Added on the fly sequence apply in float

This commit is contained in:
Xavier Arteaga 2021-04-21 14:06:27 +02:00 committed by Xavier Arteaga
parent aef18f9931
commit 243ad05dfa
2 changed files with 58 additions and 49 deletions

View File

@ -36,6 +36,8 @@ SRSRAN_API void srsran_sequence_state_init(srsran_sequence_state_t* s, uint32_t
SRSRAN_API void srsran_sequence_state_gen_f(srsran_sequence_state_t* s, float value, float* out, uint32_t length);
SRSRAN_API void srsran_sequence_state_apply_f(srsran_sequence_state_t* s, const float* in, float* out, uint32_t length);
SRSRAN_API void srsran_sequence_state_advance(srsran_sequence_state_t* s, uint32_t length);
typedef struct SRSRAN_API {

View File

@ -264,6 +264,59 @@ void srsran_sequence_state_gen_f(srsran_sequence_state_t* s, float value, float*
}
}
void srsran_sequence_state_apply_f(srsran_sequence_state_t* s, const float* in, float* out, uint32_t length)
{
uint32_t i = 0;
const float xor [2] = {+0.0F, -0.0F};
if (length >= SEQUENCE_PAR_BITS) {
for (; i < length - (SEQUENCE_PAR_BITS - 1); i += SEQUENCE_PAR_BITS) {
uint32_t c = (uint32_t)(s->x1 ^ s->x2);
uint32_t j = 0;
#ifdef LV_HAVE_SSE
for (; j < SEQUENCE_PAR_BITS - 3; j += 4) {
// Preloads bits of interest in the 4 LSB
__m128i mask = _mm_set1_epi32(c >> j);
// Masks each bit
mask = _mm_and_si128(mask, _mm_setr_epi32(1, 2, 4, 8));
// Get non zero mask
mask = _mm_cmpgt_epi32(mask, _mm_set1_epi32(0));
// And with MSB
mask = _mm_and_si128(mask, (__m128i)_mm_set1_ps(-0.0F));
// Load input
__m128 v = _mm_load_ps(in + i + j);
// Loads input and perform sign XOR
v = _mm_xor_ps((__m128)mask, v);
_mm_storeu_ps(out + i + j, v);
}
#endif
// Finish the parallel bits with generic code
for (; j < SEQUENCE_PAR_BITS; j++) {
*((uint32_t*)&out[i + j]) = *((uint32_t*)&in[i + j]) ^ *((uint32_t*)&xor[(c >> j) & 1U]);
}
// Step sequences
s->x1 = sequence_gen_LTE_pr_memless_step_par_x1(s->x1);
s->x2 = sequence_gen_LTE_pr_memless_step_par_x2(s->x2);
}
}
for (; i < length; i++) {
*((uint32_t*)&out[i]) = *((uint32_t*)&in[i]) ^ *((uint32_t*)&xor[(s->x1 ^ s->x2) & 1U]);
// Step sequences
s->x1 = sequence_gen_LTE_pr_memless_step_x1(s->x1);
s->x2 = sequence_gen_LTE_pr_memless_step_x2(s->x2);
}
}
void srsran_sequence_state_advance(srsran_sequence_state_t* s, uint32_t length)
{
uint32_t i = 0;
@ -430,56 +483,10 @@ void srsran_sequence_free(srsran_sequence_t* q)
void srsran_sequence_apply_f(const float* in, float* out, uint32_t length, uint32_t seed)
{
uint32_t x1 = sequence_x1_init; // X1 initial state is fix
uint32_t x2 = sequence_get_x2_init(seed); // loads x2 initial state
srsran_sequence_state_t seq = {};
srsran_sequence_state_init(&seq, seed);
uint32_t i = 0;
if (length >= SEQUENCE_PAR_BITS) {
for (; i < length - (SEQUENCE_PAR_BITS - 1); i += SEQUENCE_PAR_BITS) {
uint32_t c = (uint32_t)(x1 ^ x2);
uint32_t j = 0;
#ifdef LV_HAVE_SSE
for (; j < SEQUENCE_PAR_BITS - 3; j += 4) {
// Preloads bits of interest in the 4 LSB
__m128i mask = _mm_set1_epi32(c >> j);
// Masks each bit
mask = _mm_and_si128(mask, _mm_setr_epi32(1, 2, 4, 8));
// Get non zero mask
mask = _mm_cmpgt_epi32(mask, _mm_set1_epi32(0));
// And with MSB
mask = _mm_and_si128(mask, (__m128i)_mm_set1_ps(-0.0F));
// Load input
__m128 v = _mm_loadu_ps(in + i + j);
// Loads input and perform sign XOR
v = _mm_xor_ps((__m128)mask, v);
_mm_storeu_ps(out + i + j, v);
}
#endif
for (; j < SEQUENCE_PAR_BITS; j++) {
((uint32_t*)out)[i + j] = ((uint32_t*)in)[i] ^ (((c >> j) & 1U) << 31U);
}
// Step sequences
x1 = sequence_gen_LTE_pr_memless_step_par_x1(x1);
x2 = sequence_gen_LTE_pr_memless_step_par_x2(x2);
}
}
for (; i < length; i++) {
((uint32_t*)out)[i] = ((uint32_t*)in)[i] ^ (((x1 ^ x2) & 1U) << 31U);
// Step sequences
x1 = sequence_gen_LTE_pr_memless_step_x1(x1);
x2 = sequence_gen_LTE_pr_memless_step_x2(x2);
}
srsran_sequence_state_apply_f(&seq, in, out, length);
}
void srsran_sequence_apply_s(const int16_t* in, int16_t* out, uint32_t length, uint32_t seed)