RISC-V: faster DAMAX kernel for RISCV64_ZVL256B

kernel/riscv64/KERNEL.RISCV64_ZVL256B

@@ -1,5 +1,5 @@
 SAMAXKERNEL  = amax_vector.c
-DAMAXKERNEL  = amax_vector.c
+DAMAXKERNEL  = amax_zvl256b.c
 CAMAXKERNEL  = zamax_vector.c
 ZAMAXKERNEL  = zamax_vector.c
 

kernel/riscv64/amax_zvl256b.c

@@ -0,0 +1,133 @@
+/***************************************************************************
+Copyright (c) 2026, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+#include <math.h>
+
+/*
+ * Two paths, chosen from measurements on the SpacemiT X100 (VLEN 256):
+ *  - data up to STREAM_BYTES: one LMUL-8 accumulator and a single final
+ *    reduction (several times faster in cache than reducing every strip);
+ *  - larger unit-stride data, which streams from DRAM: four LMUL-1
+ *    accumulators. On the X100 small vector loads sustain a markedly higher
+ *    single-core DRAM read bandwidth than LMUL-8 loads.
+ */
+#define STREAM_BYTES (2 * 1024 * 1024)
+
+#if !defined(DOUBLE)
+#define VSETVL_M8(n)            __riscv_vsetvl_e32m8(n)
+#define VSETVL_MAX_M8           __riscv_vsetvlmax_e32m8()
+#define FLOAT_V_T_M8            vfloat32m8_t
+#define VLEV_M8                 __riscv_vle32_v_f32m8
+#define VLSEV_M8                __riscv_vlse32_v_f32m8
+#define VFABS_M8                __riscv_vfabs_v_f32m8
+#define VFMVVF_M8               __riscv_vfmv_v_f_f32m8
+#define VSETVL_M1(n)            __riscv_vsetvl_e32m1(n)
+#define VSETVL_MAX_M1           __riscv_vsetvlmax_e32m1()
+#define FLOAT_V_T_M1            vfloat32m1_t
+#define VLEV_M1                 __riscv_vle32_v_f32m1
+#define VFABS_M1                __riscv_vfabs_v_f32m1
+#define VFMVVF_M1               __riscv_vfmv_v_f_f32m1
+#define VFMVFS_M1               __riscv_vfmv_f_s_f32m1_f32
+#define ABS_FLOAT               fabsf
+#else
+#define VSETVL_M8(n)            __riscv_vsetvl_e64m8(n)
+#define VSETVL_MAX_M8           __riscv_vsetvlmax_e64m8()
+#define FLOAT_V_T_M8            vfloat64m8_t
+#define VLEV_M8                 __riscv_vle64_v_f64m8
+#define VLSEV_M8                __riscv_vlse64_v_f64m8
+#define VFABS_M8                __riscv_vfabs_v_f64m8
+#define VFMVVF_M8               __riscv_vfmv_v_f_f64m8
+#define VSETVL_M1(n)            __riscv_vsetvl_e64m1(n)
+#define VSETVL_MAX_M1           __riscv_vsetvlmax_e64m1()
+#define FLOAT_V_T_M1            vfloat64m1_t
+#define VLEV_M1                 __riscv_vle64_v_f64m1
+#define VFABS_M1                __riscv_vfabs_v_f64m1
+#define VFMVVF_M1               __riscv_vfmv_v_f_f64m1
+#define VFMVFS_M1               __riscv_vfmv_f_s_f64m1_f64
+#define ABS_FLOAT               fabs
+#endif
+#if !defined(DOUBLE)
+#define VFMAX_M8_TU              __riscv_vfmax_vv_f32m8_tu
+#define VFREDMAX_M8              __riscv_vfredmax_vs_f32m8_f32m1
+#define VFMAX_M1                 __riscv_vfmax_vv_f32m1
+#define VFMAX_M1_TU              __riscv_vfmax_vv_f32m1_tu
+#define VFREDMAX_M1              __riscv_vfredmax_vs_f32m1_f32m1
+#endif
+#if defined(DOUBLE)
+#define VFMAX_M8_TU              __riscv_vfmax_vv_f64m8_tu
+#define VFREDMAX_M8              __riscv_vfredmax_vs_f64m8_f64m1
+#define VFMAX_M1                 __riscv_vfmax_vv_f64m1
+#define VFMAX_M1_TU              __riscv_vfmax_vv_f64m1_tu
+#define VFREDMAX_M1              __riscv_vfredmax_vs_f64m1_f64m1
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    if (n <= 0 || inc_x <= 0) return(0.0);
+
+    // The result starts from 0, as in amax_vector.c.
+    FLOAT init = 0.0;
+    FLOAT_V_T_M1 v_res = VFMVVF_M1(init, VSETVL_MAX_M1);
+
+    if (inc_x == 1 && (size_t)n * sizeof(FLOAT) >= STREAM_BYTES) {
+        size_t vl1 = VSETVL_MAX_M1;
+        FLOAT_V_T_M1 a0 = VFMVVF_M1(init, vl1), a1 = a0, a2 = a0, a3 = a0;
+        BLASLONG j = 0, step = 4 * (BLASLONG)vl1;
+        for (; j + step <= n; j += step) {
+            a0 = VFMAX_M1(a0, VFABS_M1(VLEV_M1(&x[j], vl1), vl1), vl1);
+            a1 = VFMAX_M1(a1, VFABS_M1(VLEV_M1(&x[j + vl1], vl1), vl1), vl1);
+            a2 = VFMAX_M1(a2, VFABS_M1(VLEV_M1(&x[j + 2 * vl1], vl1), vl1), vl1);
+            a3 = VFMAX_M1(a3, VFABS_M1(VLEV_M1(&x[j + 3 * vl1], vl1), vl1), vl1);
+        }
+        for (size_t vl; j < n; j += vl) {
+            vl = VSETVL_M1(n - j);
+            a0 = VFMAX_M1_TU(a0, a0, VFABS_M1(VLEV_M1(&x[j], vl), vl), vl);
+        }
+        a0 = VFMAX_M1(VFMAX_M1(a0, a1, vl1), VFMAX_M1(a2, a3, vl1), vl1);
+        v_res = VFREDMAX_M1(a0, v_res, vl1);
+        return(VFMVFS_M1(v_res));
+    }
+
+    size_t vlmax = VSETVL_MAX_M8;
+    FLOAT_V_T_M8 acc = VFMVVF_M8(init, vlmax);
+    if (inc_x == 1) {
+        for (size_t vl; n > 0; n -= vl, x += vl) {
+            vl = VSETVL_M8(n);
+            acc = VFMAX_M8_TU(acc, acc, VFABS_M8(VLEV_M8(x, vl), vl), vl);
+        }
+    } else {
+        BLASLONG stride_x = inc_x * sizeof(FLOAT);
+        for (size_t vl; n > 0; n -= vl, x += vl * inc_x) {
+            vl = VSETVL_M8(n);
+            acc = VFMAX_M8_TU(acc, acc, VFABS_M8(VLSEV_M8(x, stride_x, vl), vl), vl);
+        }
+    }
+    v_res = VFREDMAX_M8(acc, v_res, vlmax);
+    return(VFMVFS_M1(v_res));
+}