Volk: A bunch of new ORC routines plus tests.

Also fixed a typo in the generic version of 16sc_magnitude_16s_a16.

6 files changed, 65 insertions, 0 deletions

diff --git a/volk/lib/qa_16sc_magnitude_16s_aligned16.cc b/volk/lib/qa_16sc_magnitude_16s_aligned16.cc
index b14610757..c8f13ff84 100644
--- a/volk/lib/qa_16sc_magnitude_16s_aligned16.cc
+++ b/volk/lib/qa_16sc_magnitude_16s_aligned16.cc
@@ -23,6 +23,7 @@ void qa_16sc_magnitude_16s_aligned16::t1() {
   std::complex<int16_t> input0[vlen] __attribute__ ((aligned (16)));
   
   int16_t output_generic[vlen] __attribute__ ((aligned (16)));
+  int16_t output_orc[vlen] __attribute__ ((aligned (16)));
   int16_t output_sse[vlen] __attribute__ ((aligned (16)));
   int16_t output_sse3[vlen] __attribute__ ((aligned (16)));
 
@@ -41,6 +42,13 @@ void qa_16sc_magnitude_16s_aligned16::t1() {
   printf("generic_time: %f\n", total);
   start = clock();
   for(int count = 0; count < ITERS; ++count) {
+    volk_16sc_magnitude_16s_aligned16_manual(output_orc, input0, vlen, "orc");
+  }
+  end = clock();
+  total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+  printf("orc_time: %f\n", total);
+  start = clock();
+  for(int count = 0; count < ITERS; ++count) {
     volk_16sc_magnitude_16s_aligned16_manual(output_sse, input0, vlen, "sse");
   }
   end = clock();
@@ -64,6 +72,7 @@ void qa_16sc_magnitude_16s_aligned16::t1() {
     //printf("%d...%d\n", output0[i], output01[i]);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse[i], 1.1);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse3[i], 1.1);
+    CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_orc[i], 1.1);
   }
 }
 
diff --git a/volk/lib/qa_16sc_magnitude_32f_aligned16.cc b/volk/lib/qa_16sc_magnitude_32f_aligned16.cc
index 2c9e48f6e..e7178863c 100644
--- a/volk/lib/qa_16sc_magnitude_32f_aligned16.cc
+++ b/volk/lib/qa_16sc_magnitude_32f_aligned16.cc
@@ -15,6 +15,7 @@ void qa_16sc_magnitude_32f_aligned16::t1() {
   std::complex<int16_t> input0[vlen] __attribute__ ((aligned (16)));
   
   float output_generic[vlen] __attribute__ ((aligned (16)));
+  float output_orc[vlen] __attribute__ ((aligned (16)));
   float output_known[vlen] __attribute__ ((aligned (16)));
 
   int16_t* inputLoad = (int16_t*)input0;
@@ -37,6 +38,14 @@ void qa_16sc_magnitude_32f_aligned16::t1() {
   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) {
+    volk_16sc_magnitude_32f_aligned16_manual(output_orc, input0, scale, vlen, "orc");
+  }
+  end = clock();
+  total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+  printf("orc_time: %f\n", total);
 
   /*
   for(int i = 0; i < 100; ++i) {
@@ -48,6 +57,7 @@ void qa_16sc_magnitude_32f_aligned16::t1() {
   for(int i = 0; i < vlen; ++i) {
     //printf("%d...%d\n", output0[i], output01[i]);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_known[i], fabs(output_generic[i])*1e-4);
+    CPPUNIT_ASSERT_DOUBLES_EQUAL(output_orc[i], output_known[i], fabs(output_generic[i])*1e-4);
   }
 }
 
@@ -63,6 +73,7 @@ void qa_16sc_magnitude_32f_aligned16::t1() {
   std::complex<int16_t> input0[vlen] __attribute__ ((aligned (16)));
   
   float output_generic[vlen] __attribute__ ((aligned (16)));
+  float output_orc[vlen] __attribute__ ((aligned (16)));
   float output_sse[vlen] __attribute__ ((aligned (16)));
   float output_sse3[vlen] __attribute__ ((aligned (16)));
 
