volk: New volk kernel for conjugate dot products with unaligned buffers.

Note: need to convert this to new naming standard.

4 files changed, 160 insertions, 0 deletions

diff --git a/volk/lib/Makefile.am b/volk/lib/Makefile.am
index 7a355e86a..beb815e63 100644
--- a/volk/lib/Makefile.am
+++ b/volk/lib/Makefile.am
@@ -93,6 +93,7 @@ libvolk_qa_la_SOURCES = \
 	qa_32fc_index_max_aligned16.cc \
 	qa_32f_index_max_aligned16.cc \
 	qa_32fc_conjugate_dot_prod_aligned16.cc \
+	qa_32fc_conjugate_dot_prod_unaligned.cc \
 	qa_16s_permute_and_scalar_add_aligned16.cc \
 	qa_16s_branch_4_state_8_aligned16.cc \
 	qa_16s_max_star_horizontal_aligned16.cc \
@@ -195,6 +196,7 @@ noinst_HEADERS = \
 	qa_32fc_index_max_aligned16.h \
 	qa_32f_index_max_aligned16.h \
 	qa_32fc_conjugate_dot_prod_aligned16.h \
+	qa_32fc_conjugate_dot_prod_unaligned.h \
 	qa_16s_permute_and_scalar_add_aligned16.h \
 	qa_16s_branch_4_state_8_aligned16.h \
 	qa_16s_max_star_horizontal_aligned16.h \
diff --git a/volk/lib/qa_32fc_conjugate_dot_prod_unaligned.cc b/volk/lib/qa_32fc_conjugate_dot_prod_unaligned.cc
new file mode 100644
index 000000000..a0680bab6
--- /dev/null
+++ b/volk/lib/qa_32fc_conjugate_dot_prod_unaligned.cc
@@ -0,0 +1,138 @@
+#include <volk/volk.h>
+#include <qa_32fc_conjugate_dot_prod_unaligned.h>
+#include <stdlib.h>
+#include <math.h>
+#include <time.h>
+
+
+#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)
+
+//test for sse
+
+#if LV_HAVE_SSE && LV_HAVE_64
+
+static float uniform() {
+  return 2.0 * ((float) rand() / RAND_MAX - 0.5);	// uniformly (-1, 1)
+}
+
+static void
+random_floats (float *buf, unsigned n)
+{
+  for (unsigned i = 0; i < n; i++)
+    buf[i] = uniform () * 32767;
+}
+
+
+void qa_32fc_conjugate_dot_prod_unaligned::t1() {
+  const int vlen = 789743;
+  
+  volk_environment_init();
+  int ret;
+
+  std::complex<float>* input;
+  std::complex<float>* taps;
+  
+  std::complex<float>* result_generic;
+  std::complex<float>* result;
+
+  ret = posix_memalign((void**)&input, 16, vlen << 3);
+  ret = posix_memalign((void**)&taps, 16, vlen << 3);
+  ret = posix_memalign((void**)&result_generic, 16, 8);
+  ret = posix_memalign((void**)&result, 16, 8);
+  
+
+  result_generic[0] = std::complex<float>(0,0);
+  result[0] = std::complex<float>(0,0);
+
+  random_floats((float*)input, vlen * 2);
+  random_floats((float*)taps, vlen * 2);
+  
+  
+
+  volk_32fc_conjugate_dot_prod_unaligned_manual(result_generic, input, taps, vlen * 8,  "generic");
+
+  
+  volk_32fc_conjugate_dot_prod_unaligned_manual(result, input, taps, vlen * 8, "sse");
+
+  printf("32fc_conjugate_dot_prod_unaligned\n");
+  printf("generic: %f +i%f ... sse: %f +i%f\n", std::real(result_generic[0]), std::imag(result_generic[0]), std::real(result[0]), std::imag(result[0]));
+
+  assertcomplexEqual(result_generic[0], result[0], ERR_DELTA);
+
+  free(input);
+  free(taps);
+  free(result_generic);
+  free(result);
+  
+}
+
+
+#elif LV_HAVE_SSE && LV_HAVE_32
+
+static float uniform() {
+  return 2.0 * ((float) rand() / RAND_MAX - 0.5);	// uniformly (-1, 1)
+}
+
+static void
+random_floats (float *buf, unsigned n)
+{
