Initial framework for benchmarking fault-only-first load

fault-first-load/ * .gitignore: Created. * Makefile: Created. * do-all.sh: Created. * template-ld.S: Created. Signed-off-by: Jeremy Bennett <[email protected]>
embecosm · Feb 5, 2025 · 23c21aa · 23c21aa
1 parent c1251ba
commit 23c21aa
Show file tree

Hide file tree

Showing 4 changed files with 302 additions and 0 deletions.
diff --git a/fault-first-load/.gitignore b/fault-first-load/.gitignore
@@ -0,0 +1,2 @@
+# Generated results
+*.csv
diff --git a/fault-first-load/Makefile b/fault-first-load/Makefile
@@ -0,0 +1,30 @@
+# Makefile to generate benchmarks for whole word load/store
+
+# Copyright (C) 2025 Embecosm Limited <www.embecosm.com>
+# Contributor Jeremy Bennett <[email protected]>
+
+# SPDX-License-Identifier: GPL-3.0-or-later
+
+# Parameters that can be set
+LD_CNT ?= 1
+LD_LOOP_CNT ?= 1000000
+VLDOP ?= vle8ff
+DOMASK ?= 0
+VMASK ?=
+ELEM ?= e8
+LMUL ?= m1
+
+# The tools and their flags
+AS=riscv64-unknown-linux-gnu-gcc
+ASFLAGS=-march=rv64gcv -DLD_CNT="$(LD_CNT)" -DLD_LOOP_CNT="$(LD_LOOP_CNT)" \
+        -DVLDOP="$(VLDOP)" -DDOMASK="$(DOMASK)" -DVMASK="$(VMASK)" \
+	-DELEM="$(ELEM)" -DLMUL="$(LMUL)"
+CFLAGS=-march=rv64gcv -O0
+LDFLAGS=-march=rv64gcv -O0
+
+$(VLDOP).exe: template-ld.S
+	$(AS) $(ASFLAGS) $^ -o $@
+
+.PHONE: clean
+clean:
+	$(RM) *.exe fault-first-load*.txt
diff --git a/fault-first-load/do-all.sh b/fault-first-load/do-all.sh
@@ -0,0 +1,185 @@
+# Run all the fault-only-first load benchmarks
+
+# Copyright (C) 2025 Embecosm Limited <www.embecosm.com>
+# Contributor Jeremy Bennett <[email protected]>
+
+# SPDX-License-Identifier: GPL-3.0-or-later
+
+set -u
+
+# Produce help message
+usage () {
+    cat <<EOF
+Usage ./do-all.sh                     : Benchmark whole word load/store.
+                   [--nloops <count>] : Number of iterations of the test
+                                        program (default 10000000)
+                   [--nstats <count>] : Number of times to repeat each test
+                                        for statistical analysis (default 10).
+                   [--base <id>]      : Commit of the QEMU version to use as
+                                        baseline
+                   [--test <id>]      : Commit of the QEMU version to use be
+                                        tested against the baseline
+                   [--help]           : Print this message and exit
+
+The QMEU are assumed to have been built without plugins enabled, with the
+binary installed in ../../install/qemu-${qid}-no-plugin/bin, where qid is the
+argument to --base or --test.
+EOF
+}
+
+# Build the exe
+# $1 Instruction
+# $2 Instruction count
+# $3 "load" or "store"
+buildit() {
+    local i=$1
+    local icount=$2
+    local lmul=$3
+    local elem=$(echo "${i}" | sed -e 's/vl//' -e 's/ff//')
+    # Single instr build
+    make clean > /dev/null 2>&1
+    make LD_CNT=${icount} LD_LOOP_CNT=${nloops} VLDOP=${i} \
+	 DOMASK=0 ELEM="${elem}" LMUL=${lmul}   ${i}.exe > /dev/null 2>&1
+}
+
+# Return the time for doing a single run. Assumes the exe has been built
+# $1 VLEN
+# $2 Instruction
+# $3 QEMU commit
+runone() {
+    local vl=$1
+    local i=$2
+    local qid=$3
+
+    local OLDPATH=${PATH}
+    PATH=../../install/qemu-${qid}-no-plugin/bin:${OLDPATH}
+    (time qemu-riscv64 -cpu rv64,v=true,vlen=${vl} ${i}.exe) \
+	> ${tmpf} 2>&1
+     local u=$(sed -n -e 's/user[[:space:]]*0m\(.\+\)s$/\1/p' < ${tmpf})
+    local s=$(sed -n -e 's/sys[[:space:]]*0m\(.\+\)s$/\1/p' < ${tmpf})
+    PATH=${OLDPATH}
+    t=$(echo "print(${u} + ${s})" | python3)
+    printf "%.3f" ${t}
+}
+
+# Run a single measurement
+# $1 VLEN
+# $2 LMUL
+# $3 Instruction
+runit() {
+    local vl=$1
+    local lmul=$2
+    local i=$3
+    local res=""
+    # We repeat until we get a meaningful result.  This should mean we skip
+    # some silly results due to chron jobs
+    while [[ "x${res}" == "x" ]]
+    do
+	# single instruction runs for base and test QEMU
+	buildit ${i} 1 ${lmul}
+	local tbase1=$(runone ${vl} ${i} ${qemubase})
+	local ttest1=$(runone ${vl} ${i} ${qemutest})
+	# 11 instruction runs for base and test
+	buildit ${i} 11 ${lmul}
+	local tbase11=$(runone ${vl} ${i} ${qemubase})
+	local ttest11=$(runone ${vl} ${i} ${qemutest})
+	# Calculate net instructions times for base and test
+	local tdbase=$(echo "print(${tbase11} - ${tbase1})" | python3)
+	local tdtest=$(echo "print(${ttest11} - ${ttest1})" | python3)
+	res=$(echo "print(${tdbase} / ${tdtest} - 1 \
+                          if (${tdbase} > 0) and (${tdtest} > 0) else '')" \
+		  | python3)
+    done
