph/iolibs/template_files/gpu/cudacpp.mk

@@ -1106,10 +1106,20 @@ else
 	@cat /proc/cpuinfo | grep "physical id" | sort -u
 endif
 	@echo ""
-ifneq ($(shell which nvidia-smi 2>/dev/null),)
-	nvidia-smi -L
+	# some systems may have <vendor>-smi installed, but it fails to execute
+	# because it cannot communicate with the hardware
+	@if nvidia-smi &>/dev/null ; then \
+		nvidia-smi -L ;\
+	else \
+		echo "No Nvidia GPU detected" ;\
+	fi
 	@echo ""
-endif
+	@if amd-smi &>/dev/null ; then \
+		amd-smi static -a ;\
+		amd-smi list ;\
+	else \
+		echo "No AMD GPU detected" ;\
+	fi
 	@echo USECCACHE=$(USECCACHE)
 ifeq ($(USECCACHE),1)
 	ccache --version | head -1

epochX/cudacpp/ee_mumu.mad/Cards/ident_card.dat

@@ -2,32 +2,32 @@ ccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc
 c      written by the UFO converter
 ccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc
 
+mass 5 mdl_MB
 
+mass 6 mdl_MT
 
+mass 15 mdl_MTA
 
+mass 23 mdl_MZ
 
+mass 25 mdl_MH
 
+sminputs 1 aEWM1
 
+sminputs 2 mdl_Gf
 
+sminputs 3 aS
 
+yukawa 5 mdl_ymb
 
+yukawa 6 mdl_ymt
 
+yukawa 15 mdl_ymtau
 
-
-
+decay 6 mdl_WT
 
 decay 23 mdl_WZ
+
 decay 24 mdl_WW
+
 decay 25 mdl_WH
-decay 6 mdl_WT
-mass 15 mdl_MTA
-mass 23 mdl_MZ
-mass 25 mdl_MH
-mass 5 mdl_MB
-mass 6 mdl_MT
-sminputs 1 aEWM1
-sminputs 2 mdl_Gf
-sminputs 3 aS
-yukawa 15 mdl_ymtau
-yukawa 5 mdl_ymb
-yukawa 6 mdl_ymt

epochX/cudacpp/ee_mumu.mad/Cards/me5_configuration.txt

@@ -235,7 +235,7 @@
 # pineappl = pineappl
 
 
-#mg5_path = /data/avalassi/GPU2023/madgraph4gpuX/MG5aMC/mg5amcnlo 
+mg5_path = /home/dmass/Development/madgraph4gpu/amd-ci/MG5aMC/mg5amcnlo 
 
 # MG5 MAIN DIRECTORY
-#mg5_path = /data/avalassi/GPU2023/madgraph4gpuX/MG5aMC/mg5amcnlo
+mg5_path = /home/dmass/Development/madgraph4gpu/amd-ci/MG5aMC/mg5amcnlo

epochX/cudacpp/ee_mumu.mad/Cards/proc_card_mg5.dat

@@ -8,7 +8,7 @@
 #*                *                       *                 *
 #*                                                          *
 #*                                                          *
-#*         VERSION 3.6.0                 2024-09-30         *
+#*         VERSION 3.6.3                 2025-06-12         *
 #*                                                          *
 #*          WARNING: UNKNOWN DEVELOPMENT VERSION.           *
 #*            WARNING: DO NOT USE FOR PRODUCTION            *
@@ -35,15 +35,17 @@ set loop_color_flows False
 set max_npoint_for_channel 0
 set default_unset_couplings 99
 set max_t_for_channel 99
+set zerowidth_tchannel True
 set nlo_mixed_expansion True
-set stdout_level DEBUG
-set zerowidth_tchannel F
-generate e+ e- > mu+ mu-
+import model sm
 define p = g u c d s u~ c~ d~ s~
 define j = g u c d s u~ c~ d~ s~
 define l+ = e+ mu+
 define l- = e- mu-
 define vl = ve vm vt
 define vl~ = ve~ vm~ vt~
-output madevent_simd ../TMPOUT/CODEGEN_mad_ee_mumu --hel_recycling=Fal\
-se --vector_size=32
+set stdout_level DEBUG
+set zerowidth_tchannel F
+generate e+ e- > mu+ mu-
+output madevent_simd ee_mumu.mad_gen --hel_recycling=False --vector_si\
+ze=32

