mirrors/gnss-sdr
1
0
Fork 0
mirror of https://github.com/gnss-sdr/gnss-sdr synced 2024-11-16 14:54:59 +00:00
Code Issues Releases Wiki Activity

Merge branch 'next' of https://github.com/gnss-sdr/gnss-sdr into next

Browse Source
This commit is contained in:
Carles Fernandez 2016-03-08 15:29:35 +01:00
commit fcae1e0ec5
5 changed files with 42 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_s32fc_x2_rotator_16ic.h
View File

@ -103,6 +103,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_a_sse3(lv_16sc_t* out
for(unsigned int number = 0; number < sse_iters; number++) for(unsigned int number = 0; number < sse_iters; number++)
{ {
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 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);
//complex 32fc multiplication b=a*two_phase_acc_reg //complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr 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 yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -149,7 +150,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_a_sse3(lv_16sc_t* out
_out += 4; _out += 4;
} }
_mm_storeu_ps((float*)two_phase_acc, two_phase_acc_reg); _mm_store_ps((float*)two_phase_acc, two_phase_acc_reg);
(*phase) = two_phase_acc[0]; (*phase) = two_phase_acc[0];
for (unsigned int i = sse_iters * 4; i < num_points; ++i) for (unsigned int i = sse_iters * 4; i < num_points; ++i)
@ -199,6 +200,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_u_sse3(lv_16sc_t* out
for(unsigned int number = 0; number < sse_iters; number++) for(unsigned int number = 0; number < sse_iters; number++)
{ {
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 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);
//complex 32fc multiplication b=a*two_phase_acc_reg //complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr 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 yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -219,7 +221,6 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_u_sse3(lv_16sc_t* out
//next two samples //next two samples
_in += 2; _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 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);
//complex 32fc multiplication b=a*two_phase_acc_reg //complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr 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 yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di

