diff options
| -rw-r--r-- | numpy/core/setup_common.py | 2 | ||||
| -rw-r--r-- | numpy/core/src/umath/scalarmath.c.src | 60 | ||||
| -rw-r--r-- | numpy/core/tests/test_scalarmath.py | 2 |
3 files changed, 30 insertions, 34 deletions
diff --git a/numpy/core/setup_common.py b/numpy/core/setup_common.py index 1143872b1..44fa0a651 100644 --- a/numpy/core/setup_common.py +++ b/numpy/core/setup_common.py @@ -183,6 +183,8 @@ OPTIONAL_FUNCTION_ATTRIBUTES = [('__attribute__((optimize("unroll-loops")))', 'attribute_optimize_unroll_loops'), ('__attribute__((optimize("O3")))', 'attribute_optimize_opt_3'), + ('__attribute__((optimize("O2")))', + 'attribute_optimize_opt_2'), ('__attribute__((nonnull (1)))', 'attribute_nonnull'), ('__attribute__((target ("avx")))', diff --git a/numpy/core/src/umath/scalarmath.c.src b/numpy/core/src/umath/scalarmath.c.src index 3bcb62f7c..42b6d2ed6 100644 --- a/numpy/core/src/umath/scalarmath.c.src +++ b/numpy/core/src/umath/scalarmath.c.src @@ -404,7 +404,7 @@ half_ctype_divmod(npy_half a, npy_half b, npy_half *out1, npy_half *out2) * #rtype = npy_float, npy_double, npy_longdouble# * #c = f,,l# */ -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_add(@type@ a, @type@ b, @type@ *out) { out->real = a.real + b.real; @@ -412,7 +412,7 @@ NPY_FINLINE int return 0; } -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_subtract(@type@ a, @type@ b, @type@ *out) { out->real = a.real - b.real; @@ -420,8 +420,13 @@ NPY_FINLINE int return 0; } -NPY_FINLINE int -@name@_ctype_multiply(@type@ a, @type@ b, @type@ *out) + +/* + * TODO: Mark as to work around FPEs not being issues on clang 12. + * This should be removed when possible. + */ +static NPY_INLINE int +@name@_ctype_multiply( @type@ a, @type@ b, @type@ *out) { out->real = a.real * b.real - a.imag * b.imag; out->imag = a.real * b.imag + a.imag * b.real; @@ -429,7 +434,7 @@ NPY_FINLINE int } /* Use the ufunc loop directly to avoid duplicating the complicated logic */ -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_divide(@type@ a, @type@ b, @type@ *out) { char *args[3] = {(char *)&a, (char *)&b, (char *)out}; @@ -441,18 +446,6 @@ NPY_FINLINE int #define @name@_ctype_true_divide @name@_ctype_divide -/* - * TODO: This function should NOT exist! We removed it from ufuncs! - */ -NPY_FINLINE int -@name@_ctype_floor_divide(@type@ a, @type@ b, @type@ *out) -{ - out->imag = 0; - return @rname@_ctype_floor_divide( - a.real * b.real + a.imag * b.imag, - b.real * b.real + b.imag * b.imag, - &out->real); -} /**end repeat**/ @@ -462,7 +455,7 @@ NPY_FINLINE int * longlong, ulonglong# */ -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_divmod(npy_@name@ a, npy_@name@ b, npy_@name@ *out, npy_@name@ *out2) { int res = @name@_ctype_floor_divide(a, b, out); @@ -530,7 +523,7 @@ half_ctype_negative(npy_half a, npy_half *out) * #name = cfloat, cdouble, clongdouble# * #type = npy_cfloat, npy_cdouble, npy_clongdouble# */ -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_negative(@type@ a, @type@ *out) { out->real = -a.real; @@ -560,7 +553,7 @@ static NPY_INLINE int * #type = npy_cfloat, npy_cdouble, npy_clongdouble# * #c = f,,l# */ -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_positive(@type@ a, @type@ *out) { out->real = a.real; @@ -568,7 +561,7 @@ NPY_FINLINE int return 0; } -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_power(@type@ a, @type@ b, @type@ *out) { *out = npy_cpow@c@(a, b); @@ -624,7 +617,7 @@ half_ctype_absolute(npy_half a, npy_half *out) * #rtype = npy_float, npy_double, npy_longdouble# * #c = f,,l# */ -NPY_FINLINE int +static NPY_INLINE int @name@_ctype_absolute(@type@ a, @rtype@ *out) { *out = npy_cabs@c@(a); @@ -1124,6 +1117,9 @@ static PyObject * } } +#if @fperr@ + npy_clear_floatstatus_barrier((char*)&arg1); +#endif if (is_forward) { arg1 = PyArrayScalar_VAL(a, @Name@); arg2 = other_val; @@ -1134,7 +1130,7 @@ static PyObject * } /* - * Prepare the actual calculation: + * Prepare the actual calculation. */ @otype@ out; #if @twoout@ @@ -1142,9 +1138,7 @@ static PyObject * PyObject *obj; #endif -#if @fperr@ - npy_clear_floatstatus_barrier((char*)&out); -#endif + /* * here we do the actual calculation with arg1 and arg2 @@ -1154,9 +1148,9 @@ static PyObject * * the following `npy_get_floatstatus_barrier`. */ #if @twoout@ - int retstatus = @name@_ctype_@oper@(arg1, arg2, (@otype@ *)&out, &out2); + int retstatus = @name@_ctype_@oper@(arg1, arg2, &out, &out2); #else - int retstatus = @name@_ctype_@oper@(arg1, arg2, (@otype@ *)&out); + int retstatus = @name@_ctype_@oper@(arg1, arg2, &out); #endif #if @fperr@ @@ -1289,6 +1283,10 @@ static PyObject * } } +#if !@isint@ + npy_clear_floatstatus_barrier((char*)&arg1); +#endif + if (is_forward) { arg1 = PyArrayScalar_VAL(a, @Name@); arg2 = other_val; @@ -1303,9 +1301,6 @@ static PyObject * */ @type@ out; -#if !@isint@ - npy_clear_floatstatus_barrier((char*)&out); -#endif /* * here we do the actual calculation with arg1 and arg2 * as a function call. @@ -1588,7 +1583,7 @@ static PyObject * * #to_ctype = , , , , , , , , , , npy_half_to_double, , , , , , # * #func = PyFloat_FromDouble*17# */ -static NPY_INLINE PyObject * +static PyObject * @name@_float(PyObject *obj) { #if @cmplx@ @@ -1638,7 +1633,6 @@ static PyObject* * Extract the other value (if it is compatible). */ int res = convert_to_@name@(other, &arg2); - switch (res) { case CONVERSION_ERROR: return NULL; /* an error occurred (should never happen) */ diff --git a/numpy/core/tests/test_scalarmath.py b/numpy/core/tests/test_scalarmath.py index 1e4450fc7..56a7c1567 100644 --- a/numpy/core/tests/test_scalarmath.py +++ b/numpy/core/tests/test_scalarmath.py @@ -949,7 +949,7 @@ ops_with_names = [ @pytest.mark.parametrize(["__op__", "__rop__", "op", "cmp"], ops_with_names) -@pytest.mark.parametrize("sctype", [np.float32, np.longdouble]) +@pytest.mark.parametrize("sctype", [np.float32, np.float64, np.longdouble]) def test_subclass_deferral(sctype, __op__, __rop__, op, cmp): """ This test covers scalar subclass deferral. Note that this is exceedingly |