From f4b9e7311a1daa4b306c9e8bebd3c956db3e6cad Mon Sep 17 00:00:00 2001 From: yagoda Date: Tue, 17 Oct 2017 15:51:27 +0000 Subject: [PATCH] adding neon support to new vector structure --- lib/include/srslte/phy/utils/simd.h | 123 +++++++++++++++++++--------- lib/src/phy/utils/vector_simd.c | 37 +++++++-- 2 files changed, 116 insertions(+), 44 deletions(-) diff --git a/lib/include/srslte/phy/utils/simd.h b/lib/include/srslte/phy/utils/simd.h index 6e4185788..e7820c307 100644 --- a/lib/include/srslte/phy/utils/simd.h +++ b/lib/include/srslte/phy/utils/simd.h @@ -33,6 +33,11 @@ #endif #include #endif /* LV_HAVE_SSE */ +#include + +#ifdef HAVE_NEON +#include +#endif /* * SSE Macros @@ -140,6 +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 */ @@ -159,7 +165,7 @@ typedef __m256 simd_f_t; typedef __m128 simd_f_t; #else /* HAVE_NEON */ #ifdef HAVE_NEON -typedef float32x4 simd_f_t; +typedef float32x4_t simd_f_t; #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -188,7 +194,7 @@ static inline simd_f_t srslte_simd_f_loadu(float *ptr) { #ifdef LV_HAVE_AVX512 return _mm512_loadu_ps(ptr); #else /* LV_HAVE_AVX512 */ - #ifdef LV_HAVE_AVX2 + #ifdef LV_HAVE_AVX2 return _mm256_loadu_ps(ptr); #else /* LV_HAVE_AVX2 */ #ifdef LV_HAVE_SSE @@ -202,7 +208,7 @@ static inline simd_f_t srslte_simd_f_loadu(float *ptr) { #endif /* LV_HAVE_AVX512 */ } -static inline void srslte_simd_f_store(float *ptr, simd_f_t simdreg) {vst1q_f32 +static inline void srslte_simd_f_store(float *ptr, simd_f_t simdreg) { #ifdef LV_HAVE_AVX512 _mm512_store_ps(ptr, simdreg); #else /* LV_HAVE_AVX512 */ @@ -281,11 +287,11 @@ 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 vrecpeq_f32(a); + return vmulq_f32(vrecpeq_f32(a), vrecpsq_f32(vrecpeq_f32(a), a)); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -302,6 +308,22 @@ 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 = &a; + float* b_ptr = &b; + simd_f_t ret; + float* c_ptr = &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 */ @@ -373,6 +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 */ @@ -421,7 +447,9 @@ static inline simd_f_t srslte_simd_f_sqrt(simd_f_t a) { return _mm_sqrt_ps(a); #else /* LV_HAVE_SSE */ #ifdef HAVE_NEON - return vrecpeq_f32(vrsqrteq_f32(a)); + 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 */ @@ -530,8 +558,8 @@ static inline simd_cf_t srslte_simd_cf_load(float *re, float *im) { ret.im = _mm_load_ps(im); #else /*HAVE_NEON*/ #ifdef HAVE_NEON - ret.val[0] = vld1q_f32(ptr); - ret.val[1] = vld1q_f32(ptr); + ret.val[0] = vld1q_f32(re); + ret.val[1] = vld1q_f32(im); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -554,8 +582,8 @@ static inline simd_cf_t srslte_simd_cf_loadu(float *re, float *im) { ret.im = _mm_loadu_ps(im); #else /*HAVE_NEON*/ #ifdef HAVE_NEON - ret.val[0] = vld1q_f32(ptr); - ret.val[1] = vld1q_f32(ptr); + ret.val[0] = vld1q_f32(re); + ret.val[1] = vld1q_f32(im); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -684,8 +712,8 @@ static inline simd_cf_t srslte_simd_cf_set1 (cf_t x) { ret.im = _mm_set1_ps(__imag__ x); #else /*HAVE_NEON*/ #ifdef HAVE_NEON - re.val[0] = vdupq_n_f32(__real__ x); - im.val[1] = vdupq_n_f32(__imag__ x); + 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 */ @@ -772,8 +800,8 @@ static inline simd_cf_t srslte_simd_cf_add (simd_cf_t a, simd_cf_t b) { ret.im = _mm_add_ps(a.im, b.im); #else /* LV_HAVE_SSE */ #ifdef HAVE_NEON - ret.val[0] = vaddq_f32(a.val[0],a.val[0]); - ret.val[1] = vaddq_f32(a.val[1],a.val[1]); + 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 */ @@ -839,8 +867,8 @@ static inline simd_cf_t srslte_simd_cf_rcp (simd_cf_t a) { 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 = vrecpeq_f32(mod2); - simd_f_t neg_a_im = vnegq_f32(vdupq_n_f32(-0.0f), a.val[1]); + 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 */ @@ -868,6 +896,7 @@ static inline simd_cf_t srslte_simd_cf_zero (void) { 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; @@ -889,9 +918,9 @@ typedef __m256 simd_sel_t; typedef __m128i simd_i_t; typedef __m128 simd_sel_t; #else /* LV_HAVE_AVX2 */ -#ifdef LV_HAVE_SSE +#ifdef HAVE_NEON typedef int32x4_t simd_i_t; -typedef __m128 simd_sel_t; +typedef int32x4_t simd_sel_t; #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -908,7 +937,7 @@ static inline simd_i_t srslte_simd_i_load(int *x) { return _mm_load_si128((__m128i*)x); #else #ifdef HAVE_NEON - return vld1_s32((int32x4_t*)x); + return