mirrors/gnss-sdr
1
0
Fork 0
mirror of https://github.com/gnss-sdr/gnss-sdr synced 2024-12-14 12:10:34 +00:00
Code Issues Releases Wiki Activity

Make prefetching more portable

Browse Source 
See https://github.com/gnuradio/volk/issues/97 and
https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html#Other-Builtins
This commit is contained in:
Carles Fernandez 2016-10-16 11:31:42 +02:00
parent bbd1c2fe7c
commit 3cce2e8340
14 changed files with 84 additions and 78 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/include/volk_gnsssdr/volk_gnsssdr_common.h
View File

@ -39,6 +39,8 @@
# define __VOLK_ATTR_EXPORT
# define __VOLK_ATTR_IMPORT
# endif
# define __VOLK_PREFETCH(addr) __builtin_prefetch(addr)
# define __VOLK_PREFETCH_LOCALITY(addr, rw, locality) __builtin_prefetch(addr, rw, locality)
#elif _MSC_VER
# define __VOLK_ATTR_ALIGNED(x) __declspec(align(x))
# define __VOLK_ATTR_UNUSED
@ -46,6 +48,8 @@
# define __VOLK_ATTR_DEPRECATED __declspec(deprecated)
# define __VOLK_ATTR_EXPORT __declspec(dllexport)
# define __VOLK_ATTR_IMPORT __declspec(dllimport)
# define __VOLK_PREFETCH(addr)
# define __VOLK_PREFETCH_LOCALITY(addr, rw, locality)
#else
# define __VOLK_ATTR_ALIGNED(x)
# define __VOLK_ATTR_UNUSED
@ -53,6 +57,8 @@
# define __VOLK_ATTR_DEPRECATED
# define __VOLK_ATTR_EXPORT
# define __VOLK_ATTR_IMPORT
# define __VOLK_PREFETCH(addr)
# define __VOLK_PREFETCH_LOCALITY(addr, rw, locality)
#endif
////////////////////////////////////////////////////////////////////////

2
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_convert_32fc.h
View File

@ -197,7 +197,7 @@ static inline void volk_gnsssdr_16ic_convert_32fc_neon(lv_32fc_t* outputVector,
for(i = 0; i < sse_iters; i++)
{
a16x4 = vld1_s16((const int16_t*)_in);
__builtin_prefetch(_in + 4);
__VOLK_PREFETCH(_in + 4);
a32x4 = vmovl_s16(a16x4);
f32x4 = vcvtq_f32_s32(a32x4);
vst1q_f32((float32_t*)_out, f32x4);

18
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_s32fc_x2_rotator_16ic.h
View File

@ -181,7 +181,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_a_sse3(lv_16sc_t* out
//next two samples
_in += 2;
a = _mm_set_ps((float)(lv_cimag(_in[1])), (float)(lv_creal(_in[1])), (float)(lv_cimag(_in[0])), (float)(lv_creal(_in[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -287,7 +287,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_a_sse3_reload(lv_16sc
//next two samples
_in += 2;
a = _mm_set_ps((float)(lv_cimag(_in[1])), (float)(lv_creal(_in[1])), (float)(lv_cimag(_in[0])), (float)(lv_creal(_in[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -344,7 +344,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_a_sse3_reload(lv_16sc
//next two samples
_in += 2;
a = _mm_set_ps((float)(lv_cimag(_in[1])), (float)(lv_creal(_in[1])), (float)(lv_cimag(_in[0])), (float)(lv_creal(_in[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -436,7 +436,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_u_sse3(lv_16sc_t* out
//next two samples
_in += 2;
a = _mm_set_ps((float)(lv_cimag(_in[1])), (float)(lv_creal(_in[1])), (float)(lv_cimag(_in[0])), (float)(lv_creal(_in[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -541,7 +541,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_u_sse3_reload(lv_16sc
//next two samples
_in += 2;
a = _mm_set_ps((float)(lv_cimag(_in[1])), (float)(lv_creal(_in[1])), (float)(lv_cimag(_in[0])), (float)(lv_creal(_in[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -598,7 +598,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_u_sse3_reload(lv_16sc
//next two samples
_in += 2;
a = _mm_set_ps((float)(lv_cimag(_in[1])), (float)(lv_creal(_in[1])), (float)(lv_cimag(_in[0])), (float)(lv_creal(_in[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -686,7 +686,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_neon(lv_16sc_t* outVe
{
/* load 4 complex numbers (int 16 bits each component) */
tmp16 = vld2_s16((int16_t*)_in);
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
_in += 4;
/* promote them to int 32 bits */
@ -822,7 +822,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_neon_reload(lv_16sc_t
{
/* load 4 complex numbers (int 16 bits each component) */
tmp16 = vld2_s16((int16_t*)_in);
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
_in += 4;
/* promote them to int 32 bits */
@ -891,7 +891,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_neon_reload(lv_16sc_t
{
/* load 4 complex numbers (int 16 bits each component) */
tmp16 = vld2_s16((int16_t*)_in);
__builtin_prefetch(_in + 8);
__VOLK_PREFETCH(_in + 8);
_in += 4;
/* promote them to int 32 bits */

