Merge 06ced6da16 into 06c09deee9

.github/workflows/c910v.yml

@@ -14,8 +14,8 @@ jobs:
     if: "github.repository == 'OpenMathLib/OpenBLAS'"
     runs-on: ubuntu-latest
     env:
-      xuetie_toolchain: https://occ-oss-prod.oss-cn-hangzhou.aliyuncs.com/resource//1698113812618
-      toolchain_file_name: Xuantie-900-gcc-linux-5.10.4-glibc-x86_64-V2.8.0-20231018.tar.gz
+      xuetie_toolchain: https://occ-oss-prod.oss-cn-hangzhou.aliyuncs.com/resource//1749714096626
+      toolchain_file_name: Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.1.0-20250522.tar.gz
     strategy:
       fail-fast: false
       matrix:
@@ -77,7 +77,7 @@ jobs:
         run: |
             wget ${xuetie_toolchain}/${toolchain_file_name}
             tar -xvf ${toolchain_file_name} -C /opt
-            export PATH="/opt/Xuantie-900-gcc-linux-5.10.4-glibc-x86_64-V2.8.0/bin:$PATH"
+            export PATH="/opt/Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.1.0/bin:$PATH"
 
             make CC='ccache ${{ matrix.triple }}-gcc -static' FC='ccache ${{ matrix.triple }}-gfortran -static' ${{ matrix.opts }} HOSTCC='ccache gcc' -j$(nproc)
 

Makefile.prebuild

@@ -52,7 +52,7 @@ TARGET_FLAGS = -mips64r6
 endif
 
 ifeq ($(TARGET), C910V)
-TARGET_FLAGS = -march=rv64gcv0p7_zfh_xtheadc -mabi=lp64d
+TARGET_FLAGS = -march=rv64gc_zfh_xtheadc_xtheadvector -mabi=lp64d
 endif
 
 ifeq ($(TARGET), CK860FV)

Makefile.riscv64

@@ -1,6 +1,6 @@
 ifeq ($(CORE), C910V)
-CCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920
-FCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920 -static
+CCOMMON_OPT += -march=rv64imafdc_zfh_xtheadc_xtheadvector -mabi=lp64d -mtune=c920
+FCOMMON_OPT += -march=rv64imafdc_zfh_xtheadc_xtheadvector -mabi=lp64d -mtune=c920 -static
 endif
 ifeq ($(CORE), x280)
 CCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh_zvl512b -mabi=lp64d 

common_macro.h

@@ -2709,7 +2709,7 @@
 #ifndef ASSEMBLER
 #if !defined(DYNAMIC_ARCH) \
   && (defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64) \
-      || defined(ARCH_LOONGARCH64) || defined(ARCH_E2K) || defined(ARCH_ALPHA))
+      || defined(ARCH_LOONGARCH64) || defined(ARCH_E2K) || defined(ARCH_ALPHA) || defined(ARCH_RISCV64))
 extern BLASLONG gemm_offset_a;
 extern BLASLONG gemm_offset_b;
 extern BLASLONG bgemm_p;

common_riscv64.h

@@ -93,13 +93,7 @@ static inline int blas_quickdivide(blasint x, blasint y){
 # include <riscv_vector.h>
 #endif
 
-#if defined( __riscv_xtheadc ) && defined( __riscv_v ) && ( __riscv_v <= 7000 )
-// t-head toolchain uses obsolete rvv intrinsics, can't build for C910V without this
-#define RISCV_0p10_INTRINSICS
-#define RISCV_RVV(x) x
-#else
 #define RISCV_RVV(x) __riscv_ ## x
-#endif
 
