diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic.h index 4ec49d645..1dc13b49a 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic.h @@ -233,6 +233,156 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_u_sse2(lv_16sc_t* out, con #endif /* LV_HAVE_SSE2 */ +#ifdef LV_HAVE_AVX2 +#include + +static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_u_axv2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) +{ + lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0); + + const unsigned int avx_iters = num_points / 8; + + const lv_16sc_t* _in_a = in_a; + const lv_16sc_t* _in_b = in_b; + lv_16sc_t* _out = out; + unsigned int i; + + if (avx_iters > 0) + { + __m256i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, realcacc, imagcacc, result; + __VOLK_ATTR_ALIGNED(32) lv_16sc_t dotProductVector[8]; + + realcacc = _mm256_setzero_si256(); + imagcacc = _mm256_setzero_si256(); + + mask_imag = _mm256_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); + mask_real = _mm256_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); + + for(unsigned int number = 0; number < avx_iters; number++) + { + a = _mm256_loadu_si256((__m256i*)_in_a); + __builtin_prefetch(_in_a + 16); + b = _mm256_loadu_si256((__m256i*)_in_b); + __builtin_prefetch(_in_b + 16); + c = _mm256_mullo_epi16(a, b); + + c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. + real = _mm256_subs_epi16(c, c_sr); + + b_sl = _mm256_slli_si256(b, 2); + a_sl = _mm256_slli_si256(a, 2); + + imag1 = _mm256_mullo_epi16(a, b_sl); + imag2 = _mm256_mullo_epi16(b, a_sl); + + imag = _mm256_adds_epi16(imag1, imag2); //with saturation arithmetic! + + realcacc = _mm256_adds_epi16(realcacc, real); + imagcacc = _mm256_adds_epi16(imagcacc, imag); + + _in_a += 8; + _in_b += 8; + } + + realcacc = _mm256_and_si256(realcacc, mask_real); + imagcacc = _mm256_and_si256(imagcacc, mask_imag); + + result = _mm256_or_si256(realcacc, imagcacc); + + _mm256_storeu_si256((__m256i*)dotProductVector, result); // Store the results back into the dot product vector + + for (i = 0; i < 8; ++i) + { + dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[i])), sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[i]))); + } + } + + for (i = 0; i < (num_points % 8); ++i) + { + lv_16sc_t tmp = (*_in_a++) * (*_in_b++); + dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(tmp)), sat_adds16i(lv_cimag(dotProduct), lv_cimag(tmp))); + } + + *_out = dotProduct; +} +#endif /* LV_HAVE_AVX2 */ + + +#ifdef LV_HAVE_AVX2 +#include + +static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_a_axv2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) +{ + lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0); + + const unsigned int avx_iters = num_points / 8; + + const lv_16sc_t* _in_a = in_a; + const lv_16sc_t* _in_b = in_b; + lv_16sc_t* _out = out; + unsigned int i; + + if (avx_iters > 0) + { + __m256i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, realcacc, imagcacc, result; + __VOLK_ATTR_ALIGNED(32) lv_16sc_t dotProductVector[8]; + + realcacc = _mm256_setzero_si256(); + imagcacc = _mm256_setzero_si256(); + + mask_imag = _mm256_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); + mask_real = _mm256_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); + + for(unsigned int number = 0; number < avx_iters; number++) + { + a = _mm256_load_si256((__m256i*)_in_a); + __builtin_prefetch(_in_a + 16); + b = _mm256_load_si256((__m256i*)_in_b); + __builtin_prefetch(_in_b + 16); + c = _mm256_mullo_epi16(a, b); + + c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. + real = _mm256_subs_epi16(c, c_sr); + + b_sl = _mm256_slli_si256(b, 2); + a_sl = _mm256_slli_si256(a, 2); + + imag1 = _mm256_mullo_epi16(a, b_sl); + imag2 = _mm256_mullo_epi16(b, a_sl); + + imag = _mm256_adds_epi16(imag1, imag2); //with saturation arithmetic! + + realcacc = _mm256_adds_epi16(realcacc, real); + imagcacc = _mm256_adds_epi16(imagcacc, imag); + + _in_a += 8; + _in_b += 8; + } + + realcacc = _mm256_and_si256(realcacc, mask_real); + imagcacc = _mm256_and_si256(imagcacc, mask_imag); + + result = _mm256_or_si256(realcacc, imagcacc); + + _mm256_store_si256((__m256i*)dotProductVector, result); // Store the results back into the dot product vector + + for (i = 0; i < 