merge

.github/workflows/linux.yaml

@@ -1,13 +1,24 @@
-name: Build benchmarks on Ubuntu
-on: [push, pull_request]
+name: Build Benchmarks on Ubuntu
+on: [push]
 jobs:
   BuildBenchmarks:
     # Only Ubuntu for now.
     runs-on: ubuntu-latest
     steps:
-      - run: sudo apt-get update
-      - run: sudo apt-get install -y build-essential ocl-icd-opencl-dev opencl-headers libnuma-dev
+      - name: Install prerequisites
+        run: sudo apt update && sudo apt -qq --assume-yes full-upgrade && sudo apt install -qq -y build-essential crossbuild-essential-arm64 gcc-riscv64-linux-gnu ocl-icd-opencl-dev opencl-headers libnuma-dev b3sum unzip
+      - name: Wild tomfoolery attempt
+        run: eval "$(/home/linuxbrew/.linuxbrew/bin/brew shellenv)" && brew install mingw-w64
       - name: Check out repository code
         uses: actions/checkout@v3
-      - name: Try to build all benchmarks with a Makefile
-        run: make -j4 all
+      - name: Build all benchmarks
+        run: eval "$(/home/linuxbrew/.linuxbrew/bin/brew shellenv)" && make ci
+      - name: Package benchmarks
+        run: make package
+      - name: b3sum
+        run: b3sum clammarks.txz
+      - name: Upload package
+        env:
+          UPLOAD_KEY: ${{ secrets.UPLOAD_KEY }}
+          UPLOAD_URL: ${{ secrets.UPLOAD_URL }}
+        run:  curl -X PUT -T clammarks.txz -H "$UPLOAD_KEY" "$UPLOAD_URL"

AsmGen/tests/MixIntVec128RfTest.cs

@@ -17,7 +17,7 @@ public MixIntVec128RfTest(int low, int high, int step)
         public override bool SupportsIsa(IUarchTest.ISA isa)
         {
             if (isa == IUarchTest.ISA.amd64) return true;
-            if (isa == IUarchTest.ISA.aarch64) return true;
+            if (isa == IUarchTest.ISA.aarch64) return false;
             return false;
         }
 

CoherencyLatency/CoherencyLatency.cpp

@@ -1,6 +1,12 @@
 #include <stdio.h>
 #include <stdint.h>
-#include <sys\timeb.h>
+
+#ifndef __MINGW32__
+    #include <sys\timeb.h>
+#else
+    #include <sys/timeb.h>
+#endif
+
 #include <intrin.h>
 #include <windows.h>
 
@@ -227,4 +233,4 @@ DWORD WINAPI ReadLatencyTestThread(LPVOID param) {
     }
 
     return current;
-}
+}

CoherencyLatency/Makefile

@@ -0,0 +1,26 @@
+include ../Common/arch_detect.mk
+
+CFLAGS = -pthread -O3
+
+all: $(TARGET)
+
+amd64:
+	$(CC) $(CFLAGS) PThreadsCoherencyLatency.c -o CoherencyLatency_amd64 $(LDFLAGS)
+
+aarch64:
+	$(CC) $(CFLAGS) PThreadsCoherencyLatency.c -o CoherencyLatency_aarch64 $(LDFLAGS)
+
+riscv64:
+	$(CC) $(CFLAGS) PThreadsCoherencyLatency.c -o CoherencyLatency_riscv64 $(LDFLAGS)
+
+w64:
+	$(CC) $(CFLAGS) CoherencyLatency.cpp -o CoherencyLatency_w64.exe $(LDFLAGS)
+
+# w64 can build with mingw 11, which isn't available on jammy
+
+ci: amd64 aarch64 riscv64
+
+clean:
+	rm -rf *.o *.zip "ocl-icd-libopencl1*" "OpenCL-SDK*" && find . -type f -executable -delete
+
+.PHONY: all ci clean

