Merge pull request #4127 from XiWeiGu/LoongArch64-CI

LoongArch64 CI

.github/workflows/loongarch64.yml

@@ -0,0 +1,110 @@
+name: loongarch64 qemu test
+
+on: [push, pull_request]
+
+jobs:
+  TEST:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - target: LOONGSONGENERIC
+            triple:  loongarch64-unknown-linux-gnu
+            opts: NO_SHARED=1 TARGET=LOONGSONGENERIC
+          - target: LOONGSON3R5
+            triple: loongarch64-unknown-linux-gnu
+            opts: NO_SHARED=1 TARGET=LOONGSON3R5
+          - target: LOONGSON2K1000
+            triple: loongarch64-unknown-linux-gnu
+            opts: NO_SHARED=1 TARGET=LOONGSON2K1000
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v3
+
+      - name: Install APT deps
+        run: |
+          sudo add-apt-repository ppa:savoury1/virtualisation
+          sudo apt-get update
+          sudo apt-get install autoconf automake autotools-dev ninja-build make ccache \
+          qemu-user-static
+
+      - name: Download and install loongarch64-toolchain
+        run: |
+          wget https://github.com/loongson/build-tools/releases/download/2022.09.06/loongarch64-clfs-7.3-cross-tools-gcc-glibc.tar.xz
+          tar -xf loongarch64-clfs-7.3-cross-tools-gcc-glibc.tar.xz -C /opt
+
+      - name: Set env
+        run: |
+          echo "LD_LIBRARY_PATH=/opt/cross-tools/target/usr/lib64:/opt/cross-tools/loongarch64-unknown-linux-gnu/lib64:$LD_LIBRARY_PATH" >> $GITHUB_ENV
+          echo "PATH=$GITHUB_WORKSPACE:/opt/cross-tools/bin:$PATH" >> $GITHUB_ENV
+
+      - name: Compilation cache
+        uses: actions/cache@v3
+        with:
+          path: ~/.ccache
+          key: ccache-${{ runner.os }}-${{ matrix.target }}-${{ github.ref }}-${{ github.sha }}
+          restore-keys: |
+            ccache-${{ runner.os }}-${{ matrix.target }}-${{ github.ref }}
+            ccache-${{ runner.os }}-${{ matrix.target }}
+
+      - name: Configure ccache
+        run: |
+          test -d ~/.ccache || mkdir -p ~/.ccache
+          echo "max_size = 300M" > ~/.ccache/ccache.conf
+          echo "compression = true" >> ~/.ccache/ccache.conf
+          ccache -s
+
+      - name: Disable utest dsdot:dsdot_n_1
+        run: |
+          echo -n > utest/test_dsdot.c
+          echo "Due to the qemu versions 7.2 causing utest cases to fail,"
+          echo "the utest dsdot:dsdot_n_1 have been temporarily disabled."
+
+      - name: Build OpenBLAS
+        run: make CC='ccache ${{ matrix.triple }}-gcc -static' FC='ccache ${{ matrix.triple }}-gfortran -static' ${{ matrix.opts }} HOSTCC='ccache gcc' -j$(nproc)
+
+      - name: Test
+        run: |
+          qemu-loongarch64-static ./utest/openblas_utest
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xscblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xdcblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xccblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xzcblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xscblat2 < ./ctest/sin2
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xdcblat2 < ./ctest/din2
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xccblat2 < ./ctest/cin2
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xzcblat2 < ./ctest/zin2
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xscblat3 < ./ctest/sin3
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xdcblat3 < ./ctest/din3
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xccblat3 < ./ctest/cin3
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./ctest/xzcblat3 < ./ctest/zin3
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/sblat1
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/dblat1
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/cblat1
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/zblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/sblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/dblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/cblat1
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/zblat1
+          rm -f ./test/?BLAT2.SUMM
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/sblat2 < ./test/sblat2.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/dblat2 < ./test/dblat2.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/cblat2 < ./test/cblat2.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/zblat2 < ./test/zblat2.dat
+          rm -f ./test/?BLAT2.SUMM
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/sblat2 < ./test/sblat2.dat
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/dblat2 < ./test/dblat2.dat
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/cblat2 < ./test/cblat2.dat
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/zblat2 < ./test/zblat2.dat
+          rm -f ./test/?BLAT3.SUMM
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/sblat3 < ./test/sblat3.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/dblat3 < ./test/dblat3.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/cblat3 < ./test/cblat3.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-loongarch64-static ./test/zblat3 < ./test/zblat3.dat
+          rm -f ./test/?BLAT3.SUMM
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/sblat3 < ./test/sblat3.dat
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/dblat3 < ./test/dblat3.dat
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/cblat3 < ./test/cblat3.dat
+          OPENBLAS_NUM_THREADS=2 qemu-loongarch64-static ./test/zblat3 < ./test/zblat3.dat

