Merge pull request #3769 from XiWeiGu/mips64-test

[WIP,Testing]: Add test for mips64

.github/workflows/mips64.yml

@@ -0,0 +1,114 @@
+name: mips64 qemu test
+
+on: [push, pull_request]
+
+jobs:
+  TEST:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - target: MIPS64_GENERIC
+            triple: mips64el-linux-gnuabi64
+            opts: NO_SHARED=1 TARGET=MIPS64_GENERIC
+          - target: SICORTEX
+            triple: mips64el-linux-gnuabi64
+            opts: NO_SHARED=1 TARGET=SICORTEX
+          - target: I6400
+            triple: mipsisa64r6el-linux-gnuabi64
+            opts: NO_SHARED=1 TARGET=I6400
+          - target: P6600
+            triple: mipsisa64r6el-linux-gnuabi64
+            opts: NO_SHARED=1 TARGET=P6600
+          - target: I6500
+            triple: mipsisa64r6el-linux-gnuabi64
+            opts: NO_SHARED=1 TARGET=I6500
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v3
+
+      - name: install build deps
+        run: |
+          sudo apt-get update
+          sudo apt-get install autoconf automake autotools-dev ninja-build make ccache \
+          gcc-${{ matrix.triple }} gfortran-${{ matrix.triple }} libgomp1-mips64el-cross
+
+      - name: checkout qemu
+        uses: actions/checkout@v3
+        with:
+          repository: qemu/qemu
+          path: qemu
+          ref: 79dfa177ae348bb5ab5f97c0915359b13d6186e2
+
+      - name: build qemu
+        run: |
+          cd qemu
+          ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=mips64el-linux-user --disable-system
+          make -j$(nproc)
+          make install
+
+      - 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: 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: |
+          export PATH=$GITHUB_WORKSPACE/qemu-install/bin/:$PATH
+          qemu-mips64el ./utest/openblas_utest
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xscblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xdcblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xccblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xzcblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xscblat2 < ./ctest/sin2
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xdcblat2 < ./ctest/din2
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xccblat2 < ./ctest/cin2
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xzcblat2 < ./ctest/zin2
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xscblat3 < ./ctest/sin3
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xdcblat3 < ./ctest/din3
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xccblat3 < ./ctest/cin3
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./ctest/xzcblat3 < ./ctest/zin3
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/sblat1
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/dblat1
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/cblat1
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/zblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/sblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/dblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/cblat1
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/zblat1
+          rm -f ./test/?BLAT2.SUMM
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/sblat2 < ./test/sblat2.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/dblat2 < ./test/dblat2.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/cblat2 < ./test/cblat2.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/zblat2 < ./test/zblat2.dat
+          rm -f ./test/?BLAT2.SUMM
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/sblat2 < ./test/sblat2.dat
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/dblat2 < ./test/dblat2.dat
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/cblat2 < ./test/cblat2.dat
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/zblat2 < ./test/zblat2.dat
+          rm -f ./test/?BLAT3.SUMM
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/sblat3 < ./test/sblat3.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/dblat3 < ./test/dblat3.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/cblat3 < ./test/cblat3.dat
+          OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 qemu-mips64el ./test/zblat3 < ./test/zblat3.dat
+          rm -f ./test/?BLAT3.SUMM
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/sblat3 < ./test/sblat3.dat
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/dblat3 < ./test/dblat3.dat
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/cblat3 < ./test/cblat3.dat
+          OPENBLAS_NUM_THREADS=2 qemu-mips64el ./test/zblat3 < ./test/zblat3.dat

kernel/mips/sdot_msa.c

@@ -39,10 +39,19 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
     FLOAT x0, x1, x2, x3, y0, y1, y2, y3;
     v4f32 vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7;
     v4f32 vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7;
+#if defined(DSDOT)
+    v2f64 dvx0, dvx1, dvx2, dvx3, dvx4, dvx5, dvx6, dvx7;
+    v2f64 dvy0, dvy1, dvy2, dvy3, dvy4, dvy5, dvy6, dvy7;
+    v2f64 dot0 = {0, 0};
+    v2f64 dot1 = {0, 0};
+    v2f64 dot2 = {0, 0};
+    v2f64 dot3 = {0, 0};
+#else
     v4f32 dot0 = {0, 0, 0, 0};
     v4f32 dot1 = {0, 0, 0, 0};
     v4f32 dot2 = {0, 0, 0, 0};
     v4f32 dot3 = {0, 0, 0, 0};
+#endif
 
     if (n < 1) return (dot);
 
@@ -83,6 +92,61 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
             x_pref += 32;
             y_pref += 32;
 
