From 485a405bab9f2ae06fb7edd19ceda084039f213f Mon Sep 17 00:00:00 2001 From: Carles Fernandez Date: Sun, 20 Mar 2016 13:11:53 +0100 Subject: [PATCH] Adding new neon kernel and solving x86 issues Managing memory with volk_gnsssdr instead of malloc and free. This seems to solve runtime problems (segmentation faults) in i386 (32 bit) architectures. --- .../volk_gnsssdr_16ic_x2_dot_prod_16ic.h | 57 +++++++- .../volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h | 125 +++++++++++++++--- ...olk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic.h | 28 +++- ...gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h | 3 +- 4 files changed, 187 insertions(+), 26 deletions(-) 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 5f760c2c7..4ec49d645 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 @@ -336,8 +336,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_vma(lv_16sc_t* out, c tmp.val[0] = vmls_s16(tmp.val[0], a_val.val[1], b_val.val[1]); tmp.val[1] = vmla_s16(tmp.val[1], a_val.val[0], b_val.val[1]); - accumulator.val[0] = vadd_s16(accumulator.val[0], tmp.val[0]); - accumulator.val[1] = vadd_s16(accumulator.val[1], tmp.val[1]); + accumulator.val[0] = vqadd_s16(accumulator.val[0], tmp.val[0]); + accumulator.val[1] = vqadd_s16(accumulator.val[1], tmp.val[1]); a_ptr += 4; b_ptr += 4; @@ -355,4 +355,57 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_vma(lv_16sc_t* out, c #endif /* LV_HAVE_NEON */ + +#ifdef LV_HAVE_NEON +#include + +static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_optvma(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) +{ + unsigned int quarter_points = num_points / 4; + unsigned int number; + + lv_16sc_t* a_ptr = (lv_16sc_t*) in_a; + lv_16sc_t* b_ptr = (lv_16sc_t*) in_b; + // for 2-lane vectors, 1st lane holds the real part, + // 2nd lane holds the imaginary part + int16x4x2_t a_val, b_val, accumulator1, accumulator2; + + __VOLK_ATTR_ALIGNED(16) lv_16sc_t accum_result[4]; + accumulator1.val[0] = vdup_n_s16(0); + accumulator1.val[1] = vdup_n_s16(0); + accumulator2.val[0] = vdup_n_s16(0); + accumulator2.val[1] = vdup_n_s16(0); + + for(number = 0; number < quarter_points; ++number) + { + a_val = vld2_s16((int16_t*)a_ptr); // a0r|a1r|a2r|a3r || a0i|a1i|a2i|a3i + b_val = vld2_s16((int16_t*)b_ptr); // b0r|b1r|b2r|b3r || b0i|b1i|b2i|b3i + __builtin_prefetch(a_ptr + 8); + __builtin_prefetch(b_ptr + 8); + + // 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]); + accumulator2.val[1] = vmla_s16(accumulator2.val[1], a_val.val[1], b_val.val[0]); + + a_ptr += 4; + b_ptr += 4; + } + + accumulator1.val[0] = vqadd_s16(accumulator1.val[0], accumulator2.val[0]); + accumulator1.val[1] = vqadd_s16(accumulator1.val[1], accumulator2.val[1]); + + vst2_s16((int16_t*)accum_result, accumulator1); + *out = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3]; + + // tail case + for(number = quarter_points * 4; number < num_points; ++number) + { + *out += (*a_ptr++) * (*b_ptr++); + } +} + +#endif /* LV_HAVE_NEON */ + #endif /*INCLUDED_volk_gnsssdr_16ic_x2_dot_prod_16ic_H*/ 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 6d27bf97a..341b9ebcb 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 @@ -62,6 +62,8 @@ #include +#include +#include #include #ifdef LV_HAVE_GENERIC @@ -120,11 +122,14 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul { __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4]; - __m128i* realcacc; - __m128i* imagcacc; + __m128i* realcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment()); + __m128i* imagcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment()); - realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 - imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + realcacc[n_vec] = _mm_setzero_si128(); + imagcacc[n_vec] = _mm_setzero_si128(); + } __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag; @@ -176,8 +181,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul } _out[n_vec] = dotProduct; } - free(realcacc); - free(imagcacc); + volk_gnsssdr_free(realcacc); + volk_gnsssdr_free(imagcacc); } for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) @@ -211,11 +216,14 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul { __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4]; - __m128i* realcacc; - __m128i* imagcacc; + __m128i* realcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment()); + __m128i* imagcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment()); - realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 - imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + realcacc[n_vec] = _mm_setzero_si128(); + imagcacc[n_vec] = _mm_setzero_si128(); + } __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag; @@ -246,7 +254,6 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul realcacc[n_vec] = _mm_adds_epi16(realcacc[n_vec], real); imagcacc[n_vec] = _mm_adds_epi16(imagcacc[n_vec], imag); - } _in_common += 4; } @@ -267,8 +274,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul } _out[n_vec] = dotProduct; } - free(realcacc); - free(imagcacc); + volk_gnsssdr_free(realcacc); + volk_gnsssdr_free(imagcacc); } for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) @@ -304,9 +311,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon(lv_16sc_t* result, int16x4x2_t a_val, b_val, c_val; - //todo dyn mem reg - int16x4x2_t* accumulator; - accumulator = (int16x4x2_t*)calloc(num_a_vectors, sizeof(int16x4x2_t)); + int16x4x2_t* accumulator = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment()); int16x4x2_t tmp_real, tmp_imag; @@ -357,7 +362,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon(lv_16sc_t* result, } _out[n_vec] = dotProduct; } - free(accumulator); + volk_gnsssdr_free(accumulator); } for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) @@ -393,8 +398,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_vma(lv_16sc_t* res int16x4x2_t a_val, b_val, tmp; - int16x4x2_t* accumulator; - accumulator = (int16x4x2_t*)malloc(num_a_vectors * sizeof(int16x4x2_t)); + int16x4x2_t* accumulator = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment()); for(int n_vec = 0; n_vec < num_a_vectors; n_vec++) { @@ -434,7 +438,88 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_vma(lv_16sc_t* res } _out[n_vec] = dotProduct; } - free(accumulator); + volk_gnsssdr_free(accumulator); + } + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + for(unsigned int n = neon_iters * 4; 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_NEON */ + + +#ifdef LV_HAVE_NEON +#include + +static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_optvma(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 neon_iters = num_points / 4; + + const lv_16sc_t** _in_a = in_a; + const lv_16sc_t* _in_common = in_common; + lv_16sc_t* _out = result; + + if (neon_iters > 0) + { + __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4]; + + int16x4x2_t a_val, b_val; + + int16x4x2_t* accumulator1 = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment()); + int16x4x2_t* accumulator2 = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment()); + + for(int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + accumulator1[n_vec].val[0] = vdup_n_s16(0); + accumulator1[n_vec].val[1] = vdup_n_s16(0); + accumulator2[n_vec].val[0] = vdup_n_s16(0); + accumulator2[n_vec].val[1] = vdup_n_s16(0); + } + + for(unsigned int number = 0; number < neon_iters; number++) + { + b_val = vld2_s16((int16_t*)_in_common); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg + __builtin_prefetch(_in_common + 8); + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + a_val = vld2_s16((int16_t*)&(_in_a[n_vec][number*4])); + + accumulator1[n_vec].val[0] = vmla_s16(accumulator1[n_vec].val[0], a_val.val[0], b_val.val[0]); + accumulator1[n_vec].val[1] = vmla_s16(accumulator1[n_vec].val[1], a_val.val[0], b_val.val[1]); + accumulator2[n_vec].val[0] = vmls_s16(accumulator2[n_vec].val[0], a_val.val[1], b_val.val[1]); + accumulator2[n_vec].val[1] = vmla_s16(accumulator2[n_vec].val[1], a_val.val[1], b_val.val[0]); + } + _in_common += 4; + } + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + accumulator1[n_vec].val[0] = vqadd_s16(accumulator1[n_vec].val[0], accumulator2[n_vec].val[0]); + accumulator1[n_vec].val[1] = vqadd_s16(accumulator1[n_vec].val[1], accumulator2[n_vec].val[1]); + } + + for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) + { + vst2_s16((int16_t*)dotProductVector, accumulator1[n_vec]); // Store the results back into the dot product vector + dotProduct = lv_cmake(0,0); + for (int i = 0; i < 4; ++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(accumulator1); + volk_gnsssdr_free(accumulator2); } for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) 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 83c207524..8857f0c0d 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 @@ -110,7 +110,7 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_a_sse2(lv_16sc_t* r #endif /* SSE2 */ -#if LV_HAVE_SSE2 && LV_HAVE_64 +#if LV_HAVE_SSE2 static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_sse2(lv_16sc_t* result, const lv_16sc_t* local_code, const lv_16sc_t* in, unsigned int num_points) { @@ -131,7 +131,7 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_sse2(lv_16sc_t* r volk_gnsssdr_free(in_a); } -#endif /* LV_HAVE_SSE2 && LV_HAVE_64 */ +#endif /* LV_HAVE_SSE2 */ #ifdef LV_HAVE_NEON @@ -180,6 +180,30 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_neon_vma(lv_16sc_t* } #endif // NEON + +#ifdef LV_HAVE_NEON + +static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_neon_optvma(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_neon_optvma(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 // NEON + #endif // INCLUDED_volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_H diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h index cebacb8e3..362ca16db 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h @@ -71,9 +71,8 @@ #include #include -#include #include -//#include +#include #ifdef LV_HAVE_GENERIC