diff options
| -rw-r--r-- | common/pffft.cpp | 53 | ||||
| -rw-r--r-- | common/pffft.h | 41 |
2 files changed, 36 insertions, 58 deletions
diff --git a/common/pffft.cpp b/common/pffft.cpp index 9d9dad23..bbfbaa49 100644 --- a/common/pffft.cpp +++ b/common/pffft.cpp @@ -380,7 +380,9 @@ force_inline void vcplxmulconj(v4sf &ar, v4sf &ai, v4sf br, v4sf bi) noexcept #endif //!PFFFT_SIMD_DISABLE /* SSE and co like 16-bytes aligned pointers */ -#define MALLOC_V4SF_ALIGNMENT 64 // with a 64-byte alignment, we are even aligned on L2 cache lines... +/* with a 64-byte alignment, we are even aligned on L2 cache lines... */ +constexpr auto V4sfAlignment = size_t(64); +constexpr auto V4sfAlignVal = std::align_val_t(V4sfAlignment); /* passf2 and passb2 has been merged here, fsign = -1 for passf2, +1 for passb2 @@ -1406,24 +1408,20 @@ void cffti1_ps(const uint n, float *wa, const al::span<uint,15> ifac) } // namespace -void *pffft_aligned_malloc(size_t nb_bytes) -{ return al_malloc(MALLOC_V4SF_ALIGNMENT, nb_bytes); } - -void pffft_aligned_free(void *p) noexcept { al_free(p); } - -int pffft_simd_size() noexcept { return SIMD_SZ; } - +/* NOLINTNEXTLINE(clang-analyzer-optin.performance.Padding) */ struct PFFFT_Setup { - 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; + 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 - al::span<v4sf> e; // N/4*3 elements + float *twiddle{}; /* N/4 elements */ + al::span<v4sf> e; /* N/4*3 elements */ + + alignas(V4sfAlignment) std::byte end; }; -PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform) +gsl::owner<PFFFT_Setup*> pffft_new_setup(unsigned int N, pffft_transform_t transform) { assert(transform == PFFFT_REAL || transform == PFFFT_COMPLEX); assert(N > 0); @@ -1436,23 +1434,25 @@ PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform) else assert((N%(SIMD_SZ*SIMD_SZ)) == 0); - const uint Ncvec = (transform == PFFFT_REAL ? N/2 : N)/SIMD_SZ; - const size_t storelen{sizeof(PFFFT_Setup) + (2_zu*Ncvec * sizeof(v4sf))}; + const uint Ncvec{(transform == PFFFT_REAL ? N/2 : N)/SIMD_SZ}; - void *store{al_calloc(MALLOC_V4SF_ALIGNMENT, storelen)}; - if(!store) return nullptr; + const size_t storelen{std::max(offsetof(PFFFT_Setup, end) + 2_zu*Ncvec*sizeof(v4sf), + sizeof(PFFFT_Setup))}; + gsl::owner<std::byte*> storage{::new(V4sfAlignVal) std::byte[storelen]{}}; + al::span extrastore{&storage[offsetof(PFFFT_Setup, end)], 2_zu*Ncvec*sizeof(v4sf)}; - PFFFT_Setup *s{::new(store) PFFFT_Setup{}}; + gsl::owner<PFFFT_Setup*> s{::new(storage) PFFFT_Setup{}}; s->N = N; s->transform = transform; - /* nb of complex simd vectors */ s->Ncvec = Ncvec; - s->e = {reinterpret_cast<v4sf*>(reinterpret_cast<char*>(s+1)), 2_zu*Ncvec}; - s->twiddle = reinterpret_cast<float*>(&s->e[2u*Ncvec*(SIMD_SZ-1)/SIMD_SZ]); + + const size_t ecount{2_zu*Ncvec*(SIMD_SZ-1)/SIMD_SZ}; + s->e = {std::launder(reinterpret_cast<v4sf*>(extrastore.data())), ecount}; + s->twiddle = std::launder(reinterpret_cast<float*>(&extrastore[ecount*sizeof(v4sf)])); if constexpr(SIMD_SZ > 1) { - auto e = std::vector<float>(2_zu*Ncvec*(SIMD_SZ-1), 0.0f); + auto e = std::vector<float>(s->e.size()*SIMD_SZ, 