 #if defined(C910V) || defined(RISCV64_ZVL256B)
 # if !defined(DOUBLE)

kernel/riscv64/dsdot_vector.c

@@ -37,114 +37,114 @@ double CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
         vfloat32m2_t vx, vy;
         unsigned int gvl = 0;
         vfloat64m1_t v_res, v_z0;
-        gvl = vsetvlmax_e64m1();
-        v_res = vfmv_v_f_f64m1(0, gvl);
-        v_z0 = vfmv_v_f_f64m1(0, gvl);
+        gvl = __riscv_vsetvlmax_e64m1();
+        v_res = __riscv_vfmv_v_f_f64m1(0, gvl);
+        v_z0 = __riscv_vfmv_v_f_f64m1(0, gvl);
 
         if(inc_x == 1 && inc_y == 1){
-                gvl = vsetvl_e64m4(n);
-                vr = vfmv_v_f_f64m4(0, gvl);
+                gvl = __riscv_vsetvl_e64m4(n);
+                vr = __riscv_vfmv_v_f_f64m4(0, gvl);
                 for(i=0,j=0; i<n/gvl; i++){
-                        vx = vle32_v_f32m2(&x[j], gvl);
-                        vy = vle32_v_f32m2(&y[j], gvl);
-                        vr = vfwmacc_vv_f64m4(vr, vx, vy, gvl);
+                        vx = __riscv_vle32_v_f32m2(&x[j], gvl);
+                        vy = __riscv_vle32_v_f32m2(&y[j], gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vr, vx, vy, gvl);
                         j += gvl;
                 }
                 if(j > 0){
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
                 }
                 //tail
                 if(j < n){
-                        gvl = vsetvl_e64m4(n-j);
-                        vx = vle32_v_f32m2(&x[j], gvl);
-                        vy = vle32_v_f32m2(&y[j], gvl);
-                        vfloat64m4_t vz = vfmv_v_f_f64m4(0, gvl);
-                        //vr = vfdot_vv_f32m2(vx, vy, gvl);
-                        vr = vfwmacc_vv_f64m4(vz, vx, vy, gvl);
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        gvl = __riscv_vsetvl_e64m4(n-j);
+                        vx = __riscv_vle32_v_f32m2(&x[j], gvl);
+                        vy = __riscv_vle32_v_f32m2(&y[j], gvl);
+                        vfloat64m4_t vz = __riscv_vfmv_v_f_f64m4(0, gvl);
+                        //vr = __riscv_vfdot_vv_f32m2(vx, vy, gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vz, vx, vy, gvl);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
                 }
         }else if(inc_y == 1){
-                gvl = vsetvl_e64m4(n);
-                vr = vfmv_v_f_f64m4(0, gvl);
+                gvl = __riscv_vsetvl_e64m4(n);
+                vr = __riscv_vfmv_v_f_f64m4(0, gvl);
                  int stride_x = inc_x * sizeof(FLOAT);
                 for(i=0,j=0; i<n/gvl; i++){
-                        vx = vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
-                        vy = vle32_v_f32m2(&y[j], gvl);
-                        vr = vfwmacc_vv_f64m4(vr, vx, vy, gvl);
+                        vx = __riscv_vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
+                        vy = __riscv_vle32_v_f32m2(&y[j], gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vr, vx, vy, gvl);
                         j += gvl;
                 }
                 if(j > 0){
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
 
                 }
                 //tail
                 if(j < n){
-                        gvl = vsetvl_e64m4(n-j);
-                        vx = vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
-                        vy = vle32_v_f32m2(&y[j], gvl);
-                        vfloat64m4_t vz = vfmv_v_f_f64m4(0, gvl);
-                        //vr = vfdot_vv_f32m2(vx, vy, gvl);
-                        vr = vfwmacc_vv_f64m4(vz, vx, vy, gvl);
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        gvl = __riscv_vsetvl_e64m4(n-j);
+                        vx = __riscv_vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
+                        vy = __riscv_vle32_v_f32m2(&y[j], gvl);
+                        vfloat64m4_t vz = __riscv_vfmv_v_f_f64m4(0, gvl);
+                        //vr = __riscv_vfdot_vv_f32m2(vx, vy, gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vz, vx, vy, gvl);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
 
