Poly octave SIMD optimizations (#358)

* Setting up Poly Octave filterbank as a struct of arrays * SSE and AVX implementations * Fixing NEON up2 implementation * Fixing up2 processing on Intel * Apply clang-format * Missing include --------- Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Chowdhury-DSP · May 18, 2024 · 6ef1720 · 6ef1720
1 parent cb19dc5
commit 6ef1720
Show file tree

Hide file tree

Showing 10 changed files with 761 additions and 238 deletions.
diff --git a/modules/cmake/RuntimeSIMDLib.cmake b/modules/cmake/RuntimeSIMDLib.cmake
@@ -1,10 +1,13 @@
 include(CheckCXXCompilerFlag)
-function(make_lib_simd_runtime name file)
+function(make_lib_simd_runtime name)
+    set(multiValueArgs SOURCES)
+    cmake_parse_arguments(ARG "" "" "${multiValueArgs}" ${ARGN})
+
     add_library(${name}_sse_or_arm STATIC)
-    target_sources(${name}_sse_or_arm PRIVATE ${file})
+    target_sources(${name}_sse_or_arm PRIVATE ${ARG_SOURCES})
 
     add_library(${name}_avx STATIC)
-    target_sources(${name}_avx PRIVATE ${file})
+    target_sources(${name}_avx PRIVATE ${ARG_SOURCES})
     target_compile_definitions(${name}_avx PRIVATE BYOD_COMPILING_WITH_AVX=1)
     if(WIN32)
         CHECK_CXX_COMPILER_FLAG("/arch:AVX" COMPILER_OPT_ARCH_AVX_SUPPORTED)

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -167,9 +167,13 @@ if (NOT(${JAI_COMPILER} STREQUAL "JAI_COMPILER-NOTFOUND"))
     target_compile_definitions(BYOD PRIVATE BYOD_BUILDING_JAI_MODULES=1)
 endif()
 
-# AVX/SSE files for accelerated neural nets
-make_lib_simd_runtime(rnn_accelerated processors/drive/neural_utils/RNNAccelerated.cpp)
-foreach(target IN ITEMS rnn_accelerated_sse_or_arm rnn_accelerated_avx)
+# AVX/SSE files for accelerated neural nets and other DSP
+make_lib_simd_runtime(dsp_accelerated
+    SOURCES
+        processors/drive/neural_utils/RNNAccelerated.cpp
+        processors/other/poly_octave/PolyOctaveV2FilterBankImpl.cpp
+)
+foreach(target IN ITEMS dsp_accelerated_sse_or_arm dsp_accelerated_avx)
     target_link_libraries(${target}
         PRIVATE
             math_approx
@@ -182,17 +186,21 @@ foreach(target IN ITEMS rnn_accelerated_sse_or_arm rnn_accelerated_avx)
             ${CMAKE_CURRENT_SOURCE_DIR}/../modules/RTNeural
             ${CMAKE_CURRENT_SOURCE_DIR}/../modules/RTNeural/modules/xsimd/include
     )
-    target_compile_definitions(${target} PRIVATE RTNEURAL_USE_XSIMD=1)
+    target_compile_definitions(${target}
+        PRIVATE
+            RTNEURAL_USE_XSIMD=1
+            _USE_MATH_DEFINES=1
+    )
     set_target_properties(${target} PROPERTIES
         POSITION_INDEPENDENT_CODE TRUE
         VISIBILITY_INLINES_HIDDEN TRUE
         C_VISIBILITY_PRESET hidden
         CXX_VISIBILITY_PRESET hidden
     )
 endforeach()
-target_compile_definitions(rnn_accelerated_sse_or_arm PRIVATE RTNEURAL_DEFAULT_ALIGNMENT=16 RTNEURAL_NAMESPACE=RTNeural_sse_arm)
-target_compile_definitions(rnn_accelerated_avx PRIVATE RTNEURAL_DEFAULT_ALIGNMENT=32 RTNEURAL_NAMESPACE=RTNeural_avx)
-target_link_libraries(BYOD PRIVATE rnn_accelerated)
+target_compile_definitions(dsp_accelerated_sse_or_arm PRIVATE RTNEURAL_DEFAULT_ALIGNMENT=16 RTNEURAL_NAMESPACE=RTNeural_sse_arm)
+target_compile_definitions(dsp_accelerated_avx PRIVATE RTNEURAL_DEFAULT_ALIGNMENT=32 RTNEURAL_NAMESPACE=RTNeural_avx)
+target_link_libraries(BYOD PRIVATE dsp_accelerated)
 
 # special flags for MSVC
 if (MSVC)

diff --git a/src/processors/other/poly_octave/PolyOctave.cpp b/src/processors/other/poly_octave/PolyOctave.cpp
@@ -1,5 +1,6 @@
 #include "PolyOctave.h"
-#include "PolyOctaveFilterBandHelpers.h"
+#include "PolyOctaveV1FilterBankImpl.h"
+#include "PolyOctaveV2FilterBankImpl.h"
 #include "processors/ParameterHelpers.h"
 
 namespace PolyOctaveTags
@@ -42,6 +43,14 @@ PolyOctave::PolyOctave (UndoManager* um)
     uiOptions.powerColour = Colour { 0xffe70510 };
     uiOptions.info.description = "A \"polyphonic octave generator\" effect.";
     uiOptions.info.authors = StringArray { "Jatin Chowdhury" };
+
+#if JUCE_INTEL
+    if (juce::SystemStats::hasAVX() && juce::SystemStats::hasFMA3())
+    {
+        juce::Logger::writeToLog ("Using Poly Octave with AVX SIMD instructions!");
+        use_avx = true;
+    }
+#endif
 }
 