+#if defined(DSDOT)
+            /* Extend single precision to double precision */
+            dvy0 = __msa_fexupr_d(vy0);
+            dvy1 = __msa_fexupr_d(vy1);
+            dvy2 = __msa_fexupr_d(vy2);
+            dvy3 = __msa_fexupr_d(vy3);
+            dvy4 = __msa_fexupr_d(vy4);
+            dvy5 = __msa_fexupr_d(vy5);
+            dvy6 = __msa_fexupr_d(vy6);
+            dvy7 = __msa_fexupr_d(vy7);
+
+            vy0 = (v4f32)__msa_fexupl_d(vy0);
+            vy1 = (v4f32)__msa_fexupl_d(vy1);
+            vy2 = (v4f32)__msa_fexupl_d(vy2);
+            vy3 = (v4f32)__msa_fexupl_d(vy3);
+            vy4 = (v4f32)__msa_fexupl_d(vy4);
+            vy5 = (v4f32)__msa_fexupl_d(vy5);
+            vy6 = (v4f32)__msa_fexupl_d(vy6);
+            vy7 = (v4f32)__msa_fexupl_d(vy7);
+
+            dvx0 = __msa_fexupr_d(vx0);
+            dvx1 = __msa_fexupr_d(vx1);
+            dvx2 = __msa_fexupr_d(vx2);
+            dvx3 = __msa_fexupr_d(vx3);
+            dvx4 = __msa_fexupr_d(vx4);
+            dvx5 = __msa_fexupr_d(vx5);
+            dvx6 = __msa_fexupr_d(vx6);
+            dvx7 = __msa_fexupr_d(vx7);
+
+            vx0 = (v4f32)__msa_fexupl_d(vx0);
+            vx1 = (v4f32)__msa_fexupl_d(vx1);
+            vx2 = (v4f32)__msa_fexupl_d(vx2);
+            vx3 = (v4f32)__msa_fexupl_d(vx3);
+            vx4 = (v4f32)__msa_fexupl_d(vx4);
+            vx5 = (v4f32)__msa_fexupl_d(vx5);
+            vx6 = (v4f32)__msa_fexupl_d(vx6);
+            vx7 = (v4f32)__msa_fexupl_d(vx7);
+
+            dot0 += (dvy0 * dvx0);
+            dot1 += (dvy1 * dvx1);
+            dot2 += (dvy2 * dvx2);
+            dot3 += (dvy3 * dvx3);
+            dot0 += (dvy4 * dvx4);
+            dot1 += (dvy5 * dvx5);
+            dot2 += (dvy6 * dvx6);
+            dot3 += (dvy7 * dvx7);
+            dot0 += ((v2f64)vy0 * (v2f64)vx0);
+            dot1 += ((v2f64)vy1 * (v2f64)vx1);
+            dot2 += ((v2f64)vy2 * (v2f64)vx2);
+            dot3 += ((v2f64)vy3 * (v2f64)vx3);
+            dot0 += ((v2f64)vy4 * (v2f64)vx4);
+            dot1 += ((v2f64)vy5 * (v2f64)vx5);
+            dot2 += ((v2f64)vy6 * (v2f64)vx6);
+            dot3 += ((v2f64)vy7 * (v2f64)vx7);
+#else
             dot0 += (vy0 * vx0);
             dot1 += (vy1 * vx1);
             dot2 += (vy2 * vx2);
@@ -91,6 +155,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
             dot1 += (vy5 * vx5);
             dot2 += (vy6 * vx6);
             dot3 += (vy7 * vx7);
+#endif
         }
 
         if (n & 31)
@@ -100,10 +165,41 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
                 LD_SP4_INC(x, 4, vx0, vx1, vx2, vx3);
                 LD_SP4_INC(y, 4, vy0, vy1, vy2, vy3);
 
+#if defined(DSDOT)
+                dvy0 = __msa_fexupr_d(vy0);
+                dvy1 = __msa_fexupr_d(vy1);
+                dvy2 = __msa_fexupr_d(vy2);
+                dvy3 = __msa_fexupr_d(vy3);
+
+                vy0 = (v4f32)__msa_fexupl_d(vy0);
+                vy1 = (v4f32)__msa_fexupl_d(vy1);
+                vy2 = (v4f32)__msa_fexupl_d(vy2);
+                vy3 = (v4f32)__msa_fexupl_d(vy3);
+
+                dvx0 = __msa_fexupr_d(vx0);
+                dvx1 = __msa_fexupr_d(vx1);
+                dvx2 = __msa_fexupr_d(vx2);
+                dvx3 = __msa_fexupr_d(vx3);
+
+                vx0 = (v4f32)__msa_fexupl_d(vx0);
+                vx1 = (v4f32)__msa_fexupl_d(vx1);
+                vx2 = (v4f32)__msa_fexupl_d(vx2);
+                vx3 = (v4f32)__msa_fexupl_d(vx3);
+
+                dot0 += (dvy0 * dvx0);
+                dot1 += (dvy1 * dvx1);
+                dot2 += (dvy2 * dvx2);
+                dot3 += (dvy3 * dvx3);
+                dot0 += ((v2f64)vy0 * (v2f64)vx0);
+                dot1 += ((v2f64)vy1 * (v2f64)vx1);
+                dot2 += ((v2f64)vy2 * (v2f64)vx2);
+                dot3 += ((v2f64)vy3 * (v2f64)vx3);
+#else
                 dot0 += (vy0 * vx0);
                 dot1 += (vy1 * vx1);
                 dot2 += (vy2 * vx2);
                 dot3 += (vy3 * vx3);
+#endif
             }
 
