Fix

argilo · Feb 1, 2024 · e2cb86f · e2cb86f
1 parent 264ab06
commit e2cb86f
Showing 1 changed file with 21 additions and 13 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
@@ -634,8 +634,8 @@ static inline void volk_8u_x4_conv_k7_r2_8u_neonspiral(unsigned char* Y,
         uint16x8_t high_bits;
         uint32x4_t paired16;
         uint64x2_t paired32;
+        uint8x16_t paired32_new;
         uint8x16_t paired64;
-        uint16x8_t paired64_new;
         a71 = ((uint8x16_t*)X);
         s18 = *(a71);
         a72 = (a71 + 2);
@@ -668,14 +668,18 @@ static inline void volk_8u_x4_conv_k7_r2_8u_neonspiral(unsigned char* Y,
         d10 = vceqq_u8(a90, m26);
         high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d9, 7));
         paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
-        paired32 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 12));
-        paired64_new = vreinterpretq_u16_u64(vsraq_n_u64(paired32, paired32, 24));
-        s20 = ((unsigned int)vgetq_lane_u16(paired64_new, 0) << 1) | ((unsigned int)vgetq_lane_u16(paired64_new, 4) << 17);
+        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 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 12));
-        paired64_new = vreinterpretq_u16_u64(vsraq_n_u64(paired32, paired32, 24));
-        s20 |= ((unsigned int)vgetq_lane_u16(paired64_new, 0)) | ((unsigned int)vgetq_lane_u16(paired64_new, 4) << 16);
+        paired32_new = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 12));
+        s20 |= ((unsigned int)vgetq_lane_u8(paired32_new, 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);
         a91 = ((unsigned int*)dec);
         a92 = (4 * i9);
         a93 = (a91 + a92);
@@ -710,14 +714,18 @@ static inline void volk_8u_x4_conv_k7_r2_8u_neonspiral(unsigned char* Y,
         d12 = vceqq_u8(a109, m30);
         high_bits = vreinterpretq_u16_u8(vshrq_n_u8(d11, 7));
         paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 6));
-        paired32 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 12));
-        paired64_new = vreinterpretq_u16_u64(vsraq_n_u64(paired32, paired32, 24));
-        s26 = ((unsigned int)vgetq_lane_u16(paired64_new, 0) << 1) | ((unsigned int)vgetq_lane_u16(paired64_new, 4) << 17);
+        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 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 12));
-        paired64_new = vreinterpretq_u16_u64(vsraq_n_u64(paired32, paired32, 24));
-        s26 |= ((unsigned int)vgetq_lane_u16(paired64_new, 0)) | ((unsigned int)vgetq_lane_u16(paired64_new, 4) << 16);
+        paired32_new = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 12));
+        s26 |= ((unsigned int)vgetq_lane_u8(paired32_new, 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);
         a110 = (a93 + 1);
         *(a110) = s26;
         s28 = vzip1q_u8(a108, a109);