dowerx/pico-radio
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Compare commits

base: dowerx:09ca162abb2f370bf4deceea0cf84b3383fe0858
Branches Tags
dowerx:master
dowerx:recorder-simd
..
compare: dowerx:125cd1e8181a03bf07e07554560fb29d4bb1114a
Branches Tags
dowerx:recorder-simd
dowerx:master

No commits in common. "09ca162abb2f370bf4deceea0cf84b3383fe0858" and "125cd1e8181a03bf07e07554560fb29d4bb1114a" have entirely different histories.
09ca162abb ... 125cd1e818

1 changed files with 14 additions and 3 deletions
Show Stats Expand all files Collapse all files

17
src/recorder/main.c
View File

@ -1,4 +1,3 @@
#include <sched.h>
#include <signal.h> #include <signal.h>
#include <stdbool.h> #include <stdbool.h>
#include <stdint.h> #include <stdint.h>
@ -49,7 +48,7 @@ __always_inline void to_float_iq(const int16_t* i_c, const int16_t* q_c, float*
q_float_hi = _mm256_mul_ps(q_float_hi, scale); q_float_hi = _mm256_mul_ps(q_float_hi, scale);
// interleave and store // interleave and store
__m256 interleaved_lo0 = _mm256_unpacklo_ps(i_float_lo, q_float_lo); __m256 interleaved_lo0 = _mm256_unpacklo_ps(i_float_lo, q_float_hi);
__m256 interleaved_lo1 = _mm256_unpackhi_ps(i_float_lo, q_float_lo); __m256 interleaved_lo1 = _mm256_unpackhi_ps(i_float_lo, q_float_lo);
__m256 interleaved_hi0 = _mm256_unpacklo_ps(i_float_hi, q_float_hi); __m256 interleaved_hi0 = _mm256_unpacklo_ps(i_float_hi, q_float_hi);
@ -59,6 +58,19 @@ __always_inline void to_float_iq(const int16_t* i_c, const int16_t* q_c, float*
_mm256_storeu_ps(&output[i * 2 + 8], _mm256_permute2f128_ps(interleaved_lo0, interleaved_lo1, 0x31)); _mm256_storeu_ps(&output[i * 2 + 8], _mm256_permute2f128_ps(interleaved_lo0, interleaved_lo1, 0x31));
_mm256_storeu_ps(&output[i * 2 + 16], _mm256_permute2f128_ps(interleaved_hi0, interleaved_hi1, 0x20)); _mm256_storeu_ps(&output[i * 2 + 16], _mm256_permute2f128_ps(interleaved_hi0, interleaved_hi1, 0x20));
_mm256_storeu_ps(&output[i * 2 + 24], _mm256_permute2f128_ps(interleaved_hi0, interleaved_hi1, 0x31)); _mm256_storeu_ps(&output[i * 2 + 24], _mm256_permute2f128_ps(interleaved_hi0, interleaved_hi1, 0x31));
// float i_output[16];
// float q_output[16];
// _mm256_storeu_ps(&i_output[i], i_float_lo);
// _mm256_storeu_ps(&i_output[i + 8], i_float_hi);
// _mm256_storeu_ps(&q_output[i], q_float_lo);
// _mm256_storeu_ps(&q_output[i + 8], q_float_hi);
// for (int j = 0; j < 16; j++) {
// output[(i + j) * 2 + 0] = i_output[j];
// output[(i + j) * 2 + 1] = q_output[j];
// }
} }
for (; i < len; ++i) { for (; i < len; ++i) {
@ -194,7 +206,6 @@ int main(int argc, char** argv) {
while (!stop) { while (!stop) {
ps2000_get_streaming_last_values(unit, &get_overview_buffers); ps2000_get_streaming_last_values(unit, &get_overview_buffers);
sched_yield();
} }
cleanup(unit); cleanup(unit);