ggml-cpu: fix fallback for RVV kernels without zvfh (#21157)

* ggml-cpu: refactor sgemm; fix rvv checks

* ggml-cpu: refactor rvv kernels; set zvfbfwma default to off

ggml/CMakeLists.txt

@@ -172,7 +172,8 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"         ON)
 option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"        ON)
 option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"      ON)
-option(GGML_RV_ZIHINTPAUSE   "ggml: enable riscv zihintpause "  ON)
+option(GGML_RV_ZIHINTPAUSE   "ggml: enable riscv zihintpause" ON)
+option(GGML_RV_ZVFBFWMA      "ggml: enable riscv zvfbfwma"    OFF)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"      OFF)
 option(GGML_VXE              "ggml: enable vxe"               ${GGML_NATIVE})
 

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -180,44 +180,49 @@ inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
 }
 #endif
 
-#if defined(__riscv_zvfh)
+#if defined(__riscv_v_intrinsic)
-template <>
+template <> inline vfloat32m1_t madd(vfloat32m1_t a, vfloat32m1_t b, vfloat32m1_t c) {
-inline vfloat32m1_t madd(vfloat16mf2_t a, vfloat16mf2_t b, vfloat32m1_t c) {
-    return __riscv_vfwmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
-}
-inline vfloat32m2_t madd(vfloat16m1_t a, vfloat16m1_t b, vfloat32m2_t c) {
-    return __riscv_vfwmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
-}
-inline vfloat32m4_t madd(vfloat16m2_t a, vfloat16m2_t b, vfloat32m4_t c) {
-    return __riscv_vfwmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
-}
-inline vfloat32m8_t madd(vfloat16m4_t a, vfloat16m4_t b, vfloat32m8_t c) {
-    return __riscv_vfwmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
-}
-inline vfloat32m1_t madd(vfloat32m1_t a, vfloat32m1_t b, vfloat32m1_t c) {
     return __riscv_vfmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
 }
-inline vfloat32m2_t madd(vfloat32m2_t a, vfloat32m2_t b, vfloat32m2_t c) {
+template <> inline vfloat32m2_t madd(vfloat32m2_t a, vfloat32m2_t b, vfloat32m2_t c) {
     return __riscv_vfmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
 }
-inline vfloat32m4_t madd(vfloat32m4_t a, vfloat32m4_t b, vfloat32m4_t c) {
+template <> inline vfloat32m4_t madd(vfloat32m4_t a, vfloat32m4_t b, vfloat32m4_t c) {
     return __riscv_vfmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
 }
-inline vfloat32m8_t madd(vfloat32m8_t a, vfloat32m8_t b, vfloat32m8_t c) {
+template <> inline vfloat32m8_t madd(vfloat32m8_t a, vfloat32m8_t b, vfloat32m8_t c) {
     return __riscv_vfmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <> inline vfloat32m1_t madd(vfloat16mf2_t a, vfloat16mf2_t b, vfloat32m1_t c) {
+    return __riscv_vfwmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+template <> inline vfloat32m2_t madd(vfloat16m1_t a, vfloat16m1_t b, vfloat32m2_t c) {
+    return __riscv_vfwmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+template <> inline vfloat32m4_t madd(vfloat16m2_t a, vfloat16m2_t b, vfloat32m4_t c) {
+    return __riscv_vfwmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+template <> inline vfloat32m8_t madd(vfloat16m4_t a, vfloat16m4_t b, vfloat32m8_t c) {
+    return __riscv_vfwmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
 #if defined(__riscv_zvfbfwma)
-inline vfloat32m1_t madd(vbfloat16mf2_t a, vbfloat16mf2_t b, vfloat32m1_t c) {
+template <> inline vfloat32m1_t madd(vbfloat16mf2_t a, vbfloat16mf2_t b, vfloat32m1_t c) {
     return __riscv_vfwmaccbf16_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
 }
-inline vfloat32m2_t madd(vbfloat16m1_t a, vbfloat16m1_t b, vfloat32m2_t c) {
+template <> inline vfloat32m2_t madd(vbfloat16m1_t a, vbfloat16m1_t b, vfloat32m2_t c) {
     return __riscv_vfwmaccbf16_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
 }
-inline vfloat32m4_t madd(vbfloat16m2_t a, vbfloat16m2_t b, vfloat32m4_t c) {
+template <> inline vfloat32m4_t madd(vbfloat16m2_t a, vbfloat16m2_t b, vfloat32m4_t c) {
     return __riscv_vfwmaccbf16_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
 }
+template <> inline vfloat32m8_t madd(vbfloat16m4_t a, vbfloat16m4_t b, vfloat32m8_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
 #endif
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -272,7 +277,7 @@ inline float hsum(__m512 x) {
 }
 #endif // __AVX512F__
 
