From aecb7f9537600dc9990b4efb2e1dea40cc3f97be Mon Sep 17 00:00:00 2001
From: Chip Kerchner <ckerchner@tenstorrent.com>
Date: Tue, 7 Oct 2025 13:14:20 +0000
Subject: [PATCH] Change signature of SBGEMV.

---
 kernel/riscv64/sbgemv_n_vector.c | 37 +++++++++++++++++++++++++++++++-
 kernel/riscv64/sbgemv_t_vector.c |  6 +++---
 2 files changed, 39 insertions(+), 4 deletions(-)

diff --git a/kernel/riscv64/sbgemv_n_vector.c b/kernel/riscv64/sbgemv_n_vector.c
index 0c6064b67..94b9488cf 100644
--- a/kernel/riscv64/sbgemv_n_vector.c
+++ b/kernel/riscv64/sbgemv_n_vector.c
@@ -32,6 +32,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define VLSEV_FLOAT             RISCV_RVV(vlse32_v_f32m8)
 #define VSEV_FLOAT              RISCV_RVV(vse32_v_f32m8)
 #define VSSEV_FLOAT             RISCV_RVV(vsse32_v_f32m8)
+#define VFMULVF_FLOAT           RISCV_RVV(vfmul_vf_f32m8)
+#define VFMVVF_FLOAT            RISCV_RVV(vfmv_v_f_f32m8)
 
 #define VSETVL(n)               RISCV_RVV(vsetvl_e16m4)(n)
 
@@ -45,7 +47,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define VFMACCVF_FLOAT          RISCV_RVV(vfwmaccbf16_vf_f32m8)
 #endif
 
-int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *y, BLASLONG inc_y)
 {
     if (n < 0) return(0);
 
@@ -55,7 +57,24 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, IFLOAT *a, BLASL
     IFLOAT_V_T va;
     FLOAT_V_T vy;
 
+    y_ptr = y;
     if (inc_y == 1) {
+        if (beta == 0.0) {
+            for (i = m; i > 0; i -= vl) {
+                vl = VSETVL(i);
+                vy = VFMVVF_FLOAT(0.0, vl);
+                VSEV_FLOAT(y_ptr, vy, vl);
+                y_ptr += vl;
+            }
+        } else if (beta != 1.0) {
+            for (i = m; i > 0; i -= vl) {
+                vl = VSETVL(i);
+                vy = VLEV_FLOAT(y_ptr, vl);
+                vy = VFMULVF_FLOAT(vy, beta, vl);
+                VSEV_FLOAT(y_ptr, vy, vl);
+                y_ptr += vl;
+            }
+        }
         for (j = 0; j < n; j++) {
             temp = (IFLOAT)(alpha * (FLOAT)(x[0]));
             y_ptr = y;
@@ -74,6 +93,22 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, IFLOAT *a, BLASL
         }
     } else {
         BLASLONG stride_y = inc_y * sizeof(FLOAT);
+        if (beta == 0.0) {
+            for (i = m; i > 0; i -= vl) {
+                vl = VSETVL(i);
+                vy = VFMVVF_FLOAT(0.0, vl);
+                VSSEV_FLOAT(y_ptr, stride_y, vy, vl);
+                y_ptr += vl * inc_y;
+            }
+        } else if (beta != 1.0) {
+            for (i = m; i > 0; i -= vl) {
+                vl = VSETVL(i);
+                vy = VLSEV_FLOAT(y_ptr, stride_y, vl);
+                vy = VFMULVF_FLOAT(vy, beta, vl);
+                VSSEV_FLOAT(y_ptr, stride_y, vy, vl);
+                y_ptr += vl * inc_y;
+            }
+        }
         for (j = 0; j < n; j++) {
             temp = (IFLOAT)(alpha * (FLOAT)(x[0]));
             y_ptr = y;
diff --git a/kernel/riscv64/sbgemv_t_vector.c b/kernel/riscv64/sbgemv_t_vector.c
index 09069caaa..7fdccee1b 100644
--- a/kernel/riscv64/sbgemv_t_vector.c
+++ b/kernel/riscv64/sbgemv_t_vector.c
@@ -54,7 +54,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define VFMVVF_FLOAT            RISCV_RVV(vfmv_v_f_f32m8)
 #define VFMVVF_FLOAT_M1         RISCV_RVV(vfmv_v_f_f32m1)
 
-int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+int CNAME(BLASLONG m, BLASLONG n, FLOAT alpha, IFLOAT *a, BLASLONG lda, IFLOAT *x, BLASLONG inc_x, FLOAT beta, FLOAT *y, BLASLONG inc_y)
 {
     BLASLONG i = 0, j = 0, k = 0;
     BLASLONG ix = 0, iy = 0;
@@ -92,7 +92,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, IFLOAT *a, BLASL
                 v_res = VFREDSUM_FLOAT(vr, v_res, gvl);
             }
             temp = (FLOAT)EXTRACT_FLOAT(v_res);
-            y[iy] += alpha * temp;
+            y[iy] = y[iy] * beta + alpha * temp;
 
             iy += inc_y;
             a_ptr += lda;
@@ -123,7 +123,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, IFLOAT *a, BLASL
                 v_res = VFREDSUM_FLOAT(vr, v_res, gvl);
             }
             temp = (FLOAT)EXTRACT_FLOAT(v_res);
-            y[iy] += alpha * temp;
+            y[iy] = y[iy] * beta + alpha * temp;
 
             iy += inc_y;
             a_ptr += lda;