Hacked together a replacement of the ellitic filters with the tabuli

CMakeLists.txt

@@ -20,6 +20,8 @@ add_library(zimtohrli_base STATIC
     cpp/zimt/elliptic.h
     cpp/zimt/filterbank.cc
     cpp/zimt/filterbank.h
+    cpp/zimt/fourier_bank.h
+    cpp/zimt/fourier_bank.cc
     cpp/zimt/loudness.cc
     cpp/zimt/loudness.h
     cpp/zimt/masking.cc

configure.sh

@@ -8,5 +8,5 @@ elif [ "${1}" == "asan" ]; then
     (cd asan_build && cmake -G Ninja -DCMAKE_C_FLAGS='-fsanitize=address -fPIC' -DCMAKE_CXX_FLAGS='-fsanitize=address -fPIC' -DCMAKE_LINKER_FLAGS_DEBUG='-fsanitize=address' -DCMAKE_BUILD_TYPE=RelWithDebInfo ..)
 else
     mkdir -p build
-    (cd build && cmake -G Ninja -DCMAKE_C_FLAGS='-fPIC' -DCMAKE_CXX_FLAGS='-fPIC' -DCMAKE_BUILD_TYPE=Release ..)
+    (cd build && cmake -G Ninja -DCMAKE_C_FLAGS='-fPIC -march=native -O3' -DCMAKE_CXX_FLAGS='-fPIC -march=native -O3' -DCMAKE_BUILD_TYPE=Release ..)
 fi

cpp/zimt/cam.cc

@@ -45,7 +45,7 @@ CamFilterbank Cam::CreateFilterbank(float sample_rate) const {
 
   int sections = -1;
   for (float left_cam = low_threshold_cam;
-       left_cam + cam_delta < high_threshold_cam; left_cam += cam_delta) {
+       left_cam + 0.5 * cam_delta < high_threshold_cam; left_cam += cam_delta) {
     float left_hz = HzFromCam(left_cam);
     float right_hz = HzFromCam(left_cam + cam_delta);
     const std::vector<BACoeffs> filter_coeffs = DigitalSOSBandPass(

cpp/zimt/cam.h

@@ -70,7 +70,7 @@ struct Cam {
   float high_threshold_hz = 20000;
   // Frequency resolution at low threshold. Default is 5 since it has
   // provided the best correlation scores with tested datasets.
-  float minimum_bandwidth_hz = 5;
+  float minimum_bandwidth_hz = 1;
   // Order (sharpness and slowness) of channel filters.
   int filter_order = 1;
   // Attenuation in dB in each filter where the filterbank filters meet.

cpp/zimt/compare.cc

@@ -300,8 +300,6 @@ int Main(int argc, char* argv[]) {
     const size_t num_downscaled_samples_a = static_cast<size_t>(
         std::ceil(static_cast<float>(file_a->Frames().shape()[1]) *
                   z.perceptual_sample_rate / z.cam_filterbank->sample_rate));
-    hwy::AlignedNDArray<float, 2> channels_a(
-        {file_a->Frames().shape()[1], z.cam_filterbank->filter.Size()});
     hwy::AlignedNDArray<float, 2> energy_channels_db_a(
         {num_downscaled_samples_a, z.cam_filterbank->filter.Size()});
     hwy::AlignedNDArray<float, 2> partial_energy_channels_db_a(
@@ -311,7 +309,7 @@ int Main(int argc, char* argv[]) {
          ++channel_index) {
       hwy::AlignedNDArray<float, 2> spectrogram(
           {num_downscaled_samples_a, z.cam_filterbank->filter.Size()});
-      z.Spectrogram(file_a->Frames()[{channel_index}], channels_a,
+      z.Spectrogram(file_a->Frames()[{channel_index}],
                     energy_channels_db_a, partial_energy_channels_db_a,
                     spectrogram);
       file_a_spectrograms.push_back(std::move(spectrogram));
@@ -322,8 +320,6 @@ int Main(int argc, char* argv[]) {
       const size_t num_downscaled_samples_b = static_cast<size_t>(
           std::ceil(static_cast<float>(file_b.Frames().shape()[1]) *
                     z.perceptual_sample_rate / z.cam_filterbank->sample_rate));
-      hwy::AlignedNDArray<float, 2> channels_b(
-          {file_b.Frames().shape()[1], z.cam_filterbank->filter.Size()});
       hwy::AlignedNDArray<float, 2> energy_channels_db_b(
           {num_downscaled_samples_b, z.cam_filterbank->filter.Size()});
       hwy::AlignedNDArray<float, 2> partial_energy_channels_db_b(
@@ -333,7 +329,7 @@ int Main(int argc, char* argv[]) {
       float sum_of_squares = 0;
       for (size_t channel_index = 0; channel_index < file_a->Info().channels;
            ++channel_index) {
-        z.Spectrogram(file_b.Frames()[{channel_index}], channels_b,
+        z.Spectrogram(file_b.Frames()[{channel_index}],
                       energy_channels_db_b, partial_energy_channels_db_b,
                       spectrogram_b);
         const float distance =