per-Bungee-instance FFT cache

kupix · Oct 19, 2024 · f7a54a5 · f7a54a5
1 parent bca568c
commit f7a54a5
Show file tree

Hide file tree

Showing 10 changed files with 33 additions and 46 deletions.
diff --git a/src/Basic.cpp b/src/Basic.cpp
@@ -36,9 +36,10 @@ bool Stretcher<Basic>::isFlushed() const
 
 Basic::Basic(SampleRates sampleRates, int channelCount, int log2SynthesisHopOverride) :
 	Timing(sampleRates, log2SynthesisHopOverride),
-	input(log2SynthesisHop, channelCount),
+	transforms(Fourier::transforms()),
+	input(log2SynthesisHop, channelCount, *transforms),
 	grains(4),
-	output(log2SynthesisHop, channelCount, maxOutputFrameCount(true), 0.25f, {1.f, 0.5f})
+	output(*transforms, log2SynthesisHop, channelCount, maxOutputFrameCount(true), 0.25f, {1.f, 0.5f})
 {
 	for (auto &grain : grains.vector)
 		grain = std::make_unique<Grain>(log2SynthesisHop, channelCount);
@@ -68,7 +69,7 @@ void Basic::analyseGrain(const float *data, std::ptrdiff_t stride)
 
 		auto log2TransformLength = input.applyAnalysisWindow(ref);
 
-		Fourier::transforms->forward(log2TransformLength, input.windowedInput, grain.transformed);
+		transforms->forward(log2TransformLength, input.windowedInput, grain.transformed);
 
 		const auto n = Fourier::binCount(grain.log2TransformLength) - 1;
 		grain.validBinCount = std::min<int>(std::ceil(n / grain.resampleOperations.output.ratio), n) + 1;
@@ -112,7 +113,7 @@ void Basic::synthesiseGrain(OutputChunk &outputChunk)
 		else
 			grain.transformed.topRows(grain.validBinCount).colwise() *= t;
 
-		Fourier::transforms->inverse(grain.log2TransformLength, output.inverseTransformed, grain.transformed);
+		transforms->inverse(grain.log2TransformLength, output.inverseTransformed, grain.transformed);
 	}
 
 	output.applySynthesisWindow(log2SynthesisHop, grains, output.synthesisWindow);

diff --git a/src/Basic.h b/src/Basic.h
@@ -9,12 +9,14 @@
 #include "Output.h"
 #include "Timing.h"
 
+#include <memory>
+
 namespace Bungee {
 
 struct Basic :
-	Fourier::Bootstrap,
 	Timing
 {
+	std::unique_ptr<Fourier::Transforms> transforms;
 	Input input;
 	Grains grains;
 	Output output;

diff --git a/src/Fourier.cpp b/src/Fourier.cpp
@@ -57,13 +57,9 @@ void Kiss::Kernel<isInverse>::inverse(int, float *t, std::complex<float> *f) con
 	kiss_fftri((kiss_fftr_cfg)implementation, (kiss_fft_cpx *)f, t);
 }
 
-Transforms *transforms{};
-
-Bootstrap::Bootstrap()
+std::unique_ptr<Transforms> transforms()
 {
-	BUNGEE_ASSERT1(!transforms);
-	static Fourier::Cache<Kiss, 16> cache;
-	BUNGEE_ASSERT1(transforms == &cache);
+	return std::make_unique<Cache<Kiss, 16>>();
 }
 