8; ++i) + { + dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[i])), sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[i]))); + } + } + + for (i = 0; i < (num_points % 8); ++i) + { + lv_16sc_t tmp = (*_in_a++) * (*_in_b++); + dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(tmp)), sat_adds16i(lv_cimag(dotProduct), lv_cimag(tmp))); + } + + *_out = dotProduct; +} +#endif /* LV_HAVE_AVX2 */ + + #ifdef LV_HAVE_NEON #include @@ -385,8 +535,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_optvma(lv_16sc_t* out // use 2 accumulators to remove inter-instruction data dependencies accumulator1.val[0] = vmla_s16(accumulator1.val[0], a_val.val[0], b_val.val[0]); - accumulator1.val[1] = vmla_s16(accumulator1.val[1], a_val.val[0], b_val.val[1]); accumulator2.val[0] = vmls_s16(accumulator2.val[0], a_val.val[1], b_val.val[1]); + accumulator1.val[1] = vmla_s16(accumulator1.val[1], a_val.val[0], b_val.val[1]); accumulator2.val[1] = vmla_s16(accumulator2.val[1], a_val.val[1], b_val.val[0]); a_ptr += 4; diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h index 341b9ebcb..0480a8532 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h @@ -292,6 +292,190 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul #endif /* LV_HAVE_SSE2 */ +#ifdef LV_HAVE_AVX2 +#include + +static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_avx2(lv_16sc_t* result, const lv_16sc_t* in_common, const lv_16sc_t** in_a, int num_a_vectors, unsigned int num_points) +{ + lv_16sc_t dotProduct = lv_cmake(0,0); + + const unsigned int sse_iters = num_points / 8; + + const lv_16sc_t** _in_a = in_a; + const lv_16sc_t* _in_common = in_common; + lv_16sc_t* _out = result; + + if (sse_iters > 0) + { + __VOLK_ATTR_ALIGNED(32) lv_16sc_t dotProductVector[8]; + + __m256i* realcacc = (__m256i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m256i), volk_gnsssdr_get_alignment()); + __m256i* imagcacc = (__m256i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m256i), volk_gnsssdr_get_alignment()); + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + realcacc[n_vec] = _mm256_setzero_si256(); + imagcacc[n_vec] = _mm256_setzero_si256(); + } + + __m256i a, b, c, c_sr, mask_imag, mask_real, real, imag; + + mask_imag = _mm256_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); + mask_real = _mm256_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); + + for(unsigned int number = 0; number < sse_iters; number++) + { + b = _mm256_load_si256((__m256i*)_in_common); + __builtin_prefetch(_in_common + 16); + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + a = _mm256_load_si256((__m256i*)&(_in_a[n_vec][number*8])); + + c = _mm256_mullo_epi16(a, b); + + c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. + real = _mm256_subs_epi16(c, c_sr); + + c_sr = _mm256_slli_si256(b, 2); // b3.r, b2.i .... + c = _mm256_mullo_epi16(a, c_sr); // a3.i*b3.r, .... + + c_sr = _mm256_slli_si256(a, 2); // a3.r, a2.i .... + imag = _mm256_mullo_epi16(b, c_sr); // b3.i*a3.r, .... + + imag = _mm256_adds_epi16(c, imag); + + realcacc[n_vec] = _mm256_adds_epi16(realcacc[n_vec], real); + imagcacc[n_vec] = _mm256_adds_epi16(imagcacc[n_vec], imag); + } + _in_common += 8; + } + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + realcacc[n_vec] = _mm256_and_si256(realcacc[n_vec], mask_real); + imagcacc[n_vec] = _mm256_and_si256(imagcacc[n_vec], mask_imag); + + a = _mm256_or_si256(realcacc[n_vec], imagcacc[n_vec]); + + _mm256_store_si256((__m256i*)dotProductVector, a); // Store the results back into the dot product vector + dotProduct = lv_cmake(0,0); + for (int i = 0; i < 8; ++i) + { + dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[i])), + sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[i]))); + } + _out[n_vec] = dotProduct; + } + volk_gnsssdr_free(realcacc); + volk_gnsssdr_free(imagcacc); + } + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + for(unsigned int n = sse_iters * 8; n < num_points; n++) + { + lv_16sc_t tmp = in_common[n] * in_a[n_vec][n]; + + _out[n_vec] = lv_cmake(sat_adds16i(lv_creal(_out[n_vec]), lv_creal(tmp)), + sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp))); + } + } +} +#endif /* LV_HAVE_AVX2 */ + + +#ifdef LV_HAVE_AVX2 +#include + +static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_avx2(lv_16sc_t* result, const lv_16sc_t* in_common, const lv_16sc_t** in_a, int