                 }
         }else if(inc_x == 1){
-                gvl = vsetvl_e64m4(n);
-                vr = vfmv_v_f_f64m4(0, gvl);
+                gvl = __riscv_vsetvl_e64m4(n);
+                vr = __riscv_vfmv_v_f_f64m4(0, gvl);
                  int stride_y = inc_y * sizeof(FLOAT);
                 for(i=0,j=0; i<n/gvl; i++){
-                        vx = vle32_v_f32m2(&x[j], gvl);
-                        vy = vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
-                        vr = vfwmacc_vv_f64m4(vr, vx, vy, gvl);
+                        vx = __riscv_vle32_v_f32m2(&x[j], gvl);
+                        vy = __riscv_vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vr, vx, vy, gvl);
                         j += gvl;
                 }
                 if(j > 0){
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
 
                 }
                 //tail
                 if(j < n){
-                        gvl = vsetvl_e64m4(n-j);
-                        vx = vle32_v_f32m2(&x[j], gvl);
-                        vy = vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
-                        vfloat64m4_t vz = vfmv_v_f_f64m4(0, gvl);
-                        //vr = vfdot_vv_f32m2(vx, vy, gvl);
-                        vr = vfwmacc_vv_f64m4(vz, vx, vy, gvl);
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        gvl = __riscv_vsetvl_e64m4(n-j);
+                        vx = __riscv_vle32_v_f32m2(&x[j], gvl);
+                        vy = __riscv_vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
+                        vfloat64m4_t vz = __riscv_vfmv_v_f_f64m4(0, gvl);
+                        //vr = __riscv_vfdot_vv_f32m2(vx, vy, gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vz, vx, vy, gvl);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
 
                 }
         }else{
-                gvl = vsetvl_e64m4(n);
-                vr = vfmv_v_f_f64m4(0, gvl);
+                gvl = __riscv_vsetvl_e64m4(n);
+                vr = __riscv_vfmv_v_f_f64m4(0, gvl);
                  int stride_x = inc_x * sizeof(FLOAT);
                  int stride_y = inc_y * sizeof(FLOAT);
                 for(i=0,j=0; i<n/gvl; i++){
-                        vx = vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
-                        vy = vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
-                        vr = vfwmacc_vv_f64m4(vr, vx, vy, gvl);
+                        vx = __riscv_vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
+                        vy = __riscv_vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vr, vx, vy, gvl);
                         j += gvl;
                 }
                 if(j > 0){
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
 
                 }
                 //tail
                 if(j < n){
-                        gvl = vsetvl_e64m4(n-j);
-                        vx = vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
-                        vy = vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
-                        vfloat64m4_t vz = vfmv_v_f_f64m4(0, gvl);
-                        //vr = vfdot_vv_f32m2(vx, vy, gvl);
-                        vr = vfwmacc_vv_f64m4(vz, vx, vy, gvl);
-                        v_res = vfredusum_vs_f64m4_f64m1(v_res, vr, v_z0, gvl);
-                        dot += (double)vfmv_f_s_f64m1_f64(v_res);
+                        gvl = __riscv_vsetvl_e64m4(n-j);
+                        vx = __riscv_vlse32_v_f32m2(&x[j*inc_x], stride_x, gvl);
+                        vy = __riscv_vlse32_v_f32m2(&y[j*inc_y], stride_y, gvl);
+                        vfloat64m4_t vz = __riscv_vfmv_v_f_f64m4(0, gvl);
+                        //vr = __riscv_vfdot_vv_f32m2(vx, vy, gvl);
+                        vr = __riscv_vfwmacc_vv_f64m4(vz, vx, vy, gvl);
+                        v_res = __riscv_vfredusum_vs_f64m4_f64m1(vr, v_z0, gvl);
+                        dot += (double)__riscv_vfmv_f_s_f64m1_f64(v_res);
 
                 }
         }