-#if defined(__riscv_zvfh)
+#if defined(__riscv_v_intrinsic)
 inline float hsum(vfloat32m1_t x) {
     return __riscv_vfmv_f_s_f32m1_f32(
         __riscv_vfredusum_vs_f32m1_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m1()));
@@ -379,19 +384,7 @@ template <> inline __m256bh load(const float *p) {
 }
 #endif
 
-#if defined(__riscv_zvfh)
+#if defined(__riscv_v_intrinsic)
-template <> inline vfloat16mf2_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16mf2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16mf2());
-}
-template <> inline vfloat16m1_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16m1(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m1());
-}
-template <> inline vfloat16m2_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16m2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m2());
-}
-template <> inline vfloat16m4_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16m4(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m4());
-}
 template <> inline vfloat32m1_t load(const float *p) {
     return __riscv_vle32_v_f32m1(p, __riscv_vsetvlmax_e32m1());
 }
@@ -406,6 +399,21 @@ template <> inline vfloat32m8_t load(const float *p) {
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <> inline vfloat16mf2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16mf2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vfloat16m1_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m1(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m1());
+}
+template <> inline vfloat16m2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m2());
+}
+template <> inline vfloat16m4_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m4(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m4());
+}
+#endif
+
 #if defined(__riscv_zvfbfwma)
 template <> inline vbfloat16mf2_t load(const ggml_bf16_t *p) {
     return __riscv_vle16_v_bf16mf2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16mf2());
@@ -416,23 +424,14 @@ template <> inline vbfloat16m1_t load(const ggml_bf16_t *p) {
 template <> inline vbfloat16m2_t load(const ggml_bf16_t *p) {
     return __riscv_vle16_v_bf16m2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m2());
 }
+template <> inline vbfloat16m4_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16m4(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m4());
+}
 #endif
 
-#if defined(__riscv_zvfh)
+#if defined(__riscv_v_intrinsic)
 template <typename T> T set_zero();
 
-template <> inline vfloat16mf2_t set_zero() {
-    return __riscv_vfmv_v_f_f16mf2(0, __riscv_vsetvlmax_e16mf2());
-}
-template <> inline vfloat16m1_t set_zero() {
-    return __riscv_vfmv_v_f_f16m1(0, __riscv_vsetvlmax_e16m1());
-}
-template <> inline vfloat16m2_t set_zero() {
-    return __riscv_vfmv_v_f_f16m2(0, __riscv_vsetvlmax_e16m2());
-}
-template <> inline vfloat16m4_t set_zero() {
-    return __riscv_vfmv_v_f_f16m4(0, __riscv_vsetvlmax_e16m4());
-}
 template <> inline vfloat32m1_t set_zero() {
     return __riscv_vfmv_v_f_f32m1(0.0f, __riscv_vsetvlmax_e32m1());
 }
@@ -449,14 +448,22 @@ template <> inline vfloat32m8_t set_zero() {
 
 #if defined(__riscv_v_intrinsic)
 template <typename T> size_t vlmax() {
-    if constexpr (std::is_same_v<T, vfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
+    if constexpr (std::is_same_v<T, vfloat32m1_t>) { return  __riscv_vsetvlmax_e32m1(); }
-    else if constexpr (std::is_same_v<T, vfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
-    else if constexpr (std::is_same_v<T, vfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
-    else if constexpr (std::is_same_v<T, vfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
-    else if constexpr (std::is_same_v<T, vfloat32m1_t>) { return  __riscv_vsetvlmax_e32m1(); }
     else if constexpr (std::is_same_v<T, vfloat32m2_t>) { return  __riscv_vsetvlmax_e32m2(); }
     else if constexpr (std::is_same_v<T, vfloat32m4_t>) { return  __riscv_vsetvlmax_e32m4(); }
     else if constexpr (std::is_same_v<T, vfloat32m8_t>) { return  __riscv_vsetvlmax_e32m8(); }
+    #if defined (__riscv_zvfh)
+    else if constexpr (std::is_same_v<T, vfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
+    else if constexpr (std::is_same_v<T, vfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
+    #endif
+    #if defined (__riscv_zvfbfwma)
+    else if constexpr (std::is_same_v<T, vbfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
+    else if constexpr (std::is_same_v<T, vbfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
+    else if constexpr (std::is_same_v<T, vbfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
+    else if constexpr (std::is_same_v<T, vbfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
+    #endif
     return 0;
 }
 #endif
@@ -3740,7 +3747,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             params->ith, params->nth};
         tb.matmul(m, n);
         return true;
-#elif defined(__riscv_zvfh)
+#elif defined(__riscv_v_intrinsic)
     #if LMUL == 1
         tinyBLAS_RVV<vfloat32m1_t, vfloat32m1_t, float, float, float> tb{ params,
             k, (const float *)A, lda,
@@ -3804,6 +3811,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             return true;
         }
 #elif defined(__riscv_zvfbfwma)
+        if (Btype == GGML_TYPE_BF16) {
             #if LMUL == 1
                 tinyBLAS_RVV<vfloat32m1_t, vbfloat16mf2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
                     k, (const ggml_bf16_t *)A, lda,
@@ -3821,6 +3829,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                     (float *)C, ldc};
             #endif
                 return tb.matmul(m, n);
+        }
 #endif
         return false;
     }