28
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic.h
View File

@ -108,9 +108,9 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_a_sse2(lv_16sc_t* out, con
{
// a[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
a = _mm_load_si128((__m128i*)_in_a); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
__builtin_prefetch(_in_a + 8);
__VOLK_PREFETCH(_in_a + 8);
b = _mm_load_si128((__m128i*)_in_b);
__builtin_prefetch(_in_b + 8);
__VOLK_PREFETCH(_in_b + 8);
c = _mm_mullo_epi16(a, b); // a3.i*b3.i, a3.r*b3.r, ....
c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
@ -188,9 +188,9 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_u_sse2(lv_16sc_t* out, con
//imaginery part -> reinterpret_cast<cv T*>(a)[2*i + 1]
// a[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
a = _mm_loadu_si128((__m128i*)_in_a); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
__builtin_prefetch(_in_a + 8);
__VOLK_PREFETCH(_in_a + 8);
b = _mm_loadu_si128((__m128i*)_in_b);
__builtin_prefetch(_in_b + 8);
__VOLK_PREFETCH(_in_b + 8);
c = _mm_mullo_epi16(a, b); // a3.i*b3.i, a3.r*b3.r, ....
c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
@ -264,9 +264,9 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_u_axv2(lv_16sc_t* out, con
for(number = 0; number < avx_iters; number++)
{
a = _mm256_loadu_si256((__m256i*)_in_a);
__builtin_prefetch(_in_a + 16);
__VOLK_PREFETCH(_in_a + 16);
b = _mm256_loadu_si256((__m256i*)_in_b);
__builtin_prefetch(_in_b + 16);
__VOLK_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.
@ -341,9 +341,9 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_a_axv2(lv_16sc_t* out, con
for(number = 0; number < avx_iters; number++)
{
a = _mm256_load_si256((__m256i*)_in_a);
__builtin_prefetch(_in_a + 16);
__VOLK_PREFETCH(_in_a + 16);
b = _mm256_load_si256((__m256i*)_in_b);
__builtin_prefetch(_in_b + 16);
__VOLK_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.
@ -416,8 +416,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon(lv_16sc_t* out, const
{
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);
__VOLK_PREFETCH(a_ptr + 8);
__VOLK_PREFETCH(b_ptr + 8);
// multiply the real*real and imag*imag to get real result
// a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r
@ -482,8 +482,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_vma(lv_16sc_t* out, c
{
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);
__VOLK_PREFETCH(a_ptr + 8);
__VOLK_PREFETCH(b_ptr + 8);
tmp.val[0] = vmul_s16(a_val.val[0], b_val.val[0]);
tmp.val[1] = vmul_s16(a_val.val[1], b_val.val[0]);
@ -536,8 +536,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_optvma(lv_16sc_t* out
{
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);
__VOLK_PREFETCH(a_ptr + 8);
__VOLK_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]);