Makefile.system

@@ -1770,6 +1770,8 @@ export TARGET_CORE
 export NO_AVX512
 export NO_AVX2
 export BUILD_BFLOAT16
+export NO_LSX
+export NO_LASX
 
 export SBGEMM_UNROLL_M
 export SBGEMM_UNROLL_N

c_check

@@ -185,6 +185,37 @@ if [ "$architecture" = "mips" ] || [ "$architecture" = "mips64" ]; then
     rm -rf "$tmpd"
 fi
 
+no_lsx=0
+no_lasx=0
+if [ "$architecture" = "loongarch64" ]; then
+    tmpd="$(mktemp -d)"
+    tmplsx="$tmpd/lsx.c"
+    codelsx='"vadd.b $vr0, $vr0, $vr0"'
+    lsx_flags='-march=loongarch64 -mlsx'
+    printf "#include <lsxintrin.h>\n\n" >> "$tmplsx"
+    printf "void main(void){ __asm__ volatile(%s);}\n" "$codelsx" >> "$tmplsx"
+    args="$lsx_flags -o $tmplsx.o $tmplsx"
+    {
+        $compiler_name $flags $args >/dev/null 2>&1
+    } || {
+        no_lsx=1
+    }
+
+    tmplasx="$tmpd/lasx.c"
+    codelasx='"xvadd.b $xr0, $xr0, $xr0"'
+    lasx_flags='-march=loongarch64 -mlasx'
+    printf "#include <lasxintrin.h>\n\n" >> "$tmplasx"
+    printf "void main(void){ __asm__ volatile(%s);}\n" "$codelasx" >> "$tmplasx"
+    args="$lasx_flags -o $tmplasx.o $tmplasx"
+    {
+        $compiler_name $flags $args >/dev/null 2>&1
+    } || {
+        no_lasx=1
+    }
+
+    rm -rf "$tmpd"
+fi
+
 case "$data" in
     *ARCH_X86_64*) architecture=x86_64 ;;
     *ARCH_X86*) architecture=x86 ;;
@@ -399,6 +430,8 @@ done
     [ "$no_avx512" -eq 1 ] && printf "NO_AVX512=1\n"
     [ "$no_avx2" -eq 1 ] && printf "NO_AVX2=1\n"
     [ "$oldgcc" -eq 1 ] && printf "OLDGCC=1\n"
+    [ "$no_lsx" -eq 1 ] && printf "NO_LSX=1\n"
+    [ "$no_lasx" -eq 1 ] && printf "NO_LASX=1\n"
 } >> "$makefile"
 
 os=`echo "$os" | tr '[[:lower:]]' '[[:upper:]]'/ `
@@ -414,6 +447,8 @@ compiler=`echo "$compiler" | tr '[[:lower:]]' '[[:upper:]]' `
     [ -n "$need_fu" ] && printf "#define FUNDERSCORE\t%s\n" "$need_fu"
     [ "$no_msa" -eq 1 ] && printf "#define NO_MSA\t1\n"
     [ "$c11_atomics" -eq 1 ] && printf "#define HAVE_C11\t1\n"
+    [ "$no_lsx" -eq 1 ] && printf "#define NO_LSX\t1\n"
+    [ "$no_lasx" -eq 1 ] && printf "#define NO_LASX\t1\n"
 } >> "$config"
 
 

c_check.pl

@@ -232,6 +232,47 @@ if (($architecture eq "mips") || ($architecture eq "mips64")) {
     }
 }
 
+$no_lsx = 0;
+$no_lasx = 0;
+if (($architecture eq "loongarch64")) {
+    eval "use File::Temp qw(tempfile)";
+    if ($@){
+	warn "could not load PERL module File::Temp, so could not check LSX and LASX capatibility";
+    } else {
+	$tmplsx = new File::Temp( SUFFIX => '.c' , UNLINK => 1 );
+	$codelsx = '"vadd.b $vr0, $vr0, $vr0"';
+	$lsx_flags = "-march=loongarch64 -mlsx";
+	print $tmplsx "#include <lsxintrin.h>\n\n";
+	print $tmplsx "void main(void){ __asm__ volatile($codelsx); }\n";
+
+	$args = "$lsx_flags -o $tmplsx.o $tmplsx";
+	my @cmd = ("$compiler_name $flags $args >/dev/null 2>/dev/null");
+	system(@cmd) == 0;
+	if ($? != 0) {
+	    $no_lsx = 1;
+	} else {
+	    $no_lsx = 0;
+	}
+	unlink("$tmplsx.o");
+
+	$tmplasx = new File::Temp( SUFFIX => '.c' , UNLINK => 1 );
+	$codelasx = '"xvadd.b $xr0, $xr0, $xr0"';
+	$lasx_flags = "-march=loongarch64 -mlasx";
+	print $tmplasx "#include <lasxintrin.h>\n\n";
+	print $tmplasx "void main(void){ __asm__ volatile($codelasx); }\n";
+
+	$args = "$lasx_flags -o $tmplasx.o $tmplasx";
+	my @cmd = ("$compiler_name $flags $args >/dev/null 2>/dev/null");
+	system(@cmd) == 0;
+	if ($? != 0) {
+	    $no_lasx = 1;
+	} else {
+	    $no_lasx = 0;
+	}
+	unlink("$tmplasx.o");
+    }
+}
+
 $architecture = x86          if ($data =~ /ARCH_X86/);
 $architecture = x86_64       if ($data =~ /ARCH_X86_64/);
 $architecture = e2k          if ($data =~ /ARCH_E2K/);
@@ -424,6 +465,8 @@ print MAKEFILE "NO_RV64GV=1\n" if $no_rv64gv eq 1;
 print MAKEFILE "NO_AVX512=1\n" if $no_avx512 eq 1;
 print MAKEFILE "NO_AVX2=1\n" if $no_avx2 eq 1;
 print MAKEFILE "OLDGCC=1\n" if $oldgcc eq 1;
+print MAKEFILE "NO_LSX=1\n" if $no_lsx eq 1;
+print MAKEFILE "NO_LASX=1\n" if $no_lasx eq 1;
 
