use cpu_features to detect Granite Rapids and newer

.github/workflows/tests_archdetect.yml

@@ -12,6 +12,8 @@ jobs:
     strategy:
       matrix:
         proc_cpuinfo:
+          - x86_64/intel/graniterapids/AWS-Rocky9-6975PC
+          - x86_64/intel/graniterapids/Azure-Ubuntu24-6973P-C
           - x86_64/intel/haswell/archspec-linux-E5-2680-v3
           - x86_64/intel/sapphirerapids/AWS-Rocky8-8488C
           - x86_64/intel/skylake_avx512/archspec-linux-6132
@@ -50,6 +52,9 @@ jobs:
           export EESSI_MACHINE_TYPE=${{matrix.proc_cpuinfo}}
           export EESSI_MACHINE_TYPE=${EESSI_MACHINE_TYPE%%/*}
           export EESSI_PROC_CPUINFO=./tests/archdetect/${{matrix.proc_cpuinfo}}.cpuinfo
+          if [[ -f ./tests/archdetect/${{matrix.proc_cpuinfo}}.cpu_features ]]; then
+              export EESSI_CPU_FEATURES_FILE=./tests/archdetect/${{matrix.proc_cpuinfo}}.cpu_features
+          fi
           # check that printing of best match works correctly
           CPU_ARCH=$(./init/eessi_archdetect.sh cpupath)
           if [[ $CPU_ARCH == "$( cat ./tests/archdetect/${{matrix.proc_cpuinfo}}.output )" ]]; then

init/arch_specs/eessi_arch_x86.spec

@@ -1,12 +1,13 @@
 # x86_64 CPU architecture specifications
 # The overview at https://github.com/InstLatx64/InstLatX64_Misc/tree/main/SIMD_Euler may be helpful in defining this list
 # Software path in EESSI 	| Vendor ID 	| List of defining CPU features
-"x86_64/intel/haswell"		"GenuineIntel"	"avx2 fma"									# Intel Haswell, Broadwell
-"x86_64/intel/skylake_avx512"	"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl"				# Intel Skylake
-"x86_64/intel/cascadelake"	"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl avx512_vnni"		# Intel Cascade Lake
-"x86_64/intel/icelake"		"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl avx512_vnni avx512_vbmi2"	# Intel Icelake Lake
-"x86_64/intel/sapphirerapids"	"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl avx512_bf16 amx_tile"	# Intel Sapphire/Emerald Rapids
-"x86_64/amd/zen2"		"AuthenticAMD"	"avx2 fma"									# AMD Rome
-"x86_64/amd/zen3"		"AuthenticAMD"	"avx2 fma vaes"									# AMD Milan, Milan-X
-"x86_64/amd/zen4"		"AuthenticAMD"	"avx2 fma vaes avx512f avx512ifma"						# AMD Genoa, Genoa-X
-"x86_64/amd/zen5"		"AuthenticAMD"	"avx2 fma vaes avx512f avx512ifma avx512_vp2intersect avx_vnni movdiri"		# AMD Turin
+"x86_64/intel/haswell"		"GenuineIntel"	"avx2 fma"										# Intel Haswell, Broadwell
+"x86_64/intel/skylake_avx512"	"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl"					# Intel Skylake
+"x86_64/intel/cascadelake"	"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl avx512_vnni"			# Intel Cascade Lake
+"x86_64/intel/icelake"		"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl avx512_vnni avx512_vbmi2"		# Intel Icelake Lake
+"x86_64/intel/sapphirerapids"	"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl avx512_bf16 amx_tile"		# Intel Sapphire/Emerald Rapids
+"x86_64/intel/graniterapids"	"GenuineIntel"	"avx2 fma avx512f avx512bw avx512cd avx512dq avx512vl avx512_bf16 amx_tile amx_fp16"	# Granite
+"x86_64/amd/zen2"		"AuthenticAMD"	"avx2 fma"										# AMD Rome
+"x86_64/amd/zen3"		"AuthenticAMD"	"avx2 fma vaes"										# AMD Milan, Milan-X
+"x86_64/amd/zen4"		"AuthenticAMD"	"avx2 fma vaes avx512f avx512ifma"							# AMD Genoa, Genoa-X
+"x86_64/amd/zen5"		"AuthenticAMD"	"avx2 fma vaes avx512f avx512ifma avx512_vp2intersect avx_vnni movdiri"			# AMD Turin

init/eessi_archdetect.sh

@@ -24,6 +24,16 @@ LOG_LEVEL="WARN"
 # Default result type is a best match
 CPUPATH_RESULT="best"
 