16
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h
View File

@ -145,7 +145,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul
{
// b[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
b = _mm_load_si128((__m128i*)_in_common); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a = _mm_load_si128((__m128i*)&(_in_a[n_vec][index*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
@ -239,7 +239,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul
{
// b[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
b = _mm_loadu_si128((__m128i*)_in_common); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a = _mm_loadu_si128((__m128i*)&(_in_a[n_vec][index*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
@ -332,7 +332,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_avx2(lv_16sc_t* resul
for(index = 0; index < sse_iters; index++)
{
b = _mm256_load_si256((__m256i*)_in_common);
__builtin_prefetch(_in_common + 16);
__VOLK_PREFETCH(_in_common + 16);
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a = _mm256_load_si256((__m256i*)&(_in_a[n_vec][index*8]));
@ -426,7 +426,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_avx2(lv_16sc_t* resul
for(index = 0; index < sse_iters; index++)
{
b = _mm256_loadu_si256((__m256i*)_in_common);
__builtin_prefetch(_in_common + 16);
__VOLK_PREFETCH(_in_common + 16);
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a = _mm256_loadu_si256((__m256i*)&(_in_a[n_vec][index*8]));
@ -518,11 +518,11 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon(lv_16sc_t* result,
for(index = 0; index < neon_iters; index++)
{
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);
__VOLK_PREFETCH(_in_common + 8);
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a_val = vld2_s16((int16_t*)&(_in_a[n_vec][index*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
//__builtin_prefetch(&_in_a[n_vec][index*4] + 8);
//__VOLK_PREFETCH(&_in_a[n_vec][index*4] + 8);
// multiply the real*real and imag*imag to get real result
// a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r
@ -604,7 +604,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_vma(lv_16sc_t* res
for(index = 0; index < neon_iters; index++)
{
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);
__VOLK_PREFETCH(_in_common + 8);
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a_val = vld2_s16((int16_t*)&(_in_a[n_vec][index*4]));
@ -684,7 +684,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_optvma(lv_16sc_t*
for(index = 0; index < neon_iters; index++)
{
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);
__VOLK_PREFETCH(_in_common + 8);
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a_val = vld2_s16((int16_t*)&(_in_a[n_vec][index*4]));

4
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_multiply_16ic.h
View File

@ -309,8 +309,8 @@ static inline void volk_gnsssdr_16ic_x2_multiply_16ic_neon(lv_16sc_t* out, const
{
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 + 4);
__builtin_prefetch(b_ptr + 4);
__VOLK_PREFETCH(a_ptr + 4);
__VOLK_PREFETCH(b_ptr + 4);
// multiply the real*real and imag*imag to get real result
// a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r

24
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h
View File

@ -248,7 +248,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3(lv_16sc_
//next two samples
_in_common += 2;
pa = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -428,7 +428,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3_reload(l
//next two samples
_in_common += 2;
pa = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -504,7 +504,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3_reload(l
//next two samples
_in_common += 2;
pa = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -651,7 +651,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
{
// Phase rotation on operand in_common starts here:
pa = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -672,7 +672,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
//next two samples
_in_common += 2;
pa = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -877,7 +877,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_avx2(lv_16sc_
//next two samples
_in_common += 2;
a = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 16);
__VOLK_PREFETCH(_in_common + 16);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -1087,7 +1087,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_avx2_reload(l
//next two samples
_in_common += 2;
a = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 16);
__VOLK_PREFETCH(_in_common + 16);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -1203,7 +1203,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_avx2_reload(l
//next two samples
_in_common += 2;
a = _mm_set_ps((float)(lv_cimag(_in_common[1])), (float)(lv_creal(_in_common[1])), (float)(lv_cimag(_in_common[0])), (float)(lv_creal(_in_common[0]))); // //load (2 byte imag, 2 byte real) x 2 into 128 bits reg
__builtin_prefetch(_in_common + 16);
__VOLK_PREFETCH(_in_common + 16);
//complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -1350,7 +1350,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon(lv_16sc_t*
{
/* load 4 complex numbers (int 16 bits each component) */
tmp16 = vld2_s16((int16_t*)_in_common);
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
_in_common += 4;
/* promote them to int 32 bits */
@ -1398,7 +1398,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon(lv_16sc_t*
for (n_vec = 0; n_vec < num_a_vectors; n_vec++)
{
a_val = vld2_s16((int16_t*)&(_in_a[n_vec][number*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
//__builtin_prefetch(&_in_a[n_vec][number*4] + 8);
//__VOLK_PREFETCH(&_in_a[n_vec][number*4] + 8);
// multiply the real*real and imag*imag to get real result
// a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r
@ -1533,7 +1533,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon_vma(lv_16s
{
/* load 4 complex numbers (int 16 bits each component) */
tmp16 = vld2_s16((int16_t*)_in_common);
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
_in_common += 4;
/* promote them to int 32 bits */
@ -1726,7 +1726,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon_optvma(lv_
{
/* load 4 complex numbers (int 16 bits each component) */
b_val = vld2_s16((int16_t*)_in_common);
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
_in_common += 4;
/* promote them to int 32 bits */

