Be less messy with PFFFT

Remove a 1-element array for an over-allocated struct array. Also add a wrapper
struct for C++.

5 files changed, 82 insertions, 56 deletions

diff --git a/alc/effects/convolution.cpp b/alc/effects/convolution.cpp
index 8db7a045..5e81f6d1 100644
--- a/alc/effects/convolution.cpp
+++ b/alc/effects/convolution.cpp
@@ -190,12 +190,6 @@ void apply_fir(al::span<float> dst, const float *RESTRICT src, const float *REST
 }
 
 
-struct PFFFTSetupDeleter {
-    void operator()(PFFFT_Setup *ptr) { pffft_destroy_setup(ptr); }
-};
-using PFFFTSetupPtr = std::unique_ptr<PFFFT_Setup,PFFFTSetupDeleter>;
-
-
 struct ConvolutionState final : public EffectState {
     FmtChannels mChannels{};
     AmbiLayout mAmbiLayout{};
@@ -207,7 +201,7 @@ struct ConvolutionState final : public EffectState {
     al::vector<std::array<float,ConvolveUpdateSamples>,16> mFilter;
     al::vector<std::array<float,ConvolveUpdateSamples*2>,16> mOutput;
 
-    PFFFTSetupPtr mFft{};
+    PFFFTSetup mFft{};
     alignas(16) std::array<float,ConvolveUpdateSize> mFftBuffer{};
     alignas(16) std::array<float,ConvolveUpdateSize> mFftWorkBuffer{};
 
@@ -270,7 +264,7 @@ void ConvolutionState::deviceUpdate(const DeviceBase *device, const BufferStorag
     static constexpr uint MaxConvolveAmbiOrder{1u};
 
     if(!mFft)
-        mFft = PFFFTSetupPtr{pffft_new_setup(ConvolveUpdateSize, PFFFT_REAL)};
+        mFft = PFFFTSetup{ConvolveUpdateSize, PFFFT_REAL};
 
     mFifoPos = 0;
     mInput.fill(0.0f);
@@ -400,7 +394,7 @@ void ConvolutionState::deviceUpdate(const DeviceBase *device, const BufferStorag
             /* Reorder backward to make it suitable for pffft_zconvolve and the
              * subsequent pffft_transform(..., PFFFT_BACKWARD).
              */
-            pffft_zreorder(mFft.get(), ffttmp.data(), al::to_address(filteriter), PFFFT_BACKWARD);
+            mFft.zreorder(ffttmp.data(), al::to_address(filteriter), PFFFT_BACKWARD);
             filteriter += ConvolveUpdateSize;
         }
     }
@@ -642,7 +636,7 @@ void ConvolutionState::process(const size_t samplesToDo,
         /* Calculate the frequency-domain response and add the relevant
          * frequency bins to the FFT history.
          */
-        pffft_transform(mFft.get(), mInput.data(), mComplexData.data() + curseg*ConvolveUpdateSize,
+        mFft.transform(mInput.data(), mComplexData.data() + curseg*ConvolveUpdateSize,
             mFftWorkBuffer.data(), PFFFT_FORWARD);
 
         const float *filter{mComplexData.data() + mNumConvolveSegs*ConvolveUpdateSize};
@@ -655,14 +649,14 @@ void ConvolutionState::process(const size_t samplesToDo,
             const float *input{&mComplexData[curseg*ConvolveUpdateSize]};
             for(size_t s{curseg};s < mNumConvolveSegs;++s)
             {
-                pffft_zconvolve_accumulate(mFft.get(), input, filter, mFftBuffer.data());
+                mFft.zconvolve_accumulate(input, filter, mFftBuffer.data());
                 input += ConvolveUpdateSize;
                 filter += ConvolveUpdateSize;
             }
             input = mComplexData.data();
             for(size_t s{0};s < curseg;++s)
             {
-                pffft_zconvolve_accumulate(mFft.get(), input, filter, mFftBuffer.data());
+                mFft.zconvolve_accumulate(input, filter, mFftBuffer.data());
                 input += ConvolveUpdateSize;
                 filter += ConvolveUpdateSize;
             }
@@ -672,8 +666,8 @@ void ConvolutionState::process(const size_t samplesToDo,
              * second-half samples (and this output's second half is
              * subsequently saved for next time).
              */
-            pffft_transform(mFft.get(), mFftBuffer.data(), mFftBuffer.data(),
-                mFftWorkBuffer.data(), PFFFT_BACKWARD);
+            mFft.transform(mFftBuffer.data(), mFftBuffer.data(), mFftWorkBuffer.data(),
+                PFFFT_BACKWARD);
 
