[hamradio-commits] [gnss-sdr] 71/236: prefetching data in the cache

[Carles Fernandez]

This is an automated email from the git hooks/post-receive script.

carles_fernandez-guest pushed a commit to branch next
in repository gnss-sdr.

commit aac79eb78a74cb27bbf40ead1a26d542edc3dc06
Author: Carles Fernandez <carles.fernandez at gmail.com>
Date:   Mon Mar 7 19:25:12 2016 +0100

    prefetching data in the cache
---
 .../kernels/volk_gnsssdr/volk_gnsssdr_16ic_s32fc_x2_rotator_16ic.h | 5 +++--
 .../volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h   | 7 +++----
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_s32fc_x2_rotator_16ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_s32fc_x2_rotator_16ic.h
index 6496b84..0858335 100644
--- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_s32fc_x2_rotator_16ic.h
+++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_s32fc_x2_rotator_16ic.h
@@ -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++)
         {
             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
             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
@@ -149,7 +150,7 @@ static inline void volk_gnsssdr_16ic_s32fc_x2_rotator_16ic_a_sse3(lv_16sc_t* out
             _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];
 
     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++)
         {
             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
             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
@@ -219,7 +221,6 @@ 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);
             //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
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 9e0218e..1578796 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
@@ -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:
             //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
+            __builtin_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
@@ -303,7 +304,6 @@ 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++)
         {
             // 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);
             //complex 32fc multiplication b=a*two_phase_acc_reg
@@ -378,7 +378,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
 
             a = _mm_or_si128(realcacc[n_vec], imagcacc[n_vec]);
 
-            _mm_storeu_si128((__m128i*)dotProductVector, a); // 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);
             for (int i = 0; i < 4; ++i)
                 {
@@ -390,10 +390,9 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
     free(realcacc);
     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];
 
-
     for(unsigned int n  = sse_iters * 4; n < num_points; n++)
         {
             tmp16 = in_common[n];

-- 
Alioth's /usr/local/bin/git-commit-notice on /srv/git.debian.org/git/pkg-hamradio/gnss-sdr.git