epochX/cudacpp/ee_mumu.mad/Cards/run_card.dat

@@ -97,11 +97,11 @@
 #*********************************************************************
 # Compilation flag. 
 #*********************************************************************   
-   -O3 -ffast-math -fbounds-check = global_flag ! build flags for all Fortran code (for a fair comparison to cudacpp; default is -O)
+   -O = global_flag ! fortran optimization flag use for the all code.
    --fast-math  = aloha_flag ! fortran optimization flag for aloha function. Suggestions: '-ffast-math'
    -O3 = matrix_flag ! fortran optimization flag for matrix.f function. Suggestions: '-O3'
-   32 = vector_size ! size of fortran arrays allocated in the multi-event API for SIMD/GPU (VECSIZE_MEMMAX)
-   512 = nb_warp ! total number of warp/frontwave
+   16 = vector_size ! size of fortran arrays allocated in the multi-event API for SIMD/GPU (VECSIZE_MEMMAX)
+   1 = nb_warp ! total number of warp/frontwave
 
 #*********************************************************************
 # Customization (custom cuts/scale/bias/...)                         *
@@ -112,6 +112,7 @@
 # Parton level cuts definition *
 #*******************************
   0.0  = dsqrt_shat ! minimal shat for full process
+  -1  = dsqrt_shatmax ! maximum shat for full process
 #                                                                    
 #
 #*********************************************************************

epochX/cudacpp/ee_mumu.mad/Cards/run_card_default.dat

@@ -112,6 +112,7 @@
 # Parton level cuts definition *
 #*******************************
   0.0  = dsqrt_shat ! minimal shat for full process
+  -1  = dsqrt_shatmax ! maximum shat for full process
 #                                                                    
 #
 #*********************************************************************

epochX/cudacpp/ee_mumu.mad/MGMEVersion.txt

@@ -1 +1 @@
-3.6.0
+3.6.3

epochX/cudacpp/ee_mumu.mad/Source/.make_opts

@@ -102,6 +102,7 @@ endif
 ifneq ($(lhapdf),)
   CXXFLAGS += $(shell $(lhapdf) --cppflags)
   alfas_functions=alfas_functions_lhapdf
+  alfas_to_clean=alfas_functions.o
   llhapdf+= $(shell $(lhapdf) --cflags --libs) -lLHAPDF
 # check if we need to activate c++11 (for lhapdf6.2)
   ifeq ($(origin CXX),default)
@@ -113,10 +114,11 @@ ifneq ($(lhapdf),)
   endif	     
 else
   alfas_functions=alfas_functions
+  alfas_to_clean=alfas_functions_lhapdf.o
   llhapdf=
 endif
 
 # Helper function to check MG5 version
 define CHECK_MG5AMC_VERSION
 python -c 'import re; from distutils.version import StrictVersion; print StrictVersion("$(MG5AMC_VERSION)") >= StrictVersion("$(1)") if re.match("^[\d\.]+$$","$(MG5AMC_VERSION)") else True;'
-endef
+endef

epochX/cudacpp/ee_mumu.mad/Source/DHELAS/FFV4_0.f

@@ -9,15 +9,15 @@ SUBROUTINE FFV4_0(F1, F2, V3, COUP,VERTEX)
       COMPLEX*16 COUP
       COMPLEX*16 F1(*)
       COMPLEX*16 F2(*)
-      COMPLEX*16 TMP1
       COMPLEX*16 TMP3
+      COMPLEX*16 TMP4
       COMPLEX*16 V3(*)
       COMPLEX*16 VERTEX
-      TMP1 = (F1(3)*(F2(5)*(V3(3)+V3(6))+F2(6)*(V3(4)+CI*(V3(5))))
+      TMP3 = (F1(3)*(F2(5)*(V3(3)+V3(6))+F2(6)*(V3(4)+CI*(V3(5))))
      $ +F1(4)*(F2(5)*(V3(4)-CI*(V3(5)))+F2(6)*(V3(3)-V3(6))))
-      TMP3 = (F1(5)*(F2(3)*(V3(3)-V3(6))-F2(4)*(V3(4)+CI*(V3(5))))
+      TMP4 = (F1(5)*(F2(3)*(V3(3)-V3(6))-F2(4)*(V3(4)+CI*(V3(5))))
      $ +F1(6)*(F2(3)*(-V3(4)+CI*(V3(5)))+F2(4)*(V3(3)+V3(6))))
-      VERTEX = COUP*(-1D0)*(+CI*(TMP1)+2D0 * CI*(TMP3))
+      VERTEX = COUP*(-1D0)*(+CI*(TMP3)+2D0 * CI*(TMP4))
       END
 