             /* The filter was attenuated, so the response is already scaled. */
             for(size_t i{0};i < ConvolveUpdateSamples;++i)
diff --git a/alc/effects/pshifter.cpp b/alc/effects/pshifter.cpp
index 0c27be30..871e866a 100644
--- a/alc/effects/pshifter.cpp
+++ b/alc/effects/pshifter.cpp
@@ -74,12 +74,6 @@ struct Windower {
 const Windower gWindow{};
 
 
-struct PFFFTSetupDeleter {
-    void operator()(PFFFT_Setup *ptr) { pffft_destroy_setup(ptr); }
-};
-using PFFFTSetupPtr = std::unique_ptr<PFFFT_Setup,PFFFTSetupDeleter>;
-
-
 struct FrequencyBin {
     float Magnitude;
     float FreqBin;
@@ -99,7 +93,7 @@ struct PshifterState final : public EffectState {
     std::array<float,StftHalfSize+1> mSumPhase;
     std::array<float,StftSize> mOutputAccum;
 
-    PFFFTSetupPtr mFft;
+    PFFFTSetup mFft;
     alignas(16) std::array<float,StftSize> mFftBuffer;
     alignas(16) std::array<float,StftSize> mFftWorkBuffer;
 
@@ -142,7 +136,7 @@ void PshifterState::deviceUpdate(const DeviceBase*, const BufferStorage*)
     std::fill(std::begin(mTargetGains),  std::end(mTargetGains),  0.0f);
 
     if(!mFft)
-        mFft = PFFFTSetupPtr{pffft_new_setup(StftSize, PFFFT_REAL)};
+        mFft = PFFFTSetup{StftSize, PFFFT_REAL};
 }
 
 void PshifterState::update(const ContextBase*, const EffectSlot *slot,
@@ -197,8 +191,8 @@ void PshifterState::process(const size_t samplesToDo,
             mFftBuffer[k] = mFIFO[src] * gWindow.mData[k];
         for(size_t src{0u}, k{StftSize-mPos};src < mPos;++src,++k)
             mFftBuffer[k] = mFIFO[src] * gWindow.mData[k];
-        pffft_transform_ordered(mFft.get(), mFftBuffer.data(), mFftBuffer.data(),
-            mFftWorkBuffer.data(), PFFFT_FORWARD);
+        mFft.transform_ordered(mFftBuffer.data(), mFftBuffer.data(), mFftWorkBuffer.data(),
+            PFFFT_FORWARD);
 
         /* Analyze the obtained data. Since the real FFT is symmetric, only
          * StftHalfSize+1 samples are needed.
@@ -296,8 +290,8 @@ void PshifterState::process(const size_t samplesToDo,
         /* Apply an inverse FFT to get the time-domain signal, and accumulate
          * for the output with windowing.
          */
-        pffft_transform_ordered(mFft.get(), mFftBuffer.data(), mFftBuffer.data(),
-            mFftWorkBuffer.data(), PFFFT_BACKWARD);
+        mFft.transform_ordered(mFftBuffer.data(), mFftBuffer.data(), mFftWorkBuffer.data(),
+            PFFFT_BACKWARD);
 
         static constexpr float scale{3.0f / OversampleFactor / StftSize};
         for(size_t dst{mPos}, k{0u};dst < StftSize;++dst,++k)
diff --git a/common/pffft.cpp b/common/pffft.cpp
index 505c9791..bf564086 100644
--- a/common/pffft.cpp
+++ b/common/pffft.cpp
@@ -1413,13 +1413,13 @@ void pffft_aligned_free(void *p) { al_free(p); }
 int pffft_simd_size() { return SIMD_SZ; }
 
 struct PFFFT_Setup {
-    uint N;
+    alignas(MALLOC_V4SF_ALIGNMENT) uint N;
     uint Ncvec; // nb of complex simd vectors (N/4 if PFFFT_COMPLEX, N/8 if PFFFT_REAL)
     std::array<uint,15> ifac;
     pffft_transform_t transform;
 
     float *twiddle; // N/4 elements
-    alignas(MALLOC_V4SF_ALIGNMENT) v4sf e[1]; // N/4*3 elements
+    al::span<v4sf> e; // N/4*3 elements
 };
 
 PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform)
@@ -1436,8 +1436,7 @@ PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform)
         assert((N%(SIMD_SZ*SIMD_SZ)) == 0);
 
     const uint Ncvec = (transform == PFFFT_REAL ? N/2 : N)/SIMD_SZ;
-    const size_t storelen{std::max(sizeof(PFFFT_Setup),
-        offsetof(PFFFT_Setup, e[0]) + (2u*Ncvec * sizeof(v4sf)))};
+    const size_t storelen{sizeof(PFFFT_Setup) + (2u*Ncvec * sizeof(v4sf))};
 
     void *store{al_calloc(MALLOC_V4SF_ALIGNMENT, storelen)};
     if(!store) return nullptr;
@@ -1447,6 +1446,7 @@ PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform)
     s->transform = transform;
     /* nb of complex simd vectors */
     s->Ncvec = Ncvec;
+    s->e = {reinterpret_cast<v4sf*>(reinterpret_cast<char*>(s+1)), 2u*Ncvec};
     s->twiddle = reinterpret_cast<float*>(&s->e[2u*Ncvec*(SIMD_SZ-1)/SIMD_SZ]);
 
     if constexpr(SIMD_SZ > 1)
@@ -1463,7 +1463,7 @@ PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform)
                 e[((i*3 + m)*2 + 1)*SIMD_SZ + j] = static_cast<float>(std::sin(A));
             }
         }
-        std::memcpy(s->e, e.data(), e.size()*sizeof(float));
+        std::memcpy(s->e.data(), e.data(), e.size()*sizeof(float));
     }
     if(transform == PFFFT_REAL)
         rffti1_ps(N/SIMD_SZ, s->twiddle, s->ifac);
@@ -1825,7 +1825,7 @@ void pffft_transform_internal(const PFFFT_Setup *setup, const v4sf *vinput, v4sf
         if(setup->transform == PFFFT_REAL)
         {
             ib = (rfftf1_ps(Ncvec*2, vinput, buff[ib], buff[!ib], setup->twiddle, setup->ifac) == buff[1]);
-            pffft_real_finalize(Ncvec, buff[ib], buff[!ib], setup->e);
+            pffft_real_finalize(Ncvec, buff[ib], buff[!ib], setup->e.data());
         }
         else
         {
@@ -1834,7 +1834,7 @@ void pffft_transform_internal(const PFFFT_Setup *setup, const v4sf *vinput, v4sf
                 uninterleave2(vinput[k*2], vinput[k*2+1], tmp[k*2], tmp[k*2+1]);
 
             ib = (cfftf1_ps(Ncvec, buff[ib], buff[!ib], buff[ib], setup->twiddle, setup->ifac, -1.0f) == buff[1]);
-            pffft_cplx_finalize(Ncvec, buff[ib], buff[!ib], setup->e);
+            pffft_cplx_finalize(Ncvec, buff[ib], buff[!ib], setup->e.data());
         }
         if(ordered)
             pffft_zreorder(setup, reinterpret_cast<float*>(buff[!ib]),
@@ -1856,12 +1856,12 @@ void pffft_transform_internal(const PFFFT_Setup *setup, const v4sf *vinput, v4sf
         }
         if(setup->transform == PFFFT_REAL)
         {
-            pffft_real_preprocess(Ncvec, vinput, buff[ib], setup->e);
+            pffft_real_preprocess(Ncvec, vinput, buff[ib], setup->e.data());
             ib = (rfftb1_ps(Ncvec*2, buff[ib], buff[0], buff[1], setup->twiddle, setup->ifac) == buff[1]);
         }
         else
         {
-            pffft_cplx_preprocess(Ncvec, vinput, buff[ib], setup->e);
+            pffft_cplx_preprocess(Ncvec, vinput, buff[ib], setup->e.data());
             ib = (cfftf1_ps(Ncvec, buff[ib], buff[0], buff[1],  setup->twiddle, setup->ifac, +1.0f) == buff[1]);
             for(size_t k{0};k < Ncvec;++k)
                 interleave2(buff[ib][k*2], buff[ib][k*2+1], buff[ib][k*2], buff[ib][k*2+1]);
diff --git a/common/pffft.h b/common/pffft.h
index 9cff9e54..b31304f6 100644
--- a/common/pffft.h
+++ b/common/pffft.h
@@ -83,23 +83,27 @@
 #include <stdint.h>
 