 ParamLayout PolyOctave::createParameterLayout()
@@ -94,9 +103,9 @@ void PolyOctave::prepare (double sampleRate, int samplesPerBlock)
     }
 
     mixOutBuffer.setSize (2, samplesPerBlock);
-    up1OutBuffer.setSize (2, 4 * samplesPerBlock + 8); // padding for SIMD
-    up2OutBuffer.setSize (2, 4 * samplesPerBlock + 8); // padding for SIMD
-    down1OutBuffer.setSize (2, 4 * samplesPerBlock + 8); // padding for SIMD
+    up1OutBuffer.setSize (2, 8 * samplesPerBlock + 32); // padding for SIMD
+    up2OutBuffer.setSize (2, 8 * samplesPerBlock + 32); // padding for SIMD
+    down1OutBuffer.setSize (2, samplesPerBlock);
 }
 
 void PolyOctave::processAudio (AudioBuffer<float>& buffer)
@@ -125,23 +134,47 @@ void PolyOctave::processAudio (AudioBuffer<float>& buffer)
     chowdsp::BufferMath::applyGainSmoothedBuffer (down1OutBuffer, downOctaveGain);
 
     // "up1" processing
-    for (auto [ch, data_in, data_out] : chowdsp::buffer_iters::zip_channels (std::as_const (buffer), up1OutBuffer))
+    for (const auto& [ch, data_in, data_out] : chowdsp::buffer_iters::zip_channels (std::as_const (buffer), up1OutBuffer))
     {
-        poly_octave_v2::process<1> (octaveUpFilterBank[ch],
-                                    data_in.data(),
-                                    data_out.data(),
-                                    numSamples);
+#if JUCE_INTEL
+        if (use_avx)
+        {
+            poly_octave_v2::process_avx<1> (octaveUpFilterBank[ch],
+                                            data_in.data(),
+                                            data_out.data(),
+                                            numSamples);
+        }
+        else
+#endif
+        {
+            poly_octave_v2::process<1> (octaveUpFilterBank[ch],
+                                        data_in.data(),
+                                        data_out.data(),
+                                        numSamples);
+        }
     }
     upOctaveGain.process (numSamples);
     chowdsp::BufferMath::applyGainSmoothedBuffer (up1OutBuffer, upOctaveGain);
 
     // "up2" processing
-    for (auto [ch, data_in, data_out] : chowdsp::buffer_iters::zip_channels (std::as_const (buffer), up2OutBuffer))
+    for (const auto& [ch, data_in, data_out] : chowdsp::buffer_iters::zip_channels (std::as_const (buffer), up2OutBuffer))
     {
-        poly_octave_v2::process<2> (octaveUp2FilterBank[ch],
-                                    data_in.data(),
-                                    data_out.data(),
-                                    numSamples);
+#if JUCE_INTEL
+        if (use_avx)
+        {
+            poly_octave_v2::process_avx<2> (octaveUp2FilterBank[ch],
+                                            data_in.data(),
+                                            data_out.data(),
+                                            numSamples);
+        }
+        else
+#endif
+        {
+            poly_octave_v2::process<2> (octaveUp2FilterBank[ch],
+                                        data_in.data(),
+                                        data_out.data(),
+                                        numSamples);
+        }
     }
     up2OctaveGain.process (numSamples);
     chowdsp::BufferMath::applyGainSmoothedBuffer (up2OutBuffer, up2OctaveGain);

diff --git a/src/processors/other/poly_octave/PolyOctave.h b/src/processors/other/poly_octave/PolyOctave.h
@@ -1,7 +1,8 @@
 #pragma once
 
 #include "DelayPitchShifter.h"
-#include "PolyOctaveFilterBankTypes.h"
+#include "PolyOctaveV1FilterBank.h"
+#include "PolyOctaveV2FilterBank.h"
 #include "processors/BaseProcessor.h"
 
 class PolyOctave : public BaseProcessor
@@ -61,5 +62,9 @@ class PolyOctave : public BaseProcessor
     juce::AudioBuffer<float> up2OutBuffer;
     juce::AudioBuffer<float> down1OutBuffer;
 
+#if JUCE_INTEL
+    bool use_avx = false;
+#endif
+
     JUCE_DECLARE_NON_COPYABLE_WITH_LEAK_DETECTOR (PolyOctave)
 };
diff --git a/src/processors/other/poly_octave/PolyOctaveFilterBankTypes.h b/src/processors/other/poly_octave/PolyOctaveFilterBankTypes.h
diff --git a/src/processors/other/poly_octave/PolyOctaveV1FilterBank.h b/src/processors/other/poly_octave/PolyOctaveV1FilterBank.h
@@ -0,0 +1,13 @@
+#pragma once
+
+#include <pch.h>
+
+namespace poly_octave_v1
+{
+using float_2 = xsimd::batch<double>;
+struct ComplexERBFilterBank
+{
+    static constexpr size_t numFilterBands = 44;
+    std::array<chowdsp::IIRFilter<4, float_2>, numFilterBands / float_2::size> erbFilterReal, erbFilterImag;
+};
+} // namespace poly_octave_v1