volk_64u_byteswap_neonv8 incorrect results about volk HOT 5 CLOSED

On 10/24/22 10:20, jsallay wrote: I am using this kernel on an nvidia jetson and noticed that I am getting incorrect results. To debug, I made a simple program that shows the issue. |#include <arm_neon.h> #include <cstddef> #include <cstdio> int main() { uint64_t data[] = {0x0706050403020100, 0x1716151413121110, 0x2726252423222120, 0x3736353433323130}; uint8x16x2_t input; uint8x16_t idx = { 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8 }; for (size_t i = 0; i < 32; i++) printf("%02x ", ((char*)(&data))[i]); printf("\n"); input = vld2q_u8((uint8_t*)data); input.val[0] = vqtbl1q_u8(input.val[0], idx); input.val[1] = vqtbl1q_u8(input.val[1], idx); vst2q_u8((uint8_t*)data, input); for (size_t i = 0; i < 32; i++) printf("%02x ", ((char*)(&data))[i]); printf("\n"); } | I've tried this program out on an armv8 nivdia jetson nx and using QEMU in docker. In both cases, the output does not match what is expected. Doing some digging, it appears that the load instruction isn't doing what I would expect: |vld2q_u8 Load multiple 2-element structures to two registers. This instruction loads multiple 2-element structures from memory and writes the result to the two SIMD&FP registers, with de-interleaving. | The instruction is loading in the 32-bytes, but assuming that they are interleaved and moving the memory around before we do the shuffle instruction. Switching the load and store to: |// input = vld2q_u8((uint8_t*)data); Doesn't work input = vld1q_u8_x2((uint8_t*)data); // vst2q_u8((uint8_t*)data, input); Doesn't work vst1q_u8_x2((uint8_t*)data, input); | Appears to solve the problem. — Reply to this email directly, view it on GitHub <#605>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAA5HEZXYPLTV3HB3XBK6ELWE2LJ7ANCNFSM6AAAAAARNA2ZWI>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>