fix for reals?

argilo · Feb 1, 2024 · 48d113d · 48d113d
1 parent a289514
commit 48d113d
Showing 1 changed file with 16 additions and 16 deletions.
diff --git a/kernels/volk/volk_8u_x4_conv_k7_r2_8u.h b/kernels/volk/volk_8u_x4_conv_k7_r2_8u.h
@@ -668,17 +668,17 @@ static inline void volk_8u_x4_conv_k7_r2_8u_neonspiral(unsigned char* Y,
         high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d9, 7));
         paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
         paired32_new = vreinterpretq_u8_u32(vsraq_n_u32(paired16, paired16, 12));
-        s20 = ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 1)
-            | ((unsigned int)vgetq_lane_u8(paired32_new, 4) << 9)
-            | ((unsigned int)vgetq_lane_u8(paired32_new, 8) << 17)
-            | ((unsigned int)vgetq_lane_u8(paired32_new, 12) << 25);
-        high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d10, 7));
-        paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
-        paired32_new = vreinterpretq_u8_u32(vsraq_n_u32(paired16, paired16, 12));
-        s20 |= ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 0)
+        s20 = ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 0)
             | ((unsigned int)vgetq_lane_u8(paired32_new, 4) << 8)
             | ((unsigned int)vgetq_lane_u8(paired32_new, 8) << 16)
             | ((unsigned int)vgetq_lane_u8(paired32_new, 12) << 24);
+        high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d10, 7));
+        paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
+        paired32_new = vreinterpretq_u8_u32(vsraq_n_u32(paired16, paired16, 12));
+        s20 |= ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 1)
+            | ((unsigned int)vgetq_lane_u8(paired32_new, 4) << 9)
+            | ((unsigned int)vgetq_lane_u8(paired32_new, 8) << 17)
+            | ((unsigned int)vgetq_lane_u8(paired32_new, 12) << 25);
         a91 = ((unsigned int*)dec);
         a92 = (4 * i9);
         a93 = (a91 + a92);
@@ -715,17 +715,17 @@ printf("%02x%02x%02x%02x", dec[i9*16 + 0], dec[i9*16 + 1], dec[i9*16 + 2], dec[i
         high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d11, 7));
         paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
         paired32_new = vreinterpretq_u8_u32(vsraq_n_u32(paired16, paired16, 12));
-        s26 = ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 1)
-            | ((unsigned int)vgetq_lane_u8(paired32_new, 4) << 9)
-            | ((unsigned int)vgetq_lane_u8(paired32_new, 8) << 17)
-            | ((unsigned int)vgetq_lane_u8(paired32_new, 12) << 25);
-        high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d12, 7));
-        paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
-        paired32_new = vreinterpretq_u8_u32(vsraq_n_u32(paired16, paired16, 12));
-        s26 |= ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 0)
+        s26 = ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 0)
             | ((unsigned int)vgetq_lane_u8(paired32_new, 4) << 8)
             | ((unsigned int)vgetq_lane_u8(paired32_new, 8) << 16)
             | ((unsigned int)vgetq_lane_u8(paired32_new, 12) << 24);
+        high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d12, 7));
+        paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
+        paired32_new = vreinterpretq_u8_u32(vsraq_n_u32(paired16, paired16, 12));
+        s26 |= ((unsigned int)vgetq_lane_u8(paired32_new, 0) << 1)
+            | ((unsigned int)vgetq_lane_u8(paired32_new, 4) << 9)
+            | ((unsigned int)vgetq_lane_u8(paired32_new, 8) << 17)
+            | ((unsigned int)vgetq_lane_u8(paired32_new, 12) << 25);
         a110 = (a93 + 1);
         *(a110) = s26;
 printf("%02x%02x%02x%02x", dec[i9*16 + 4], dec[i9*16 + 5], dec[i9*16 + 6], dec[i9*16 + 7]);