Common/arch_detect.mk

@@ -0,0 +1,28 @@
+TARGET ?= amd64
+
+ifeq ($(OS),Windows_NT)
+    TARGET = w64
+else
+    UNAME_M := $(shell uname -m)
+    ifeq ($(UNAME_M),x86_64)
+        TARGET = amd64
+    endif
+    ifeq ($(UNAME_M),aarch64)
+        TARGET = aarch64
+    endif
+    ifeq ($(UNAME_M),riscv64)
+        TARGET = riscv64
+    endif
+    UNAME_S := $(shell uname -s)
+    ifeq ($(UNAME_S),Darwin)
+    TARGET = darwin
+    endif
+endif
+
+amd64: CC = x86_64-linux-gnu-gcc
+amd64_numa: CC = x86_64-linux-gnu-gcc
+aarch64: CC := aarch64-linux-gnu-gcc
+aarch64_numa: CC = aarch64-linux-gnu-gcc
+riscv64: CC = riscv64-linux-gnu-gcc
+w64: CC = x86_64-w64-mingw32-gcc
+darwin: CC = clang

Common/ci_gpumemlatency.sh

@@ -0,0 +1,35 @@
+#!/bin/sh
+
+make_all () {
+	make amd64
+	make clean-obj
+	LDFLAGS="-lm -L ocl-icd-arm64/usr/lib/aarch64-linux-gnu -lOpenCL" make aarch64
+	make clean-obj
+	LDFLAGS="-lm -L ocl-icd-riscv64/usr/lib/riscv64-linux-gnu -lOpenCL" make riscv64
+	make clean-obj
+	CPPFLAGS="-I OpenCL-SDK-${OCL_VER}-Win-x64/include" LDFLAGS="-lm -L OpenCL-SDK-${OCL_VER}-Win-x64/lib -lOpenCL" make w64
+	make clean-obj
+}
+
+linux_deps () {
+	for ARCH in arm64 riscv64; do
+	if ! grep -q $ARCH /etc/apt/sources.list; then
+		echo "deb [arch=${ARCH}] http://ports.ubuntu.com/ubuntu-ports $(lsb_release -c -s) universe" | sudo tee -a /etc/apt/sources.list
+		echo "deb-src [arch=${ARCH}] http://ports.ubuntu.com/ubuntu-ports $(lsb_release -c -s) universe" | sudo tee -a /etc/apt/sources.list
+		sudo apt update
+	fi
+		apt-get download "ocl-icd-libopencl1:${ARCH}"
+		find . -type f -name "*${ARCH}*.deb" -exec dpkg-deb -x {} "ocl-icd-${ARCH}" \;
+done
+cp ocl-icd-arm64/usr/lib/aarch64-linux-gnu/libOpenCL.so.1 ocl-icd-arm64/usr/lib/aarch64-linux-gnu/libOpenCL.so
+cp ocl-icd-riscv64/usr/lib/riscv64-linux-gnu/libOpenCL.so.1 ocl-icd-riscv64/usr/lib/riscv64-linux-gnu/libOpenCL.so
+}
+
+w64_deps () {
+	curl -fssLO "https://github.com/KhronosGroup/OpenCL-SDK/releases/download/${OCL_VER}/OpenCL-SDK-${OCL_VER}-Win-x64.zip"
+	unzip "OpenCL-SDK-${OCL_VER}-Win-x64.zip"
+}
+
+linux_deps
+w64_deps
+make_all

Common/ci_package.sh

@@ -0,0 +1,17 @@
+#!/bin/sh
+
+PKG="clammarks-$(git rev-parse --short HEAD)"
+rm -rf "$PKG" "clammarks.txz"
+mkdir -p "$PKG"
+
+for TARGET in "amd64" "aarch64" "riscv64" "w64"; do
+	mkdir "$PKG/$TARGET"
+	for COMPONENT in CoherencyLatency MemoryLatency MemoryBandwidth InstructionRate Meshsim CoreClockChecker GpuMemLatency; do
+		find "$COMPONENT" -type f -name "*$TARGET*" -executable -exec cp {} "$PKG/$TARGET" \;
+	done
+	find "GpuMemLatency" -type f -name "*.cl" -exec cp {} "$PKG/$TARGET" \;
+done
+
+cp "LICENSE" "$PKG"
+
+tar caf "clammarks.txz" "$PKG"

CoreClockChecker/BoostClockChecker_arm.s

@@ -1,7 +1,12 @@
 .text
 .global clktsctest
 
+.global _clktsctest
+
+.balign 4
+
 /* x0 = iterations, return elapsed TSC in x0 */
+_clktsctest:
 clktsctest:
   sub sp, sp, #0x40
   stp x10, x11, [sp, #0x10]