 #ifdef __cplusplus
+#include <cstddef>
+#include <utility>
+
 extern "C" {
 #endif
 
 /* opaque struct holding internal stuff (precomputed twiddle factors) this
  * struct can be shared by many threads as it contains only read-only data.
  */
-typedef struct PFFFT_Setup PFFFT_Setup;
-
-#ifndef PFFFT_COMMON_ENUMS
-#define PFFFT_COMMON_ENUMS
+struct PFFFT_Setup;
 
 /* direction of the transform */
-typedef enum { PFFFT_FORWARD, PFFFT_BACKWARD } pffft_direction_t;
+enum pffft_direction_t { PFFFT_FORWARD, PFFFT_BACKWARD };
 
 /* type of transform */
-typedef enum { PFFFT_REAL, PFFFT_COMPLEX } pffft_transform_t;
+enum pffft_transform_t { PFFFT_REAL, PFFFT_COMPLEX };
 
+#ifndef __cplusplus
+typedef struct PFFFT_Setup PFFFT_Setup;
+typedef enum pffft_direction_t pffft_direction_t;
+typedef enum pffft_transform_t pffft_transform_t;
 #endif
 
 /**
@@ -187,6 +191,47 @@ int pffft_simd_size();
 
 #ifdef __cplusplus
 }
+
+struct PFFFTSetup {
+    PFFFT_Setup *mSetup{};
+
+    PFFFTSetup() = default;
+    PFFFTSetup(const PFFFTSetup&) = delete;
+    PFFFTSetup(PFFFTSetup&& rhs) : mSetup{rhs.mSetup} { rhs.mSetup = nullptr; }
+    explicit PFFFTSetup(std::nullptr_t) { }
+    explicit PFFFTSetup(unsigned int n, pffft_transform_t transform)
+        : mSetup{pffft_new_setup(n, transform)}
+    { }
+    ~PFFFTSetup() { if(mSetup) pffft_destroy_setup(mSetup); }
+
+    PFFFTSetup& operator=(const PFFFTSetup&) = delete;
+    PFFFTSetup& operator=(PFFFTSetup&& rhs)
+    {
+        if(mSetup)
+            pffft_destroy_setup(mSetup);
+        mSetup = std::exchange(rhs.mSetup, nullptr);
+        return *this;
+    }
+
+    void transform(const float *input, float *output, float *work, pffft_direction_t direction) const
+    { pffft_transform(mSetup, input, output, work, direction); }
+
+    void transform_ordered(const float *input, float *output, float *work,
+        pffft_direction_t direction) const
+    { pffft_transform_ordered(mSetup, input, output, work, direction); }
+
+    void zreorder(const float *input, float *output, pffft_direction_t direction) const
+    { pffft_zreorder(mSetup, input, output, direction); }
+
+    void zconvolve_scale_accumulate(const float *dft_a, const float *dft_b, float *dft_ab,
+        float scaling) const
+    { pffft_zconvolve_scale_accumulate(mSetup, dft_a, dft_b, dft_ab, scaling); }
+
+    void zconvolve_accumulate(const float *dft_a, const float *dft_b, float *dft_ab) const
+    { pffft_zconvolve_accumulate(mSetup, dft_a, dft_b, dft_ab); }
+
+    [[nodiscard]] operator bool() const noexcept { return mSetup != nullptr; }
+};
 #endif
 