0.0f); for(size_t k{0};k < s->Ncvec;++k) { const size_t i{k / SIMD_SZ}; @@ -1486,10 +1486,11 @@ PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform) } -void pffft_destroy_setup(PFFFT_Setup *s) noexcept +void pffft_destroy_setup(gsl::owner<PFFFT_Setup*> s) noexcept { std::destroy_at(s); - al_free(s); + auto storage = reinterpret_cast<gsl::owner<std::byte*>>(s); + ::operator delete[](storage, V4sfAlignVal); } #if !defined(PFFFT_SIMD_DISABLE) diff --git a/common/pffft.h b/common/pffft.h index 5ef03820..cf356524 100644 --- a/common/pffft.h +++ b/common/pffft.h @@ -79,15 +79,12 @@ #ifndef PFFFT_H #define PFFFT_H -#include <stddef.h> // for size_t -#include <stdint.h> - -#ifdef __cplusplus #include <cstddef> +#include <cstdint> #include <utility> -extern "C" { -#endif +#include "almalloc.h" + /* opaque struct holding internal stuff (precomputed twiddle factors) this * struct can be shared by many threads as it contains only read-only data. @@ -100,20 +97,14 @@ enum pffft_direction_t { PFFFT_FORWARD, PFFFT_BACKWARD }; /* type of transform */ 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 - /** * Prepare for performing transforms of size N -- the returned PFFFT_Setup * structure is read-only so it can safely be shared by multiple concurrent * threads. */ [[gnu::malloc]] -PFFFT_Setup *pffft_new_setup(unsigned int N, pffft_transform_t transform); -void pffft_destroy_setup(PFFFT_Setup *setup) noexcept; +gsl::owner<PFFFT_Setup*> pffft_new_setup(unsigned int N, pffft_transform_t transform); +void pffft_destroy_setup(gsl::owner<PFFFT_Setup*> setup) noexcept; /** * Perform a Fourier transform. The z-domain data is stored in the most @@ -179,28 +170,14 @@ void pffft_zconvolve_scale_accumulate(const PFFFT_Setup *setup, const float *dft */ void pffft_zconvolve_accumulate(const PFFFT_Setup *setup, const float *dft_a, const float *dft_b, float *dft_ab); -/** - * The float buffers must have the correct alignment (16-byte boundary on intel - * and powerpc). This function may be used to obtain such correctly aligned - * buffers. - */ -[[gnu::alloc_size(1), gnu::malloc]] -void *pffft_aligned_malloc(size_t nb_bytes); -void pffft_aligned_free(void *ptr) noexcept; - -/* Return 4 or 1 depending if vectorization was enable when building pffft.cpp. */ -int pffft_simd_size() noexcept; - -#ifdef __cplusplus -} struct PFFFTSetup { - PFFFT_Setup *mSetup{}; + gsl::owner<PFFFT_Setup*> mSetup{}; PFFFTSetup() = default; PFFFTSetup(const PFFFTSetup&) = delete; PFFFTSetup(PFFFTSetup&& rhs) noexcept : mSetup{rhs.mSetup} { rhs.mSetup = nullptr; } - explicit PFFFTSetup(std::nullptr_t) { } + explicit PFFFTSetup(std::nullptr_t) noexcept { } explicit PFFFTSetup(unsigned int n, pffft_transform_t transform) : mSetup{pffft_new_setup(n, transform)} { } @@ -211,7 +188,8 @@ struct PFFFTSetup { { if(mSetup) pffft_destroy_setup(mSetup); - mSetup = std::exchange(rhs.mSetup, nullptr); + mSetup = rhs.mSetup; + rhs.mSetup = nullptr; return *this; } @@ -234,6 +212,5 @@ struct PFFFTSetup { [[nodiscard]] operator bool() const noexcept { return mSetup != nullptr; } }; -#endif #endif // PFFFT_H |