5g-team
/
srsRAN_4G
mirror of https://github.com/pvnis/srsRAN_4G.git
2
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #75 from softwareradiosystems/next_novolk_update

Browse Source 
Next novolk update
master
Xavier Arteaga 7 years ago committed by GitHub
parent c14393b24f 3292f9c269
commit 02680ed2aa
4 changed files with 394 additions and 36 deletions
Show Stats Download Patch File Download Diff File

320
lib/include/srslte/phy/utils/simd.h
Unescape Escape View File

@ -33,6 +33,11 @@
#endif
#include <immintrin.h>
#endif /* LV_HAVE_SSE */
#include <stdio.h>
#ifdef HAVE_NEON
#include <arm_neon.h>
#endif
/*
* SSE Macros
@ -121,7 +126,17 @@
#define SRSLTE_SIMD_C16_SIZE 8
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
#define SRSLTE_SIMD_F_SIZE 4
#define SRSLTE_SIMD_CF_SIZE 4
#define SRSLTE_SIMD_I_SIZE 4
#define SRSLTE_SIMD_S_SIZE 8
#define SRSLTE_SIMD_C16_SIZE 8
#else /* LV_HAVE_NEON */
#define SRSLTE_SIMD_F_SIZE 0
#define SRSLTE_SIMD_CF_SIZE 0
@ -130,6 +145,7 @@
#define SRSLTE_SIMD_S_SIZE 0
#define SRSLTE_SIMD_C16_SIZE 0
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -147,6 +163,10 @@ typedef __m256 simd_f_t;
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
typedef __m128 simd_f_t;
#else /* HAVE_NEON */
#ifdef HAVE_NEON
typedef float32x4_t simd_f_t;
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -161,6 +181,10 @@ static inline simd_f_t srslte_simd_f_load(float *ptr) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_load_ps(ptr);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vld1q_f32(ptr);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -175,6 +199,10 @@ static inline simd_f_t srslte_simd_f_loadu(float *ptr) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_loadu_ps(ptr);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vld1q_f32(ptr);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -189,6 +217,10 @@ static inline void srslte_simd_f_store(float *ptr, simd_f_t simdreg) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
_mm_store_ps(ptr, simdreg);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
vst1q_f32(ptr, simdreg);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -203,6 +235,10 @@ static inline void srslte_simd_f_storeu(float *ptr, simd_f_t simdreg) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
_mm_storeu_ps(ptr, simdreg);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
vst1q_f32(ptr, simdreg);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -217,6 +253,10 @@ static inline simd_f_t srslte_simd_f_set1(float x) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_set1_ps(x);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vdupq_n_f32(x);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -231,6 +271,10 @@ static inline simd_f_t srslte_simd_f_mul(simd_f_t a, simd_f_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_mul_ps(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vmulq_f32(a,b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -243,8 +287,12 @@ static inline simd_f_t srslte_simd_f_rcp(simd_f_t a) {
#ifdef LV_HAVE_AVX2
return _mm256_rcp_ps(a);
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
#ifdef LV_HAVE_SSE