+  for (unsigned i = 0; i < n; i++)
+    buf[i] = uniform () * 32767;
+}
+
+
+void qa_32fc_conjugate_dot_prod_unaligned::t1() {
+  const int vlen = 789743;
+  
+  volk_environment_init();
+  int ret;
+
+  std::complex<float>* input;
+  std::complex<float>* taps;
+  
+  std::complex<float>* result_generic;
+  std::complex<float>* result;
+
+  ret = posix_memalign((void**)&input, 16, vlen << 3);
+  ret = posix_memalign((void**)&taps, 16, vlen << 3);
+  ret = posix_memalign((void**)&result_generic, 16, 8);
+  ret = posix_memalign((void**)&result, 16, 8);
+  
+
+  result_generic[0] = std::complex<float>(0,0);
+  result[0] = std::complex<float>(0,0);
+
+  random_floats((float*)input, vlen * 2);
+  random_floats((float*)taps, vlen * 2);
+  
+  
+
+  volk_32fc_conjugate_dot_prod_unaligned_manual(result_generic, input, taps, vlen * 8,  "generic");
+
+  
+  volk_32fc_conjugate_dot_prod_unaligned_manual(result, input, taps, vlen * 8, "sse_32");
+
+  printf("32fc_conjugate_dot_prod_unaligned\n");
+  printf("generic: %f +i%f ... sse: %f +i%f\n", std::real(result_generic[0]), std::imag(result_generic[0]), std::real(result[0]), std::imag(result[0]));
+
+  assertcomplexEqual(result_generic[0], result[0], ERR_DELTA);
+
+  free(input);
+  free(taps);
+  free(result_generic);
+  free(result);
+  
+}
+
+
+#else
+
+void qa_32fc_conjugate_dot_prod_unaligned::t1() {
+  printf("sse not available... no test performed\n");
+}
+
+#endif /*LV_HAVE_SSE*/
diff --git a/volk/lib/qa_32fc_conjugate_dot_prod_unaligned.h b/volk/lib/qa_32fc_conjugate_dot_prod_unaligned.h
new file mode 100644
index 000000000..7aead53a1
--- /dev/null
+++ b/volk/lib/qa_32fc_conjugate_dot_prod_unaligned.h
@@ -0,0 +1,18 @@
+#ifndef INCLUDED_QA_32FC_CONJUGATE_DOT_PROD_UNALIGNED_H
+#define INCLUDED_QA_32FC_CONJUGATE_DOT_PROD_UNALIGNED_H
+
+#include <cppunit/extensions/HelperMacros.h>
+#include <cppunit/TestCase.h>
+
+class qa_32fc_conjugate_dot_prod_unaligned : public CppUnit::TestCase {
+
+  CPPUNIT_TEST_SUITE (qa_32fc_conjugate_dot_prod_unaligned);
+  CPPUNIT_TEST (t1);
+  CPPUNIT_TEST_SUITE_END ();
+
+ private:
+  void t1 ();
+};
+
+
+#endif /* INCLUDED_QA_32FC_CONJUGATE_DOT_PROD_UNALIGNED_H */
diff --git a/volk/lib/qa_volk.cc b/volk/lib/qa_volk.cc
index c3c27b69b..98d3e9728 100644
--- a/volk/lib/qa_volk.cc
+++ b/volk/lib/qa_volk.cc
@@ -34,6 +34,7 @@
 #include <qa_32fc_index_max_aligned16.h>
 #include <qa_32f_index_max_aligned16.h>
 #include <qa_32fc_conjugate_dot_prod_aligned16.h>
+#include <qa_32fc_conjugate_dot_prod_unaligned.h>
 #include <qa_16s_permute_and_scalar_add_aligned16.h>
 #include <qa_16s_branch_4_state_8_aligned16.h>
 #include <qa_16s_max_star_horizontal_aligned16.h>
@@ -127,6 +128,7 @@ qa_volk::suite()
   s->addTest(qa_32fc_index_max_aligned16::suite());
   s->addTest(qa_32f_index_max_aligned16::suite());
   s->addTest(qa_32fc_conjugate_dot_prod_aligned16::suite());
+  s->addTest(qa_32fc_conjugate_dot_prod_unaligned::suite());
   s->addTest(qa_16s_permute_and_scalar_add_aligned16::suite());
   s->addTest(qa_16s_branch_4_state_8_aligned16::suite());
   s->addTest(qa_16s_max_star_horizontal_aligned16::suite());