num_a_vectors, unsigned int num_points) +{ + lv_16sc_t dotProduct = lv_cmake(0,0); + + const unsigned int sse_iters = num_points / 8; + + const lv_16sc_t** _in_a = in_a; + const lv_16sc_t* _in_common = in_common; + lv_16sc_t* _out = result; + + if (sse_iters > 0) + { + __VOLK_ATTR_ALIGNED(32) lv_16sc_t dotProductVector[8]; + + __m256i* realcacc = (__m256i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m256i), volk_gnsssdr_get_alignment()); + __m256i* imagcacc = (__m256i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m256i), volk_gnsssdr_get_alignment()); + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + realcacc[n_vec] = _mm256_setzero_si256(); + imagcacc[n_vec] = _mm256_setzero_si256(); + } + + __m256i a, b, c, c_sr, mask_imag, mask_real, real, imag; + + mask_imag = _mm256_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); + mask_real = _mm256_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); + + for(unsigned int number = 0; number < sse_iters; number++) + { + b = _mm256_loadu_si256((__m256i*)_in_common); + __builtin_prefetch(_in_common + 16); + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + a = _mm256_loadu_si256((__m256i*)&(_in_a[n_vec][number*8])); + + c = _mm256_mullo_epi16(a, b); + + c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. + real = _mm256_subs_epi16(c, c_sr); + + c_sr = _mm256_slli_si256(b, 2); // b3.r, b2.i .... + c = _mm256_mullo_epi16(a, c_sr); // a3.i*b3.r, .... + + c_sr = _mm256_slli_si256(a, 2); // a3.r, a2.i .... + imag = _mm256_mullo_epi16(b, c_sr); // b3.i*a3.r, .... + + imag = _mm256_adds_epi16(c, imag); + + realcacc[n_vec] = _mm256_adds_epi16(realcacc[n_vec], real); + imagcacc[n_vec] = _mm256_adds_epi16(imagcacc[n_vec], imag); + } + _in_common += 8; + } + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + realcacc[n_vec] = _mm256_and_si256(realcacc[n_vec], mask_real); + imagcacc[n_vec] = _mm256_and_si256(imagcacc[n_vec], mask_imag); + + a = _mm256_or_si256(realcacc[n_vec], imagcacc[n_vec]); + + _mm256_store_si256((__m256i*)dotProductVector, a); // Store the results back into the dot product vector + dotProduct = lv_cmake(0,0); + for (int i = 0; i < 8; ++i) + { + dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[i])), + sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[i]))); + } + _out[n_vec] = dotProduct; + } + volk_gnsssdr_free(realcacc); + volk_gnsssdr_free(imagcacc); + } + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + for(unsigned int n = sse_iters * 8; n < num_points; n++) + { + lv_16sc_t tmp = in_common[n] * in_a[n_vec][n]; + + _out[n_vec] = lv_cmake(sat_adds16i(lv_creal(_out[n_vec]), lv_creal(tmp)), + sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp))); + } + } +} +#endif /* LV_HAVE_AVX2 */ + + #ifdef LV_HAVE_NEON #include diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic.h index 8857f0c0d..e68c5ac9a 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic.h @@ -134,6 +134,54 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_sse2(lv_16sc_t* r #endif /* LV_HAVE_SSE2 */ +#if LV_HAVE_AVX2 + +static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_a_avx2(lv_16sc_t* result, const lv_16sc_t* local_code, const lv_16sc_t* in, unsigned int num_points) +{ + int num_a_vectors = 3; + lv_16sc_t** in_a = (lv_16sc_t**)volk_gnsssdr_malloc(sizeof(lv_16sc_t*) * num_a_vectors, volk_gnsssdr_get_alignment()); + for(unsigned int n = 0; n < num_a_vectors; n++) + { + in_a[n] = (lv_16sc_t*)volk_gnsssdr_malloc(sizeof(lv_16sc_t)*num_points, volk_gnsssdr_get_alignment()); + memcpy((lv_16sc_t*)in_a[n], (lv_16sc_t*)in, sizeof(lv_16sc_t)*num_points); + } + + volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_avx2(result, local_code, (const lv_16sc_t**) in_a, num_a_vectors, num_points); + + for(unsigned int n = 0; n < num_a_vectors; n++) + { + volk_gnsssdr_free(in_a[n]); + } + volk_gnsssdr_free(in_a); +} + +#endif /* LV_HAVE_AVX2 */ + + +#if LV_HAVE_AVX2 + +static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_avx2(lv_16sc_t* result, const lv_16sc_t* local_code, const lv_16sc_t* in, unsigned int num_points) +{ + int num_a_vectors = 3; + lv_16sc_t** in_a = (lv_16sc_t**)volk_gnsssdr_malloc(sizeof(lv_16sc_t*) * num_a_vectors, volk_gnsssdr_get_alignment()); + for(unsigned int n = 0; n < num_a_vectors; n++) + { + in_a[n] = (lv_16sc_t*)volk_gnsssdr_malloc(sizeof(lv_16sc_t)*num_points, volk_gnsssdr_get_alignment()); + memcpy((lv_16sc_t*)in_a[n], (lv_16sc_t*)in, sizeof(lv_16sc_t)*num_points); + } + + volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_avx2(result, local_code, (const lv_16sc_t**) in_a, num_a_vectors, num_points); + + for(unsigned int n = 0; n < num_a_vectors; n++) + { + volk_gnsssdr_free(in_a[n]); + } + volk_gnsssdr_free(in_a); +} + +#endif /* LV_HAVE_AVX2 */ + + #ifdef LV_HAVE_NEON static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_neon(lv_16sc_t* result, const lv_16sc_t* local_code, const lv_16sc_t* in, unsigned int num_points) diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_multiply_16ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_multiply_16ic.h index 848330453..50df6d4cf 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_multiply_16ic.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_multiply_16ic.h @@ -81,7 +81,7 @@ static inline void volk_gnsssdr_16ic_x2_multiply_16ic_generic(lv_16sc_t* result, static inline void volk_gnsssdr_16ic_x2_multiply_16ic_a_sse2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) { const unsigned int sse_iters = num_points / 4; - __m128i a,b,c, c_sr, mask_imag, mask_real, real, imag, imag1,imag2, b_sl, a_sl, result; + __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, result; mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); mask_real = _mm_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); @@ -134,7 +134,7 @@ static inline void volk_gnsssdr_16ic_x2_multiply_16ic_a_sse2(lv_16sc_t* out, con static inline void volk_gnsssdr_16ic_x2_multiply_16ic_u_sse2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) { const unsigned int sse_iters = num_points / 4; - __m128i a,b,c, c_sr, mask_imag, mask_real, real, imag, imag1,imag2, b_sl, a_sl, result; + __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1,imag2, b_sl, a_sl, result; mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); mask_real = _mm_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); @@ -181,6 +181,115 @@ static inline void volk_gnsssdr_16ic_x2_multiply_16ic_u_sse2(lv_16sc_t* out, con #endif /* LV_HAVE_SSE2 */ +#ifdef LV_HAVE_AVX2 +#include + +static inline void volk_gnsssdr_16ic_x2_multiply_16ic_u_avx2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) +{ + unsigned int number = 0; + const unsigned int avx2_points = num_points / 8; + + const lv_16sc_t* _in_a = in_a; + const lv_16sc_t* _in_b = in_b; + lv_16sc_t* _out = out; + + __m256i a, b, c, c_sr, real, imag, imag1, imag2, b_sl, a_sl, result; + + const __m256i mask_imag = _mm256_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); + const __m256i mask_real = _mm256_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); + + for(;number < avx2_points; number++) + { + a = _mm256_loadu_si256((__m256i*)_in_a); // Load the ar + ai, br + bi as ar,ai,br,bi + b = _mm256_loadu_si256((__m256i*)_in_b); // Load the cr + ci, dr + di as cr,ci,dr,di + c = _mm256_mullo_epi16(a, b); + + c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. + real = _mm256_subs_epi16(c, c_sr); + real = _mm256_and_si256(real, mask_real); // a3.r*b3.r-a3.i*b3.i , 0, a3.r*b3.r- a3.i*b3.i + + b_sl = _mm256_slli_si256(b, 2); // b3.r, b2.i .... + a_sl = _mm256_slli_si256(a, 2); // a3.r, a2.i .... + + imag1 = _mm256_mullo_epi16(a, b_sl); // a3.i*b3.r, .... + imag2 = _mm256_mullo_epi16(b, a_sl); // b3.i*a3.r, .... + + imag = _mm256_adds_epi16(imag1, imag2); + imag = _mm256_and_si256(imag, mask_imag); // a3.i*b3.r+b3.i*a3.r, 0, ... + + result = _mm256_or_si256(real, imag); + + _mm256_storeu_si256((__m256i*)_out, result); + + _in_a += 8; + _in_b += 8; + _out += 8; + } + + number = avx2_points * 8; + for(;number < num_points; number++) + { + *_out++ = (*_in_a++) * (*_in_b++); + } +} +#endif /* LV_HAVE_AVX2 */ + + +#ifdef LV_HAVE_AVX2 +#include + +static inline void volk_gnsssdr_16ic_x2_multiply_16ic_a_avx2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) +{ + unsigned int number = 0; + const unsigned int avx2_points = num_points / 8; + + const lv_16sc_t* _in_a = in_a; + const lv_16sc_t* _in_b = in_b; + lv_16sc_t* _out = out; + + __m256i a, b, c, c_sr, real, imag, imag1, imag2, b_sl, a_sl, result; + + const __m256i mask_imag = _mm256_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); + const __m256i mask_real = _mm256_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); + + for(;number < avx2_points; number++) + { + a = _mm256_load_si256((__m256i*)_in_a); // Load the ar + ai, br + bi as ar,ai,br,bi + b = _mm256_load_si256((__m256i*)_in_b); // Load the cr + ci, dr + di as cr,ci,dr,di + c = _mm256_mullo_epi16(a, b); + + c_sr = _mm256_srli_si256(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. + real = _mm256_subs_epi16(c, c_sr); + real = _mm256_and_si256(real, mask_real); // a3.r*b3.r-a3.i*b3.i , 0, a3.r*b3.r- a3.i*b3.i + + b_sl = _mm256_slli_si256(b, 2); // b3.r, b2.i .... + a_sl = _mm256_slli_si256(a, 2); // a3.r, a2.i .... + + imag1 = _mm256_mullo_epi16(a, b_sl); // a3.i*b3.r, .... + imag2 = _mm256_mullo_epi16(b, a_sl); // b3.i*a3.r, .... + + imag = _mm256_adds_epi16(imag1, imag2); + imag = _mm256_and_si256(imag, mask_imag); // a3.i*b3.r+b3.i*a3.r, 0, ... + + result = _mm256_or_si256(real, imag); + + _mm256_store_si256((__m256i*)_out, result); + + _in_a += 8; + _in_b += 8; + _out += 8; + } + + number = avx2_points * 8; + for(;number < num_points; number++) + { + *_out++ = (*_in_a++) * (*_in_b++); + } +} +#endif /* LV_HAVE_AVX2 */ + + + #ifdef LV_HAVE_NEON #include diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_16ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_16ic.h index bd14fa310..cfd3d00fc 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_16ic.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_16ic.h @@ -69,20 +69,22 @@ static inline void volk_gnsssdr_32fc_convert_16ic_u_sse2(lv_16sc_t* outputVector float* inputVectorPtr = (float*)inputVector; int16_t* outputVectorPtr = (int16_t*)outputVector; + float aux; - float min_val = SHRT_MIN; - float max_val = SHRT_MAX; + const float min_val = (float)SHRT_MIN; + const float max_val = (float)SHRT_MAX; __m128 inputVal1, inputVal2; __m128i intInputVal1, intInputVal2; __m128 ret1, ret2; - __m128 vmin_val = _mm_set_ps1(min_val); - __m128 vmax_val = _mm_set_ps1(max_val); + const __m128 vmin_val = _mm_set_ps1(min_val); + const __m128 vmax_val = _mm_set_ps1(max_val); for(unsigned int i = 0; i < sse_iters; i++) { inputVal1 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal2 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; + __builtin_prefetch(inputVectorPtr + 8); // Clip ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val); @@ -99,11 +101,12 @@ static inline void volk_gnsssdr_32fc_convert_16ic_u_sse2(lv_16sc_t* outputVector for(unsigned int i = sse_iters * 8; i < num_points * 2; i++) { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - *outputVectorPtr++ = (int16_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int16_t)rintf(aux); } } #endif /* LV_HAVE_SSE2 */ @@ -118,20 +121,22 @@ static inline void volk_gnsssdr_32fc_convert_16ic_u_sse(lv_16sc_t* outputVector, float* inputVectorPtr = (float*)inputVector; int16_t* outputVectorPtr = (int16_t*)outputVector; + float aux; - float min_val = SHRT_MIN; - float max_val = SHRT_MAX; + const float min_val = (float)SHRT_MIN; + const float max_val = (float)SHRT_MAX; __m128 inputVal1, inputVal2; __m128i intInputVal1, intInputVal2; // is __m128i defined in xmmintrin.h? __m128 ret1, ret2; - __m128 vmin_val = _mm_set_ps1(min_val); - __m128 vmax_val = _mm_set_ps1(max_val); + const __m128 vmin_val = _mm_set_ps1(min_val); + const __m128 vmax_val = _mm_set_ps1(max_val); for(unsigned int i = 0;i < sse_iters; i++) { inputVal1 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal2 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; + __builtin_prefetch(inputVectorPtr + 8); // Clip ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val); @@ -148,11 +153,12 @@ static inline void volk_gnsssdr_32fc_convert_16ic_u_sse(lv_16sc_t* outputVector, for(unsigned int i = sse_iters * 8; i < num_points*2; i++) { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - *outputVectorPtr++ = (int16_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int16_t)rintf(aux); } } #endif /* LV_HAVE_SSE */ @@ -167,20 +173,22 @@ static inline void volk_gnsssdr_32fc_convert_16ic_a_sse2(lv_16sc_t* outputVector float* inputVectorPtr = (float*)inputVector; int16_t* outputVectorPtr = (int16_t*)outputVector; + float aux; - float min_val = SHRT_MIN; - float max_val = SHRT_MAX; + const