14
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_xn_resampler_16ic_xn.h
View File

@ -395,8 +395,8 @@ static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_a_avx(lv_16sc_t** resu
indexn = n0;
for(n = 0; n < avx_iters; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__builtin_prefetch(&local_code_chip_index[8], 1, 3);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__VOLK_PREFETCH_LOCALITY(&local_code_chip_index[8], 1, 3);
aux = _mm256_mul_ps(code_phase_step_chips_reg, indexn);
aux = _mm256_add_ps(aux, aux2);
// floor
@ -472,8 +472,8 @@ static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_u_avx(lv_16sc_t** resu
indexn = n0;
for(n = 0; n < avx_iters; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__builtin_prefetch(&local_code_chip_index[8], 1, 3);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__VOLK_PREFETCH_LOCALITY(&local_code_chip_index[8], 1, 3);
aux = _mm256_mul_ps(code_phase_step_chips_reg, indexn);
aux = _mm256_add_ps(aux, aux2);
// floor
@ -554,8 +554,8 @@ static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_neon(lv_16sc_t** resul
indexn = n0;
for(n = 0; n < neon_iters; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][4 * n + 3], 1, 0);
__builtin_prefetch(&local_code_chip_index[4]);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][4 * n + 3], 1, 0);
__VOLK_PREFETCH(&local_code_chip_index[4]);
aux = vmulq_f32(code_phase_step_chips_reg, indexn);
aux = vaddq_f32(aux, aux2);
@ -588,7 +588,7 @@ static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_neon(lv_16sc_t** resul
}
for(n = neon_iters * 4; n < num_points; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][n], 1, 0);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][n], 1, 0);
// resample code for current tap
local_code_chip_index_ = (int)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
//Take into account that in multitap correlators, the shifts can be negative!

10
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32f_sincos_32fc.h
View File

@ -100,7 +100,7 @@ static inline void volk_gnsssdr_32f_sincos_32fc_u_sse4_1(lv_32fc_t* out, const f
for(;number < quarterPoints; number++)
{
aVal = _mm_loadu_ps(aPtr);
__builtin_prefetch(aPtr + 8);
__VOLK_PREFETCH(aPtr + 8);
s = _mm_sub_ps(aVal, _mm_and_ps(_mm_mul_ps(aVal, ftwos), _mm_cmplt_ps(aVal, fzeroes)));
q = _mm_cvtps_epi32(_mm_floor_ps(_mm_mul_ps(s, m4pi)));
r = _mm_add_epi32(q, _mm_and_si128(q, ones));
@ -194,7 +194,7 @@ static inline void volk_gnsssdr_32f_sincos_32fc_a_sse4_1(lv_32fc_t* out, const f
for(;number < quarterPoints; number++)
{
aVal = _mm_load_ps(aPtr);
__builtin_prefetch(aPtr + 8);
__VOLK_PREFETCH(aPtr + 8);
s = _mm_sub_ps(aVal, _mm_and_ps(_mm_mul_ps(aVal, ftwos), _mm_cmplt_ps(aVal, fzeroes)));
q = _mm_cvtps_epi32(_mm_floor_ps(_mm_mul_ps(s, m4pi)));
r = _mm_add_epi32(q, _mm_and_si128(q, ones));
@ -292,7 +292,7 @@ static inline void volk_gnsssdr_32f_sincos_32fc_a_sse2(lv_32fc_t* out, const flo
for(;number < sse_iters; number++)
{
x = _mm_load_ps(aPtr);
__builtin_prefetch(aPtr + 8);
__VOLK_PREFETCH(aPtr + 8);
sign_bit_sin = x;
/* take the absolute value */
@ -445,7 +445,7 @@ static inline void volk_gnsssdr_32f_sincos_32fc_u_sse2(lv_32fc_t* out, const flo
for(;number < sse_iters; number++)
{
x = _mm_loadu_ps(aPtr);
__builtin_prefetch(aPtr + 8);
__VOLK_PREFETCH(aPtr + 8);
sign_bit_sin = x;
/* take the absolute value */
@ -640,7 +640,7 @@ static inline void volk_gnsssdr_32f_sincos_32fc_neon(lv_32fc_t* out, const float
for(;number < neon_iters; number++)
{
x = vld1q_f32(aPtr);
__builtin_prefetch(aPtr + 8);
__VOLK_PREFETCH(aPtr + 8);
sign_mask_sin = vcltq_f32(x, vdupq_n_f32(0));
x = vabsq_f32(x);

