ENH: AVX support for exp/log for strided float32 arrays #13581
Conversation
(1) AVX2 and AVX512F supports gather_ps instruction to load strided data into a register. Rather than resort to scalar, these provide good benefit when the input array is strided. (2) Added some tests to validate AVX based algorithms for exp and log. The tests compare the output of float32 against glibc's float64 implementation and verify maxulp error.
|
|
Seems like a random CI failure nothing to do with the patch: |
numpy/core/src/umath/simd.inc.src
Outdated
| @vtype@ x = @isa@_masked_load(load_mask, ip); | ||
| num_lanes); | ||
| @vtype@ x; | ||
| if (stride == 1) |
There was a problem hiding this comment.
always put { } around loops and conditions also if it is only one line
There was a problem hiding this comment.
sorry should have been more specific the numpy style for conditionals and loops is:
if () {
}
else {
}
There was a problem hiding this comment.
ah sorry about that, should be more consistent now.
| strides = np.random.randint(low=-100, high=100, size=100) | ||
| sizes = np.random.randint(low=1, high=2000, size=100) | ||
| for ii in sizes: | ||
| x_f32 = np.float32(np.random.uniform(low=0.01,high=88.1,size=ii)) |
There was a problem hiding this comment.
denormal floats are excluded here. how is the accuracy for these?
There was a problem hiding this comment.
These tests are meant to be like a sanity check and are not comprehensive at all. It will be slow to test for a large sample of float32's. But the MAXULP error of 2.6 and 3.9 hold even for denormals (this is something I validated separately by enumerating all float32 numbers).
|
looks good. I get very different performance numbers on my AMD Ryzen 7 1700X with glibc 2.27
on an old intel i5-4310M with the same os it is significantly faster
|
7fc34bb to
3251618
Compare
numpy/core/src/umath/simd.inc.src
Outdated
| const npy_int num_lanes = @BYTES@/sizeof(npy_float); | ||
| npy_int indexarr[16]; | ||
| for (npy_int ii = 0; ii < 16; ii++) |
There was a problem hiding this comment.
also add braces
for () {
}
numpy/core/src/umath/simd.inc.src
Outdated
| const npy_int num_lanes = @BYTES@/sizeof(npy_float); | ||
| npy_float xmax = 88.72283935546875f; | ||
| npy_float xmin = -87.3365478515625f; | ||
| npy_int indexarr[16]; | ||
| for (npy_int ii = 0; ii < 16; ii++) |
There was a problem hiding this comment.
also add braces
for () {
}
3251618 to
59b2a1d
Compare
|
As a point of interest, are these functions currently working correctly in master? |
Yes, this patch was only addressing performance issues for strided arrays. |
|
It seems there is an issue with clang 7.0, see #13586 |
(1) AVX2 and AVX512F supports gather_ps instruction to load strided data
into a register. Rather than resort to scalar, these provide good
benefit when the input array is strided.
(2) Added some tests to validate AVX based algorithms for exp and log.
The tests compare the output of float32 against glibc's float64
implementation and verify maxulp error.