@@ -81,6 +92,14 @@ void qa_16sc_magnitude_32f_aligned16::t1() {
   printf("generic_time: %f\n", total);
   start = clock();
   for(int count = 0; count < ITERS; ++count) {
+    volk_16sc_magnitude_32f_aligned16_manual(output_orc, input0, 32768.0, vlen, "orc");
+  }
+  end = clock();
+  total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+  printf("orc_time: %f\n", total);
+
+  start = clock();
+  for(int count = 0; count < ITERS; ++count) {
     volk_16sc_magnitude_32f_aligned16_manual(output_sse, input0, 32768.0, vlen, "sse");
   }
   end = clock();
@@ -104,6 +123,7 @@ void qa_16sc_magnitude_32f_aligned16::t1() {
     //printf("%d...%d\n", output0[i], output01[i]);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse[i], fabs(output_generic[i])*1e-4);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse3[i], fabs(output_generic[i])*1e-4);
+    CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_orc[i], fabs(output_generic[i])*1e-4);
   }
 }
 
diff --git a/volk/lib/qa_32f_divide_aligned16.cc b/volk/lib/qa_32f_divide_aligned16.cc
index f104e0443..b2c2ecf9a 100644
--- a/volk/lib/qa_32f_divide_aligned16.cc
+++ b/volk/lib/qa_32f_divide_aligned16.cc
@@ -88,6 +88,7 @@ void qa_32f_divide_aligned16::t1() {
   
   float output0[vlen] __attribute__ ((aligned (16)));
   float output01[vlen] __attribute__ ((aligned (16)));
+  float output02[vlen] __attribute__ ((aligned (16)));
 
   for(int i = 0; i < vlen; ++i) {   
     input0[i] = ((float) (rand() - (RAND_MAX/2))) / static_cast<float>((RAND_MAX/2));
@@ -104,6 +105,13 @@ void qa_32f_divide_aligned16::t1() {
   printf("generic_time: %f\n", total);
   start = clock();
   for(int count = 0; count < ITERS; ++count) {
+    volk_32f_divide_aligned16_manual(output02, input0, input1, vlen, "orc");
+  }
+  end = clock();
+  total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+  printf("orc_time: %f\n", total);
+  start = clock();
+  for(int count = 0; count < ITERS; ++count) {
     volk_32f_divide_aligned16_manual(output01, input0, input1, vlen, "sse");
   }
   end = clock();
@@ -117,6 +125,7 @@ void qa_32f_divide_aligned16::t1() {
   for(int i = 0; i < vlen; ++i) {
     //printf("%d...%d\n", output0[i], output01[i]);
     CPPUNIT_ASSERT_EQUAL(output0[i], output01[i]);
+    CPPUNIT_ASSERT_EQUAL(output0[i], output02[i]);
   }
 }
 
diff --git a/volk/lib/qa_32fc_magnitude_16s_aligned16.cc b/volk/lib/qa_32fc_magnitude_16s_aligned16.cc
index a4be1616b..c3e65866b 100644
--- a/volk/lib/qa_32fc_magnitude_16s_aligned16.cc
+++ b/volk/lib/qa_32fc_magnitude_16s_aligned16.cc
@@ -23,6 +23,7 @@ void qa_32fc_magnitude_16s_aligned16::t1() {
   std::complex<float> input0[vlen] __attribute__ ((aligned (16)));
   
   int16_t output_generic[vlen] __attribute__ ((aligned (16)));
+  int16_t output_orc[vlen] __attribute__ ((aligned (16)));
   int16_t output_sse[vlen] __attribute__ ((aligned (16)));
   int16_t output_sse3[vlen] __attribute__ ((aligned (16)));
 