+    printf "%.3f" ${res}
+}
+
+ldinstr="vle8ff  \
+	 vle16ff \
+	 vle32ff \
+	 vle64ff"
+
+vlenlist="128 256 512 1024"
+
+lmullist="m1 m2 m4 m8"
+
+tmpf=$(mktemp fault-first-load-XXXXXX.txt)
+
+ldresf="ldres.csv"
+
+# Defaults for variables
+nloops=10000000
+nstats=10
+qemubase="6528013b5f"
+qemutest="db95037b42"
+
+# Parse command line options
+set +u
+until
+  opt="$1"
+  case "${opt}" in
+      --nloops)
+	  shift
+	  nloops="$1"
+	  ;;
+      --nstats)
+	  shift
+	  nstats="$1"
+	  ;;
+      --base)
+	  shift
+	  qemubase="$1"
+	  ;;
+      --test)
+	  shift
+	  qemutest="$1"
+	  ;;
+      --help)
+	  usage
+	  exit 0
+	  ;;
+      ?*)
+	  echo "Unknown argument '$1'"
+	  usage
+	  exit 1
+	  ;;
+      *)
+	  ;;
+  esac
+[ "x${opt}" = "x" ]
+do
+  shift
+done
+set -u
+
+# All the load instructions
+echo "Fault only first load instructions"
+printf '"vlen","lmul","ldop"' > ${ldresf}
+for n in $(seq ${nstats})
+do
+    printf ',"run %d"' ${n} >> ${ldresf}
+done
+printf '\n' >> ${ldresf}
+
+for vl in ${vlenlist}
+do
+    for lmul in ${lmullist}
+    do
+	for i in ${ldinstr}
+	do
+	    printf '"%d","%s","%s"' ${vl} ${lmul} ${i} >> ${ldresf}
+	    printf "VLEN = %4d, LMUL = %2s: %-7s " ${vl} ${lmul} ${i}
+	    for n in $(seq ${nstats})
+	    do
+		r=$(runit ${vl} ${lmul} ${i})
+		printf ',"%.3f"' ${r} >> ${ldresf}
+		printf "."
+	    done
+	    printf '\n' >> ${ldresf}
+	    printf "\n"
+	done
+    done
+done
+
+# Tidy up
+make clean > /dev/null 2>&1
+rm -f ${tmpf}
diff --git a/fault-first-load/template-ld.S b/fault-first-load/template-ld.S
@@ -0,0 +1,85 @@
+// A generic framework for benchmarking whole word load
+//
+// Copyright (C) 2025 Embecosm Limited
+// Contributor: Jeremy Bennett <[email protected]>
+//
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+	// Test LD_CNT loads LD_LOOP_CNT times
+	.text
+	.globl	main
+	.align	2
+	.type	symbol,@function
+main:
+	.cfi_startproc
+
+	csrr	t1, vlenb
+#if DOMASK
+	la	a0, testmask		// Location of mask to load
+	vsetvli	t2, t1, e8, m1, ta, ma
+	vle8.v	v0, (a0)
+#endif
+	la	a0, testlddata		// Location of values to load
+	li	t0, LD_LOOP_CNT		// Number of times around the loop
+
+.L0:
+	vsetvli t2, t1, ELEM, LMUL, ta, ma
+	.rept	LD_CNT
+#if DOMASK
+	VLDOP.v	v1, (a0), v0.t
+#else
+	VLDOP.v	v1, (a0)
+#endif
+	.endr
+	addi	t0, t0, -1
+	bnez	t0, .L0
+
+	mv	a0, zero
+	ret
+
+	.cfi_endproc
+	.size	main,.-main
+
+	.data
+	// Mask to set.
+#if DOMASK
+	// Mask if needed.
+testmask:
+	.8byte	VMASK
+#endif
+	// Data to load.  Maximum is for 8 registers of 1024 bits, so 1Kibytes.
+testlddata:
+	.rept	4
+	.byte	0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07
+	.byte	0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
+	.byte	0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17
+	.byte	0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f
+	.byte	0x20, 0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27
+	.byte	0x28, 0x29, 0x2a, 0x2b, 0x2c, 0x2d, 0x2e, 0x2f
+	.byte	0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37
+	.byte	0x38, 0x39, 0x3a, 0x3b, 0x3c, 0x3d, 0x3e, 0x3f
+	.byte	0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47
+	.byte	0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f
+	.byte	0x50, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57
+	.byte	0x58, 0x59, 0x5a, 0x5b, 0x5c, 0x5d, 0x5e, 0x5f
+	.byte	0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67
+	.byte	0x68, 0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f
+	.byte	0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77
+	.byte	0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f
+	.byte	0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87
+	.byte	0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f
+	.byte	0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97
+	.byte	0x98, 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f
+	.byte	0xa0, 0xa1, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7
+	.byte	0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf
+	.byte	0xb0, 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7
+	.byte	0xb8, 0xb9, 0xba, 0xbb, 0xbc, 0xbd, 0xbe, 0xbf
+	.byte	0xc0, 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc7
+	.byte	0xc8, 0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf
+	.byte	0xd0, 0xd1, 0xd2, 0xd3, 0xd4, 0xd5, 0xd6, 0xd7
+	.byte	0xd8, 0xd9, 0xda, 0xdb, 0xdc, 0xdd, 0xde, 0xdf
+	.byte	0xe0, 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7
+	.byte	0xe8, 0xe9, 0xea, 0xeb, 0xec, 0xed, 0xee, 0xef
+	.byte	0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7
+	.byte	0xf8, 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff
+	.endr