 } // namespace Bungee::Fourier
diff --git a/src/Fourier.h b/src/Fourier.h
@@ -11,7 +11,6 @@
 #include <complex>
 #include <limits>
 #include <memory>
-#include <mutex>
 #include <vector>
 
 namespace Bungee {
@@ -68,14 +67,12 @@ inline void resize(int log2TransformLength, int channelCount, T &array, int extr
 struct Transforms
 {
 	virtual ~Transforms() {}
-	virtual void prepareForward(int log2Length) = 0;
-	virtual void prepareInverse(int log2Length) = 0;
+	virtual void prepareForward(int log2TransformLength) = 0;
+	virtual void prepareInverse(int log2TransformLength) = 0;
 	virtual void forward(int log2TransformLength, const Eigen::Ref<const Eigen::ArrayXXf> &t, Eigen::Ref<Eigen::ArrayXXcf> f) const = 0;
 	virtual void inverse(int log2TransformLength, Eigen::Ref<Eigen::ArrayXXf> t, const Eigen::Ref<const Eigen::ArrayXXcf> &f) const = 0;
 };
 
-extern Transforms *transforms;
-
 // General case when an FFT implementation has different states for forward and reverse transforms of same size.
 template <class F, class I>
 struct KernelPair
@@ -129,27 +126,19 @@ struct Cache :
 {
 	typedef KernelPair<typename K::Forward, typename K::Inverse> Entry;
 	typedef std::array<Entry, log2MaxSize + 1> Table;
-	std::mutex preparationMutex;
 
 	Table table;
 
-	Cache()
-	{
-		transforms = this;
-	}
-
-	void prepareForward(int log2Length) override
+	void prepareForward(int log2TransformLength) override
 	{
-		std::scoped_lock lock(preparationMutex);
-		if (!table[log2Length].forward())
-			table[log2Length].forward(new typename K::Forward(log2Length));
+		if (!table[log2TransformLength].forward())
+			table[log2TransformLength].forward(new typename K::Forward(log2TransformLength));
 	}
 
-	void prepareInverse(int log2Length) override
+	void prepareInverse(int log2TransformLength) override
 	{
-		std::scoped_lock lock(preparationMutex);
-		if (!table[log2Length].inverse())
-			table[log2Length].inverse(new typename K::Inverse(log2Length));
+		if (!table[log2TransformLength].inverse())
+			table[log2TransformLength].inverse(new typename K::Inverse(log2TransformLength));
 	}
 
 	void forward(int log2TransformLength, const Eigen::Ref<const Eigen::ArrayXXf> &t, Eigen::Ref<Eigen::ArrayXXcf> f) const override
@@ -177,9 +166,6 @@ struct Cache :
 	}
 };
 
-struct Bootstrap
-{
-	Bootstrap();
-};
+std::unique_ptr<Transforms> transforms();
 
 } // namespace Bungee::Fourier
diff --git a/src/Input.cpp b/src/Input.cpp
@@ -14,12 +14,12 @@ static constexpr float pi = std::numbers::pi_v<float>;
 static constexpr float gain = (3 * pi) / (3 * pi + 8);
 } // namespace
 
-Input::Input(int log2SynthesisHop, int channelCount) :
-	analysisWindowBasic(Window::fromFrequencyDomainCoefficients(log2SynthesisHop + 3, gain / (8 << log2SynthesisHop), {1.f, 0.5f})),
+Input::Input(int log2SynthesisHop, int channelCount, Fourier::Transforms &transforms) :
+	analysisWindowBasic(Window::fromFrequencyDomainCoefficients(transforms, log2SynthesisHop + 3, gain / (8 << log2SynthesisHop), {1.f, 0.5f})),
 	windowedInput{(8 << log2SynthesisHop), channelCount}
 {
 	windowedInput.setZero();
-	Fourier::transforms->prepareForward(log2SynthesisHop + 3);
+	transforms.prepareForward(log2SynthesisHop + 3);
 }
 
 int Input::applyAnalysisWindow(const Eigen::Ref<const Eigen::ArrayXXf> &input)

diff --git a/src/Input.h b/src/Input.h
@@ -4,6 +4,7 @@
 #pragma once
 
 #include "Assert.h"
+#include "Fourier.h"
 
 #include <Eigen/Dense>
 
@@ -14,7 +15,7 @@ struct Input
 	Eigen::ArrayXf analysisWindowBasic;
 	Eigen::ArrayXXf windowedInput;
 