return _mm_rcp_ps(a);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vmulq_f32(vrecpeq_f32(a), vrecpsq_f32(vrecpeq_f32(a), a));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -260,11 +308,30 @@ static inline simd_f_t srslte_simd_f_addsub(simd_f_t a, simd_f_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_addsub_ps(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON // CURRENTLY USES GENERIC IMPLEMENTATION FOR NEON
float* a_ptr = (float*) &a;
float* b_ptr = (float*) &b;
simd_f_t ret;
float* c_ptr = (float*) &ret;
for(int i = 0; i<4;i++){
if(i%2==0){
c_ptr[i] = a_ptr[i] - b_ptr[i];
}else{
c_ptr[i] = a_ptr[i] + b_ptr[i];
}
}
return ret;
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
}
static inline simd_f_t srslte_simd_f_sub(simd_f_t a, simd_f_t b) {
#ifdef LV_HAVE_AVX512
return _mm512_sub_ps(a, b);
@ -274,6 +341,10 @@ static inline simd_f_t srslte_simd_f_sub(simd_f_t a, simd_f_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_sub_ps(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vsubq_f32(a, b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -288,6 +359,10 @@ static inline simd_f_t srslte_simd_f_add(simd_f_t a, simd_f_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_add_ps(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vaddq_f32(a, b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -302,6 +377,10 @@ static inline simd_f_t srslte_simd_f_zero (void) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_setzero_ps();
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vdupq_n_f32(0);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -316,6 +395,10 @@ static inline simd_f_t srslte_simd_f_swap(simd_f_t a) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_shuffle_ps(a, a, 0b10110001);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vcombine_f32(vrev64_f32(vget_low_f32(a)), vrev64_f32(vget_high_f32(a)));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -344,6 +427,10 @@ static inline simd_f_t srslte_simd_f_hadd(simd_f_t a, simd_f_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_hadd_ps(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vcombine_f32( vpadd_f32( vget_low_f32(a), vget_high_f32(a) ), vpadd_f32( vget_low_f32(b), vget_high_f32(b) ) );
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -358,6 +445,12 @@ static inline simd_f_t srslte_simd_f_sqrt(simd_f_t a) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_sqrt_ps(a);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
float32x4_t sqrt_reciprocal = vrsqrteq_f32(a);
sqrt_reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(a,sqrt_reciprocal), sqrt_reciprocal),sqrt_reciprocal);
return vmulq_f32(a,sqrt_reciprocal);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -368,10 +461,15 @@ static inline simd_f_t srslte_simd_f_sqrt(simd_f_t a) {
#if SRSLTE_SIMD_CF_SIZE
#ifdef HAVE_NEON
typedef float32x4x2_t simd_cf_t;
#else
typedef struct {
simd_f_t re;
simd_f_t im;
} simd_cf_t;
#endif
/* Complex Single precission Floating point functions */
static inline simd_cf_t srslte_simd_cfi_load(cf_t *ptr) {
@ -399,6 +497,10 @@ static inline simd_cf_t srslte_simd_cfi_load(cf_t *ptr) {
__m128 i2 = _mm_load_ps((float*)(ptr + 2));
ret.re = _mm_shuffle_ps(i1, i2, _MM_SHUFFLE(2,0,2,0));
ret.im = _mm_shuffle_ps(i1, i2, _MM_SHUFFLE(3,1,3,1));
#else
#ifdef HAVE_NEON
ret = vld2q_f32((float*)(ptr));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -431,6 +533,10 @@ static inline simd_cf_t srslte_simd_cfi_loadu(cf_t *ptr) {
__m128 i2 = _mm_loadu_ps((float*)(ptr + 2));
ret.re = _mm_shuffle_ps(i1, i2, _MM_SHUFFLE(2,0,2,0));
