ggml-cpu: add floating-point conversion kernels

ggml/src/ggml-cpu/ggml-cpu.c

@@ -3296,12 +3296,30 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
         __m128 y_vec = _mm_cvtph_ps(x_vec);
         _mm_storeu_ps(y + i, y_vec);
     }
-#elif defined(__riscv_zvfh)
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfhmin)
-    for (int vl; i < n; i += vl) {
+    // calculate step size
-        vl = __riscv_vsetvl_e16m1(n - i);
+    const int epr = __riscv_vsetvlmax_e16m4();
-        vfloat16m1_t vx = __riscv_vle16_v_f16m1((_Float16 *)&x[i], vl);
+    const int step = epr * 2;
-        vfloat32m2_t vy = __riscv_vfwcvt_f_f_v_f32m2(vx, vl);
+    const int np = (n & ~(step - 1));
-        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+
+    // unroll by 2
+    for (; i < np; i += step) {
+        vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, epr);
+        vfloat32m8_t ay0 = __riscv_vfwcvt_f_f_v_f32m8(ax0, epr);
+        __riscv_vse32_v_f32m8(y + i, ay0, epr);
+
+        vfloat16m4_t ax1 = __riscv_vle16_v_f16m4((const _Float16*)x + i + epr, epr);
+        vfloat32m8_t ay1 = __riscv_vfwcvt_f_f_v_f32m8(ax1, epr);
+        __riscv_vse32_v_f32m8(y + i + epr, ay1, epr);
+    }
+
+    // leftovers
+    int vl;
+    for (i = np; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m4(n - i);
+        vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, vl);
+        vfloat32m8_t ay0 = __riscv_vfwcvt_f_f_v_f32m8(ax0, vl);
+        __riscv_vse32_v_f32m8(y + i, ay0, vl);
     }
 #endif
 
@@ -3347,6 +3365,31 @@ void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) {
                                         (const __m128i *)(x + i))),
                                 16)));
     }
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfmin)
+    // calculate step size
+    const int epr = __riscv_vsetvlmax_e16m4();
+    const int step = epr * 2;
+    const int np = (n & ~(step - 1));
+
+    // unroll by 2
+    for (; i < np; i += step) {
+        vbfloat16m4_t ax0 = __riscv_vle16_v_bf16m4((const __bf16*)x + i, epr);
+        vfloat32m8_t ay0 = __riscv_vfwcvtbf16_f_f_v_f32m8(ax0, epr);
+        __riscv_vse32_v_f32m8(y + i, ay0, epr);
+
+        vbfloat16m4_t ax1 = __riscv_vle16_v_bf16m4((const __bf16*)x + i + epr, epr);
+        vfloat32m8_t ay1 = __riscv_vfwcvtbf16_f_f_v_f32m8(ax1, epr);
+        __riscv_vse32_v_f32m8(y + i + epr, ay1, epr);
+    }
+
+    // leftovers
+    int vl;
+    for (i = np; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m4(n - i);
+        vbfloat16m4_t ax0 = __riscv_vle16_v_bf16m4((const __bf16*)x + i, vl);
+        vfloat32m8_t ay0 = __riscv_vfwcvtbf16_f_f_v_f32m8(ax0, vl);
+        __riscv_vse32_v_f32m8(y + i, ay0, vl);
+    }
 #endif
     for (; i < n; i++) {
         y[i] = GGML_BF16_TO_FP32(x[i]);