summaryrefslogtreecommitdiff
path: root/volk/lib/qa_8sc_multiply_conjugate_32fc_aligned16.cc
diff options
context:
space:
mode:
authorEric Blossom2010-12-21 13:31:58 -0800
committerEric Blossom2010-12-21 13:31:58 -0800
commit75b650ef2506fe5c607517a3d5188b0705f60fde (patch)
tree99cb6fd6adc3e91edc832cbaee236d2fe30fa3fd /volk/lib/qa_8sc_multiply_conjugate_32fc_aligned16.cc
parent1984aab345857296564e174395cf3af1375e1469 (diff)
parented78ba5d9999bbe50507373a1aa2877ef0da64c6 (diff)
downloadgnuradio-75b650ef2506fe5c607517a3d5188b0705f60fde.tar.gz
gnuradio-75b650ef2506fe5c607517a3d5188b0705f60fde.tar.bz2
gnuradio-75b650ef2506fe5c607517a3d5188b0705f60fde.zip
Merge branch 'next' into guile.
Passes distcheck. * next: (32 commits) volk: Fix for popcnt's 64/32-bit issues. Using a copy of config.guess and config.sub instead of sym links. Including time header to qa files. Changed python env variable to more globally usable version. gr_uhd: Quick fix for make distcheck failures if UHD is not installed. PFB resampler: fix it this way to avoid the signed/unsigned warning. PFB resampler: fixes bug where filter could be looking past the number of inputs. resampler PFB: Sets relative rate when rate is changed. Create method to set rate on pfb_arb_resamp after it has been created. Allow it to be called from GRC. volk: fix for running 32-bit OS on 64-bit processor. System is correctly identified as 32-bit and compiles with the correct flags. volk: changing the path variables again. This works on my various systems tested. Using abs_ path names failed on Ubuntu 8.04 32-bit. volk: May be a hack, but it was required for my 32-bit Fedora 13 to work. volk: Removing unnecessary shell script; last commit takes care if its functions. volk: Fixing build system to handle making volk_mktables, volk_tables.h, and volk_config.h instead of a standalone shell script. volk: readding 16sc_magnitude_32f_sse with fix for SSE hadd_ps error. volk: Adding a few more generic-only test cases. volk: adding generic QA test for 16sc_magnitude_32f. volk: modified the configure scripts to output which architectures it will be building based on the configure tests. uhd: update notes in grc blocks for addressing scheme volk: Removing erroneous SSE function that actually usese an SSE3 intrin (mm_hadd_ps). ...
Diffstat (limited to 'volk/lib/qa_8sc_multiply_conjugate_32fc_aligned16.cc')
-rw-r--r--volk/lib/qa_8sc_multiply_conjugate_32fc_aligned16.cc87
1 files changed, 87 insertions, 0 deletions
diff --git a/volk/lib/qa_8sc_multiply_conjugate_32fc_aligned16.cc b/volk/lib/qa_8sc_multiply_conjugate_32fc_aligned16.cc
new file mode 100644
index 000000000..4c707446e
--- /dev/null
+++ b/volk/lib/qa_8sc_multiply_conjugate_32fc_aligned16.cc
@@ -0,0 +1,87 @@
+#include <volk/volk_runtime.h>
+#include <volk/volk.h>
+#include <qa_8sc_multiply_conjugate_32fc_aligned16.h>
+#include <stdlib.h>
+#include <math.h>
+#include <ctime>
+
+#define assertcomplexEqual(expected, actual, delta) \
+ CPPUNIT_ASSERT_DOUBLES_EQUAL (std::real(expected), std::real(actual), fabs(std::real(expected)) * delta); \
+ CPPUNIT_ASSERT_DOUBLES_EQUAL (std::imag(expected), std::imag(actual), fabs(std::imag(expected))* delta);
+
+#define ERR_DELTA (1e-4)
+
+#ifndef LV_HAVE_SSE4_1
+
+void qa_8sc_multiply_conjugate_32fc_aligned16::t1() {
+ printf("sse4.1 not available... no test performed\n");
+}
+
+#else
+
+void qa_8sc_multiply_conjugate_32fc_aligned16::t1() {
+
+
+ volk_runtime_init();
+
+ const int vlen = 2046;
+ const int ITERS = 100000;
+
+ volk_environment_init();
+ int ret;
+ clock_t start, end;
+ double total;
+ std::complex<int8_t>* input;
+ std::complex<int8_t>* taps;
+
+ std::complex<float>* result_generic;
+ std::complex<float>* result_sse4_1;
+ int i;
+ int8_t* inputInt8_T;
+ int8_t* tapsInt8_T;
+
+ ret = posix_memalign((void**)&input, 16, vlen*2*sizeof(int8_t));
+ ret = posix_memalign((void**)&taps, 16, vlen*2*sizeof(int8_t));
+ ret = posix_memalign((void**)&result_generic, 16, vlen*2*sizeof(float));
+ ret = posix_memalign((void**)&result_sse4_1, 16, vlen*2*sizeof(float));
+
+
+ inputInt8_T = (int8_t*)input;
+ tapsInt8_T = (int8_t*)taps;
+ for(int i = 0; i < vlen*2; ++i) {
+ inputInt8_T[i] =((int8_t)((((float) (rand() - (RAND_MAX/2))) / static_cast<float>((RAND_MAX/2))) * 128.0));
+ tapsInt8_T[i] =((int8_t)((((float) (rand() - (RAND_MAX/2))) / static_cast<float>((RAND_MAX/2))) * 128.0));
+ }
+
+ printf("8sc_multiply_conjugate_32fc_aligned16\n");
+
+ start = clock();
+ for(int count = 0; count < ITERS; ++count) {
+ volk_8sc_multiply_conjugate_32fc_aligned16_manual(result_generic, (const std::complex<int8_t>*)input, (const std::complex<int8_t>*)taps, 32768.0, vlen, "generic");
+ }
+ end = clock();
+ total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+ printf("generic_time: %f\n", total);
+
+
+ start = clock();
+ for(int count = 0; count < ITERS; ++count) {
+ get_volk_runtime()->volk_8sc_multiply_conjugate_32fc_aligned16(result_sse4_1, (const std::complex<int8_t>*)input, (const std::complex<int8_t>*)taps, 32768.0, vlen);
+ }
+ end = clock();
+ total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+ printf("sse4_1_time: %f\n", total);
+
+ for(i = 0; i < vlen; i++){
+ //printf("%d %d+%di %d+%di -> %e+%ei %e+%ei\n", i, std::real(input[i]), std::imag(input[i]), std::real(taps[i]), std::imag(taps[i]), std::real(result_generic[i]), std::imag(result_generic[i]), std::real(result_sse4_1[i]), std::imag(result_sse4_1[i]));
+ assertcomplexEqual(result_generic[i], result_sse4_1[i], ERR_DELTA);
+ }
+
+ free(input);
+ free(taps);
+ free(result_generic);
+ free(result_sse4_1);
+
+}
+
+#endif /*LV_HAVE_SSE4_1*/