             if (n & 8)
@@ -111,8 +207,27 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
                 LD_SP2_INC(x, 4, vx0, vx1);
                 LD_SP2_INC(y, 4, vy0, vy1);
 
+#if defined(DSDOT)
+                dvy0 = __msa_fexupr_d(vy0);
+                dvy1 = __msa_fexupr_d(vy1);
+
+                vy0 = (v4f32)__msa_fexupl_d(vy0);
+                vy1 = (v4f32)__msa_fexupl_d(vy1);
+
+                dvx0 = __msa_fexupr_d(vx0);
+                dvx1 = __msa_fexupr_d(vx1);
+
+                vx0 = (v4f32)__msa_fexupl_d(vx0);
+                vx1 = (v4f32)__msa_fexupl_d(vx1);
+
+                dot0 += (dvy0 * dvx0);
+                dot1 += (dvy1 * dvx1);
+                dot0 += ((v2f64)vy0 * (v2f64)vx0);
+                dot1 += ((v2f64)vy1 * (v2f64)vx1);
+#else
                 dot0 += (vy0 * vx0);
                 dot1 += (vy1 * vx1);
+#endif
             }
 
             if (n & 4)
@@ -120,7 +235,16 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
                 vx0 = LD_SP(x); x += 4;
                 vy0 = LD_SP(y); y += 4;
 
+#if defined(DSDOT)
+                dvy0 = __msa_fexupr_d(vy0);
+                vy0 = (v4f32)__msa_fexupl_d(vy0);
+                dvx0 = __msa_fexupr_d(vx0);
+                vx0 = (v4f32)__msa_fexupl_d(vx0);
+                dot0 += (dvy0 * dvx0);
+                dot0 += ((v2f64)vy0 * (v2f64)vx0);
+#else
                 dot0 += (vy0 * vx0);
+#endif
             }
 
             if (n & 2)
@@ -128,8 +252,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
                 LD_GP2_INC(x, 1, x0, x1);
                 LD_GP2_INC(y, 1, y0, y1);
 
+#if defined(DSDOT)
+                dot += ((double)y0 * (double)x0);
+                dot += ((double)y1 * (double)x1);
+#else
                 dot += (y0 * x0);
                 dot += (y1 * x1);
+#endif
             }
 
             if (n & 1)
@@ -137,7 +266,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
                 x0 = *x;
                 y0 = *y;
 
+#if defined(DSDOT)
+                dot += ((double)y0 * (double)x0);
+#else
                 dot += (y0 * x0);
+#endif
             }
         }
 
@@ -145,8 +278,10 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 
         dot += dot0[0];
         dot += dot0[1];
+#if !defined(DSDOT)
         dot += dot0[2];
         dot += dot0[3];
+#endif
     }
     else
     {
@@ -155,10 +290,17 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
             LD_GP4_INC(x, inc_x, x0, x1, x2, x3);
             LD_GP4_INC(y, inc_y, y0, y1, y2, y3);
 
+#if defined(DSDOT)
+            dot += ((double)y0 * (double)x0);
+            dot += ((double)y1 * (double)x1);
+            dot += ((double)y2 * (double)x2);
+            dot += ((double)y3 * (double)x3);
+#else
             dot += (y0 * x0);
             dot += (y1 * x1);
             dot += (y2 * x2);
             dot += (y3 * x3);
+#endif
         }
 
         if (n & 2)
@@ -166,8 +308,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
             LD_GP2_INC(x, inc_x, x0, x1);
             LD_GP2_INC(y, inc_y, y0, y1);
 
+#if defined(DSDOT)
+            dot += ((double)y0 * (double)x0);
+            dot += ((double)y1 * (double)x1);
+#else
             dot += (y0 * x0);
             dot += (y1 * x1);
+#endif
         }
 
         if (n & 1)
@@ -175,7 +322,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
             x0 = *x;
             y0 = *y;
 
+#if defined(DSDOT)
+            dot += ((double)y0 * (double)x0);
+#else
             dot += (y0 * x0);
+#endif
         }
     }