vld1q_s32((int*)x); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -924,6 +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 */ @@ -963,7 +996,7 @@ static inline simd_i_t srslte_simd_i_add(simd_i_t a, simd_i_t b) { #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ #endif /* LV_HAVE_AVX512 */ -}vcgtq_f32 +} static inline simd_sel_t srslte_simd_f_max(simd_f_t a, simd_f_t b) { #ifdef LV_HAVE_AVX512 @@ -992,6 +1025,25 @@ 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 = &a; + int* b_ptr = &b; + simd_i_t ret; + int* sel = &selector; + + int* c_ptr = &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 */ @@ -1030,7 +1082,7 @@ static inline simd_s_t srslte_simd_s_load(int16_t *ptr) { return _mm_load_si128((__m128i*) ptr); #else /* LV_HAVE_SSE */ #ifdef HAVE_NEON - return vld1q_s16((int16x8_t*) ptr); + return vld1q_s16(ptr); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -1048,7 +1100,7 @@ static inline simd_s_t srslte_simd_s_loadu(int16_t *ptr) { return _mm_loadu_si128((__m128i*) ptr) #else /* LV_HAVE_SSE */ #ifdef HAVE_NEON - return vld1q_s16((int16x8_t*) ptr); + return vld1q_s16(ptr); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -1066,7 +1118,7 @@ static inline void srslte_simd_s_store(int16_t *ptr, simd_s_t simdreg) { _mm_store_si128((__m128i*) ptr, simdreg); #else /* LV_HAVE_SSE */ #ifdef HAVE_NEON - vst1q_s16((int16x8_t*) ptr, simdreg); + vst1q_s16( ptr, simdreg); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -1084,13 +1136,12 @@ static inline void srslte_simd_s_storeu(int16_t *ptr, simd_s_t simdreg) { _mm_storeu_si128((__m128i*) ptr, simdreg); #else /* LV_HAVE_SSE */ #ifdef HAVE_NEON - vst1q_s16((int16x8_t*) ptr, simdreg); + vst1q_s16(ptr, simdreg); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ #endif /* LV_HAVE_AVX512 */ } -vdupq_n_s16 static inline simd_s_t srslte_simd_s_zero(void) { #ifdef LV_HAVE_AVX512 return _mm512_setzero_si512(); @@ -1107,7 +1158,7 @@ static inline simd_s_t srslte_simd_s_zero(void) { #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ #endif /* LV_HAVE_AVX512 */ -}vmulq_s16 +} static inline simd_s_t srslte_simd_s_mul(simd_s_t a, simd_s_t b) { #ifdef LV_HAVE_AVX512 @@ -1294,7 +1345,7 @@ static inline void srslte_simd_c16i_store(c16_t *ptr, simd_c16_t simdreg) { _mm_store_si128((__m128i *) (ptr + 8), _mm_blend_epi16(re_sw, simdreg.im.m128, 0b10101010)); #else /*HAVE_NEON*/ #ifdef HAVE_NEON - vst2q_f32((float*)(ptr) ,simdreg); + vst2q_s16((int16_t*)(ptr) ,simdreg.m128); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -1314,7 +1365,7 @@ static inline void srslte_simd_c16i_storeu(c16_t *ptr, simd_c16_t simdreg) { _mm_storeu_si128((__m128i *) (ptr + 8), _mm_blend_epi16(re_sw, simdreg.im.m128, 0b10101010)); #else /*HAVE_NEON*/ #ifdef HAVE_NEON - vst2q_f32((float*)(ptr) ,simdreg); + vst2q_s16((int16_t*)(ptr) ,simdreg.m128); #endif /* HAVE_NEON */ #endif /* LV_HAVE_SSE */ #endif /* LV_HAVE_AVX2 */ @@ -1330,8 +1381,8 @@ static inline void srslte_simd_c16_store(int16_t *re, int16_t *im, simd_c16_t si _mm_store_si128((__m128i *) im, simdreg.im.m128); #else #ifdef HAVE_NEON - vst1q_f32((int16_t *) re, simdreg.m128.val[0]); - vst1q_f32((int16_t *) im, simdreg.m128.val[1]); + 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 */ @@ -1347,15 +1398,13 @@ static inline void srslte_simd_c16_storeu(int16_t *re, int16_t *im, simd_c16_t s _mm_storeu_si128((__m128i *) im, simdreg.im.m128); #else #ifdef HAVE_NEON - vst1q_f32((int16_t *) re, simdreg.m128.val[0]); - vst1q_f32((int16_t *) im, simdreg.m128.val[1]); + 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 */ } - -//TODO static inline simd_c16_t srslte_simd_c16_prod (simd_c16_t a, simd_c16_t b) { simd_c16_t ret; #ifdef LV_HAVE_AVX2 @@ -1385,14 +1434,14 @@ static inline simd_c16_t srslte_simd_c16_add (simd_c16_t a, simd_c16_t b) { ret.im.m128 = _mm_add_epi16(a.im.m128, b.im.m128); #else #ifdef HAVE_NEON - ret.m128.val[0] = vaddq_s32(a.m127.val[0],a.m127.val[0]); - ret.m128.val[1] = vaddq_s32(a.m127.val[1],a.m127.val[1]); + 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; } -vdupq_n_s16 + static inline simd_c16_t srslte_simd_c16_zero (void) { simd_c16_t ret; #ifdef LV_HAVE_AVX2 diff --git a/lib/src/phy/utils/vector_simd.c b/lib/src/phy/utils/vector_simd.c index 0294bd1af..ab281a653 100644 --- a/lib/src/phy/utils/vector_simd.c +++ b/lib/src/phy/utils/vector_simd.c @@ -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);