ret.im = _mm_shuffle_ps(i1, i2, _MM_SHUFFLE(3,1,3,1));
#else
#ifdef HAVE_NEON
ret = vld2q_f32((float*)(ptr));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -450,6 +556,11 @@ static inline simd_cf_t srslte_simd_cf_load(float *re, float *im) {
#ifdef LV_HAVE_SSE
ret.re = _mm_load_ps(re);
ret.im = _mm_load_ps(im);
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
ret.val[0] = vld1q_f32(re);
ret.val[1] = vld1q_f32(im);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -469,6 +580,11 @@ static inline simd_cf_t srslte_simd_cf_loadu(float *re, float *im) {
#ifdef LV_HAVE_SSE
ret.re = _mm_loadu_ps(re);
ret.im = _mm_loadu_ps(im);
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
ret.val[0] = vld1q_f32(re);
ret.val[1] = vld1q_f32(im);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -497,6 +613,10 @@ static inline void srslte_simd_cfi_store(cf_t *ptr, simd_cf_t simdreg) {
#ifdef LV_HAVE_SSE
_mm_store_ps((float*)(ptr), _mm_unpacklo_ps(simdreg.re, simdreg.im));
_mm_store_ps((float*)(ptr + 2), _mm_unpackhi_ps(simdreg.re, simdreg.im));
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
vst2q_f32((float*)(ptr), simdreg);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -524,6 +644,10 @@ static inline void srslte_simd_cfi_storeu(cf_t *ptr, simd_cf_t simdreg) {
#ifdef LV_HAVE_SSE
_mm_storeu_ps((float*)(ptr), _mm_unpacklo_ps(simdreg.re, simdreg.im));
_mm_storeu_ps((float*)(ptr + 2), _mm_unpackhi_ps(simdreg.re, simdreg.im));
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
vst2q_f32((float*)(ptr), simdreg);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -541,6 +665,11 @@ static inline void srslte_simd_cf_store(float *re, float *im, simd_cf_t simdreg)
#ifdef LV_HAVE_SSE
_mm_store_ps((float *) re, simdreg.re);
_mm_store_ps((float *) im, simdreg.im);
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
vst1q_f32((float *) re, simdreg.val[0]);
vst1q_f32((float *) im, simdreg.val[1]);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -558,6 +687,11 @@ static inline void srslte_simd_cf_storeu(float *re, float *im, simd_cf_t simdreg
#ifdef LV_HAVE_SSE
_mm_storeu_ps((float *) re, simdreg.re);
_mm_storeu_ps((float *) im, simdreg.im);
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
vst1q_f32((float *) re, simdreg.val[0]);
vst1q_f32((float *) im, simdreg.val[1]);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -576,6 +710,11 @@ static inline simd_cf_t srslte_simd_cf_set1 (cf_t x) {
#ifdef LV_HAVE_SSE
ret.re = _mm_set1_ps(__real__ x);
ret.im = _mm_set1_ps(__imag__ x);
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
ret.val[0] = vdupq_n_f32(__real__ x);
ret.val[1] = vdupq_n_f32(__imag__ x);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -601,6 +740,13 @@ static inline simd_cf_t srslte_simd_cf_prod (simd_cf_t a, simd_cf_t b) {
_mm_mul_ps(a.im, b.im));
ret.im = _mm_add_ps(_mm_mul_ps(a.re, b.im),
_mm_mul_ps(a.im, b.re));
#else
#ifdef HAVE_NEON
ret.val[0] = vsubq_f32(vmulq_f32(a.val[0],b.val[0]),
vmulq_f32(a.val[1],b.val[1]));
ret.val[1] = vaddq_f32(vmulq_f32(a.val[0],b.val[1]),
vmulq_f32(a.val[1],b.val[0]));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -626,6 +772,13 @@ static inline simd_cf_t srslte_simd_cf_conjprod (simd_cf_t a, simd_cf_t b) {
_mm_mul_ps(a.im, b.im));
ret.im = _mm_sub_ps(_mm_mul_ps(a.im, b.re),
_mm_mul_ps(a.re, b.im));
#else
#ifdef HAVE_NEON
