ggml-cpu: add rvv ggml_quantize_mat_4x8 for q8_0

Co-authored-by: Rehan Qasim <rehan.qasim@10xengineers.ai>

Author: taimur-10x

ggml/src/ggml-cpu/arch/riscv/repack.cpp

@@ -24,6 +24,94 @@
 
 #define UNUSED GGML_UNUSED
 
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+#if defined(__riscv_v_intrinsic)
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+    const size_t vl_calc = __riscv_vsetvl_e32m8(QK8_0);
+    const size_t vl_save = __riscv_vsetvl_e64m2(4);
+    vfloat32m1_t v_scalar_zero = __riscv_vfmv_s_f_f32m1(0.0f, __riscv_vsetvl_e32m1(1));
+
+    for (int i = 0; i < nb; i++) {
+        const float *x_block_base = x + i * QK8_0;
+        vint8m2_t q_r0, q_r1, q_r2, q_r3;
+        {
+            vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 0 * k, vl_calc);
+            vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
+            vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
+            float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
+
+            float d = amax / 127.0f;
+            y[i].d[0] = GGML_CPU_FP32_TO_FP16(d);
+
+            float id = d ? 1.0f / d : 0.0f;
+            vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
+            vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
+            q_r0 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
+        }
+        asm volatile ("" ::: "memory");
+
+        {
+            vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 1 * k, vl_calc);
+            vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
+            vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
+            float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
+
+            float d = amax / 127.0f;
+            y[i].d[1] = GGML_CPU_FP32_TO_FP16(d);
+            float id = d ? 1.0f / d : 0.0f;
+
+            vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
+            vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
+            q_r1 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
+        }
+        asm volatile ("" ::: "memory");
+        {
+            vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 2 * k, vl_calc);
+            vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
+            vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
+            float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
+
+            float d = amax / 127.0f;
+            y[i].d[2] = GGML_CPU_FP32_TO_FP16(d);
+            float id = d ? 1.0f / d : 0.0f;
+
+            vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
+            vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
+            q_r2 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
+        }
+        asm volatile ("" ::: "memory");
+        {
+            vfloat32m8_t v_src = __riscv_vle32_v_f32m8(x_block_base + 3 * k, vl_calc);
+            vfloat32m8_t v_abs = __riscv_vfabs_v_f32m8(v_src, vl_calc);
+            vfloat32m1_t v_max = __riscv_vfredmax_vs_f32m8_f32m1(v_abs, v_scalar_zero, vl_calc);
+            float amax = __riscv_vfmv_f_s_f32m1_f32(v_max);
+
+            float d = amax / 127.0f;
+            y[i].d[3] = GGML_CPU_FP32_TO_FP16(d);
+            float id = d ? 1.0f / d : 0.0f;
+
+            vfloat32m8_t v_scaled = __riscv_vfmul_vf_f32m8(v_src, id, vl_calc);
+            vint16m4_t v_i16 = __riscv_vfncvt_x_f_w_i16m4_rm(v_scaled, 4, vl_calc);
+            q_r3 = __riscv_vncvt_x_x_w_i8m2(v_i16, vl_calc);
+        }
+        vint64m2_t v_q64_r0 = __riscv_vreinterpret_v_i8m2_i64m2(q_r0);
+        vint64m2_t v_q64_r1 = __riscv_vreinterpret_v_i8m2_i64m2(q_r1);
+        vint64m2_t v_q64_r2 = __riscv_vreinterpret_v_i8m2_i64m2(q_r2);
+        vint64m2_t v_q64_r3 = __riscv_vreinterpret_v_i8m2_i64m2(q_r3);
+        vint64m2x4_t v_quant_tuple = __riscv_vcreate_v_i64m2x4(v_q64_r0, v_q64_r1, v_q64_r2, v_q64_r3);
+        __riscv_vsseg4e64_v_i64m2x4((int64_t*)y[i].qs, v_quant_tuple, vl_save);
+    }
+#else
+    UNUSED(nb);
+    UNUSED(y);
+    ggml_quantize_mat_q8_0_4x4_generic(x, vy, k);
+#endif
+}
+
 void ggml_gemv_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;
     const int nb = n / qk;