10
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_convert_16ic.h
View File

@ -84,7 +84,7 @@ static inline void volk_gnsssdr_32fc_convert_16ic_u_sse2(lv_16sc_t* outputVector
{
inputVal1 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4;
inputVal2 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4;
__builtin_prefetch(inputVectorPtr + 8);
__VOLK_PREFETCH(inputVectorPtr + 8);
// Clip
ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val);
@ -137,7 +137,7 @@ static inline void volk_gnsssdr_32fc_convert_16ic_u_sse(lv_16sc_t* outputVector,
{
inputVal1 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4;
inputVal2 = _mm_loadu_ps((float*)inputVectorPtr); inputVectorPtr += 4;
__builtin_prefetch(inputVectorPtr + 8);
__VOLK_PREFETCH(inputVectorPtr + 8);
// Clip
ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val);
@ -190,7 +190,7 @@ static inline void volk_gnsssdr_32fc_convert_16ic_a_sse2(lv_16sc_t* outputVector
{
inputVal1 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4;
inputVal2 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4;
__builtin_prefetch(inputVectorPtr + 8);
__VOLK_PREFETCH(inputVectorPtr + 8);
// Clip
ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val);
@ -241,7 +241,7 @@ static inline void volk_gnsssdr_32fc_convert_16ic_a_sse(lv_16sc_t* outputVector,
{
inputVal1 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4;
inputVal2 = _mm_load_ps((float*)inputVectorPtr); inputVectorPtr += 4;
__builtin_prefetch(inputVectorPtr + 8);
__VOLK_PREFETCH(inputVectorPtr + 8);
// Clip
ret1 = _mm_max_ps(_mm_min_ps(inputVal1, vmax_val), vmin_val);
@ -296,7 +296,7 @@ static inline void volk_gnsssdr_32fc_convert_16ic_neon(lv_16sc_t* outputVector,
{
a = vld1q_f32((const float32_t*)(inputVectorPtr)); inputVectorPtr += 4;
b = vld1q_f32((const float32_t*)(inputVectorPtr)); inputVectorPtr += 4;
__builtin_prefetch(inputVectorPtr + 8);
__VOLK_PREFETCH(inputVectorPtr + 8);
ret1 = vmaxq_f32(vminq_f32(a, max_val), min_val);
ret2 = vmaxq_f32(vminq_f32(b, max_val), min_val);

10
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_x2_rotator_dot_prod_32fc_xn.h
View File

@ -207,7 +207,7 @@ static inline void volk_gnsssdr_32fc_x2_rotator_dot_prod_32fc_xn_u_sse3(lv_32fc_
{
// Phase rotation on operand in_common starts here:
a = _mm_loadu_ps((float*)_in_common);
// __builtin_prefetch(_in_common + 4);
// __VOLK_PREFETCH(_in_common + 4);
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg);
tmp1 = _mm_mul_ps(a, yl);
@ -316,7 +316,7 @@ static inline void volk_gnsssdr_32fc_x2_rotator_dot_prod_32fc_xn_a_sse3(lv_32fc_
{
// Phase rotation on operand in_common starts here:
a = _mm_load_ps((float*)_in_common);
// __builtin_prefetch(_in_common + 4);
// __VOLK_PREFETCH(_in_common + 4);
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm_movehdup_ps(two_phase_acc_reg);
tmp1 = _mm_mul_ps(a, yl);
@ -435,7 +435,7 @@ static inline void volk_gnsssdr_32fc_x2_rotator_dot_prod_32fc_xn_u_avx(lv_32fc_t
{
// Phase rotation on operand in_common starts here:
a = _mm256_loadu_ps((float*)_in_common);
__builtin_prefetch(_in_common + 16);
__VOLK_PREFETCH(_in_common + 16);
yl = _mm256_moveldup_ps(four_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm256_movehdup_ps(four_phase_acc_reg);
tmp1 = _mm256_mul_ps(a, yl);
@ -562,7 +562,7 @@ static inline void volk_gnsssdr_32fc_x2_rotator_dot_prod_32fc_xn_a_avx(lv_32fc_t
{
// Phase rotation on operand in_common starts here:
a = _mm256_load_ps((float*)_in_common);
__builtin_prefetch(_in_common + 16);
__VOLK_PREFETCH(_in_common + 16);
yl = _mm256_moveldup_ps(four_phase_acc_reg); // Load yl with cr,cr,dr,dr
yh = _mm256_movehdup_ps(four_phase_acc_reg);
tmp1 = _mm256_mul_ps(a, yl);
@ -697,7 +697,7 @@ static inline void volk_gnsssdr_32fc_x2_rotator_dot_prod_32fc_xn_neon(lv_32fc_t*
{
/* load 4 complex numbers (float 32 bits each component) */
b_val = vld2q_f32((float32_t*)_in_common);
__builtin_prefetch(_in_common + 8);
__VOLK_PREFETCH(_in_common + 8);
_in_common += 4;
/* complex multiplication of four complex samples (float 32 bits each component) */