ret.val[0] = vaddq_f32(vmulq_f32(a.val[0],b.val[0]),
vmulq_f32(a.val[1],b.val[1]));
ret.val[1] = vsubq_f32(vmulq_f32(a.val[1],b.val[0]),
vmulq_f32(a.val[0],b.val[1]));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -645,6 +798,11 @@ static inline simd_cf_t srslte_simd_cf_add (simd_cf_t a, simd_cf_t b) {
#ifdef LV_HAVE_SSE
ret.re = _mm_add_ps(a.re, b.re);
ret.im = _mm_add_ps(a.im, b.im);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
ret.val[0] = vaddq_f32(a.val[0],b.val[0]);
ret.val[1] = vaddq_f32(a.val[1],b.val[1]);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -665,6 +823,11 @@ static inline simd_cf_t srslte_simd_cf_mul (simd_cf_t a, simd_f_t b) {
#ifdef LV_HAVE_SSE
ret.re = _mm_mul_ps(a.re, b);
ret.im = _mm_mul_ps(a.im, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
ret.val[0] = vmulq_f32(a.val[0],b);
ret.val[1] = vmulq_f32(a.val[1],b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -699,6 +862,16 @@ static inline simd_cf_t srslte_simd_cf_rcp (simd_cf_t a) {
simd_f_t neg_a_im = _mm_xor_ps(_mm_set1_ps(-0.0f), a.im);
ret.re = _mm_mul_ps(a.re, rcp);
ret.im = _mm_mul_ps(neg_a_im, rcp);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
simd_f_t a2re = vmulq_f32(a.val[0], a.val[0]);
simd_f_t a2im = vmulq_f32(a.val[1], a.val[1]);
simd_f_t mod2 = vaddq_f32(a2re, a2im);
simd_f_t rcp = vmulq_f32(vrecpeq_f32(mod2), vrecpsq_f32(vrecpeq_f32(mod2), mod2));
simd_f_t neg_a_im = vnegq_f32(a.val[1]);
ret.val[0] = vmulq_f32(a.val[0], rcp);
ret.val[1] = vmulq_f32(neg_a_im, rcp);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -718,7 +891,12 @@ static inline simd_cf_t srslte_simd_cf_zero (void) {
#ifdef LV_HAVE_SSE
ret.re = _mm_setzero_ps();
ret.im = _mm_setzero_ps();
#endif /* LV_HAVE_SSE */
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
ret.val[0] = vdupq_n_f32(0);
ret.val[1] = vdupq_n_f32(0);
#endif /* HAVE_NEON */
#endif /* HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
return ret;
@ -739,6 +917,11 @@ typedef __m256 simd_sel_t;
#ifdef LV_HAVE_SSE
typedef __m128i simd_i_t;
typedef __m128 simd_sel_t;
#else /* LV_HAVE_AVX2 */
#ifdef HAVE_NEON
typedef int32x4_t simd_i_t;
typedef int32x4_t simd_sel_t;
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -752,6 +935,10 @@ static inline simd_i_t srslte_simd_i_load(int *x) {
#else
#ifdef LV_HAVE_SSE
return _mm_load_si128((__m128i*)x);
#else
#ifdef HAVE_NEON
return vld1q_s32((int*)x);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -766,6 +953,10 @@ static inline void srslte_simd_i_store(int *x, simd_i_t reg) {
#else
#ifdef LV_HAVE_SSE
_mm_store_si128((__m128i*)x, reg);
#else
#ifdef HAVE_NEON
vst1q_s32((int*)x, reg);
#endif /*HAVE_NEON*/
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -780,6 +971,10 @@ static inline simd_i_t srslte_simd_i_set1(int x) {
#else
#ifdef LV_HAVE_SSE
return _mm_set1_epi32(x);
#else
#ifdef HAVE_NEON
return vdupq_n_s32(x);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -794,6 +989,10 @@ static inline simd_i_t srslte_simd_i_add(simd_i_t a, simd_i_t b) {
#else
#ifdef LV_HAVE_SSE
return _mm_add_epi32(a, b);
#else
#ifdef HAVE_NEON
return vaddq_s32(a, b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -808,6 +1007,10 @@ static inline simd_sel_t srslte_simd_f_max(simd_f_t a, simd_f_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return (simd_sel_t) _mm_cmpgt_ps(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return (simd_sel_t) vcgtq_f32(a, b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -822,6 +1025,23 @@ static inline simd_i_t srslte_simd_i_select(simd_i_t a, simd_i_t b, simd_sel_t s
