diff options
Diffstat (limited to 'numpy/core/src/umath/loops_autovec.dispatch.c.src')
| -rw-r--r-- | numpy/core/src/umath/loops_autovec.dispatch.c.src | 287 |
1 files changed, 287 insertions, 0 deletions
diff --git a/numpy/core/src/umath/loops_autovec.dispatch.c.src b/numpy/core/src/umath/loops_autovec.dispatch.c.src new file mode 100644 index 000000000..bdbfa0f86 --- /dev/null +++ b/numpy/core/src/umath/loops_autovec.dispatch.c.src @@ -0,0 +1,287 @@ +/*@targets + ** $maxopt $autovec baseline + ** sse2 avx2 + ** neon + ** vsx2 + ** vx + **/ +#define _UMATHMODULE +#define _MULTIARRAYMODULE +#define NPY_NO_DEPRECATED_API NPY_API_VERSION + +#include "simd/simd.h" +#include "loops_utils.h" +#include "loops.h" +// Provides the various *_LOOP macros +#include "fast_loop_macros.h" + +/* + ***************************************************************************** + ** INTEGER LOOPS + ***************************************************************************** + */ +/* + * Arithmetic bit shift operations. + * + * Intel hardware masks bit shift values, so large shifts wrap around + * and can produce surprising results. The special handling ensures that + * behavior is independent of compiler or hardware. + * TODO: We could implement consistent behavior for negative shifts, + * which is undefined in C. + */ +#define INT_left_shift_needs_clear_floatstatus +#define UINT_left_shift_needs_clear_floatstatus + +/**begin repeat + * #TYPE = BYTE, UBYTE, SHORT, USHORT, INT, UINT, + * LONG, ULONG, LONGLONG, ULONGLONG# + * #type = npy_byte, npy_ubyte, npy_short, npy_ushort, npy_int, npy_uint, + * npy_long, npy_ulong, npy_longlong, npy_ulonglong# + * #ftype = npy_float, npy_float, npy_float, npy_float, npy_double, npy_double, + * npy_double, npy_double, npy_double, npy_double# + * #SIGNED = 1, 0, 1, 0, 1, 0, 1, 0, 1, 0# + * #c = hh,uhh,h,uh,,u,l,ul,ll,ull# + */ + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_positive) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = +in); +} + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_square) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = in * in); +} + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_reciprocal) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(data)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = 1.0 / in); +} + +/**begin repeat1 + * Arithmetic + * #kind = add, subtract, multiply# + * #OP = +, -, *# + */ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + if (IS_BINARY_REDUCE) { + BINARY_REDUCE_LOOP_FAST(@type@, io1 @OP@= in2); + } + else { + BINARY_LOOP_FAST(@type@, @type@, *out = in1 @OP@ in2); + } +} +/**end repeat1**/ + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_left_shift) +(char **args, npy_intp const *dimensions, npy_intp const *steps, + void *NPY_UNUSED(func)) +{ + BINARY_LOOP_FAST(@type@, @type@, *out = npy_lshift@c@(in1, in2)); +#ifdef @TYPE@_left_shift_needs_clear_floatstatus + // For some reason, our macOS CI sets an "invalid" flag here, but only + // for some types. + npy_clear_floatstatus_barrier((char*)dimensions); +#endif +} + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_right_shift) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ +#ifndef NPY_DO_NOT_OPTIMIZE_@TYPE@_right_shift + BINARY_LOOP_FAST(@type@, @type@, *out = npy_rshift@c@(in1, in2)); +#else + BINARY_LOOP { + @type@ in1 = *(@type@ *)ip1; + @type@ in2 = *(@type@ *)ip2; + *(@type@ *)op1 = npy_rshift@c@(in1, in2); + } +#endif +} +/**end repeat**/ + +/* + ***************************************************************************** + ** UNSIGNED INTEGER LOOPS + ***************************************************************************** + */ +/**begin repeat + * #TYPE = UBYTE, USHORT, UINT, ULONG, ULONGLONG# + * #type = npy_ubyte, npy_ushort, npy_uint, npy_ulong, npy_ulonglong# + * #c = u,u,u,ul,ull# + */ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_absolute) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = in); +} + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_sign) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = in > 0 ? 