diff options
| -rw-r--r-- | numpy/core/src/umath/loops.c.src | 3 | ||||
| -rw-r--r-- | numpy/core/src/umath/simd.inc.src | 51 |
2 files changed, 54 insertions, 0 deletions
diff --git a/numpy/core/src/umath/loops.c.src b/numpy/core/src/umath/loops.c.src index 59d144569..a99bdb9d2 100644 --- a/numpy/core/src/umath/loops.c.src +++ b/numpy/core/src/umath/loops.c.src @@ -561,6 +561,9 @@ NPY_NO_EXPORT void BOOL_@kind@(char **args, npy_intp *dimensions, npy_intp *steps, void *NPY_UNUSED(func)) { if(IS_BINARY_REDUCE) { + if (run_reduce_simd_@kind@_BOOL(args, dimensions, steps)) { + return; + } BINARY_REDUCE_LOOP(npy_bool) { const npy_bool in2 = *(npy_bool *)ip2; io1 = io1 @OP@ in2; diff --git a/numpy/core/src/umath/simd.inc.src b/numpy/core/src/umath/simd.inc.src index 05f8710e5..dc3b6ad8e 100644 --- a/numpy/core/src/umath/simd.inc.src +++ b/numpy/core/src/umath/simd.inc.src @@ -197,6 +197,23 @@ run_binary_simd_@kind@_BOOL(char **args, npy_intp *dimensions, npy_intp *steps) return 0; } + +static void +sse2_reduce_@kind@_BOOL(npy_bool * op, npy_bool * ip, npy_intp n); + +static NPY_INLINE int +run_reduce_simd_@kind@_BOOL(char **args, npy_intp *dimensions, npy_intp *steps) +{ +#if defined HAVE_EMMINTRIN_H + if (sizeof(npy_bool) == 1 && IS_BLOCKABLE_REDUCE(sizeof(npy_bool), 16)) { + sse2_reduce_@kind@_BOOL((npy_bool*)args[0], (npy_bool*)args[1], + dimensions[0]); + return 1; + } +#endif + return 0; +} + /**end repeat**/ /**begin repeat @@ -518,10 +535,12 @@ sse2_@kind@_@TYPE@(@type@ * ip, @type@ * op, const npy_intp n) * # kind = logical_or, logical_and# * # and = 0, 1# * # op = ||, &&# + * # sc = !=, ==# * # vpre = _mm*2# * # vsuf = si128*2# * # vtype = __m128i*2# * # type = npy_bool*2# + * # vload = _mm_load_si128*2# * # vloadu = _mm_loadu_si128*2# * # vstore = _mm_store_si128*2# */ @@ -568,6 +587,38 @@ sse2_binary_@kind@_BOOL(npy_bool * op, npy_bool * ip1, npy_bool * ip2, npy_intp } } + +static void +sse2_reduce_@kind@_BOOL(npy_bool * op, npy_bool * ip, const npy_intp n) +{ + const @vtype@ zero = @vpre@_setzero_@vsuf@(); + LOOP_BLOCK_ALIGN_VAR(ip, npy_bool, 16) { + *op = *op @op@ ip[i]; + if (*op @sc@ 0) { + return; + } + } + LOOP_BLOCKED(npy_bool, 16) { + @vtype@ v = @vload@((@vtype@*)&ip[i]); + v = @vpre@_cmpeq_epi8(v, zero); +#if @and@ + if ((@vpre@_movemask_epi8(v) != 0)) { + *op = 0; +#else + if ((@vpre@_movemask_epi8(v) != 0xFFFF)) { + *op = 1; +#endif + return; + } + } + LOOP_BLOCKED_END { + *op = *op @op@ ip[i]; + if (*op @sc@ 0) { + return; + } + } +} + /**end repeat**/ /**begin repeat |