#else
#ifdef LV_HAVE_SSE
return (__m128i) _mm_blendv_ps((__m128)a, (__m128)b, selector);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON // CURRENTLY USES GENERIC IMPLEMENTATION FOR NEON
int* a_ptr = (int*) &a;
int* b_ptr = (int*) &b;
simd_i_t ret;
int* sel = (int*) &selector;
int* c_ptr = (int*) &ret;
for(int i = 0;i<4;i++)
{
if(sel[i] == -1){
c_ptr[i] = b_ptr[i];
}else{
c_ptr[i] = a_ptr[i];
}
}
return ret;
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -841,6 +1061,10 @@ typedef __m256i simd_s_t;
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
typedef __m128i simd_s_t;
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
typedef int16x8_t simd_s_t;
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -854,6 +1078,10 @@ static inline simd_s_t srslte_simd_s_load(int16_t *ptr) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_load_si128((__m128i*) ptr);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vld1q_s16(ptr);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -868,6 +1096,10 @@ static inline simd_s_t srslte_simd_s_loadu(int16_t *ptr) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_loadu_si128((__m128i*) ptr);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vld1q_s16(ptr);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -882,6 +1114,10 @@ static inline void srslte_simd_s_store(int16_t *ptr, simd_s_t simdreg) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
_mm_store_si128((__m128i*) ptr, simdreg);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
vst1q_s16( ptr, simdreg);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -896,11 +1132,14 @@ static inline void srslte_simd_s_storeu(int16_t *ptr, simd_s_t simdreg) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
_mm_storeu_si128((__m128i*) ptr, simdreg);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
vst1q_s16(ptr, simdreg);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
}
static inline simd_s_t srslte_simd_s_zero(void) {
#ifdef LV_HAVE_AVX512
return _mm512_setzero_si512();
@ -910,6 +1149,10 @@ static inline simd_s_t srslte_simd_s_zero(void) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_setzero_si128();
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vdupq_n_s16(0);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -924,6 +1167,10 @@ static inline simd_s_t srslte_simd_s_mul(simd_s_t a, simd_s_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_mullo_epi16(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vmulq_s16(a, b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -938,6 +1185,10 @@ static inline simd_s_t srslte_simd_s_add(simd_s_t a, simd_s_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_add_epi16(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vaddq_s16(a, b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -952,6 +1203,10 @@ static inline simd_s_t srslte_simd_s_sub(simd_s_t a, simd_s_t b) {
#else /* LV_HAVE_AVX2 */
#ifdef LV_HAVE_SSE
return _mm_sub_epi16(a, b);
#else /* LV_HAVE_SSE */
#ifdef HAVE_NEON
return vsubq_s16(a, b);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -962,8 +1217,9 @@ static inline simd_s_t srslte_simd_s_sub(simd_s_t a, simd_s_t b) {
