diff options
| author | Matti Picus <matti.picus@gmail.com> | 2020-06-12 10:34:42 +0300 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2020-06-12 10:34:42 +0300 |
| commit | 9eabe2f87dbe087e878983bdd69edd8a9084c358 (patch) | |
| tree | 13ae82c9ab6c631f88dea9c305d3ae4eb55a85eb | |
| parent | a89f3ebaec7441f4ba5e30eb07206c2a7269778e (diff) | |
| parent | 02a7aee4adeb22ecc3a166811989ab9c5d5960f6 (diff) | |
| download | numpy-9eabe2f87dbe087e878983bdd69edd8a9084c358.tar.gz | |
Merge pull request #16572 from xiegengxin/fix-sin-cos
BUG: fix sin/cos bug when input is strided array
| -rw-r--r-- | numpy/core/src/umath/simd.inc.src | 8 | ||||
| -rw-r--r-- | numpy/core/tests/test_umath.py | 10 |
2 files changed, 10 insertions, 8 deletions
diff --git a/numpy/core/src/umath/simd.inc.src b/numpy/core/src/umath/simd.inc.src index 48e89915c..e6414e29e 100644 --- a/numpy/core/src/umath/simd.inc.src +++ b/numpy/core/src/umath/simd.inc.src @@ -2698,17 +2698,17 @@ static NPY_GCC_OPT_3 NPY_GCC_TARGET_@ISA@ void /* process elements using glibc for large elements */ if (my_trig_op == npy_compute_cos) { - for (int ii = 0; iglibc_mask != 0; ii++) { + for (int ii = 0, jj = 0; iglibc_mask != 0; ii++, jj += stride) { if (iglibc_mask & 0x01) { - op[ii] = npy_cosf(ip[ii]); + op[ii] = npy_cosf(ip[jj]); } iglibc_mask = iglibc_mask >> 1; } } else { - for (int ii = 0; iglibc_mask != 0; ii++) { + for (int ii = 0, jj = 0; iglibc_mask != 0; ii++, jj += stride) { if (iglibc_mask & 0x01) { - op[ii] = npy_sinf(ip[ii]); + op[ii] = npy_sinf(ip[jj]); } iglibc_mask = iglibc_mask >> 1; } diff --git a/numpy/core/tests/test_umath.py b/numpy/core/tests/test_umath.py index 91acd6ac3..f836af168 100644 --- a/numpy/core/tests/test_umath.py +++ b/numpy/core/tests/test_umath.py @@ -890,15 +890,17 @@ class TestAVXFloat32Transcendental: sizes = np.arange(2,100) for ii in sizes: x_f32 = np.float32(np.random.uniform(low=0.01,high=88.1,size=ii)) + x_f32_large = x_f32.copy() + x_f32_large[3:-1:4] = 120000.0 exp_true = np.exp(x_f32) log_true = np.log(x_f32) - sin_true = np.sin(x_f32) - cos_true = np.cos(x_f32) + sin_true = np.sin(x_f32_large) + cos_true = np.cos(x_f32_large) for jj in strides: assert_array_almost_equal_nulp(np.exp(x_f32[::jj]), exp_true[::jj], nulp=2) assert_array_almost_equal_nulp(np.log(x_f32[::jj]), log_true[::jj], nulp=2) - assert_array_almost_equal_nulp(np.sin(x_f32[::jj]), sin_true[::jj], nulp=2) - assert_array_almost_equal_nulp(np.cos(x_f32[::jj]), cos_true[::jj], nulp=2) + assert_array_almost_equal_nulp(np.sin(x_f32_large[::jj]), sin_true[::jj], nulp=2) + assert_array_almost_equal_nulp(np.cos(x_f32_large[::jj]), cos_true[::jj], nulp=2) class TestLogAddExp(_FilterInvalids): def test_logaddexp_values(self): |