+# map CPU flag names from cpu_features to the ones used in /proc/cpuinfo
+declare -A cpu_flags_name_map=(
+    [fma3]=fma
+    [avx512bitalg]=avx512_bitalg
+    [avx512vbmi2]=avx512_vbmi2
+    [avx512vnni]=avx512_vnni
+    [avx512vpopcntdq]=avx512_vpopcntdq
+    [sha]=sha_ni
+)
+
 timestamp () {
     date "+%Y-%m-%d %H:%M:%S"
 }
@@ -70,6 +80,50 @@ get_cpuinfo(){
     grep -i "$cpuinfo_pattern" ${EESSI_PROC_CPUINFO:-/proc/cpuinfo} | tail -n 1 | sed "s/$cpuinfo_pattern//i"
 }
 
+# CPU specification of host system as reported by cpu_features
+get_cpu_features(){
+    # Return the value from list_cpu_features for the matching key
+    # 1: string with key pattern
+
+    [ -z "$1" ] && log "ERROR" "get_cpu_features: missing key pattern in argument list"
+    cpu_features_pattern="^${1}\s*:\s*"
+
+    if [ ! -z ${EESSI_CPU_FEATURES_FILE} ]; then
+        cpu_features_output=$(cat ${EESSI_CPU_FEATURES_FILE})
+    elif command -v "list_cpu_features" >/dev/null 2>&1; then
+        cpu_features_output=$(list_cpu_features)
+    else
+        log "DEBUG" "cpu_features cannot be found"
+        return 1
+    fi
+
+    # /proc/cpuinfo uses space-separated flags and uses different names for some flags,
+    # so we reformat and rename things a bit here to make it easier to do the comparisons
+    cpu_features_reformatted=$(
+        while IFS= read -r line; do
+            if [[ $line =~ ^(flags[[:space:]]*:[[:space:]]*)(.*)$ ]]; then
+                prefix=${BASH_REMATCH[1]}
+                flags=${BASH_REMATCH[2]}
+
+                new_flags=""
+                IFS=',' read -ra flag_array <<< "$flags"
+
+                for flag in "${flag_array[@]}"; do
+                    new_flags+=" ${cpu_flags_name_map[$flag]:-$flag}"
+                done
+
+                printf '%s%s\n' "$prefix" "${new_flags# }"
+            else
+                printf '%s\n' "$line"
+            fi
+        done <<< "${cpu_features_output}"
+    )
+
+    # case insensitive match of key pattern and delete key pattern from result
+    echo "${cpu_features_reformatted}" | grep -i "$cpu_features_pattern" | tail -n 1 | sed "s/$cpu_features_pattern//i"
+    return 0
+}
+
 check_allinfirst(){
     # Return true if all given arguments after the first are found in the first one
     # 1: reference string of space separated values
@@ -168,26 +222,42 @@ cpupath(){
             # each flag in this CPU specification must be found in the list of flags of the host
             check_allinfirst "${cpu_flags[*]}" ${arch_spec[2]} && best_arch_match=${arch_spec[0]} && \
                 all_arch_matches="$best_arch_match:$all_arch_matches" && \
-                log "DEBUG" "cpupath: host CPU best match updated to $best_arch_match" 
+                log "DEBUG" "cpupath: host CPU best match updated to $best_arch_match"
         fi
     done
 
     # Some Intel microarchitectures are flag-indistinguishable from an older one because their
     # new features are not exposed in /proc/cpuinfo. Granite Rapids (Xeon 6) shows the exact same
     # visible flags as Sapphire/Emerald Rapids (its extras like amx_fp16 are hidden by the kernel),
-    # so the flag match above lands on 'sapphirerapids'. Refine using the CPU model number - the
-    # only reliable discriminator on Linux. If no dedicated graniterapids subdir is shipped yet,
-    # downstream subdir resolution falls back to the next entry in the chain, so prepending is safe.
+    # so the flag match above lands on 'sapphirerapids'.
+    # Refine using the list_cpu_features tool from Google's cpu_features if available,
+    # and otherwise use the CPU model number.
     if [ "${best_arch_match}" == "x86_64/intel/sapphirerapids" ]; then
         local cpu_family=$(get_cpuinfo "cpu[ _]family")
         local cpu_model=$(get_cpuinfo "model")
         log "DEBUG" "cpupath: refining Sapphire Rapids match (family='$cpu_family', model='$cpu_model')"
+        if cpu_features_flags=$(get_cpu_features "$cpu_flag_tag"); then
+            log "DEBUG" "Flags reported by list_cpu_features: ${cpu_features_flags}"
+            # Now that we have a (possibly) changed set of flags, reset the results to their defaults and
+            # reiterate over the supported CPU specifications to find the best match for host CPU.
+            # Order of the specifications matters, the last one to match will be selected
+            local best_arch_match="$machine_type/generic"
+            local all_arch_matches=$best_arch_match
+            for arch in "${cpu_arch_spec[@]}"; do
+                eval "arch_spec=$arch"
+                if [ "${cpu_vendor}x" == "${arch_spec[1]}x" ]; then
+                    # each flag in this CPU specification must be found in the list of flags of the host
+                    check_allinfirst "${cpu_features_flags[*]}" ${arch_spec[2]} && best_arch_match=${arch_spec[0]} && \
+                        all_arch_matches="$best_arch_match:$all_arch_matches" && \
+                        log "DEBUG" "cpupath: host CPU best match updated to $best_arch_match"
+                fi
+            done
         # Intel family 6 model numbers below come from the kernel's authoritative table
         # arch/x86/include/asm/intel-family.h (what the kernel itself uses for model dispatch):
         #   INTEL_GRANITERAPIDS_X = IFM(6, 0xAD) -> family 6, model 173 (Granite Rapids-SP/AP)
         #   INTEL_GRANITERAPIDS_D = IFM(6, 0xAE) -> family 6, model 174 (Granite Rapids-D)
         # (cf. INTEL_SAPPHIRERAPIDS_X = 0x8F/143, INTEL_EMERALDRAPIDS_X = 0xCF/207)
-        if [ "${cpu_family}" == "6" ] && { [ "${cpu_model}" == "173" ] || [ "${cpu_model}" == "174" ]; }; then
+        elif [ "${cpu_family}" == "6" ] && { [ "${cpu_model}" == "173" ] || [ "${cpu_model}" == "174" ]; }; then
             best_arch_match="x86_64/intel/graniterapids"
             all_arch_matches="$best_arch_match:$all_arch_matches"
             log "DEBUG" "cpupath: model $cpu_model identifies Granite Rapids; best match upgraded to $best_arch_match"

