Moar

argilo · Feb 1, 2024 · 264ab06 · 264ab06
1 parent 210e5ae
commit 264ab06
Showing 1 changed file with 14 additions and 21 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
@@ -624,8 +624,7 @@ static inline void volk_8u_x4_conv_k7_r2_8u_neonspiral(unsigned char* Y,
         short int s26, s27;
         unsigned int s20;
         unsigned char *a74, *a80, *b6;
-        short int *a110, *a111;
-        unsigned int *a91, *a93, *a94;
+        unsigned int *a110, *a91, *a93, *a94;
         uint8x16_t *a102, *a112, *a113, *a71, *a72, *a77, *a83, *a95, *a96, *a97, *a98,
             *a99;
         uint8x16_t a105, a86;
@@ -676,8 +675,8 @@ static inline void volk_8u_x4_conv_k7_r2_8u_neonspiral(unsigned char* Y,
         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_u8(paired64_new, 4) << 16);
-        a91 = ((short int*)dec);
+        s20 |= ((unsigned int)vgetq_lane_u16(paired64_new, 0)) | ((unsigned int)vgetq_lane_u16(paired64_new, 4) << 16);
+        a91 = ((unsigned int*)dec);
         a92 = (4 * i9);
         a93 = (a91 + a92);
         *(a93) = s20;
@@ -709,24 +708,18 @@ static inline void volk_8u_x4_conv_k7_r2_8u_neonspiral(unsigned char* Y,
         d11 = vceqq_u8(a108, m28);
         a109 = vminq_u8(m30, m29);
         d12 = vceqq_u8(a109, m30);
-        // s26 = _mm_movemask_epi8(vzip1q_u8(d11, d12));
-        high_bits = vreinterpretq_u16_u8(vshrq_n_u8(vzip1q_u8(d11, d12), 7));
-        paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 7));
-        paired32 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 14));
-        paired64 = vreinterpretq_u8_u64(vsraq_n_u64(paired32, paired32, 28));
-        s26 = vgetq_lane_u8(paired64, 0) | ((int)vgetq_lane_u8(paired64, 8) << 8);
-
-        a110 = (a93 + 2);
+        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);
+        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);
+        a110 = (a93 + 1);
         *(a110) = s26;
-        // s27 = _mm_movemask_epi8(vzip2q_u8(d11, d12));
-        high_bits = vreinterpretq_u16_u8(vshrq_n_u8(vzip2q_u8(d11, d12), 7));
-        paired16 = vreinterpretq_u32_u16(vsraq_n_u16(high_bits, high_bits, 7));
-        paired32 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 14));
-        paired64 = vreinterpretq_u8_u64(vsraq_n_u64(paired32, paired32, 28));
-        s27 = vgetq_lane_u8(paired64, 0) | ((int)vgetq_lane_u8(paired64, 8) << 8);
-
-        a111 = (a93 + 3);
-        *(a111) = s27;
         s28 = vzip1q_u8(a108, a109);
         s29 = vzip2q_u8(a108, a109);
         a112 = (a95 + 2);