CoreClockChecker/Makefile

@@ -1,8 +1,23 @@
+include ../Common/arch_detect.mk
+
+CFLAGS = -O3
+LDFLAGS = -lm
+
+all: $(TARGET)
+
 amd64:
-	x86_64-linux-gnu-gcc -pthread -O3 CoreClockChecker.c CoreClockChecker_x86.s -o CoreClockChecker -lm
-win64:
-	x86_64-w64-mingw32-gcc -pthread -O3 CoreClockChecker.c CoreClockChecker_x86.s -o CoreClockChecker -lm
-boostclockchecker:
-	gcc -O3 BoostClockChecker.c BoostClockChecker.s -o BoostClockChecker
-win64boostclockchecker:
-	x86_64-w64-mingw32-gcc BoostClockChecker.c BoostClockChecker.s -o BoostClockChecker.exe
+	$(CC) $(CFLAGS) -pthread CoreClockChecker.c CoreClockChecker_x86.s -o CoreClockChecker_amd64 $(LDFLAGS)
+	$(CC) $(CFLAGS) BoostClockChecker.c BoostClockChecker_x86.s -o BoostClockChecker_amd64 $(LDFLAGS)
+
+aarch64:
+	$(CC) $(CFLAGS) BoostClockChecker.c BoostClockChecker_arm.s -o BoostClockChecker_aarch64 $(LDFLAGS)
+
+w64:
+	$(CC) $(CFLAGS) BoostClockChecker.c BoostClockChecker_x86.s -o BoostClockChecker_w64.exe $(LDFLAGS)
+
+ci: amd64 aarch64 w64
+
+clean:
+	rm -f *.o && find . -type f -executable -delete
+
+.PHONY: all ci clean

GpuMemLatency/Makefile

@@ -1,20 +1,52 @@
-UNAME_S := $(shell uname -s)
-CC = gcc
-CFLAGS = -O3
-DEPS = ../common/timings.h
-OBJ = opencltest.o latency_test.o bw_test.o common.o atomic_test.o instruction_rate.o timing.o
+include ../Common/arch_detect.mk
+
+OCL_VER = v2023.04.17
+CI_SCRIPT = ../Common/ci_gpumemlatency.sh
 
-ifeq ($(UNAME_S),Darwin)
-	LDFLAGS += -framework OpenCL
-else
-	LDFLAGS += -lOpenCL
+CFLAGS = -O3 -I ../Common
+DEPS = ../Common/timings.h
+OBJ = opencltest.o latency_test.o bw_test.o common.o atomic_test.o instruction_rate.o timing.o
+LDFLAGS ?= -lm -lOpenCL
+ifeq ($(TARGET), Darwin)
+    LDFLAGS = -lm -framework OpenCL
 endif
 
-opencltest: $(OBJ)
-	gcc $(CFLAGS) $^ -o $@ -lm $(LDFLAGS)
+all: $(TARGET)
+
+GpuMemLatency: $(OBJ)
+	$(CC) $(CPPFLAGS) $(CFLAGS) $^ -o $@ $(LDFLAGS)
 
 %.o: %.c $(DEPS)
 	$(CC) $(CFLAGS) -c -o $@ $<
 
 timing.o:
-	$(CC) $(CFLAGS) -c ../common/timing.c -o timing.o
+	$(CC) $(CFLAGS) -c ../Common/timing.c -o timing.o
+
+amd64: $(OBJ)
+	$(CC) $(CFLAGS) $^ -o GpuMemLatency_amd64 $(LDFLAGS)
+
+aarch64: $(OBJ)
+	$(CC) $(CFLAGS) $^ -o GpuMemLatency_aarch64 $(LDFLAGS)
+
+riscv64: $(OBJ)
+	$(CC) $(CFLAGS) $^ -o GpuMemLatency_riscv64 $(LDFLAGS)
+
+w64: $(OBJ)
+	$(CC) $(CFLAGS) $^ -o GpuMemLatency_w64.exe $(LDFLAGS)
+
+darwin: $(OBJ)
+	$(CC) $(CFLAGS) $^ -o GpuMemLatency_darwin $(LDFLAGS)
+
+ci: clean
+	@OCL_VER=$(OCL_VER) sh $(CI_SCRIPT)
+
+clean-ci:
+	rm -rf "*.deb" "*.zip" "ocl-icd-*" "OpenCL-SDK-*"
+
+clean-obj: 
+	rm -f *.o
+
+clean: clean-ci clean-obj
+	find . -type f -executable -delete
+
+.PHONY: all ci clean-ci clean-obj clean