ggml-cpu: fix RVV checks in quants and repacking (llama/20682)

* ggml-cpu: refactor quants.c; add rvv check

* ggml-cpu: refactor; disable generic fallback

diff --git a/src/ggml-cpu/arch/riscv/quants.c b/src/ggml-cpu/arch/riscv/quants.c
index 826055dd9a42c0b4b745df2c9800c647af2a4cf3..d7e9ba4634846dff73c307235d6fb20177366d46 100644 (file)
--- a/src/ggml-cpu/arch/riscv/quants.c
+++ b/src/ggml-cpu/arch/riscv/quants.c
@@ -115,10 +115,10 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
 void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_K == 0);
-    block_q8_K * y_blocks = (block_q8_K *)y;
     size_t nb = k / QK_K;
 
 #if defined(__riscv_v_intrinsic)
+    block_q8_K * y_blocks = (block_q8_K *)y;
     const size_t vlmax_f32m8 = __riscv_vsetvlmax_e32m8();
 
     for (size_t i = 0; i < nb; i++) {
@@ -2052,6 +2052,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_iq1_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
@@ -2147,6 +2148,7 @@ static void ggml_vec_dot_iq1_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
 
     *s = sumf;
 }
+#endif
 
 void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -2163,6 +2165,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_iq1_m_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
@@ -2269,6 +2272,7 @@ static void ggml_vec_dot_iq1_m_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
 
     *s = sumf;
 }
+#endif
 
 void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -2285,6 +2289,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static const uint8_t sign_gather_indices_arr[64] = {
     0,0,0,0,0,0,0,0, 1,1,1,1,1,1,1,1, 2,2,2,2,2,2,2,2, 3,3,3,3,3,3,3,3,
     4,4,4,4,4,4,4,4, 5,5,5,5,5,5,5,5, 6,6,6,6,6,6,6,6, 7,7,7,7,7,7,7,7
@@ -2488,6 +2493,7 @@ static void ggml_vec_dot_iq2_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
     }
     *s = 0.125f * sumf;
 }
+#endif
 
 void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -2507,7 +2513,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
-#if defined(__riscv_v)
+#if defined(__riscv_v_intrinsic)
 static const int8_t keven_signs_q2xs[1024] = {
      1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
      1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
@@ -2542,7 +2548,6 @@ static const int8_t keven_signs_q2xs[1024] = {
      1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
      1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
 };
-#endif
 
 static void ggml_vec_dot_iq2_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -2618,6 +2623,7 @@ static void ggml_vec_dot_iq2_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_
     }
     *s = 0.125f * sumf;
 }
+#endif
 
 void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -2634,6 +2640,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_iq2_xxs_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
@@ -2818,6 +2825,7 @@ static void ggml_vec_dot_iq2_xxs_q8_K_vl256(int n, float * GGML_RESTRICT s, size
     }
     *s = 0.125f * sumf;
 }
+#endif
 
 void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -2830,10 +2838,11 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
             break;
     }
 #else
-    ggml_vec_dot_iq2_xxs_q8_K(n, s, bs, vx, bx, vy, by, nrc);
+    ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_iq3_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
     UNUSED(nrc);
@@ -2928,6 +2937,7 @@ static void ggml_vec_dot_iq3_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
     }
     *s = sumf;
 }
+#endif
 
 void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -2944,6 +2954,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_iq3_xxs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
@@ -3036,6 +3047,7 @@ static void ggml_vec_dot_iq3_xxs_q8_K_vl256(int n, float * GGML_RESTRICT s, size
     }
     *s = 0.25f * sumf;
 }
+#endif
 
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -3052,6 +3064,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_iq4_nl_q8_0_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
@@ -3161,6 +3174,7 @@ static void ggml_vec_dot_iq4_nl_q8_0_vl256(int n, float * GGML_RESTRICT s, size_
 
     *s = sumf;
 }
+#endif
 
 void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -3177,6 +3191,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_iq4_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
@@ -3190,7 +3205,6 @@ static void ggml_vec_dot_iq4_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_
 
     const int nb = n / QK_K;
 
-#if defined __riscv_v_intrinsic
     const vint8m4_t values = __riscv_vle8_v_i8m4(kvalues_iq4nl, 16);
     float sumf = 0;
     int acc[4];
@@ -3252,14 +3266,8 @@ static void ggml_vec_dot_iq4_xs_q8_K_vl256(int n, float * GGML_RESTRICT s, size_
     }
 
     *s = sumf;
-
-#else
-    UNUSED(x);
-    UNUSED(y);
-    UNUSED(nb);
-    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
-#endif
 }
+#endif
 
 void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -3276,6 +3284,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_tq1_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
@@ -3381,6 +3390,7 @@ static void ggml_vec_dot_tq1_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
 
     *s = sumf;
 }
+#endif
 
 void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -3397,6 +3407,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_tq2_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
     assert(nrc == 1);
@@ -3467,6 +3478,7 @@ static void ggml_vec_dot_tq2_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t
 
     *s = sumf;
 }
+#endif
 
 void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -3483,6 +3495,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+#if defined __riscv_v_intrinsic
 static void ggml_vec_dot_mxfp4_q8_0_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
     UNUSED(nrc);
@@ -3592,6 +3605,7 @@ static void ggml_vec_dot_mxfp4_q8_0_vl256(int n, float * GGML_RESTRICT s, size_t
 
