| 1 | This patch replaces #ifndef POCKETFFT_NO_VECTORS by #if POCKETFFT_NO_VECTORS.
|
| 2 | It also makes it the default, as SIMD instructions are not that well-suited
|
| 3 | for substitutes.
|
| 4 |
|
| 5 | diff --git a/pocketfft_hdronly.h b/pocketfft_hdronly.h
|
| 6 | index d75ada6..b2d0a23 100644
|
| 7 | |
| 8 | |
| 9 | @@ -39,6 +39,10 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
| 10 | #ifndef POCKETFFT_HDRONLY_H
|
| 11 | #define POCKETFFT_HDRONLY_H
|
| 12 |
|
| 13 | +#ifndef POCKETFFT_NO_VECTORS
|
| 14 | +#define POCKETFFT_NO_VECTORS 1
|
| 15 | +#endif
|
| 16 | +
|
| 17 | #ifndef __cplusplus
|
| 18 | #error This file is C++ and requires a C++ compiler.
|
| 19 | #endif
|
| 20 | @@ -106,29 +110,29 @@ constexpr bool FORWARD = true,
|
| 21 | BACKWARD = false;
|
| 22 |
|
| 23 | // only enable vector support for gcc>=5.0 and clang>=5.0
|
| 24 | -#ifndef POCKETFFT_NO_VECTORS
|
| 25 | -#define POCKETFFT_NO_VECTORS
|
| 26 | +#if !(POCKETFFT_NO_VECTORS)
|
| 27 | +#define POCKETFFT_NO_VECTORS 1
|
| 28 | #if defined(__INTEL_COMPILER)
|
| 29 | // do nothing. This is necessary because this compiler also sets __GNUC__.
|
| 30 | #elif defined(__clang__)
|
| 31 | // AppleClang has their own version numbering
|
| 32 | #ifdef __apple_build_version__
|
| 33 | # if (__clang_major__ > 9) || (__clang_major__ == 9 && __clang_minor__ >= 1)
|
| 34 | -# undef POCKETFFT_NO_VECTORS
|
| 35 | +#define POCKETFFT_NO_VECTORS 0
|
| 36 | # endif
|
| 37 | #elif __clang_major__ >= 5
|
| 38 | -# undef POCKETFFT_NO_VECTORS
|
| 39 | +#define POCKETFFT_NO_VECTORS 0
|
| 40 | #endif
|
| 41 | #elif defined(__GNUC__)
|
| 42 | #if __GNUC__>=5
|
| 43 | -#undef POCKETFFT_NO_VECTORS
|
| 44 | +#define POCKETFFT_NO_VECTORS 0
|
| 45 | #endif
|
| 46 | #endif
|
| 47 | #endif
|
| 48 |
|
| 49 | template<typename T> struct VLEN { static constexpr size_t val=1; };
|
| 50 |
|
| 51 | -#ifndef POCKETFFT_NO_VECTORS
|
| 52 | +#if !(POCKETFFT_NO_VECTORS)
|
| 53 | #if (defined(__AVX512F__))
|
| 54 | template<> struct VLEN<float> { static constexpr size_t val=16; };
|
| 55 | template<> struct VLEN<double> { static constexpr size_t val=8; };
|
| 56 | @@ -145,7 +149,7 @@ template<> struct VLEN<double> { static constexpr size_t val=2; };
|
| 57 | template<> struct VLEN<float> { static constexpr size_t val=4; };
|
| 58 | template<> struct VLEN<double> { static constexpr size_t val=2; };
|
| 59 | #else
|
| 60 | -#define POCKETFFT_NO_VECTORS
|
| 61 | +#define POCKETFFT_NO_VECTORS 1
|
| 62 | #endif
|
| 63 | #endif
|
| 64 |
|
| 65 | @@ -180,7 +184,7 @@ template<typename T> class arr
|
| 66 | T *p;
|
| 67 | size_t sz;
|
| 68 |
|
| 69 | -#if defined(POCKETFFT_NO_VECTORS)
|
| 70 | +#if POCKETFFT_NO_VECTORS
|
| 71 | static T *ralloc(size_t num)
|
| 72 | {
|
| 73 | if (num==0) return nullptr;
|
| 74 | @@ -3026,7 +3030,7 @@ class rev_iter
|
| 75 | template<typename T> struct VTYPE {};
|
| 76 | template <typename T> using vtype_t = typename VTYPE<T>::type;
|
| 77 |
|
| 78 | -#ifndef POCKETFFT_NO_VECTORS
|
| 79 | +#if !(POCKETFFT_NO_VECTORS)
|
| 80 | template<> struct VTYPE<float>
|
| 81 | {
|
| 82 | using type = float __attribute__ ((vector_size (VLEN<float>::val*sizeof(float))));
|
| 83 | @@ -3139,7 +3143,7 @@ POCKETFFT_NOINLINE void general_nd(const cndarr<T> &in, ndarr<T> &out,
|
| 84 | auto storage = alloc_tmp<T0>(in.shape(), len, sizeof(T));
|
| 85 | const auto &tin(iax==0? in : out);
|
| 86 | multi_iter<vlen> it(tin, out, axes[iax]);
|
| 87 | -#ifndef POCKETFFT_NO_VECTORS
|
| 88 | +#if !(POCKETFFT_NO_VECTORS)
|
| 89 | if (vlen>1)
|
| 90 | while (it.remaining()>=vlen)
|
| 91 | {
|
| 92 | @@ -3245,7 +3249,7 @@ template<typename T> POCKETFFT_NOINLINE void general_r2c(
|
| 93 | constexpr auto vlen = VLEN<T>::val;
|
| 94 | auto storage = alloc_tmp<T>(in.shape(), len, sizeof(T));
|
| 95 | multi_iter<vlen> it(in, out, axis);
|
| 96 | -#ifndef POCKETFFT_NO_VECTORS
|
| 97 | +#if !(POCKETFFT_NO_VECTORS)
|
| 98 | if (vlen>1)
|
| 99 | while (it.remaining()>=vlen)
|
| 100 | {
|
| 101 | @@ -3300,7 +3304,7 @@ template<typename T> POCKETFFT_NOINLINE void general_c2r(
|
| 102 | constexpr auto vlen = VLEN<T>::val;
|
| 103 | auto storage = alloc_tmp<T>(out.shape(), len, sizeof(T));
|
| 104 | multi_iter<vlen> it(in, out, axis);
|
| 105 | -#ifndef POCKETFFT_NO_VECTORS
|
| 106 | +#if !(POCKETFFT_NO_VECTORS)
|
| 107 | if (vlen>1)
|
| 108 | while (it.remaining()>=vlen)
|
| 109 | {
|