@@ -41,6 +42,13 @@ void qa_32fc_magnitude_16s_aligned16::t1() {
   printf("generic_time: %f\n", total);
   start = clock();
   for(int count = 0; count < ITERS; ++count) {
+    volk_32fc_magnitude_16s_aligned16_manual(output_orc, input0, 32768.0, vlen, "orc");
+  }
+  end = clock();
+  total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+  printf("orc_time: %f\n", total);
+  start = clock();
+  for(int count = 0; count < ITERS; ++count) {
     volk_32fc_magnitude_16s_aligned16_manual(output_sse, input0, 32768.0, vlen, "sse");
   }
   end = clock();
@@ -64,6 +72,7 @@ void qa_32fc_magnitude_16s_aligned16::t1() {
     //printf("%d...%d\n", output0[i], output01[i]);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse[i], 1.1);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse3[i], 1.1);
+    CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_orc[i], 1.1);
   }
 }
 
diff --git a/volk/lib/qa_32fc_magnitude_32f_aligned16.cc b/volk/lib/qa_32fc_magnitude_32f_aligned16.cc
index d69ada408..6a1d46c7a 100644
--- a/volk/lib/qa_32fc_magnitude_32f_aligned16.cc
+++ b/volk/lib/qa_32fc_magnitude_32f_aligned16.cc
@@ -23,6 +23,7 @@ void qa_32fc_magnitude_32f_aligned16::t1() {
   std::complex<float> input0[vlen] __attribute__ ((aligned (16)));
   
   float output_generic[vlen] __attribute__ ((aligned (16)));
+  float output_orc[vlen] __attribute__ ((aligned (16)));
   float output_sse[vlen] __attribute__ ((aligned (16)));
   float output_sse3[vlen] __attribute__ ((aligned (16)));
 
@@ -41,6 +42,13 @@ void qa_32fc_magnitude_32f_aligned16::t1() {
   printf("generic_time: %f\n", total);
   start = clock();
   for(int count = 0; count < ITERS; ++count) {
+    volk_32fc_magnitude_32f_aligned16_manual(output_orc, input0, vlen, "orc");
+  }
+  end = clock();
+  total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+  printf("orc_time: %f\n", total);
+  start = clock();
+  for(int count = 0; count < ITERS; ++count) {
     volk_32fc_magnitude_32f_aligned16_manual(output_sse, input0, vlen, "sse");
   }
   end = clock();
@@ -64,6 +72,7 @@ void qa_32fc_magnitude_32f_aligned16::t1() {
     //printf("%d...%d\n", output0[i], output01[i]);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse[i], fabs(output_generic[i])*1e-4);
     CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_sse3[i], fabs(output_generic[i])*1e-4);
+    CPPUNIT_ASSERT_DOUBLES_EQUAL(output_generic[i], output_orc[i], fabs(output_generic[i])*1e-4);
   }
 }
 
diff --git a/volk/lib/qa_32s_or_aligned16.cc b/volk/lib/qa_32s_or_aligned16.cc
index e09dfb91c..9ea5283a6 100644
--- a/volk/lib/qa_32s_or_aligned16.cc
+++ b/volk/lib/qa_32s_or_aligned16.cc
@@ -25,6 +25,7 @@ void qa_32s_or_aligned16::t1() {
   
   int32_t output0[vlen] __attribute__ ((aligned (16)));
   int32_t output01[vlen] __attribute__ ((aligned (16)));
+  int32_t output02[vlen] __attribute__ ((aligned (16)));
 
   for(int i = 0; i < vlen; ++i) {   
     input0[i] = ((int32_t) (rand() - (RAND_MAX/2)));
@@ -41,6 +42,13 @@ void qa_32s_or_aligned16::t1() {
   printf("generic_time: %f\n", total);
   start = clock();
   for(int count = 0; count < ITERS; ++count) {
+    volk_32s_or_aligned16_manual(output02, input0, input1, vlen, "orc");
+  }
+  end = clock();
+  total = (double)(end-start)/(double)CLOCKS_PER_SEC;
+  printf("orc_time: %f\n", total);
+  start = clock();
+  for(int count = 0; count < ITERS; ++count) {
     volk_32s_or_aligned16_manual(output01, input0, input1, vlen, "sse");
   }
   end = clock();
@@ -54,6 +62,7 @@ void qa_32s_or_aligned16::t1() {
   for(int i = 0; i < vlen; ++i) {
     //printf("%d...%d\n", output0[i], output01[i]);
     CPPUNIT_ASSERT_EQUAL(output0[i], output01[i]);
+    CPPUNIT_ASSERT_EQUAL(output0[i], output02[i]);
   }
 }