1
0
mirror of https://github.com/gnss-sdr/gnss-sdr synced 2024-12-14 12:10:34 +00:00

Still some bugs to fix in 16sc dot product. All fixed now.

This commit is contained in:
Javier Arribas 2016-01-18 10:43:09 +01:00
parent fd2af02aec
commit 4ba75a3fbe

View File

@ -40,7 +40,6 @@
#include <volk_gnsssdr/volk_gnsssdr_common.h> #include <volk_gnsssdr/volk_gnsssdr_common.h>
#include <volk_gnsssdr/volk_gnsssdr_complex.h> #include <volk_gnsssdr/volk_gnsssdr_complex.h>
#include <volk_gnsssdr/saturation_arithmetic.h> #include <volk_gnsssdr/saturation_arithmetic.h>
//#include <stdio.h>
#ifdef LV_HAVE_GENERIC #ifdef LV_HAVE_GENERIC
@ -56,11 +55,9 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_generic(lv_16sc_t* result,
result[0] = lv_cmake((int16_t)0, (int16_t)0); result[0] = lv_cmake((int16_t)0, (int16_t)0);
for (unsigned int n = 0; n < num_points; n++) for (unsigned int n = 0; n < num_points; n++)
{ {
//r*a.r - i*a.i, i*a.r + r*a.i
lv_16sc_t tmp = in_a[n] * in_b[n]; lv_16sc_t tmp = in_a[n] * in_b[n];
result[0] += lv_cmake(sat_adds16i(lv_creal(tmp), lv_creal(tmp)), sat_adds16i(lv_cimag(tmp), lv_cimag(tmp) )); result[0] = lv_cmake(sat_adds16i(lv_creal(result[0]), lv_creal(tmp)), sat_adds16i(lv_cimag(result[0]), lv_cimag(tmp) ));
} }
//printf("generic result = %i,%i", lv_creal(result[0]),lv_cimag(result[0]));
} }
#endif /*LV_HAVE_GENERIC*/ #endif /*LV_HAVE_GENERIC*/
@ -70,74 +67,73 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_generic(lv_16sc_t* result,
#include <emmintrin.h> #include <emmintrin.h>
static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_a_sse2(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points) static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_a_sse2(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);
lv_16sc_t dotProduct = lv_cmake((int16_t)0, (int16_t)0); const unsigned int sse_iters = num_points / 4;
const unsigned int sse_iters = num_points / 4; const lv_16sc_t* _in_a = in_a;
const lv_16sc_t* _in_b = in_b;
lv_16sc_t* _out = out;
const lv_16sc_t* _in_a = in_a; if (sse_iters > 0)
const lv_16sc_t* _in_b = in_b; {
lv_16sc_t* _out = out; __m128i a,b,c, c_sr, mask_imag, mask_real, real, imag, imag1,imag2, b_sl, a_sl, realcacc, imagcacc, result;
if (sse_iters > 0) realcacc = _mm_setzero_si128();
{ imagcacc = _mm_setzero_si128();
__m128i a,b,c, c_sr, mask_imag, mask_real, real, imag, imag1,imag2, b_sl, a_sl, realcacc, imagcacc, result;
realcacc = _mm_setzero_si128(); mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0);
imagcacc = _mm_setzero_si128(); mask_real = _mm_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255);
mask_imag = _mm_set_epi8(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); for(unsigned int number = 0; number < sse_iters; number++)
mask_real = _mm_set_epi8(0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255); {
//std::complex<T> memory structure: real part -> reinterpret_cast<cv T*>(a)[2*i]
//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
b = _mm_loadu_si128((__m128i*)_in_b);
c = _mm_mullo_epi16 (a, b); // a3.i*b3.i, a3.r*b3.r, ....
for(unsigned int number = 0; number < sse_iters; number++) 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);
//std::complex<T> memory structure: real part -> reinterpret_cast<cv T*>(a)[2*i]
//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
b = _mm_loadu_si128((__m128i*)_in_b);
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. b_sl = _mm_slli_si128(b, 2); // b3.r, b2.i ....
real = _mm_subs_epi16 (c,c_sr); a_sl = _mm_slli_si128(a, 2); // a3.r, a2.i ....
b_sl = _mm_slli_si128(b, 2); // b3.r, b2.i .... imag1 = _mm_mullo_epi16(a, b_sl); // a3.i*b3.r, ....
a_sl = _mm_slli_si128(a, 2); // a3.r, a2.i .... imag2 = _mm_mullo_epi16(b, a_sl); // b3.i*a3.r, ....
imag1 = _mm_mullo_epi16(a, b_sl); // a3.i*b3.r, .... imag = _mm_adds_epi16(imag1, imag2); //with saturation aritmetic!
imag2 = _mm_mullo_epi16(b, a_sl); // b3.i*a3.r, ....
imag = _mm_adds_epi16(imag1, imag2); //with saturation aritmetic! realcacc = _mm_adds_epi16 (realcacc, real);
imagcacc = _mm_adds_epi16 (imagcacc, imag);
realcacc = _mm_adds_epi16 (realcacc, real); _in_a += 4;
imagcacc = _mm_adds_epi16 (imagcacc, imag); _in_b += 4;
}
_in_a += 4; realcacc = _mm_and_si128 (realcacc, mask_real);
_in_b += 4; imagcacc = _mm_and_si128 (imagcacc, mask_imag);
}
realcacc = _mm_and_si128 (realcacc, mask_real); result = _mm_or_si128 (realcacc, imagcacc);
imagcacc = _mm_and_si128 (imagcacc, mask_imag);
result = _mm_or_si128 (realcacc, imagcacc); __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
__VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4]; _mm_storeu_si128((__m128i*)dotProductVector,result); // Store the results back into the dot product vector
_mm_storeu_si128((__m128i*)dotProductVector,result); // Store the results back into the dot product vector 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])));
}
}
for (int i = 0; i < 4; ++i) for (unsigned int i = 0; i < (num_points % 4); ++i)
{ {
dotProduct += lv_cmake(sat_adds16i(lv_creal(dotProductVector[i]), lv_creal(dotProductVector[i])), sat_adds16i(lv_cimag(dotProductVector[i]), lv_cimag(dotProductVector[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)));
} }
for (unsigned int i = sse_iters * 4; i < num_points; ++i) *_out = dotProduct;
{
lv_16sc_t tmp = (*_in_a++) * (*_in_b++);
dotProduct += lv_cmake( sat_adds16i(lv_creal(tmp), lv_creal(tmp)), sat_adds16i(lv_cimag(tmp), lv_cimag(tmp)));
}
*_out = dotProduct;
} }
#endif /* LV_HAVE_SSE2 */ #endif /* LV_HAVE_SSE2 */