50
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h
View File

@ -92,26 +92,22 @@ 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]; __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
//todo dyn mem reg
__m128i* realcacc; __m128i* realcacc;
__m128i* imagcacc; __m128i* imagcacc;
realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 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 imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
__m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, results; __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag;
mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); 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); mask_real = _mm_set_epi8(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++) for(unsigned int number = 0; number < sse_iters; number++)
{ {
//std::complex<T> memory structure: real part -> reinterpret_cast<cv T*>(a)[2*i] // b[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
//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]
b = _mm_load_si128((__m128i*)_in_common); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg 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);
for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
{ {
a = _mm_load_si128((__m128i*)&(_in_a[n_vec][number*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg a = _mm_load_si128((__m128i*)&(_in_a[n_vec][number*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
@ -121,13 +117,13 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul
c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
real = _mm_subs_epi16(c, c_sr); real = _mm_subs_epi16(c, c_sr);
b_sl = _mm_slli_si128(b, 2); // b3.r, b2.i .... c_sr = _mm_slli_si128(b, 2); // b3.r, b2.i ....
a_sl = _mm_slli_si128(a, 2); // a3.r, a2.i .... c = _mm_mullo_epi16(a, c_sr); // a3.i*b3.r, ....
imag1 = _mm_mullo_epi16(a, b_sl); // a3.i*b3.r, .... c_sr = _mm_slli_si128(a, 2); // a3.r, a2.i ....
imag2 = _mm_mullo_epi16(b, a_sl); // b3.i*a3.r, .... imag = _mm_mullo_epi16(b, c_sr); // b3.i*a3.r, ....
imag = _mm_adds_epi16(imag1, imag2); imag = _mm_adds_epi16(c, imag);
realcacc[n_vec] = _mm_adds_epi16(realcacc[n_vec], real); realcacc[n_vec] = _mm_adds_epi16(realcacc[n_vec], real);
imagcacc[n_vec] = _mm_adds_epi16(imagcacc[n_vec], imag); imagcacc[n_vec] = _mm_adds_epi16(imagcacc[n_vec], imag);
@ -141,9 +137,9 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul
realcacc[n_vec] = _mm_and_si128(realcacc[n_vec], mask_real); realcacc[n_vec] = _mm_and_si128(realcacc[n_vec], mask_real);
imagcacc[n_vec] = _mm_and_si128(imagcacc[n_vec], mask_imag); imagcacc[n_vec] = _mm_and_si128(imagcacc[n_vec], mask_imag);
results = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]); a = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]);
_mm_store_si128((__m128i*)dotProductVector, results); // Store the results back into the dot product vector _mm_store_si128((__m128i*)dotProductVector, a); // Store the results back into the dot product vector
dotProduct = lv_cmake(0,0); dotProduct = lv_cmake(0,0);
for (int i = 0; i < 4; ++i) for (int i = 0; i < 4; ++i)
{ {
@ -166,7 +162,6 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul
sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp))); sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp)));
} }
} }
} }
#endif /* LV_HAVE_SSE2 */ #endif /* LV_HAVE_SSE2 */
@ -195,26 +190,22 @@ 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]; __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
//todo dyn mem reg
__m128i* realcacc; __m128i* realcacc;
__m128i* imagcacc; __m128i* imagcacc;
realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 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 imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
__m128i a,b,c, c_sr, mask_imag, mask_real, real, imag, imag1,imag2, b_sl, a_sl, results; __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag;
mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); 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); mask_real = _mm_set_epi8(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++) for(unsigned int number = 0; number < sse_iters; number++)
{ {
//std::complex<T> memory structure: real part -> reinterpret_cast<cv T*>(a)[2*i] // b[127:0]=[a3.i,a3.r,a2.i,a2.r,a1.i,a1.r,a0.i,a0.r]
//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]
b = _mm_loadu_si128((__m128i*)_in_common); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg 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);
for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
{ {
a = _mm_loadu_si128((__m128i*)&(_in_a[n_vec][number*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg a = _mm_loadu_si128((__m128i*)&(_in_a[n_vec][number*4])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
@ -224,13 +215,13 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul
c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst. c_sr = _mm_srli_si128(c, 2); // Shift a right by imm8 bytes while shifting in zeros, and store the results in dst.
real = _mm_subs_epi16(c, c_sr); real = _mm_subs_epi16(c, c_sr);
b_sl = _mm_slli_si128(b, 2); // b3.r, b2.i .... c_sr = _mm_slli_si128(b, 2); // b3.r, b2.i ....
a_sl = _mm_slli_si128(a, 2); // a3.r, a2.i .... c = _mm_mullo_epi16(a, c_sr); // a3.i*b3.r, ....
imag1 = _mm_mullo_epi16(a, b_sl); // a3.i*b3.r, .... c_sr = _mm_slli_si128(a, 2); // a3.r, a2.i ....
imag2 = _mm_mullo_epi16(b, a_sl); // b3.i*a3.r, .... imag = _mm_mullo_epi16(b, c_sr); // b3.i*a3.r, ....
imag = _mm_adds_epi16(imag1, imag2); imag = _mm_adds_epi16(c, imag);
realcacc[n_vec] = _mm_adds_epi16(realcacc[n_vec], real); realcacc[n_vec] = _mm_adds_epi16(realcacc[n_vec], real);
imagcacc[n_vec] = _mm_adds_epi16(imagcacc[n_vec], imag); imagcacc[n_vec] = _mm_adds_epi16(imagcacc[n_vec], imag);
@ -244,9 +235,9 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul
realcacc[n_vec] = _mm_and_si128(realcacc[n_vec], mask_real); realcacc[n_vec] = _mm_and_si128(realcacc[n_vec], mask_real);
imagcacc[n_vec] = _mm_and_si128(imagcacc[n_vec], mask_imag); imagcacc[n_vec] = _mm_and_si128(imagcacc[n_vec], mask_imag);
results = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]); a = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]);
_mm_storeu_si128((__m128i*)dotProductVector, results); // Store the results back into the dot product vector _mm_store_si128((__m128i*)dotProductVector, a); // Store the results back into the dot product vector
dotProduct = lv_cmake(0,0); dotProduct = lv_cmake(0,0);
for (int i = 0; i < 4; ++i) for (int i = 0; i < 4; ++i)
{ {
@ -269,7 +260,6 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul
sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp))); sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp)));
} }
} }
} }
#endif /* LV_HAVE_SSE2 */ #endif /* LV_HAVE_SSE2 */

5
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic.h
View File

@ -85,6 +85,9 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_a_sse2(lv_16sc_t* r
#endif // SSE2 #endif // SSE2
#define WORKAROUND 1
#ifdef WORKAROUND
#ifdef LV_HAVE_SSE2 #ifdef 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) 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)
@ -105,7 +108,7 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_sse2(lv_16sc_t* r
} }
volk_gnsssdr_free(in_a); volk_gnsssdr_free(in_a);
} }
#endif
#endif // SSE2 #endif // SSE2
#ifdef LV_HAVE_NEON #ifdef LV_HAVE_NEON