1 : 0); +} + +/**begin repeat1 + * Arithmetic + * #kind = bitwise_and, bitwise_or, bitwise_xor# + * #OP = &, |, ^# + */ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + if (IS_BINARY_REDUCE) { + BINARY_REDUCE_LOOP_FAST(@type@, io1 @OP@= in2); + } + else { + BINARY_LOOP_FAST(@type@, @type@, *out = in1 @OP@ in2); + } +} +/**end repeat1**/ + +/**begin repeat1 + * #kind = logical_and, logical_or# + * #OP = &&, ||# + */ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + /* + * gcc vectorization of this is not good (PR60575) but manual integer + * vectorization is too tedious to be worthwhile + */ + BINARY_LOOP_FAST(@type@, npy_bool, *out = in1 @OP@ in2); +} +/**end repeat1**/ + +NPY_FINLINE npy_bool @TYPE@_logical_xor_(@type@ in1, @type@ in2) +{ return (!!in1) != (!!in2); } + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_logical_xor) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + BINARY_LOOP_FAST(@type@, npy_bool, *out = @TYPE@_logical_xor_(in1, in2)); +} + +/**begin repeat1 + * #kind = isnan, isinf, isfinite# + * #func = npy_isnan, npy_isinf, npy_isfinite# + * #val = NPY_FALSE, NPY_FALSE, NPY_TRUE# + **/ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + /* + * The (void)in; suppresses an unused variable warning raised by gcc and allows + * us to re-use this macro even though we do not depend on in + */ + UNARY_LOOP_FAST(@type@, npy_bool, (void)in; *out = @val@); +} +/**end repeat1**/ + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_conjugate) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = in); +} + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_logical_not) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, npy_bool, *out = !in); +} + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_invert) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = ~in); +} +/**end repeat**/ + +/* + ***************************************************************************** + ** SIGNED! INTEGER LOOPS + ***************************************************************************** + */ + +/**begin repeat + * #TYPE = BYTE, SHORT, INT, LONG, LONGLONG# + * #type = npy_byte, npy_short, npy_int, npy_long, npy_longlong# + * #c = ,,,l,ll# + */ + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_absolute) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = (in >= 0) ? in : -in); +} + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_sign) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(@type@, @type@, *out = in > 0 ? 1 : (in < 0 ? -1 : 0)); +} + +/**begin repeat1 + * #kind = conjugate, invert, isnan, isinf, isfinite, + * logical_and, logical_or, logical_xor, logical_not, + * bitwise_and, bitwise_or, bitwise_xor# + **/ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_@kind@) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *func) +{ + NPY_CPU_DISPATCH_CURFX(U@TYPE@_@kind@)(args, dimensions, steps, func); +} +/**end repeat1**/ +/**end repeat**/ + +/* + ***************************************************************************** + ** BOOLEAN LOOPS ** + ***************************************************************************** + */ +/**begin repeat + * #kind = isnan, isinf, isfinite# + * #func = npy_isnan, npy_isinf, npy_isfinite# + * #val = NPY_FALSE, NPY_FALSE, NPY_TRUE# + **/ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(BOOL_@kind@) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *func) +{ + NPY_CPU_DISPATCH_CURFX(UBYTE_@kind@)(args, dimensions, steps, func); +} +/**end repeat**/ + +/* + ***************************************************************************** + ** HALF-FLOAT LOOPS ** + ***************************************************************************** + */ + +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(HALF_absolute) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func)) +{ + UNARY_LOOP_FAST(npy_half, npy_half, *out = in&0x7fffu); +} + +/* + ***************************************************************************** + ** DATETIME LOOPS ** + ***************************************************************************** + */ + +/**begin repeat + * #type = npy_datetime, npy_timedelta# + * #TYPE = DATETIME, TIMEDELTA# + */ +NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_isinf) +(char **args, npy_intp const *dimensions, npy_intp const *steps, void *func) +{ + NPY_CPU_DISPATCH_CURFX(ULONGLONG_isinf)(args, dimensions, steps, func); +} +/**end repeat**/ |