float min_val = (float)SHRT_MIN; + const float max_val = (float)SHRT_MAX; __m128 inputVal1, inputVal2; __m128i intInputVal1, intInputVal2; __m128 ret1, ret2; - __m128 vmin_val = _mm_set_ps1(min_val); - __m128 vmax_val = _mm_set_ps1(max_val); + const __m128 vmin_val = _mm_set_ps1(min_val); + const __m128 vmax_val = _mm_set_ps1(max_val); for(unsigned int i = 0; i < sse_iters; i++) { inputVal1 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal2 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4; + __builtin_prefetch(inputVectorPtr + 8); // Clip ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val); @@ -197,11 +205,12 @@ static inline void volk_gnsssdr_32fc_convert_16ic_a_sse2(lv_16sc_t* outputVector for(unsigned int i = sse_iters * 8; i < num_points * 2; i++) { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - *outputVectorPtr++ = (int16_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int16_t)rintf(aux); } } #endif /* LV_HAVE_SSE2 */ @@ -212,24 +221,25 @@ static inline void volk_gnsssdr_32fc_convert_16ic_a_sse2(lv_16sc_t* outputVector static inline void volk_gnsssdr_32fc_convert_16ic_a_sse(lv_16sc_t* outputVector, const lv_32fc_t* inputVector, unsigned int num_points) { - const unsigned int sse_iters = num_points/4; + const unsigned int sse_iters = num_points / 4; + const float min_val = (float)SHRT_MIN; + const float max_val = (float)SHRT_MAX; + float aux; float* inputVectorPtr = (float*)inputVector; int16_t* outputVectorPtr = (int16_t*)outputVector; - float min_val = SHRT_MIN; - float max_val = SHRT_MAX; - __m128 inputVal1, inputVal2; __m128i intInputVal1, intInputVal2; __m128 ret1, ret2; - __m128 vmin_val = _mm_set_ps1(min_val); - __m128 vmax_val = _mm_set_ps1(max_val); + const __m128 vmin_val = _mm_set_ps1(min_val); + const __m128 vmax_val = _mm_set_ps1(max_val); for(unsigned int i = 0;i < sse_iters; i++) { inputVal1 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal2 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4; + __builtin_prefetch(inputVectorPtr + 8); // Clip ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val); @@ -246,11 +256,12 @@ static inline void volk_gnsssdr_32fc_convert_16ic_a_sse(lv_16sc_t* outputVector, for(unsigned int i = sse_iters * 8; i < num_points * 2; i++) { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - *outputVectorPtr++ = (int16_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int16_t)rintf(aux); } } #endif /* LV_HAVE_SSE */ @@ -266,8 +277,12 @@ static inline void volk_gnsssdr_32fc_convert_16ic_neon(lv_16sc_t* outputVector, float32_t* inputVectorPtr = (float32_t*)inputVector; int16_t* outputVectorPtr = (int16_t*)outputVector; - float32x4_t min_val = vmovq_n_f32(SHRT_MIN); - float32x4_t max_val = vmovq_n_f32(SHRT_MAX); + const float min_val_f = (float)SHRT_MIN; + const float max_val_f = (float)SHRT_MAX; + float32_t aux; + + const float32x4_t min_val = vmovq_n_f32(min_val_f); + const float32x4_t max_val = vmovq_n_f32(max_val_f); float32x4_t half = vdupq_n_f32(0.5f); float32x4_t ret1, ret2, a, b, sign, PlusHalf, Round; @@ -275,10 +290,11 @@ static inline void volk_gnsssdr_32fc_convert_16ic_neon(lv_16sc_t* outputVector, int16x4_t intInputVal1, intInputVal2; int16x8_t res; - for(unsigned int i = 0;i < neon_iters; i++) + for(unsigned int i = 0; i < neon_iters; i++) { a = vld1q_f32((const float32_t*)(inputVectorPtr)); inputVectorPtr += 4; b = vld1q_f32((const float32_t*)(inputVectorPtr)); inputVectorPtr += 4; + __builtin_prefetch(inputVectorPtr + 8); ret1 = vmaxq_f32(vminq_f32(a, max_val), min_val); ret2 = vmaxq_f32(vminq_f32(b, max_val), min_val); @@ -304,11 +320,12 @@ static inline void volk_gnsssdr_32fc_convert_16ic_neon(lv_16sc_t* outputVector, for(unsigned int i = neon_iters * 8; i < num_points * 2; i++) { - if(inputVectorPtr[i] > SHRT_MAX) - inputVectorPtr[i] = (int16_t)SHRT_MAX; - else if(inputVectorPtr[i] < SHRT_MIN) - inputVectorPtr[i] = (int16_t)SHRT_MIN; - *outputVectorPtr++ = (int16_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++; + if(aux > max_val_f) + aux = max_val_f; + else if(aux < min_val_f) + aux = min_val_f; + *outputVectorPtr++ = (int16_t)rintf(aux); } } @@ -321,16 +338,18 @@ static inline void volk_gnsssdr_32fc_convert_16ic_generic(lv_16sc_t* outputVecto { float* inputVectorPtr = (float*)inputVector; int16_t* outputVectorPtr = (int16_t*)outputVector; - float min_val = SHRT_MIN; - float max_val = SHRT_MAX; + const float min_val = (float)SHRT_MIN; + const float max_val = (float)SHRT_MAX; + float aux; - for(unsigned int i = 0; i < num_points*2; i++) + for(unsigned int i = 0; i < num_points * 2; i++) { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - outputVectorPtr[i] = (int16_t)rintf(inputVectorPtr[i]); + aux = *inputVectorPtr++; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int16_t)rintf(aux); } } #endif /* LV_HAVE_GENERIC */ diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_8ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_8ic.h index 7071214cf..6f1de0c05 100755 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_8ic.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_8ic.h @@ -62,34 +62,62 @@ #include "volk_gnsssdr/volk_gnsssdr_complex.h" +#ifdef LV_HAVE_GENERIC + +static inline void volk_gnsssdr_32fc_convert_8ic_generic(lv_8sc_t* outputVector, const lv_32fc_t* inputVector, unsigned int num_points) +{ + float* inputVectorPtr = (float*)inputVector; + int8_t* outputVectorPtr = (int8_t*)outputVector; + const float min_val = (float)SCHAR_MIN; + const float max_val = (float)SCHAR_MAX; + float aux; + + for(unsigned int i = 0; i < num_points * 2; i++) + { + aux = *inputVectorPtr++ * max_val; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int8_t)rintf(aux); + } +} +#endif /* LV_HAVE_GENERIC */ + + #ifdef LV_HAVE_SSE2 #include static inline void volk_gnsssdr_32fc_convert_8ic_u_sse2(lv_8sc_t* outputVector, const lv_32fc_t* inputVector, unsigned int num_points) { - unsigned i = 0; - const unsigned int sse_iters = num_points * 2 / 16; + const unsigned int sse_iters = num_points / 8; float* inputVectorPtr = (float*)inputVector; int8_t* outputVectorPtr = (int8_t*)outputVector; - float min_val = SCHAR_MIN; - float max_val = SCHAR_MAX; + const float min_val = (float)SCHAR_MIN; + const float max_val = (float)SCHAR_MAX; + float aux; __m128 inputVal1, inputVal2, inputVal3, inputVal4; __m128i intInputVal1, intInputVal2, intInputVal3, intInputVal4; __m128i int8InputVal; __m128 ret1, ret2, ret3, ret4; - __m128 vmin_val = _mm_set_ps1(min_val); - __m128 vmax_val = _mm_set_ps1(max_val); + const __m128 vmin_val = _mm_set_ps1(min_val); + const __m128 vmax_val = _mm_set_ps1(max_val); - for(;i < sse_iters; i++) + for(unsigned int i = 0; i < sse_iters; i++) { inputVal1 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal2 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal3 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal4 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4; + inputVal1 = _mm_mul_ps(inputVal1, vmax_val); + inputVal2 = _mm_mul_ps(inputVal2, vmax_val); + inputVal3 = _mm_mul_ps(inputVal3, vmax_val); + inputVal4 = _mm_mul_ps(inputVal4, vmax_val); + // Clip ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val); ret2 = _mm_max_ps(_mm_min_ps(inputVal2, vmax_val), vmin_val); @@ -109,39 +137,19 @@ static inline void volk_gnsssdr_32fc_convert_8ic_u_sse2(lv_8sc_t* outputVector, outputVectorPtr += 16; } - for(i = sse_iters * 16; i < num_points * 2; i++) + for(unsigned int i = sse_iters * 16; i < num_points * 2; i++) { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - *outputVectorPtr++ = (int8_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++ * max_val; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int8_t)rintf(aux); } } #endif /* LV_HAVE_SSE2 */ -#ifdef LV_HAVE_GENERIC - -static inline void volk_gnsssdr_32fc_convert_8ic_generic(lv_8sc_t* outputVector, const lv_32fc_t* inputVector, unsigned int num_points) -{ - float* inputVectorPtr = (float*)inputVector; - int8_t* outputVectorPtr = (int8_t*)outputVector; - float min_val = SCHAR_MIN; - float max_val = SCHAR_MAX; - - for(unsigned int i = 0; i < num_points * 2; i++) - { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - outputVectorPtr[i] = (int8_t)rintf(inputVectorPtr[i]); - } -} -#endif /* LV_HAVE_GENERIC */ - - #ifdef LV_HAVE_SSE2 #include @@ -152,15 +160,16 @@ static inline void volk_gnsssdr_32fc_convert_8ic_a_sse2(lv_8sc_t* outputVector, float* inputVectorPtr = (float*)inputVector; int8_t* outputVectorPtr = (int8_t*)outputVector; - float min_val = SCHAR_MIN; - float max_val = SCHAR_MAX; + const float min_val = (float)SCHAR_MIN; + const