tests/archdetect/x86_64/intel/graniterapids/AWS-Rocky9-6975PC.all.output

@@ -0,0 +1 @@
+x86_64/intel/graniterapids:x86_64/intel/sapphirerapids:x86_64/intel/icelake:x86_64/intel/cascadelake:x86_64/intel/skylake_avx512:x86_64/intel/haswell:x86_64/generic

tests/archdetect/x86_64/intel/graniterapids/AWS-Rocky9-6975PC.cpu_features

@@ -0,0 +1,8 @@
+arch            : x86
+brand           : Intel(R) Xeon(R) 6975P-C
+family          :   6 (0x06)
+model           : 173 (0xAD)
+stepping        :   1 (0x01)
+uarch           : X86_UNKNOWN
+flags           : adx,aes,amx_bf16,amx_fp16,amx_int8,amx_tile,avx,avx2,avx512_bf16,avx512_fp16,avx512_second_fma,avx512bitalg,avx512bw,avx512cd,avx512dq,avx512f,avx512ifma,avx512vbmi,avx512vbmi2,avx512vl,avx512vnni,avx512vpopcntdq,avx_vnni,bmi1,bmi2,clflushopt,clfsh,clwb,cx16,cx8,erms,f16c,fma3,fpu,fs_rep_cmpsb_scasb,fs_rep_stosb,fz_rep_movsb,gfni,lzcnt,mmx,movbe,movdir64b,movdiri,pclmulqdq,popcnt,rdrnd,rdseed,sha,ss,sse,sse2,sse3,sse4_1,sse4_2,ssse3,tsc,vaes,vpclmulqdq
+cache_info      : {"level":1,"cache_type":"data","cache_size":49152,"ways":12,"line_size":64,"tlb_entries":64,"partitioning":1},{"level":1,"cache_type":"instruction","cache_size":65536,"ways":16,"line_size":64,"tlb_entries":64,"partitioning":1},{"level":2,"cache_type":"unified","cache_size":2097152,"ways":16,"line_size":64,"tlb_entries":2048,"partitioning":1},{"level":3,"cache_type":"unified","cache_size":503316480,"ways":16,"line_size":64,"tlb_entries":491520,"partitioning":1}

tests/archdetect/x86_64/intel/graniterapids/AWS-Rocky9-6975PC.cpuinfo

@@ -0,0 +1,26 @@
+processor	: 0
+vendor_id	: GenuineIntel
+cpu family	: 6
+model		: 173
+model name	: Intel(R) Xeon(R) 6975P-C
+stepping	: 1
+microcode	: 0x1000434
+cpu MHz		: 2700.000
+cache size	: 491520 KB
+physical id	: 0
+siblings	: 16
+core id		: 0
+cpu cores	: 8
+apicid		: 0
+initial apicid	: 0
+fpu		: yes
+fpu_exception	: yes
+cpuid level	: 36
+wp		: yes
+flags		: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd ida arat avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid bus_lock_detect cldemote movdiri movdir64b md_clear serialize amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
+bugs		: spectre_v1 spectre_v2 spec_store_bypass swapgs bhi spectre_v2_user
+bogomips	: 5400.00
+clflush size	: 64
+cache_alignment	: 64
+address sizes	: 48 bits physical, 48 bits virtual
+power management:

tests/archdetect/x86_64/intel/graniterapids/AWS-Rocky9-6975PC.output

@@ -0,0 +1 @@
+x86_64/intel/graniterapids