#if SRSLTE_SIMD_C16_SIZE
typedef struct {
typedef
#ifdef LV_HAVE_AVX512
struct {
union {
__m512i m512;
int16_t i16[32];
@ -974,6 +1230,7 @@ typedef struct {
} im;
#else /* LV_HAVE_AVX512 */
#ifdef LV_HAVE_AVX2
struct {
union {
__m256i m256;
int16_t i16[16];
@ -984,6 +1241,7 @@ typedef struct {
} im;
#else
#ifdef LV_HAVE_SSE
struct {
union {
__m128i m128;
int16_t i16[8];
@ -992,6 +1250,12 @@ typedef struct {
__m128i m128;
int16_t i16[8];
} im;
#else
#ifdef HAVE_NEON
union {
int16x8x2_t m128;
int16_t i16[16];
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -1017,6 +1281,10 @@ static inline simd_c16_t srslte_simd_c16i_load(c16_t *ptr) {
__m128i in2 = _mm_load_si128((__m128i*)(ptr + 8));
ret.re.m128 = _mm_blend_epi16(in1,_mm_shufflelo_epi16(_mm_shufflehi_epi16(in2, 0b10100000), 0b10100000), 0b10101010);
ret.im.m128 = _mm_blend_epi16(_mm_shufflelo_epi16(_mm_shufflehi_epi16(in1, 0b11110101), 0b11110101),in2, 0b10101010);
#else /* LV_HAVE_SSE*/
#ifdef HAVE_NEON
ret.m128 = vld2q_s16((int16_t*)(ptr));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */
@ -1032,6 +1300,11 @@ static inline simd_c16_t srslte_simd_c16_load(int16_t *re, int16_t *im) {
#ifdef LV_HAVE_SSE
ret.re.m128 = _mm_load_si128((__m128i*)(re));
ret.im.m128 = _mm_load_si128((__m128i*)(im));
#else /* LV_HAVE_SSE*/
#ifdef HAVE_NEON
ret.m128.val[0] = vld1q_s16((int16_t*)(re));
ret.m128.val[1] = vld1q_s16((int16_t*)(im));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
return ret;
@ -1046,6 +1319,11 @@ static inline simd_c16_t srslte_simd_c16_loadu(int16_t *re, int16_t *im) {
#ifdef LV_HAVE_SSE
ret.re.m128 = _mm_loadu_si128((__m128i*)(re));
ret.im.m128 = _mm_loadu_si128((__m128i*)(im));
#else /* LV_HAVE_SSE*/
#ifdef HAVE_NEON
ret.m128.val[0] = vld1q_s16((int16_t*)(re));
ret.m128.val[1] = vld1q_s16((int16_t*)(im));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
return ret;
@ -1063,6 +1341,10 @@ static inline void srslte_simd_c16i_store(c16_t *ptr, simd_c16_t simdreg) {
__m128i im_sw = _mm_shufflelo_epi16(_mm_shufflehi_epi16(simdreg.im.m128, 0b10110001), 0b10110001);
_mm_store_si128((__m128i *) (ptr), _mm_blend_epi16(simdreg.re.m128, im_sw, 0b10101010));
_mm_store_si128((__m128i *) (ptr + 8), _mm_blend_epi16(re_sw, simdreg.im.m128, 0b10101010));
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
vst2q_s16((int16_t*)(ptr) ,simdreg.m128);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
}
@ -1079,6 +1361,10 @@ static inline void srslte_simd_c16i_storeu(c16_t *ptr, simd_c16_t simdreg) {
__m128i im_sw = _mm_shufflelo_epi16(_mm_shufflehi_epi16(simdreg.im.m128, 0b10110001), 0b10110001);
_mm_storeu_si128((__m128i *) (ptr), _mm_blend_epi16(simdreg.re.m128, im_sw, 0b10101010));
_mm_storeu_si128((__m128i *) (ptr + 8), _mm_blend_epi16(re_sw, simdreg.im.m128, 0b10101010));
#else /*HAVE_NEON*/
#ifdef HAVE_NEON
vst2q_s16((int16_t*)(ptr) ,simdreg.m128);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
}
@ -1091,6 +1377,11 @@ static inline void srslte_simd_c16_store(int16_t *re, int16_t *im, simd_c16_t si
#ifdef LV_HAVE_SSE
_mm_store_si128((__m128i *) re, simdreg.re.m128);
_mm_store_si128((__m128i *) im, simdreg.im.m128);
#else
#ifdef HAVE_NEON
vst1q_s16((int16_t *) re, simdreg.m128.val[0]);
vst1q_s16((int16_t *) im, simdreg.m128.val[1]);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
}
@ -1103,6 +1394,11 @@ static inline void srslte_simd_c16_storeu(int16_t *re, int16_t *im, simd_c16_t s