-	Input(int log2SynthesisHop, int channelCount);
+	Input(int log2SynthesisHop, int channelCount, Fourier::Transforms &transforms);
 
 	int applyAnalysisWindow(const Eigen::Ref<const Eigen::ArrayXXf> &input);
 };

diff --git a/src/Output.cpp b/src/Output.cpp
@@ -7,12 +7,12 @@
 
 namespace Bungee {
 
-Output::Output(int log2SynthesisHop, int channelCount, int maxOutputChunkSize, float windowGain, std::initializer_list<float> windowCoefficients) :
-	synthesisWindow{Window::fromFrequencyDomainCoefficients(log2SynthesisHop + 2, windowGain, windowCoefficients)},
+Output::Output(Fourier::Transforms &transforms, int log2SynthesisHop, int channelCount, int maxOutputChunkSize, float windowGain, std::initializer_list<float> windowCoefficients) :
+	synthesisWindow{Window::fromFrequencyDomainCoefficients(transforms, log2SynthesisHop + 2, windowGain, windowCoefficients)},
 	inverseTransformed(8 << log2SynthesisHop, channelCount),
 	bufferResampled(maxOutputChunkSize, channelCount)
 {
-	Fourier::transforms->prepareInverse(log2SynthesisHop + 3);
+	transforms.prepareInverse(log2SynthesisHop + 3);
 }
 
 void Output::applySynthesisWindow(int log2SynthesisHop, Grains &grains, const Eigen::Ref<const Eigen::ArrayXf> &window)

diff --git a/src/Output.h b/src/Output.h
@@ -25,7 +25,7 @@ struct Output
 	float resampleOffset = 0.f;
 	Window::DispatchApply dispatchApply;
 
-	Output(int log2SynthesisHop, int channelCount, int maxOutputChunkSize, float windowGain, std::initializer_list<float> windowCoefficients);
+	Output(Fourier::Transforms &transforms, int log2SynthesisHop, int channelCount, int maxOutputChunkSize, float windowGain, std::initializer_list<float> windowCoefficients);
 
 	void applySynthesisWindow(int log2SynthesisHop, Grains &grains, const Eigen::Ref<const Eigen::ArrayXf> &window);
 

diff --git a/src/Window.cpp b/src/Window.cpp
@@ -11,7 +11,7 @@
 
 namespace Bungee::Window {
 
-Eigen::ArrayXf fromFrequencyDomainCoefficients(int log2Size, float gain, std::initializer_list<float> coefficients)
+Eigen::ArrayXf fromFrequencyDomainCoefficients(Fourier::Transforms &transforms, int log2Size, float gain, std::initializer_list<float> coefficients)
 {
 	Eigen::ArrayXcf frequencyDomain(Fourier::binCount(log2Size));
 
@@ -22,8 +22,8 @@ Eigen::ArrayXf fromFrequencyDomainCoefficients(int log2Size, float gain, std::in
 	frequencyDomain.bottomRows(frequencyDomain.rows() - i).setZero();
 
 	Eigen::ArrayXf window(1 << log2Size);
-	Fourier::transforms->prepareInverse(log2Size);
-	Fourier::transforms->inverse(log2Size, window, frequencyDomain);
+	transforms.prepareInverse(log2Size);
+	transforms.inverse(log2Size, window, frequencyDomain);
 	return window;
 }
 

diff --git a/src/Window.h b/src/Window.h
@@ -5,14 +5,15 @@
 
 #include "Assert.h"
 #include "Dispatch.h"
+#include "Fourier.h"
 
 #include <Eigen/Dense>
 
 #include <initializer_list>
 
 namespace Bungee::Window {
 
-Eigen::ArrayXf fromFrequencyDomainCoefficients(int log2Size, float gain, std::initializer_list<float> coefficients);
+Eigen::ArrayXf fromFrequencyDomainCoefficients(Fourier::Transforms &transforms, int log2Size, float gain, std::initializer_list<float> coefficients);
 
 struct Apply
 {