float max_val = (float)SCHAR_MAX; + float aux; __m128 inputVal1, inputVal2, inputVal3, inputVal4; __m128i intInputVal1, intInputVal2, intInputVal3, intInputVal4; __m128i int8InputVal; __m128 ret1, ret2, ret3, ret4; - __m128 vmin_val = _mm_set_ps1(min_val); - __m128 vmax_val = _mm_set_ps1(max_val); + const __m128 vmin_val = _mm_set_ps1(min_val); + const __m128 vmax_val = _mm_set_ps1(max_val); for(unsigned int i = 0; i < sse_iters; i++) { @@ -169,6 +178,11 @@ static inline void volk_gnsssdr_32fc_convert_8ic_a_sse2(lv_8sc_t* outputVector, inputVal3 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4; inputVal4 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4; + inputVal1 = _mm_mul_ps(inputVal1, vmax_val); + inputVal2 = _mm_mul_ps(inputVal2, vmax_val); + inputVal3 = _mm_mul_ps(inputVal3, vmax_val); + inputVal4 = _mm_mul_ps(inputVal4, vmax_val); + // Clip ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val); ret2 = _mm_max_ps(_mm_min_ps(inputVal2, vmax_val), vmin_val); @@ -190,11 +204,12 @@ static inline void volk_gnsssdr_32fc_convert_8ic_a_sse2(lv_8sc_t* outputVector, for(unsigned int i = sse_iters * 16; i < num_points * 2; i++) { - if(inputVectorPtr[i] > max_val) - inputVectorPtr[i] = max_val; - else if(inputVectorPtr[i] < min_val) - inputVectorPtr[i] = min_val; - *outputVectorPtr++ = (int8_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++ * max_val; + if(aux > max_val) + aux = max_val; + else if(aux < min_val) + aux = min_val; + *outputVectorPtr++ = (int8_t)rintf(aux); } } #endif /* LV_HAVE_SSE2 */ @@ -209,9 +224,12 @@ static inline void volk_gnsssdr_32fc_convert_8ic_neon(lv_8sc_t* outputVector, co float32_t* inputVectorPtr = (float32_t*)inputVector; int8_t* outputVectorPtr = (int8_t*)outputVector; + const float32_t max_val_f = (float32_t)SCHAR_MAX; + const float32_t min_val_f = (float32_t)SCHAR_MIN; + float32_t aux; - const float32x4_t min_val = vmovq_n_f32((float32_t)SCHAR_MIN); - const float32x4_t max_val = vmovq_n_f32((float32_t)SCHAR_MAX); + const float32x4_t min_val = vmovq_n_f32(min_val_f); + const float32x4_t max_val = vmovq_n_f32(max_val_f); const float32x4_t half = vdupq_n_f32(0.5f); @@ -225,6 +243,7 @@ static inline void volk_gnsssdr_32fc_convert_8ic_neon(lv_8sc_t* outputVector, co for(unsigned int i = 0; i < neon_iters; i++) { a = vld1q_f32((const float32_t*)inputVectorPtr); inputVectorPtr += 4; + a = vmulq_f32(a, max_val); ret1 = vmaxq_f32(vminq_f32(a, max_val), min_val); sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(ret1), 31))); PlusHalf = vaddq_f32(ret1, half); @@ -233,6 +252,7 @@ static inline void volk_gnsssdr_32fc_convert_8ic_neon(lv_8sc_t* outputVector, co intInputVal1 = vqmovn_s32(toint_a); a = vld1q_f32((const float32_t*)inputVectorPtr); inputVectorPtr += 4; + a = vmulq_f32(a, max_val); ret1 = vmaxq_f32(vminq_f32(a, max_val), min_val); sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(ret1), 31))); PlusHalf = vaddq_f32(ret1, half); @@ -244,6 +264,7 @@ static inline void volk_gnsssdr_32fc_convert_8ic_neon(lv_8sc_t* outputVector, co res8_1 = vqmovn_s16(pack16_8_1); a = vld1q_f32((const float32_t*)inputVectorPtr); inputVectorPtr += 4; + a = vmulq_f32(a, max_val); ret1 = vmaxq_f32(vminq_f32(a, max_val), min_val); sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(ret1), 31))); PlusHalf = vaddq_f32(ret1, half); @@ -252,6 +273,7 @@ static inline void volk_gnsssdr_32fc_convert_8ic_neon(lv_8sc_t* outputVector, co intInputVal1 = vqmovn_s32(toint_a); a = vld1q_f32((const float32_t*)inputVectorPtr); inputVectorPtr += 4; + a = vmulq_f32(a, max_val); ret1 = vmaxq_f32(vminq_f32(a, max_val), min_val); sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(ret1), 31))); PlusHalf = vaddq_f32(ret1, half); @@ -270,11 +292,12 @@ static inline void volk_gnsssdr_32fc_convert_8ic_neon(lv_8sc_t* outputVector, co for(unsigned int i = neon_iters * 16; i < num_points * 2; i++) { - if(inputVectorPtr[i] > (float32_t)SCHAR_MAX) - inputVectorPtr[i] = (float32_t)SCHAR_MAX; - else if(inputVectorPtr[i] < (float32_t)SCHAR_MIN) - inputVectorPtr[i] = (float32_t)SCHAR_MIN; - *outputVectorPtr++ = (int8_t)rintf(*inputVectorPtr++); + aux = *inputVectorPtr++ * max_val_f; + if(aux > max_val_f) + aux = max_val_f; + else if(aux < min_val_f) + aux = min_val_f; + *outputVectorPtr++ = (int8_t)rintf(aux); } }