maki/openairinterface5g
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

aarch64 optimizations for precoding, more complete timing reporting in

Browse Source 
nr_dlsim
This commit is contained in:
Raymond Knopp
2025-10-09 16:01:09 +02:00
committed by Laurent THOMAS
parent e44d498cba
commit 264f7cb817
1 changed files with 121 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
openair1/PHY/MODULATION/nr_modulation.c
+121
View File
@@ -774,6 +774,42 @@ static inline __attribute__((always_inline)) __m256i cmac_prec256(__m256i y, __m
return _mm256_adds_epi16(y, produ);
}
#endif
#ifdef __aarch64__
static inline __attribute__((always_inline)) int16x8_t cmac0_prec128(int16x8_t x, int16x8_t wr, int16x8_t wi) {
//
int16x8_t xr = vuzp1q_s16(x, x); // even lanes
int16x8_t xi = vuzp2q_s16(x, x); // odd lanes
// real = ar*br - ai*bi (Q15 scaling via high-half doubling muls)
int16x8_t real = vqdmulhq_s16(xr, wr); // ≈ round((2*xr*wr)/2^16)
real = vqrdmlshq_s16(real, xi, wi); // real -= round((2*xi*wi)/2^16)
//
// imag = ar*bi + ai*br
int16x8_t imag = vqdmulhq_s16(xr, wi);
imag = vqrdmlahq_s16(imag, xi, wr); // imag += round((2*xi*wr)/2^16)
//
// Re-interleave [real, imag]
int16x8x2_t produ = vzipq_s16(real, imag);
return produ.val[0];
}
static inline __attribute__((always_inline)) int16x8_t cmac_prec128(int16x8_t y, int16x8_t x, int16x8_t wr, int16x8_t wi) {
// De-interleave real/imag
int16x8_t xr = vuzp1q_s16(x, x); // even lanes
int16x8_t xi = vuzp2q_s16(x, x); // odd lanes
//
// real = ar*br - ai*bi (Q15 scaling via high-half doubling muls)
int16x8_t real = vqdmulhq_s16(xr, wr); // ≈ round((2*ar*br)/2^16)
real = vqrdmlshq_s16(real, xi, wi); // real -= round((2*ai*bi)/2^16)
//
// imag = ar*bi + ai*br
int16x8_t imag = vqdmulhq_s16(xr, wi);
imag = vqrdmlahq_s16(imag, xi, wr); // imag += round((2*ai*br)/2^16)
//
// Re-interleave [real, imag]
int16x8x2_t produ = vzipq_s16(real, imag);
return vaddq_s16(y,produ.val[0]);
}
#endif
void nr_layer_precoder_simd(const int n_layers,
const int symSz,
const c16_t txdataF_res_mapped[n_layers][symSz],
@@ -967,6 +1003,91 @@ void nr_layer_precoder_simd(const int n_layers,
#endif
#ifdef __aarch64__
const uint32_t re_cnt_align8 = re_cnt & ~3;
if (n_layers==2) {
prec_weight.r = pmi_pdu->weights[0][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[0][ant].precoder_weight_Im;
const int16x8_t wr0 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi0 = vdupq_n_s16(prec_weight.i);
prec_weight.r = pmi_pdu->weights[1][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[1][ant].precoder_weight_Im;
const int16x8_t wr1 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi1 = vdupq_n_s16(prec_weight.i);
int16x8_t y;
for (; sc < sc_offset + (re_cnt_align8); sc += sizeof(int16x8_t) / sizeof(prec_weight)) {
const int16x8_t x0 = vld1q_s16((const int16_t*)&txdataF_res_mapped[0][sc]);
const int16x8_t x1 = vld1q_s16((const int16_t*)&txdataF_res_mapped[1][sc]);
// Accumulate the product
y = cmac0_prec128(x0,wr0,wi0);
y = cmac_prec128(y,x1,wr1,wi1);
// Store the result to txdataF
*((int16x8_t*)&txdataF_precoded[sc])= y;
}
}
if (n_layers==3) {
prec_weight.r = pmi_pdu->weights[0][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[0][ant].precoder_weight_Im;
const int16x8_t wr0 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi0 = vdupq_n_s16(prec_weight.i);
prec_weight.r = pmi_pdu->weights[1][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[1][ant].precoder_weight_Im;
const int16x8_t wr1 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi1 = vdupq_n_s16(prec_weight.i);
prec_weight.r = pmi_pdu->weights[2][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[2][ant].precoder_weight_Im;
const int16x8_t wr2 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi2 = vdupq_n_s16(prec_weight.i);
int16x8_t y;
for (; sc < sc_offset + (re_cnt_align8); sc += sizeof(int16x8_t) / sizeof(prec_weight)) {
const int16x8_t x0 = vld1q_s16((const int16_t*)&txdataF_res_mapped[0][sc]);
const int16x8_t x1 = vld1q_s16((const int16_t*)&txdataF_res_mapped[1][sc]);
const int16x8_t x2 = vld1q_s16((const int16_t*)&txdataF_res_mapped[2][sc]);
// Accumulate the product
y = cmac0_prec128(x0,wr0,wi0);
y = cmac_prec128(y,x1,wr1,wi1);
y = cmac_prec128(y,x2,wr2,wi2);
// Store the result to txdataF
*((int16x8_t*)&txdataF_precoded[sc])= y;
}
}
if (n_layers==4) {
prec_weight.r = pmi_pdu->weights[0][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[0][ant].precoder_weight_Im;
const int16x8_t wr0 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi0 = vdupq_n_s16(prec_weight.i);
prec_weight.r = pmi_pdu->weights[1][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[1][ant].precoder_weight_Im;
const int16x8_t wr1 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi1 = vdupq_n_s16(prec_weight.i);
prec_weight.r = pmi_pdu->weights[2][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[2][ant].precoder_weight_Im;
const int16x8_t wr2 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi2 = vdupq_n_s16(prec_weight.i);
prec_weight.r = pmi_pdu->weights[3][ant].precoder_weight_Re;
prec_weight.i = pmi_pdu->weights[3][ant].precoder_weight_Im;
const int16x8_t wr3 = vdupq_n_s16(prec_weight.r);
const int16x8_t wi3 = vdupq_n_s16(prec_weight.i);
int16x8_t y;
for (; sc < sc_offset + (re_cnt_align8); sc += sizeof(int16x8_t) / sizeof(prec_weight)) {
const int16x8_t x0 = vld1q_s16((const int16_t*)&txdataF_res_mapped[0][sc]);
const int16x8_t x1 = vld1q_s16((const int16_t*)&txdataF_res_mapped[1][sc]);
const int16x8_t x2 = vld1q_s16((const int16_t*)&txdataF_res_mapped[2][sc]);
const int16x8_t x3 = vld1q_s16((const int16_t*)&txdataF_res_mapped[3][sc]);
// Accumulate the product
y = cmac0_prec128(x0,wr0,wi0);
y = cmac_prec128(y,x1,wr1,wi1);
y = cmac_prec128(y,x2,wr2,wi2);
y = cmac_prec128(y,x3,wr3,wi3);
// Store the result to txdataF
*((int16x8_t*)&txdataF_precoded[sc])= y;
}
}
#else
// Matrix multiplication for 4 elements of the result (sizeof(simde__m128i) / sizeof(c16_t) = 4)
simde__m128i y = simde_mm_set1_epi16(0); // Y = W[0]*X[0] + W[1]*X[1] + ... + W[nrOfLayers-1]*X[nrOfLayers-1]