ggml/src/ggml-cpu/vec.h

@@ -126,7 +126,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         const int ggml_f16_epr = sve_register_length / 16; // running when 16
         const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
 
-        const int np = (n & ~(ggml_f16_step - 1));
+        int np = (n & ~(ggml_f16_step - 1));
 
         svfloat16_t sum_00 = svdup_n_f16(0.0f);
         svfloat16_t sum_01 = svdup_n_f16(0.0f);
@@ -224,8 +224,9 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         }
         GGML_F16x_VEC_REDUCE(sumf[0], sum_00, sum_01, sum_02, sum_03);
         GGML_F16x_VEC_REDUCE(sumf[1], sum_10, sum_11, sum_12, sum_13);
-
+        np = n;
-    #elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
+    #elif defined(__riscv_v_intrinsic)
+        #if defined(__riscv_zvfh)
             size_t vl = __riscv_vsetvlmax_e32m4();
 
             // initialize accumulators to all zeroes
@@ -237,7 +238,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
             // calculate step size
             const size_t epr = __riscv_vsetvlmax_e16m2();
             const size_t step = epr * 2;
-        const int np = (n & ~(step - 1));
+            int np = (n & ~(step - 1));
 
             // unroll by 2 along the row dimension
             for (int i = 0; i < np; i += step) {
@@ -288,7 +289,10 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
                                         acc1_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
             sumf[0] = __riscv_vfmv_f_s_f32m1_f32(redsum0);
             sumf[1] = __riscv_vfmv_f_s_f32m1_f32(redsum1);
-
+            np = n;
+        #else
+            const int np = 0;
+        #endif
     #else
         const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -313,21 +317,17 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
             GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
         }
-
+    #endif
-        // leftovers
+#else
+    // scalar path
+    const int np = 0;
+#endif
+    // scalar and leftovers
     for (int i = np; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
             sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
     }
-    #endif
-#else
-    for (int i = 0; i < n; ++i) {
-        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
-        }
-    }
-#endif
 
     for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
         s[i] = (float)sumf[i];
@@ -532,7 +532,8 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
         svst1_f16(pg, (__fp16 *)(y + np2), hy);
     }
     np = n;
-#elif defined(__riscv_zvfh) // implies __riscv_v_intrinsic
+#elif defined(__riscv_v_intrinsic) // implies __riscv_v_intrinsic
+    #if defined (__riscv_zvfh)
         const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
         const _Float16 scale = *(const _Float16*)(&s);
 
@@ -566,6 +567,10 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
             __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
         }
         np = n;
+    #else
+        // fall to scalar path
+        const int np = 0;
+    #endif
 #elif defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -584,10 +589,11 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
         }
     }
 #else
+    // scalar path
     const int np = 0;
 #endif
 
-    // leftovers
+    // scalar and leftovers
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
@@ -785,7 +791,7 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
     const int ggml_f16_step = 2 * ggml_f16_epr;
 
     GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
-    const int np = (n & ~(ggml_f16_step - 1));
+    int np = (n & ~(ggml_f16_step - 1));
     svfloat16_t ay1, ay2;
 
     for (int i = 0; i < np; i += ggml_f16_step) {
@@ -805,14 +811,16 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
         svfloat16_t out = svmul_f16_m(pg, hy, vx);
         svst1_f16(pg, (__fp16 *)(y + np), out);
     }
-#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
+    np = n;
+#elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_zvfh)
         const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
         const _Float16 scale = *(const _Float16*)(&s);
 
         // calculate step size
         const int epr = __riscv_vsetvlmax_e16m4();
         const int step = epr * 2;
-    const int np = (n & ~(step - 1));
+        int np = (n & ~(step - 1));
 
         // unroll by 2
         for (int i = 0; i < np; i += step) {
@@ -835,6 +843,11 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
             ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, vl);
             __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
         }
+        np = n;
+    #else
+        // fall to scalar path
+        const int np = 0;
+    #endif
 #elif defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -850,17 +863,14 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
             GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
     }
-
+#else
-    // leftovers
+    // scalar path
+    const int np = 0;
+#endif
+    // scalar and leftovers
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
     }
-#else
-    // scalar
-    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-    }
-#endif
 }
 
 inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s);   }