[Libclc-dev] [PATCH] math: Add fdim implementation

Jan Vesely via Libclc-dev libclc-dev at lists.llvm.org
Thu Mar 3 15:57:51 PST 2016


On Tue, 2016-03-01 at 20:23 -0600, Aaron Watry via Libclc-dev wrote:
> Based on the amd-builtin, but explicitly vectorized for all sizes
> (not just
> float4), and includes a vectorized double implementation.

I'm not a big fan of copying bit magic from amd-builtins. In this case
it only avoids branch with the same number of instructions (amdgcn,
there are few more scalar instructions so the branch version might end
up being faster).

that said. I think the patch is OK with few comments to make it more
friendly for quick eyeballing. the code follows 'naive' implementation
pretty closely. with few comments:

Reviewed-by: Jan Vesely <jan.vesely at rutgers.edu>

> 
> Passes piglit (float) tests on pitcairn.
> 
> Signed-off-by: Aaron Watry <awatry at gmail.com>
> ---
> I did test the double implementation on my pitcairn as well, but
> those
> tests aren't in piglit.
> 
> I just copied/pasted the generated float tests, enabled the fp64
> pragma,
> and did s/float/double in the test file.
> 
> I hadn't planned on sending those to piglit unless someone really
> wants them.

I agree, extending generators to generate double variants would be
preferable to introducing individual generated tests.

Jan

PS: sorry for a bit of a rant. I have attached my testing files if you
are interested.