 

epochX/cudacpp/ee_mumu.mad/Source/DHELAS/FFV4_3.f

@@ -13,7 +13,7 @@ SUBROUTINE FFV4_3(F1, F2, COUP, M3, W3,V3)
       REAL*8 OM3
       REAL*8 P3(0:3)
       COMPLEX*16 TMP2
-      COMPLEX*16 TMP4
+      COMPLEX*16 TMP5
       COMPLEX*16 V3(6)
       REAL*8 W3
       COMPLEX*16 DENOM
@@ -27,18 +27,18 @@ SUBROUTINE FFV4_3(F1, F2, COUP, M3, W3,V3)
       P3(3) = -DIMAG(V3(1))
       TMP2 = (F1(3)*(F2(5)*(P3(0)+P3(3))+F2(6)*(P3(1)+CI*(P3(2))))
      $ +F1(4)*(F2(5)*(P3(1)-CI*(P3(2)))+F2(6)*(P3(0)-P3(3))))
-      TMP4 = (F1(5)*(F2(3)*(P3(0)-P3(3))-F2(4)*(P3(1)+CI*(P3(2))))
+      TMP5 = (F1(5)*(F2(3)*(P3(0)-P3(3))-F2(4)*(P3(1)+CI*(P3(2))))
      $ +F1(6)*(F2(3)*(-P3(1)+CI*(P3(2)))+F2(4)*(P3(0)+P3(3))))
       DENOM = COUP/(P3(0)**2-P3(1)**2-P3(2)**2-P3(3)**2 - M3 * (M3 -CI
      $ * W3))
-      V3(3)= DENOM*(-2D0 * CI)*(OM3*-1D0/2D0 * P3(0)*(TMP2+2D0*(TMP4))
+      V3(3)= DENOM*(-2D0 * CI)*(OM3*-1D0/2D0 * P3(0)*(TMP2+2D0*(TMP5))
      $ +(+1D0/2D0*(F1(3)*F2(5)+F1(4)*F2(6))+F1(5)*F2(3)+F1(6)*F2(4)))
-      V3(4)= DENOM*(-2D0 * CI)*(OM3*-1D0/2D0 * P3(1)*(TMP2+2D0*(TMP4))
+      V3(4)= DENOM*(-2D0 * CI)*(OM3*-1D0/2D0 * P3(1)*(TMP2+2D0*(TMP5))
      $ +(-1D0/2D0*(F1(3)*F2(6)+F1(4)*F2(5))+F1(5)*F2(4)+F1(6)*F2(3)))
-      V3(5)= DENOM*2D0 * CI*(OM3*1D0/2D0 * P3(2)*(TMP2+2D0*(TMP4))+(
+      V3(5)= DENOM*2D0 * CI*(OM3*1D0/2D0 * P3(2)*(TMP2+2D0*(TMP5))+(
      $ +1D0/2D0 * CI*(F1(3)*F2(6))-1D0/2D0 * CI*(F1(4)*F2(5))-CI*(F1(5)
      $ *F2(4))+CI*(F1(6)*F2(3))))
-      V3(6)= DENOM*2D0 * CI*(OM3*1D0/2D0 * P3(3)*(TMP2+2D0*(TMP4))+(
+      V3(6)= DENOM*2D0 * CI*(OM3*1D0/2D0 * P3(3)*(TMP2+2D0*(TMP5))+(
      $ +1D0/2D0*(F1(3)*F2(5))-1D0/2D0*(F1(4)*F2(6))-F1(5)*F2(3)+F1(6)
      $ *F2(4)))
       END

epochX/cudacpp/ee_mumu.mad/Source/alfas_functions.f

@@ -188,6 +188,10 @@ SUBROUTINE NEWTON1(T,A_IN,A_OUT,NLOOP,NF)
 
       A_OUT=A_IN/(1D0+A_IN*B0(NF)*T)
       IF (NLOOP .EQ. 1) RETURN
+      if (1D0+A_IN*B0(NF)*T.le.0d0)THEN
+          A_OUT = 9d98
+          RETURN
+      ENDIF
       A_OUT=A_IN/(1D0+B0(NF)*A_IN*T+C1(NF)*A_IN*LOG(1D0+A_IN*B0(NF)*T))
       IF (A_OUT .LT. 0D0) AS=0.3D0
  30   AS=A_OUT

epochX/cudacpp/ee_mumu.mad/Source/cuts.inc

@@ -37,7 +37,7 @@ C
       REAL*8 misset,missetmax,ptheavy
       REAL*8 ptllmin,ptllmax
       integer maxjetflavor
-      REAl*8 dsqrt_shat
+      REAl*8 dsqrt_shat,dsqrt_shatmax
 
