Merge pull request #219 from anarkiwi/view

vkfft comparison test.

.github/workflows/test.yaml

@@ -48,4 +48,4 @@ jobs:
         fetch-depth: 0
     - name: build_test
       run: |
-        sudo apt-get -y update && sudo add-apt-repository ppa:gnuradio/gnuradio-releases && bin/apt_get.sh && bin/install-volk.sh v3.1.0 && bin/build_test.sh && bin/test_grc310.sh
+        sudo apt-get -y update && sudo add-apt-repository ppa:gnuradio/gnuradio-releases && bin/apt_get.sh && bin/install-volk.sh v3.1.0 && TEST_VKFFT=0 bin/build_test.sh && bin/test_grc310.sh

lib/vkfft_short_impl.cc

@@ -230,12 +230,11 @@ vkfft_short_impl::vkfft_short_impl(std::size_t fft_batch_size, std::size_t nfft,
 }
 
 void vkfft_short_impl::init_converter_() {
-  uhd::convert::id_type id;
-  id.input_format = "sc16_chdr";
-  id.num_inputs = 1;
-  id.output_format = "fc32";
-  id.num_outputs = 1;
-  _converter = uhd::convert::get_converter(id, 0)();
+  _id.input_format = "sc16_chdr";
+  _id.num_inputs = 1;
+  _id.output_format = "fc32";
+  _id.num_outputs = 1;
+  _converter = uhd::convert::get_converter(_id, 0)();
   _converter->set_scalar(1 / 32767.);
 }
 
@@ -254,7 +253,7 @@ int vkfft_short_impl::work(int noutput_items,
 
   for (int i = 0; i < noutput_items / fft_batch_size_;
        ++i, in_buffer_index += vlen * 2, out_buffer_index += vlen) {
-    _converter->conv(&in[in_buffer_index], &buffer[0], vlen * 2);
+    _converter->conv(&in[in_buffer_index], &buffer[0], vlen);
     vkfft_offload((char *)&buffer[0], (char *)&out[out_buffer_index]);
   }
 

lib/vkfft_short_impl.h

@@ -220,6 +220,7 @@ class vkfft_short_impl : public vkfft_short, base_impl {
   boost::scoped_array<gr_complex> input_buffer_;
   uhd::convert::converter::sptr _converter;
   void init_converter_();
+  uhd::convert::id_type _id;
 
 public:
   vkfft_short_impl(std::size_t fft_batch_size, std::size_t nfft, bool shift);

python/iqtlabs/CMakeLists.txt

@@ -37,7 +37,6 @@ add_custom_target(
 )
 if(VKFFT)
 GR_ADD_TEST(qa_vkfft ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/qa_vkfft.py)
-GR_ADD_TEST(qa_vkfft_short ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/qa_vkfft_short.py)
 endif()
 GR_ADD_TEST(qa_tuneable_test_source ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/qa_tuneable_test_source.py)
 GR_ADD_TEST(qa_retune_fft ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/qa_retune_fft.py)

python/iqtlabs/qa_iq_inference.py

@@ -361,6 +361,5 @@ def test_instance(self):
                 last_rx_freq = rx_freq
 
 
-
 if __name__ == "__main__":
     gr_unittest.run(qa_iq_inference)

python/iqtlabs/qa_vkfft.py

@@ -208,37 +208,49 @@
 from gnuradio import blocks, fft, iqtlabs, gr, gr_unittest
 
 
-def run_fft(fft_block, points, fft_roll, input_items):
+def run_fft(fft_block, short, points, fft_roll, input_items):
+    tb = gr.top_block()
     src1 = blocks.vector_source_c(input_items, vlen=points)
     dst1 = blocks.vector_sink_c(vlen=points)
-    tb = gr.top_block()
-    tb.connect(src1, fft_block)
-    tb.connect(fft_block, dst1)
+    if short:
+        v2s = blocks.vector_to_stream(gr.sizeof_gr_complex, points)
+        cis = blocks.complex_to_interleaved_short(False, 32767)
+        s2v = blocks.stream_to_vector(gr.sizeof_short, points * 2)
+        test_blocks = [src1, v2s, cis, s2v, fft_block, dst1]
+    else:
+        test_blocks = [src1, fft_block, dst1]
+    for i, b in enumerate(test_blocks, start=0):
+        if i:
+            tb.connect(test_blocks[i - 1], b)
     tb.run()
-    data = dst1.data()
     tb.stop()
     tb.wait()
+    data = dst1.data()
     del tb
     return data
 
 
 class qa_vkfft(gr_unittest.TestCase):
-    def test_instance(self):
+    def round_complex(self, l, n=3):
+        return [complex(round(x.real, n), round(x.imag, n)) for x in l]
+
+    def run_comparison(self, vkfft_block, short):
         for fft_roll in (True, False):
             fft_batch_size = 1
             points = 8
 
             batch_input_items = []
             for i in range(1, fft_batch_size + 1):
                 batch_input_items.append(
-                    1j * np.arange(points)
+                    1j * np.arange(start=0, stop=(points * 0.1), step=0.1)
                 )  # pytype: disable=wrong-arg-types
 
             sw_data = []
             for i in range(1, fft_batch_size + 1):
                 sw_data.extend(
                     run_fft(
                         fft.fft_vcc(points, True, [], fft_roll, 1),
+                        0,
                         points,
                         fft_roll,
                         batch_input_items[i - 1],
@@ -249,16 +261,25 @@ def test_instance(self):
             for i in range(1, fft_batch_size + 1):
                 vkfft_data.extend(
                     run_fft(
-                        iqtlabs.vkfft(fft_batch_size, points, fft_roll),
+                        vkfft_block(fft_batch_size, points, fft_roll),
+                        short,
                         points,
                         fft_roll,
                         batch_input_items[i - 1],
                     )
                 )
 
-            if os.getenv("TEST_VKFFT", 0):
-                self.assertEqual(vkfft_data, sw_data)
+            self.assertEqual(
+                self.round_complex(vkfft_data), self.round_complex(sw_data)
+            )
+
+    def test_vkfft(self):
+        self.run_comparison(iqtlabs.vkfft, False)
+
+    def test_short_vkfft(self):
+        self.run_comparison(iqtlabs.vkfft_short, True)
 
 
 if __name__ == "__main__":
-    gr_unittest.run(qa_vkfft)
+    if os.getenv("TEST_VKFFT", 0):
+        gr_unittest.run(qa_vkfft)