 $os           =~ tr/[a-z]/[A-Z]/;
 $architecture =~ tr/[a-z]/[A-Z]/;
@@ -437,6 +480,8 @@ print CONFFILE "#define __64BIT__\t1\n"  if $binformat eq bin64;
 print CONFFILE "#define FUNDERSCORE\t$need_fu\n" if $need_fu ne "";
 print CONFFILE "#define HAVE_MSA\t1\n"  if $have_msa eq 1;
 print CONFFILE "#define HAVE_C11\t1\n" if $c11_atomics eq 1;
+print CONFFILE "#define NO_LSX\t1\n" if $no_lsx eq 1;
+print CONFFILE "#define NO_LASX\t1\n" if $no_lasx eq 1;
 
 
 if ($os eq "LINUX") {

cpuid_loongarch64.c

@@ -32,6 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 **********************************************************************************/
 
 #include <stdint.h>
+#include <sys/auxv.h>
 
 /*  If LASX extension instructions supported,
  *  using core LOONGSON3R5
@@ -46,9 +47,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CPU_LOONGSON3R5    1
 #define CPU_LOONGSON2K1000 2
 
-#define LOONGARCH_CFG2  0x02
-#define LOONGARCH_LASX  1<<7
-#define LOONGARCH_LSX   1<<6
+#define LA_HWCAP_LSX    (1<<4)
+#define LA_HWCAP_LASX   (1<<5)
 
 static char *cpuname[] = {
   "LOONGSONGENERIC",
@@ -64,17 +64,11 @@ static char *cpuname_lower[] = {
 
 int detect(void) {
 #ifdef __linux
-  uint32_t reg = 0;
+  int flag  = (int)getauxval(AT_HWCAP);
 
-  __asm__ volatile (
-    "cpucfg %0, %1 \n\t"
-    : "+&r"(reg)
-    : "r"(LOONGARCH_CFG2)
-  );
-
-  if (reg & LOONGARCH_LASX)
+  if (flag & LA_HWCAP_LASX)
     return CPU_LOONGSON3R5;
-  else if (reg & LOONGARCH_LSX)
+  else if (flag & LA_HWCAP_LSX)
     return CPU_LOONGSON2K1000;
   else
     return CPU_GENERIC;

kernel/loongarch64/KERNEL.LOONGSON3R5

@@ -1,3 +1,4 @@
+ifndef NO_LASX
 DGEMMKERNEL    = dgemm_kernel_16x4.S
 DGEMMINCOPY    = dgemm_ncopy_16.S
 DGEMMITCOPY    = dgemm_tcopy_16.S
@@ -7,6 +8,7 @@ DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
 DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
 DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
 DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+endif
 
 DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
 DTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c

param.h

@@ -2845,15 +2845,21 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define GEMM_DEFAULT_OFFSET_B 0
 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
 
-#define SGEMM_DEFAULT_UNROLL_N 8
+#if defined(NO_LASX)
+#define DGEMM_DEFAULT_UNROLL_N 8
+#define DGEMM_DEFAULT_UNROLL_M 2
+#else
 #define DGEMM_DEFAULT_UNROLL_N 4
+#define DGEMM_DEFAULT_UNROLL_M 16
+#endif
+
+#define SGEMM_DEFAULT_UNROLL_N 8
 #define QGEMM_DEFAULT_UNROLL_N 2
 #define CGEMM_DEFAULT_UNROLL_N 4
 #define ZGEMM_DEFAULT_UNROLL_N 4
 #define XGEMM_DEFAULT_UNROLL_N 1
 
 #define SGEMM_DEFAULT_UNROLL_M 2
-#define DGEMM_DEFAULT_UNROLL_M 16
 #define QGEMM_DEFAULT_UNROLL_M 2
 #define CGEMM_DEFAULT_UNROLL_M 1
 #define ZGEMM_DEFAULT_UNROLL_M 1