       COMMON /to_min_max_cuts/
      &     PTJmax,PTBmax,PTAmax,PTLmax,
@@ -60,7 +60,7 @@ C
      &     ht2max,ht3max,ht4max,
      &     htjmin,htjmax,ihtmin,ihtmax,
      &     misset,missetmax,ptheavy,
-     &     ptllmin,ptllmax,dsqrt_shat,
+     &     ptllmin,ptllmax,dsqrt_shat,dsqrt_shatmax,
      &     maxjetflavor
 
 C

epochX/cudacpp/ee_mumu.mad/Source/make_opts

@@ -1,12 +1,11 @@
 DEFAULT_CPP_COMPILER=g++
-DEFAULT_F2PY_COMPILER=f2py3
-DEFAULT_F_COMPILER=gfortran
-GLOBAL_FLAG=-O3 -ffast-math -fbounds-check
 MACFLAG=
+STDLIB=-lstdc++
+STDLIB_FLAG=
+DEFAULT_F2PY_COMPILER=f2py
+DEFAULT_F_COMPILER=gfortran
 MG5AMC_VERSION=SpecifiedByMG5aMCAtRunTime
 PYTHIA8_PATH=NotInstalled
-STDLIB_FLAG=
-STDLIB=-lstdc++
 #end_of_make_opts_variables
 
 BIASLIBDIR=../../../lib/
@@ -103,6 +102,7 @@ endif
 ifneq ($(lhapdf),)
 CXXFLAGS += $(shell $(lhapdf) --cppflags)
 alfas_functions=alfas_functions_lhapdf
+alfas_to_clean=alfas_functions.o
 llhapdf+= $(shell $(lhapdf) --cflags --libs) -lLHAPDF
 # check if we need to activate c++11 (for lhapdf6.2)
 ifeq ($(origin CXX),default)
@@ -114,6 +114,7 @@ endif
 endif
 else
 alfas_functions=alfas_functions
+alfas_to_clean=alfas_functions_lhapdf.o
 llhapdf=
 endif
 

epochX/cudacpp/ee_mumu.mad/Source/makefile

@@ -37,10 +37,12 @@ all: $(LIBRARIES) $(LIBDIR)libdhelas.$(libext) $(LIBDIR)libpdf.$(libext) $(LIBDI
 $(LIBDIR)libdsample.$(libext): $(DSAMPLE)
 	$(call CREATELIB, $@, $^)
 $(LIBDIR)libgeneric.$(libext): $(GENERIC)
+	rm -f $@ 2>/dev/null
 	$(call CREATELIB, $@, $^)
+	rm -f $(alfas_to_clean) 2>/dev/null
 $(LIBDIR)libdhelas.$(libext): DHELAS
 	cd DHELAS; make; cd ..
-$(LIBDIR)libpdf.$(libext): PDF make_opts
+$(LIBDIR)libpdf.$(libext): PDF $(alfas_functions).o
 	cd PDF; make; cd ..
 ifneq (,$(filter edff chff, $(pdlabel1) $(pdlabel2)))
 $(LIBDIR)libgammaUPC.$(libext): PDF/gammaUPC
@@ -73,6 +75,7 @@ $(BINDIR)gensudgrid: $(GENSUDGRID) $(LIBDIR)libpdf.$(libext) $(LIBDIR)libgammaUP
 # Dependencies
 
 dsample.o: DiscreteSampler.o dsample.f genps.inc StringCast.o vector.inc
+pawgraph.o: vector.inc
 DiscreteSampler.o: StringCast.o
 invarients.o: invarients.f genps.inc
 gen_ximprove.o: gen_ximprove.f run_config.inc run_card.inc 

epochX/cudacpp/ee_mumu.mad/Source/vector.inc

@@ -30,6 +30,8 @@ C
       INTEGER WARP_SIZE
       PARAMETER (WARP_SIZE=32)
       INTEGER NB_WARP
-      PARAMETER (NB_WARP=512)
+      PARAMETER (NB_WARP=1)
       INTEGER VECSIZE_MEMMAX
-      PARAMETER (VECSIZE_MEMMAX=16384)
+      PARAMETER (VECSIZE_MEMMAX=32)
+
+