diff options
Diffstat (limited to 'numpy/core')
| -rw-r--r-- | numpy/core/include/numpy/npy_math.h | 9 | ||||
| -rw-r--r-- | numpy/core/src/umath/loops.c.src | 6 | ||||
| -rw-r--r-- | numpy/core/src/umath/simd.inc.src | 17 |
3 files changed, 20 insertions, 12 deletions
diff --git a/numpy/core/include/numpy/npy_math.h b/numpy/core/include/numpy/npy_math.h index 7831dd3d7..126b861bf 100644 --- a/numpy/core/include/numpy/npy_math.h +++ b/numpy/core/include/numpy/npy_math.h @@ -178,6 +178,15 @@ NPY_INPLACE npy_longlong npy_gcdll(npy_longlong a, npy_longlong b); NPY_INPLACE npy_longlong npy_lcmll(npy_longlong a, npy_longlong b); /* + * avx function has a common API for both sin & cos. This enum is used to + * distinguish between the two + */ +typedef enum { + npy_compute_sin, + npy_compute_cos +} NPY_TRIG_OP; + +/* * C99 double math funcs */ NPY_INPLACE double npy_sin(double x); diff --git a/numpy/core/src/umath/loops.c.src b/numpy/core/src/umath/loops.c.src index 80bd87875..2028a5712 100644 --- a/numpy/core/src/umath/loops.c.src +++ b/numpy/core/src/umath/loops.c.src @@ -1645,17 +1645,17 @@ FLOAT_@func@_@isa@(char **args, npy_intp *dimensions, npy_intp *steps, void *NPY /**begin repeat1 * #func = cos, sin# + * #enum = npy_compute_cos, npy_compute_sin# * #scalarf = npy_cosf, npy_sinf# */ NPY_NO_EXPORT NPY_GCC_OPT_3 void FLOAT_@func@_@isa@(char **args, npy_intp *dimensions, npy_intp *steps, void *NPY_UNUSED(data)) { - char str[] = "@func@"; - if (!run_unary_@isa@_sincos_FLOAT(args, dimensions, steps, str)) { + if (!run_unary_@isa@_sincos_FLOAT(args, dimensions, steps, @enum@)) { UNARY_LOOP { #if defined @CHK@ && defined NPY_HAVE_SSE2_INTRINSICS - @ISA@_sincos_FLOAT((npy_float *)op1, (npy_float *)ip1, 1, steps[0], str); + @ISA@_sincos_FLOAT((npy_float *)op1, (npy_float *)ip1, 1, steps[0], @enum@); #else const npy_float in1 = *(npy_float *)ip1; *(npy_float *)op1 = @scalarf@(in1); diff --git a/numpy/core/src/umath/simd.inc.src b/numpy/core/src/umath/simd.inc.src index 43a0ddaad..4073ad47a 100644 --- a/numpy/core/src/umath/simd.inc.src +++ b/numpy/core/src/umath/simd.inc.src @@ -165,15 +165,15 @@ run_unary_@isa@_@func@_FLOAT(char **args, npy_intp *dimensions, npy_intp *steps) #if defined @CHK@ && defined NPY_HAVE_SSE2_INTRINSICS static NPY_INLINE void -@ISA@_sincos_FLOAT(npy_float *, npy_float *, const npy_intp n, const npy_intp steps, char*); +@ISA@_sincos_FLOAT(npy_float *, npy_float *, const npy_intp n, const npy_intp steps, NPY_TRIG_OP); #endif static NPY_INLINE int -run_unary_@isa@_sincos_FLOAT(char **args, npy_intp *dimensions, npy_intp *steps, char* mychar) +run_unary_@isa@_sincos_FLOAT(char **args, npy_intp *dimensions, npy_intp *steps, NPY_TRIG_OP my_trig_op) { #if defined @CHK@ && defined NPY_HAVE_SSE2_INTRINSICS if (IS_OUTPUT_BLOCKABLE_UNARY(sizeof(npy_float), @REGISTER_SIZE@)) { - @ISA@_sincos_FLOAT((npy_float*)args[1], (npy_float*)args[0], dimensions[0], steps[0], mychar); + @ISA@_sincos_FLOAT((npy_float*)args[1], (npy_float*)args[0], dimensions[0], steps[0], my_trig_op); return 1; } else @@ -1488,14 +1488,12 @@ static NPY_GCC_OPT_3 NPY_GCC_TARGET_@ISA@ void npy_float * ip, const npy_intp array_size, const npy_intp steps, - char* operation) + NPY_TRIG_OP my_trig_op) { const npy_intp stride = steps/sizeof(npy_float); const npy_int num_lanes = @BYTES@/sizeof(npy_float); - npy_int compute_cos = 1; npy_float large_number = 71476.0625f; - if (*operation == 's') { - compute_cos = 0; + if (my_trig_op == npy_compute_sin) { large_number = 117435.992f; } @@ -1575,8 +1573,9 @@ static NPY_GCC_OPT_3 NPY_GCC_TARGET_@ISA@ void sin_invf5, sin_invf3, zero_f); iquadrant = _mm@vsize@_cvtps_epi32(quadrant); - if (compute_cos) + if (my_trig_op == npy_compute_cos) { iquadrant = _mm@vsize@_add_epi32(iquadrant, ones); + } /* blend sin and cos based on the quadrant */ sine_mask = @isa@_should_calculate_sine(iquadrant, ones, zeros); @@ -1591,7 +1590,7 @@ static NPY_GCC_OPT_3 NPY_GCC_TARGET_@ISA@ void } /* process elements using glibc for large elements */ - if (compute_cos) { + if (my_trig_op == npy_compute_cos) { for (int ii = 0; iglibc_mask != 0; ii++) { if (iglibc_mask & 0x01) { op[ii] = npy_cosf(ip[ii]); |