 #endif // PFFFT_H
diff --git a/core/uhjfilter.cpp b/core/uhjfilter.cpp
index 28999e09..e507d705 100644
--- a/core/uhjfilter.cpp
+++ b/core/uhjfilter.cpp
@@ -20,11 +20,6 @@ UhjQualityType UhjEncodeQuality{UhjQualityType::Default};
 
 namespace {
 
-struct PFFFTSetupDeleter {
-    void operator()(PFFFT_Setup *ptr) { pffft_destroy_setup(ptr); }
-};
-using PFFFTSetupPtr = std::unique_ptr<PFFFT_Setup,PFFFTSetupDeleter>;
-
 /* Convolution is implemented using a segmented overlap-add method. The filter
  * response is broken up into multiple segments of 128 samples, and each
  * segment has an FFT applied with a 256-sample buffer (the latter half left
@@ -57,13 +52,11 @@ struct SegmentedFilter {
     static_assert(N >= sFftLength);
     static_assert((N % sSampleLength) == 0);
 
-    PFFFTSetupPtr mFft;
+    PFFFTSetup mFft;
     alignas(16) std::array<float,sFftLength*sNumSegments> mFilterData;
 
-    SegmentedFilter()
+    SegmentedFilter() : mFft{sFftLength, PFFFT_REAL}
     {
-        mFft = PFFFTSetupPtr{pffft_new_setup(sFftLength, PFFFT_REAL)};
-
         using complex_d = std::complex<double>;
         constexpr size_t fft_size{N};
         constexpr size_t half_size{fft_size / 2};
@@ -113,7 +106,7 @@ struct SegmentedFilter {
                 fftTmp[i*2 + 1] = static_cast<float>((i == 0) ? fftBuffer2[sSampleLength].real()
                     : fftBuffer2[i].imag()) / float{sFftLength};
             }
-            pffft_zreorder(mFft.get(), fftTmp.data(), filter, PFFFT_BACKWARD);
+            mFft.zreorder(fftTmp.data(), filter, PFFFT_BACKWARD);
             filter += sFftLength;
         }
     }
@@ -246,7 +239,7 @@ void UhjEncoder<N>::encode(float *LeftOut, float *RightOut,
         std::copy_n(mWXInOut.begin(), sSegmentSize, input);
         std::fill_n(input+sSegmentSize, sSegmentSize, 0.0f);
 
-        pffft_transform(Filter.mFft.get(), input, input, mWorkData.data(), PFFFT_FORWARD);
+        Filter.mFft.transform(input, input, mWorkData.data(), PFFFT_FORWARD);
 
         /* Convolve each input segment with its IR filter counterpart (aligned
          * in time, from newest to oldest).
@@ -255,14 +248,14 @@ void UhjEncoder<N>::encode(float *LeftOut, float *RightOut,
         const float *filter{Filter.mFilterData.data()};
         for(size_t s{curseg};s < sNumSegments;++s)
         {
-            pffft_zconvolve_accumulate(Filter.mFft.get(), input, filter, mFftBuffer.data());
+            Filter.mFft.zconvolve_accumulate(input, filter, mFftBuffer.data());
             input += sFftLength;
             filter += sFftLength;
         }
         input = mWXHistory.data();
         for(size_t s{0};s < curseg;++s)
         {
-            pffft_zconvolve_accumulate(Filter.mFft.get(), input, filter, mFftBuffer.data());
+            Filter.mFft.zconvolve_accumulate(input, filter, mFftBuffer.data());
             input += sFftLength;
             filter += sFftLength;
         }
@@ -270,8 +263,8 @@ void UhjEncoder<N>::encode(float *LeftOut, float *RightOut,
         /* Convert back to samples, writing to the output and storing the extra
          * for next time.
          */
-        pffft_transform(Filter.mFft.get(), mFftBuffer.data(), mFftBuffer.data(),
-            mWorkData.data(), PFFFT_BACKWARD);
+        Filter.mFft.transform(mFftBuffer.data(), mFftBuffer.data(), mWorkData.data(),
+            PFFFT_BACKWARD);
 
         for(size_t i{0};i < sSegmentSize;++i)
             mWXInOut[i] = mFftBuffer[i] + mWXInOut[sSegmentSize+i];