hexagon: remove the need for vdelta in reduce-multiply-x8

Author: max-krasnyansky

ggml/src/ggml-hexagon/htp/matmul-ops.c

@@ -253,48 +253,22 @@ static inline HVX_Vector hvx_vec_rmpy_x8(HVX_Vector_x8 x, HVX_Vector_x8 y) {
     HVX_Vector r6 = Q6_Vw_vrmpy_VbVb(x.v[6], y.v[6]);
     HVX_Vector r7 = Q6_Vw_vrmpy_VbVb(x.v[7], y.v[7]);
 
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(r4, r0, 16);
-    HVX_VectorPair p1 = Q6_W_vshuff_VVR(r5, r1, 16);
-    HVX_VectorPair p2 = Q6_W_vshuff_VVR(r6, r2, 16);
-    HVX_VectorPair p3 = Q6_W_vshuff_VVR(r7, r3, 16);
-
+    HVX_VectorPair p0 = Q6_W_vdeal_VVR(r1, r0, -4);
+    HVX_VectorPair p1 = Q6_W_vdeal_VVR(r3, r2, -4);
+    HVX_VectorPair p2 = Q6_W_vdeal_VVR(r5, r4, -4);
+    HVX_VectorPair p3 = Q6_W_vdeal_VVR(r7, r6, -4);
     r0 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p0), Q6_V_hi_W(p0));
     r1 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p1), Q6_V_hi_W(p1));
     r2 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p2), Q6_V_hi_W(p2));
     r3 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p3), Q6_V_hi_W(p3));
 
-    p0 = Q6_W_vshuff_VVR(r2, r0, 8);
-    p1 = Q6_W_vshuff_VVR(r3, r1, 8);
-    r0 = Q6_Vw_vadd_VwVw(Q6_V_hi_W(p0), Q6_V_lo_W(p0));
-    r1 = Q6_Vw_vadd_VwVw(Q6_V_hi_W(p1), Q6_V_lo_W(p1));
-
-    p0 = Q6_W_vshuff_VVR(r1, r0, 4);
-    r0 = Q6_Vw_vadd_VwVw(Q6_V_hi_W(p0), Q6_V_lo_W(p0));
-
-    static const uint8_t vrd[128] __attribute__((aligned(128))) = {
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-        0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40,
-        0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54, 0x54,
-        0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14, 0x14,
-        0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28, 0x28,
-        0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68, 0x68,
-        0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C, 0x7C,
-        0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C, 0x3C,
-    };
-
-    static const uint8_t vd[128] __attribute__((aligned(128))) = {
-        0x00, 0x00, 0x00, 0x00, 0x10, 0x10, 0x10, 0x10, 0x20, 0x20, 0x20, 0x20, 0x30, 0x30, 0x30, 0x30,
-        0x24, 0x24, 0x24, 0x24, 0x34, 0x34, 0x34, 0x34, 0x04, 0x04, 0x04, 0x04, 0x14, 0x14, 0x14, 0x14,
-        0x08, 0x08, 0x08, 0x08, 0x18, 0x18, 0x18, 0x18, 0x28, 0x28, 0x28, 0x28, 0x38, 0x38, 0x38, 0x38,
-        0x2C, 0x2C, 0x2C, 0x2C, 0x3C, 0x3C, 0x3C, 0x3C, 0x0C, 0x0C, 0x0C, 0x0C, 0x1C, 0x1C, 0x1C, 0x1C,
-        0x30, 0x30, 0x30, 0x30, 0x20, 0x20, 0x20, 0x20, 0x10, 0x10, 0x10, 0x10, 0x00, 0x00, 0x00, 0x00,
-        0x14, 0x14, 0x14, 0x14, 0x04, 0x04, 0x04, 0x04, 0x34, 0x34, 0x34, 0x34, 0x24, 0x24, 0x24, 0x24,
-        0x38, 0x38, 0x38, 0x38, 0x28, 0x28, 0x28, 0x28, 0x18, 0x18, 0x18, 0x18, 0x08, 0x08, 0x08, 0x08,
-        0x1C, 0x1C, 0x1C, 0x1C, 0x0C, 0x0C, 0x0C, 0x0C, 0x3C, 0x3C, 0x3C, 0x3C, 0x2C, 0x2C, 0x2C, 0x2C,
-    };
-
-    r0 = Q6_V_vrdelta_VV(r0, *(HVX_Vector *) vrd);
-    r0 = Q6_V_vdelta_VV(r0, *(HVX_Vector *) vd);
+    p0 = Q6_W_vdeal_VVR(r1, r0, -4);
+    p1 = Q6_W_vdeal_VVR(r3, r2, -4);
+    r0 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p0), Q6_V_hi_W(p0));
+    r1 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p1), Q6_V_hi_W(p1));
+
+    p0 = Q6_W_vdeal_VVR(r1, r0, -4);
+    r0 = Q6_Vw_vadd_VwVw(Q6_V_lo_W(p0), Q6_V_hi_W(p0));
 
     return r0;
 }