#ifdef LV_HAVE_SSE
_mm_storeu_si128((__m128i *) re, simdreg.re.m128);
_mm_storeu_si128((__m128i *) im, simdreg.im.m128);
#else
#ifdef HAVE_NEON
vst1q_s16((int16_t *) re, simdreg.m128.val[0]);
vst1q_s16((int16_t *) im, simdreg.m128.val[1]);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
}
@ -1134,6 +1430,11 @@ static inline simd_c16_t srslte_simd_c16_add (simd_c16_t a, simd_c16_t b) {
#ifdef LV_HAVE_SSE
ret.re.m128 = _mm_add_epi16(a.re.m128, b.re.m128);
ret.im.m128 = _mm_add_epi16(a.im.m128, b.im.m128);
#else
#ifdef HAVE_NEON
ret.m128.val[0] = vaddq_s16(a.m128.val[0],a.m128.val[0]);
ret.m128.val[1] = vaddq_s16(a.m128.val[1],a.m128.val[1]);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
return ret;
@ -1148,6 +1449,11 @@ static inline simd_c16_t srslte_simd_c16_zero (void) {
#ifdef LV_HAVE_SSE
ret.re.m128 = _mm_setzero_si128();
ret.im.m128 = _mm_setzero_si128();
#else
#ifdef HAVE_NEON
ret.m128.val[0] = vdupq_n_s16(0);
ret.m128.val[1] = vdupq_n_s16(0);
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
return ret;
@ -1182,6 +1488,12 @@ static inline simd_s_t srslte_simd_convert_2f_s(simd_f_t a, simd_f_t b) {
__m128i ai = _mm_cvttps_epi32(a);
__m128i bi = _mm_cvttps_epi32(b);
return _mm_packs_epi32(ai, bi);
#else
#ifdef HAVE_NEON
int32x4_t ai = vcvtq_s32_f32(a);
int32x4_t bi = vcvtq_s32_f32(b);
return (simd_s_t)vcombine_s16(vqmovn_s32(ai), vqmovn_s32(bi));
#endif /* HAVE_NEON */
#endif /* LV_HAVE_SSE */
#endif /* LV_HAVE_AVX2 */
#endif /* LV_HAVE_AVX512 */

3
lib/src/phy/ue/ue_dl.c
Unescape Escape View File

@ -603,7 +603,8 @@ int srslte_ue_dl_decode_mbsfn(srslte_ue_dl_t * q,
grant.sf_type = SRSLTE_SF_MBSFN;
grant.nof_tb = 1;
grant.mcs[0].idx = 2;
grant.tb_en[0] = true;
grant.tb_en[1] = false;
grant.nof_prb = q->pmch.cell.nof_prb;
srslte_dl_fill_ra_mcs(&grant.mcs[0], grant.nof_prb);
srslte_softbuffer_rx_reset_tbs(q->softbuffers[0], (uint32_t) grant.mcs[0].tbs);

30
lib/src/phy/utils/test/vector_test.c
Unescape Escape View File

@ -63,12 +63,12 @@ bool verbose = false;
strncpy(func_name, #X, 32);\
CODE;\
passed = (mse < MAX_MSE);\
printf("%32s (%5d) ... %7.1f MSamp/s ... %3s Passed\n", func_name, block_size, \
(double) block_size*NOF_REPETITIONS/ *timing, passed?"":"Not");\
printf("%32s (%5d) ... %7.1f MSamp/s ... %3s Passed (%.6f)\n", func_name, block_size, \
(double) block_size*NOF_REPETITIONS/ *timing, passed?"":"Not", mse);\
return passed;\
}
#define MALLOC(TYPE, NAME) TYPE *NAME = malloc(sizeof(TYPE)*block_size)
#define MALLOC(TYPE, NAME) TYPE *NAME = srslte_vec_malloc(sizeof(TYPE)*block_size)
static double elapsed_us(struct timeval *ts_start, struct timeval *ts_end) {
@ -507,7 +507,7 @@ TEST(srslte_vec_abs_cf,
for (int i = 0; i < block_size; i++) {
gold = sqrtf(crealf(x[i]) * crealf(x[i]) + cimagf(x[i])*cimagf(x[i]));
mse += cabsf(gold - z[i]);
mse += cabsf(gold - z[i])/block_size;
}
free(x);
@ -771,12 +771,27 @@ int main(int argc, char **argv) {
size_count++;
}
char fname[68];
FILE *f = NULL;
void * p = popen("(date +%g%m%d && hostname) | tr '\\r\\n' '__'", "r");
if (p) {
fgets(fname, 64, p);
strncpy(fname + strnlen(fname, 64) - 1, ".tsv", 4);
f = fopen(fname, "w");
if (f) printf("Saving benchmark results in '%s'\n", fname);
}
pclose(p);
printf("\n");
printf("%32s |", "Subroutine/MSps");
if (f) fprintf(f, "Subroutine/MSps Vs Vector size\t");
for (int i = 0; i < size_count; i++) {