27
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h
View File

@ -115,7 +115,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3(lv_16sc_
realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 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 imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
__m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, results; __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl;
mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); 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); mask_real = _mm_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255);
@ -140,6 +140,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3(lv_16sc_
// Phase rotation on operand in_common starts here: // Phase rotation on operand in_common starts here:
//printf("generic phase %i: %f,%f\n", n*4,lv_creal(*phase),lv_cimag(*phase)); //printf("generic phase %i: %f,%f\n", n*4,lv_creal(*phase),lv_cimag(*phase));
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 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);
//complex 32fc multiplication b=a*two_phase_acc_reg //complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr 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 yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -200,8 +201,8 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3(lv_16sc_
imag = _mm_adds_epi16(imag1, imag2); imag = _mm_adds_epi16(imag1, imag2);
realcacc[n_vec] = _mm_adds_epi16 (realcacc[n_vec], real); realcacc[n_vec] = _mm_adds_epi16(realcacc[n_vec], real);
imagcacc[n_vec] = _mm_adds_epi16 (imagcacc[n_vec], imag); imagcacc[n_vec] = _mm_adds_epi16(imagcacc[n_vec], imag);
} }
} }
@ -210,9 +211,9 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3(lv_16sc_
realcacc[n_vec] = _mm_and_si128(realcacc[n_vec], mask_real); realcacc[n_vec] = _mm_and_si128(realcacc[n_vec], mask_real);
imagcacc[n_vec] = _mm_and_si128(imagcacc[n_vec], mask_imag); imagcacc[n_vec] = _mm_and_si128(imagcacc[n_vec], mask_imag);
results = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]); a = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]);
_mm_store_si128((__m128i*)dotProductVector, results); // Store the results back into the dot product vector _mm_store_si128((__m128i*)dotProductVector, a); // Store the results back into the dot product vector
dotProduct = lv_cmake(0,0); dotProduct = lv_cmake(0,0);
for (int i = 0; i < 4; ++i) for (int i = 0; i < 4; ++i)
{ {
@ -280,7 +281,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0 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 imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
__m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl, results; __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl;
mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); 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); mask_real = _mm_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255);
@ -303,8 +304,8 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
for(unsigned int number = 0; number < sse_iters; number++) for(unsigned int number = 0; number < sse_iters; number++)
{ {
// Phase rotation on operand in_common starts here: // 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 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);
//complex 32fc multiplication b=a*two_phase_acc_reg //complex 32fc multiplication b=a*two_phase_acc_reg
yl = _mm_moveldup_ps(two_phase_acc_reg); // Load yl with cr,cr,dr,dr 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 yh = _mm_movehdup_ps(two_phase_acc_reg); // Load yh with ci,ci,di,di
@ -372,12 +373,12 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
{ {
realcacc[n_vec] = _mm_and_si128 (realcacc[n_vec], mask_real); realcacc[n_vec] = _mm_and_si128(realcacc[n_vec], mask_real);
imagcacc[n_vec] = _mm_and_si128 (imagcacc[n_vec], mask_imag); imagcacc[n_vec] = _mm_and_si128(imagcacc[n_vec], mask_imag);
results = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]); a = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]);
_mm_storeu_si128((__m128i*)dotProductVector, results); // Store the results back into the dot product vector _mm_store_si128((__m128i*)dotProductVector, a); // Store the results back into the dot product vector
dotProduct = lv_cmake(0,0); dotProduct = lv_cmake(0,0);
for (int i = 0; i < 4; ++i) for (int i = 0; i < 4; ++i)
{ {
@ -389,7 +390,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
free(realcacc); free(realcacc);
free(imagcacc); free(imagcacc);
_mm_storeu_ps((float*)two_phase_acc, two_phase_acc_reg); _mm_store_ps((float*)two_phase_acc, two_phase_acc_reg);
(*phase) = two_phase_acc[0]; (*phase) = two_phase_acc[0];
for(unsigned int n = sse_iters * 4; n < num_points; n++) for(unsigned int n = sse_iters * 4; n < num_points; n++)
@ -527,7 +528,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon(lv_16sc_t*
for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) for (int 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 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); //__builtin_prefetch(&_in_a[n_vec][number*4] + 8);
// multiply the real*real and imag*imag to get real result // multiply the real*real and imag*imag to get real result
// a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r // a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r

2
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dotprodxnpuppet_16ic.h
View File

@ -116,7 +116,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dotprodxnpuppet_16ic_u_sse3(lv_1
for(unsigned int n = 0; n < num_a_vectors; n++) 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()); in_a[n] = (lv_16sc_t*)volk_gnsssdr_malloc(sizeof(lv_16sc_t) * num_points, volk_gnsssdr_get_alignment());
memcpy(in_a[n], in, sizeof(lv_16sc_t) * num_points); memcpy((lv_16sc_t*)in_a[n], (lv_16sc_t*)in, sizeof(lv_16sc_t) * num_points);
} }
volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(result, local_code, phase_inc[0], phase, (const lv_16sc_t**) in_a, num_a_vectors, num_points); volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(result, local_code, phase_inc[0], phase, (const lv_16sc_t**) in_a, num_a_vectors, num_points);