> 
>  generic/include/clc/clc.h         |  1 +
>  generic/include/clc/math/fdim.h   |  2 ++
>  generic/include/clc/math/fdim.inc |  1 +
>  generic/lib/SOURCES               |  1 +
>  generic/lib/math/fdim.cl          | 10 ++++++
>  generic/lib/math/fdim.inc         | 65
> +++++++++++++++++++++++++++++++++++++++
>  6 files changed, 80 insertions(+)
>  create mode 100644 generic/include/clc/math/fdim.h
>  create mode 100644 generic/include/clc/math/fdim.inc
>  create mode 100644 generic/lib/math/fdim.cl
>  create mode 100644 generic/lib/math/fdim.inc
> 
> diff --git a/generic/include/clc/clc.h b/generic/include/clc/clc.h
> index b106923..333fec9 100644
> --- a/generic/include/clc/clc.h
> +++ b/generic/include/clc/clc.h
> @@ -52,6 +52,7 @@
>  #include <clc/math/exp10.h>
>  #include <clc/math/exp2.h>
>  #include <clc/math/fabs.h>
> +#include <clc/math/fdim.h>
>  #include <clc/math/floor.h>
>  #include <clc/math/fma.h>
>  #include <clc/math/fmax.h>
> diff --git a/generic/include/clc/math/fdim.h
> b/generic/include/clc/math/fdim.h
> new file mode 100644
> index 0000000..45115b0
> --- /dev/null
> +++ b/generic/include/clc/math/fdim.h
> @@ -0,0 +1,2 @@
> +#define __CLC_BODY <clc/math/fdim.inc>
> +#include <clc/math/gentype.inc>
> diff --git a/generic/include/clc/math/fdim.inc
> b/generic/include/clc/math/fdim.inc
> new file mode 100644
> index 0000000..ca3467e
> --- /dev/null
> +++ b/generic/include/clc/math/fdim.inc
> @@ -0,0 +1 @@
> +_CLC_OVERLOAD _CLC_DECL __CLC_GENTYPE fdim(__CLC_GENTYPE a,
> __CLC_GENTYPE b);
> diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
> index facb58b..db069ab 100644
> --- a/generic/lib/SOURCES
> +++ b/generic/lib/SOURCES
> @@ -82,6 +82,7 @@ math/exp.cl
>  math/exp_helper.cl
>  math/exp2.cl
>  math/exp10.cl
> +math/fdim.cl
>  math/fmax.cl
>  math/fmin.cl
>  math/fmod.cl
> diff --git a/generic/lib/math/fdim.cl b/generic/lib/math/fdim.cl
> new file mode 100644
> index 0000000..cf25a4a
> --- /dev/null
> +++ b/generic/lib/math/fdim.cl
> @@ -0,0 +1,10 @@
> +#include <clc/clc.h>
> +
> +#include "math.h"
> +
> +#ifdef cl_khr_fp64
> +#pragma OPENCL EXTENSION cl_khr_fp64 : enable
> +#endif
> +
> +#define __CLC_BODY <fdim.inc>
> +#include <clc/math/gentype.inc>
> diff --git a/generic/lib/math/fdim.inc b/generic/lib/math/fdim.inc
> new file mode 100644
> index 0000000..ecfc25c
> --- /dev/null
> +++ b/generic/lib/math/fdim.inc
> @@ -0,0 +1,65 @@
> +/*
> + * Copyright (c) 2014 Advanced Micro Devices, Inc.
> + * Copyright (c) 2016 Aaron Watry
> + *
> + * Permission is hereby granted, free of charge, to any person
> obtaining a copy
> + * of this software and associated documentation files (the
> "Software"), to deal
> + * in the Software without restriction, including without limitation
> the rights
> + * to use, copy, modify, merge, publish, distribute, sublicense,
> and/or sell
> + * copies of the Software, and to permit persons to whom the
> Software is
> + * furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be
> included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
> MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT
> SHALL THE
> + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
> OTHER
> + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> ARISING FROM,
> + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
> DEALINGS IN
> + * THE SOFTWARE.
> + */
> +#if __CLC_FPSIZE == 32
> +#ifdef __CLC_SCALAR
> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fdim(__CLC_GENTYPE x,
> __CLC_GENTYPE y) {
> +    int n = -(isnan(x) | isnan(y)) & QNANBITPATT_SP32;
> +    int r = -(x > y) & as_int(x - y);
> +    return as_float(n | r);
> +}
> +#define __CLC_FDIM_VEC(width) \
> +_CLC_OVERLOAD _CLC_DEF float##width fdim(float##width x,
> float##width y) { \
> +    int##width n = ~((x == x) & (y == y)) & QNANBITPATT_SP32; \
> +    int##width r = (x > y) & as_int##width(x - y); \
> +    return as_float##width(n | r); \
> +}
> +__CLC_FDIM_VEC(2)
> +__CLC_FDIM_VEC(3)
> +__CLC_FDIM_VEC(4)
> +__CLC_FDIM_VEC(8)
> +__CLC_FDIM_VEC(16)
> +#undef __CLC_FDIM_VEC
> +#endif
> +#endif
> +
> +#if __CLC_FPSIZE == 64
> +#ifdef __CLC_SCALAR
> +_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE fdim(__CLC_GENTYPE x, private
> __CLC_GENTYPE y) {
> +    long n = -(isnan(x) | isnan(y)) & QNANBITPATT_DP64;
> +    long r = -(x > y) & as_long(x - y);
> +    return as_double(n | r);
> +}
> +#define __CLC_FDIM_VEC(width) \
> +_CLC_OVERLOAD _CLC_DEF double##width fdim(double##width x,
> double##width y) { \
> +    long##width n = ~((x == x) & (y == y)) & QNANBITPATT_DP64; \
> +    long##width r = (x > y) & as_long##width(x - y); \
> +    return as_double##width(n | r); \
> +}
> +__CLC_FDIM_VEC(2)
> +__CLC_FDIM_VEC(3)
> +__CLC_FDIM_VEC(4)
> +__CLC_FDIM_VEC(8)
> +__CLC_FDIM_VEC(16)
> +#undef __CLC_FDIM_VEC
> +#endif
> +#endif
-- 
Jan Vesely <jan.vesely at rutgers.edu>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: fdim_test.tgz
Type: application/x-compressed-tar
Size: 3126 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/libclc-dev/attachments/20160303/69ac721f/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part
URL: <http://lists.llvm.org/pipermail/libclc-dev/attachments/20160303/69ac721f/attachment.sig>


More information about the Libclc-dev mailing list