     *s = sumf;
 }
+#endif
 
 void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
 #if defined __riscv_v_intrinsic
@@ -3604,6 +3618,6 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
             break;
     }
 #else
-    return ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+    ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }

diff --git a/src/ggml-cpu/arch/riscv/repack.cpp b/src/ggml-cpu/arch/riscv/repack.cpp
index cd5807879ea91729dc0c73efdac1b984807b3131..c37488cae5456b29818676541f6b4916a1a2a046 100644 (file)
--- a/src/ggml-cpu/arch/riscv/repack.cpp
+++ b/src/ggml-cpu/arch/riscv/repack.cpp
@@ -107,8 +107,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
     }
 #else
     UNUSED(nb);
-    UNUSED(y);
-    ggml_quantize_mat_q8_0_4x4_generic(x, vy, k);
+    ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
 #endif
 }
 
@@ -203,6 +202,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+#if defined __riscv_zvfh
 void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -222,7 +222,6 @@ void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined __riscv_v_intrinsic
     const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
         const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb);
@@ -256,9 +255,6 @@ void ggml_gemv_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
         __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
     }
-    return;
-#endif
-    ggml_gemv_q4_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -280,7 +276,6 @@ void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined __riscv_v_intrinsic
     const block_q8_K * a_ptr = (const block_q8_K *) vy;
 
     for (int x = 0; x < nc / ncols_interleaved; x++) {
@@ -392,9 +387,6 @@ void ggml_gemv_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
         __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
     }
-    return;
-#endif
-    ggml_gemv_q4_K_16x1_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -416,7 +408,6 @@ void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined __riscv_v_intrinsic
     const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16);
     const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
@@ -451,9 +442,6 @@ void ggml_gemv_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
 
         __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
     }
-    return;
-#endif
-    ggml_gemv_iq4_nl_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -476,7 +464,6 @@ void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
     UNUSED(blocklen);
     UNUSED(bs);
 
-#if defined __riscv_v_intrinsic
     const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
     for (int x = 0; x < nc / ncols_interleaved; x++) {
         const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb);
@@ -505,9 +492,6 @@ void ggml_gemv_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
         __riscv_vse32_v_f32m2(s + x * 16, sumf, 16);
     }
-    return;
-#endif
-    ggml_gemv_q8_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -679,9 +663,9 @@ void ggml_gemv_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
         } // End K-Block
         __riscv_vse32_v_f32m2(s + col_tile, v_sumf, vl);
-
     }
 }
+#endif
 
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -909,6 +893,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+#if defined __riscv_zvfh
 void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -929,7 +914,6 @@ void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined __riscv_v_intrinsic
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
@@ -994,9 +978,6 @@ void ggml_gemm_q4_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
             __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
         }
     }
-    return;
-#endif
-    ggml_gemm_q4_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1019,7 +1000,6 @@ void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined __riscv_v_intrinsic
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
@@ -1267,9 +1247,6 @@ void ggml_gemm_q4_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
             __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
         }
     }
-    return;
-#endif
-    ggml_gemm_q4_K_16x1_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1292,7 +1269,6 @@ void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined __riscv_v_intrinsic
     const vint8mf2_t values = __riscv_vle8_v_i8mf2(kvalues_iq4nl, 16);
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
@@ -1355,9 +1331,6 @@ void ggml_gemm_iq4_nl_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const
             __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
         }
     }
-    return;
-#endif
-    ggml_gemm_iq4_nl_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1380,7 +1353,6 @@ void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if defined __riscv_v_intrinsic
     for (int y = 0; y < nr / 4; y++) {
         const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
         for (int x = 0; x < nc / ncols_interleaved; x++) {
@@ -1429,9 +1401,6 @@ void ggml_gemm_q8_0_16x1_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
             __riscv_vse32_v_f32m2(s + (y * 4 + 3) * bs + x * 16, sumf_3, 16);
         }
     }
-    return;
-#endif
-    ggml_gemm_q8_0_16x1_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1731,3 +1700,4 @@ void ggml_gemm_q2_K_16x1_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
         }
     }
 }
+#endif

diff --git a/src/ggml-cpu/repack.cpp b/src/ggml-cpu/repack.cpp
index 6b76ab3bfb11c56e4f53dd9585d99177499114c5..f18758f16bb62040333418bab06ce22efe826b20 100644 (file)
--- a/src/ggml-cpu/repack.cpp
+++ b/src/ggml-cpu/repack.cpp
@@ -1365,6 +1365,7 @@ void ggml_gemv_q8_0_4x8_q8_0_generic(int                        n,
     }
 }
 
+// Only enable these for RISC-V.
 #if defined __riscv_zvfh
 void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -1568,6 +1569,7 @@ void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     assert(nc % 16 == 0);
 
     UNUSED(bs);
+    UNUSED(nr);
 
     const int nb = n / QK_K;
     const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx;
@@ -2381,6 +2383,7 @@ void ggml_gemm_q8_0_4x8_q8_0_generic(int                        n,
     }
 }
 
+// Only enable these for RISC-V.
 #if defined __riscv_zvfh
 void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;