Optimize ggml_vec_dot_q4_K_q8_K for LoongArch ASX

Author: MQ-mengqing

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

@@ -6569,11 +6569,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     *s = vec_extract(vsumf0, 0);
 
 #elif defined __loongarch_asx
-    GGML_UNUSED(kmask1);
-    GGML_UNUSED(kmask2);
-    GGML_UNUSED(kmask3);
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
     __m128 acc_m = (__m128)__lsx_vldi(0);
@@ -6593,33 +6588,34 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+        const __m128i mins_and_scales128 = lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i mins128 = __lsx_vexth_h_b(mins_and_scales128);
+        const __m128i scales128 = __lsx_vsllwil_h_b(mins_and_scales128, 0);
 
         const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
         const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
-        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
+        const __m128i prod = lsx_madd_h(mins128, q8s);
         acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
 
-        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
-        const __m256i scales = lasx_insertf128(sc128, sc128);
+        const __m256i scales = lasx_insertf128(scales128, scales128);
 
         __m256i sumi = __lasx_xvldi(0);
 
         for (int j = 0; j < QK_K/64; ++j) {
 
-            const __m256i scale_l = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_h = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+            const __m256i scale_l = lasx_xvrepl128vei_h(scales, 2 * j + 0);
+            const __m256i scale_h = lasx_xvrepl128vei_h(scales, 2 * j + 1);
 
             const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
-            const __m256i q4l = __lasx_xvand_v(q4bits, m4);
-            const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4);
+            const __m256i q4l = __lasx_xvandi_b(q4bits, 0xf);
+            const __m256i q4h = __lasx_xvsrli_b(q4bits, 4);
 
             const __m256i q8l = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16l = lasx_maddubs_h(q4l, q8l);
+            __m256i p16l = lasx_madd_h_b(q4l, q8l);
             p16l = lasx_madd_h(scale_l, p16l);
 
             const __m256i q8h = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16h = lasx_maddubs_h(q4h, q8h);
+            __m256i p16h = lasx_madd_h_b(q4h, q8h);
             p16h = lasx_madd_h(scale_h, p16h);
             const __m256i sumj = __lasx_xvadd_w(p16l, p16h);