printf(" %7d", sizes[i]);
if (f) fprintf(f, "%d\t", sizes[i]);
}
printf(" |\n");
if (f) fprintf(f, "\n");
for (int j = 0; j < 32; j++) {
printf("-");
@ -789,12 +804,19 @@ int main(int argc, char **argv) {
for (int i = 0; i < func_count; i++) {
printf("%32s | ", func_names[i]);
if (f) fprintf(f, "%s\t", func_names[i]);
for (int j = 0; j < size_count; j++) {
printf(" %s%7.1f\x1b[0m", (passed[i][j])?"":"\x1B[31m", (double) NOF_REPETITIONS*(double)sizes[j]/timmings[i][j]);
if (f) fprintf(f, "%.1f\t", (double) NOF_REPETITIONS*(double)sizes[j]/timmings[i][j]);
all_passed &= passed[i][j];
}
printf(" |\n");
if (f) fprintf(f, "\n");
}
if (f) fclose(f);
return (all_passed)?SRSLTE_SUCCESS:SRSLTE_ERROR;
}

37
lib/src/phy/utils/vector_simd.c
Unescape Escape View File

@ -751,10 +751,37 @@ void srslte_vec_div_fff_simd(float *x, float *y, float *z, int len) {
}
}
int srslte_vec_sc_prod_ccc_simd2(cf_t *x, cf_t h, cf_t *z, int len)
{
int i = 0;
const unsigned int loops = len / 4;
#ifdef HAVE_NEON
simd_cf_t h_vec;
h_vec.val[0] = srslte_simd_f_set1(__real__ h);
h_vec.val[1] = srslte_simd_f_set1(__imag__ h);
for (; i < loops; i++) {
simd_cf_t in = srslte_simd_cfi_load(&x[i*4]);
simd_cf_t temp = srslte_simd_cf_prod(in, h_vec);
srslte_simd_cfi_store(&z[i*4], temp);
}
#endif
i = loops * 4;
return i;
}
void srslte_vec_sc_prod_ccc_simd(cf_t *x, cf_t h, cf_t *z, int len) {
int i = 0;
#if SRSLTE_SIMD_F_SIZE
#ifdef HAVE_NEON
i = srslte_vec_sc_prod_ccc_simd2(x, h, z, len);
#else
const simd_f_t hre = srslte_simd_f_set1(__real__ h);
const simd_f_t him = srslte_simd_f_set1(__imag__ h);
@ -766,8 +793,8 @@ void srslte_vec_sc_prod_ccc_simd(cf_t *x, cf_t h, cf_t *z, int len) {
simd_f_t sw = srslte_simd_f_swap(temp);
simd_f_t m2 = srslte_simd_f_mul(him, sw);
simd_f_t r = srslte_simd_f_addsub(m1, m2);
srslte_simd_f_store((float *) &z[i], r);
}
} else {
for (; i < len - SRSLTE_SIMD_F_SIZE / 2 + 1; i += SRSLTE_SIMD_F_SIZE / 2) {
@ -782,10 +809,11 @@ void srslte_vec_sc_prod_ccc_simd(cf_t *x, cf_t h, cf_t *z, int len) {
}
}
#endif
#endif
for (; i < len; i++) {
z[i] = x[i] * h;
}
}
void srslte_vec_sc_prod_fff_simd(float *x, float h, float *z, int len) {
@ -831,7 +859,6 @@ void srslte_vec_abs_cf_simd(cf_t *x, float *z, int len) {
simd_f_t z1 = srslte_simd_f_hadd(mul1, mul2);
z1 = srslte_simd_f_sqrt(z1);
srslte_simd_f_store(&z[i], z1);
}
} else {
@ -966,9 +993,7 @@ uint32_t srslte_vec_max_fi_simd(float *x, int len) {
if (SRSLTE_IS_ALIGNED(x)) {
for (; i < len - SRSLTE_SIMD_I_SIZE + 1; i += SRSLTE_SIMD_I_SIZE) {
simd_f_t a = srslte_simd_f_load(&x[i]);
simd_sel_t res = srslte_simd_f_max(a, simd_max_values);
simd_max_indexes = srslte_simd_i_select(simd_max_indexes, simd_indexes, res);
simd_max_values = (simd_f_t) srslte_simd_i_select((simd_i_t) simd_max_values, (simd_i_t) a, res);
simd_indexes = srslte_simd_i_add(simd_indexes, simd_inc);
@ -976,9 +1001,7 @@ uint32_t srslte_vec_max_fi_simd(float *x, int len) {
} else {
for (; i < len - SRSLTE_SIMD_I_SIZE + 1; i += SRSLTE_SIMD_I_SIZE) {
simd_f_t a = srslte_simd_f_loadu(&x[i]);
simd_sel_t res = srslte_simd_f_max(a, simd_max_values);
simd_max_indexes = srslte_simd_i_select(simd_max_indexes, simd_indexes, res);
simd_max_values = (simd_f_t) srslte_simd_i_select((simd_i_t) simd_max_values, (simd_i_t) a, res);
simd_indexes = srslte_simd_i_add(simd_indexes, simd_inc);

Write Preview
Loading…
Cancel
Save