14
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_xn_resampler_32fc_xn.h
View File

@ -394,8 +394,8 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_a_avx(lv_32fc_t** resu
indexn = n0;
for(n = 0; n < avx_iters; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__builtin_prefetch(&local_code_chip_index[8], 1, 3);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__VOLK_PREFETCH_LOCALITY(&local_code_chip_index[8], 1, 3);
aux = _mm256_mul_ps(code_phase_step_chips_reg, indexn);
aux = _mm256_add_ps(aux, aux2);
// floor
@ -471,8 +471,8 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_u_avx(lv_32fc_t** resu
indexn = n0;
for(n = 0; n < avx_iters; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__builtin_prefetch(&local_code_chip_index[8], 1, 3);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][8 * n + 7], 1, 0);
__VOLK_PREFETCH_LOCALITY(&local_code_chip_index[8], 1, 3);
aux = _mm256_mul_ps(code_phase_step_chips_reg, indexn);
aux = _mm256_add_ps(aux, aux2);
// floor
@ -555,8 +555,8 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_neon(lv_32fc_t** resul
indexn = n0;
for(n = 0; n < neon_iters; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][4 * n + 3], 1, 0);
__builtin_prefetch(&local_code_chip_index[4]);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][4 * n + 3], 1, 0);
__VOLK_PREFETCH(&local_code_chip_index[4]);
aux = vmulq_f32(code_phase_step_chips_reg, indexn);
aux = vaddq_f32(aux, aux2);
@ -589,7 +589,7 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_neon(lv_32fc_t** resul
}
for(n = neon_iters * 4; n < num_points; n++)
{
__builtin_prefetch(&_result[current_correlator_tap][n], 1, 0);
__VOLK_PREFETCH_LOCALITY(&_result[current_correlator_tap][n], 1, 0);
// resample code for current tap
local_code_chip_index_ = (int)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
//Take into account that in multitap correlators, the shifts can be negative!

2
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_8ic_conjugate_8ic.h
View File

@ -303,7 +303,7 @@ static inline void volk_gnsssdr_8ic_conjugate_8ic_neon(lv_8sc_t* cVector, const
for (i = 0; i < sse_iters; ++i)
{
a_val = vld2_s8((const int8_t*)a);
__builtin_prefetch(a + 16);
__VOLK_PREFETCH(a + 16);
a_val.val[1] = vneg_s8(a_val.val[1]);
vst2_s8((int8_t*)c, a_val);
a += 8;

4
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_8ic_x2_dot_prod_8ic.h
View File

@ -457,8 +457,8 @@ static inline void volk_gnsssdr_8ic_x2_dot_prod_8ic_neon(lv_8sc_t* result, const
{
a_val = vld2_s8((const int8_t*)a);
b_val = vld2_s8((const int8_t*)b);
__builtin_prefetch(a + 16);
__builtin_prefetch(b + 16);
__VOLK_PREFETCH(a + 16);
__VOLK_PREFETCH(b + 16);
// multiply the real*real and imag*imag to get real result
tmp_real.val[0] = vmul_s8(a_val.val[0], b_val.val[0]);