[Mesa-dev] [PATCH 2/3][RFC v2] mesa/main/x86: Add sse2 streaming clamping
Juha-Pekka Heikkila
juhapekka.heikkila at gmail.com
Wed Nov 5 01:00:01 PST 2014
On 04.11.2014 21:46, Patrick Baggett wrote:
>
>
> On Tue, Nov 4, 2014 at 6:05 AM, Juha-Pekka Heikkila
> <juhapekka.heikkila at gmail.com <mailto:juhapekka.heikkila at gmail.com>> wrote:
>
> Signed-off-by: Juha-Pekka Heikkila <juhapekka.heikkila at gmail.com
> <mailto:juhapekka.heikkila at gmail.com>>
> ---
> src/mesa/Makefile.am | 8 +++
> src/mesa/main/x86/sse2_clamping.c | 103
> ++++++++++++++++++++++++++++++++++++++
> src/mesa/main/x86/sse2_clamping.h | 49 ++++++++++++++++++
> 3 files changed, 160 insertions(+)
> create mode 100644 src/mesa/main/x86/sse2_clamping.c
> create mode 100644 src/mesa/main/x86/sse2_clamping.h
>
> diff --git a/src/mesa/Makefile.am b/src/mesa/Makefile.am
> index e71bccb..5d3c6f5 100644
> --- a/src/mesa/Makefile.am
> +++ b/src/mesa/Makefile.am
> @@ -111,6 +111,10 @@ if SSE41_SUPPORTED
> ARCH_LIBS += libmesa_sse41.la <http://libmesa_sse41.la>
> endif
>
> +if SSE2_SUPPORTED
> +ARCH_LIBS += libmesa_sse2.la <http://libmesa_sse2.la>
> +endif
> +
> MESA_ASM_FILES_FOR_ARCH =
>
> if HAVE_X86_ASM
> @@ -154,6 +158,10 @@ libmesa_sse41_la_SOURCES = \
> main/streaming-load-memcpy.c
> libmesa_sse41_la_CFLAGS = $(AM_CFLAGS) -msse4.1
>
> +libmesa_sse2_la_SOURCES = \
> + main/x86/sse2_clamping.c
> +libmesa_sse2_la_CFLAGS = $(AM_CFLAGS) -msse2
> +
> pkgconfigdir = $(libdir)/pkgconfig
> pkgconfig_DATA = gl.pc
>
> diff --git a/src/mesa/main/x86/sse2_clamping.c
> b/src/mesa/main/x86/sse2_clamping.c
> new file mode 100644
> index 0000000..7df1c85
> --- /dev/null
> +++ b/src/mesa/main/x86/sse2_clamping.c
> @@ -0,0 +1,103 @@
> +/*
> + * Copyright © 2014 Intel Corporation
> + *
> + * 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 (including
> the next
> + * paragraph) 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.
> + *
> + * Authors:
> + * Juha-Pekka Heikkila <juhapekka.heikkila at gmail.com
> <mailto:juhapekka.heikkila at gmail.com>>
> + *
> + */
> +
> +#ifdef __SSE2__
> +#include "main/macros.h"
> +#include "main/x86/sse2_clamping.h"
> +#include <emmintrin.h>
> +
> +/**
> + * Clamp four float values to [min,max]
> + */
> +static inline void
> +_mesa_clamp_float_rgba(GLfloat src[4], GLfloat result[4], const
> float min,
> + const float max)
> +{
> + __m128 operand, minval, maxval;
> +
> + operand = _mm_loadu_ps(src);
> + minval = _mm_set1_ps(min);
> + maxval = _mm_set1_ps(max);
> + operand = _mm_max_ps(operand, minval);
> + operand = _mm_min_ps(operand, maxval);
> + _mm_storeu_ps(result, operand);
> +}
> +
> +
> +/* Clamp n amount float rgba pixels to [min,max] using SSE2
>
>
> Conceptually, _mesa_streaming_clamp_float_rgba() is clamping a
> contiguous array of floats to some min/max value. The fact that they are
> pixels is somewhat incidental when looking at it from a stream
> perspective. It looks like the code is more or less just operating on
> n*4 floats. Given that, a more efficient implementation would check
> alignment and then use aligned loads and streaming stores. It doesn't
> really matter if you straddle pixel boundaries as long as each float is
> operated on. I'm not sure how much effort you want to put into this
> though. :)
>
I was thinking about aligned versus unaligned loads and stores when Matt
commented about it on my first rfc set but didn't do anything about it
yet. I don't know how big difference there could really be gained in
real world, never tested. Google just brought answers saying "it's
substantially faster" but no numbers or real comparisons. It could be
good combination with what Siavash suggested on out of order execution.
Of all this sse stuff this clamping is the thing why I started to write
this patch, other stuff is "collateral damage". I guess I try to rewrite
this part entirely with the new suggestions to see how the performance
changes. Currently this clamping is run on many 16bpp textures when
they're uploaded so its worthy of some attention anyhow.
/Juha-Pekka
>
> + */
> +void
> +_mesa_streaming_clamp_float_rgba(const GLuint n, GLfloat rgba_src[][4],
> + GLfloat rgba_dst[][4], const
> GLfloat min,
> + const GLfloat max)
> +{
> + int i;
> +
> + for (i = 0; i < n; i++) {
> + _mesa_clamp_float_rgba(rgba_src[i], rgba_dst[i], min, max);
> + }
> +}
> +
> +
> +/* Clamp n amount float rgba pixels to [min,max] using SSE2 and apply
> + * scaling and mapping to components.
> + *
> + * this replace handling of [RGBA] channels:
> + * rgba_temp[RCOMP] = CLAMP(rgba[i][RCOMP], 0.0F, 1.0F);
> + * rgba[i][RCOMP] = rMap[F_TO_I(rgba_temp[RCOMP] * scale[RCOMP])];
> + */
> +void
> +_mesa_clamp_float_rgba_scale_and_map(const GLuint n, GLfloat
> rgba_src[][4],
> + GLfloat rgba_dst[][4], const
> GLfloat min,
> + const GLfloat max,
> + const GLfloat scale[4],
> + const GLfloat* rMap, const
> GLfloat* gMap,
> + const GLfloat* bMap, const
> GLfloat* aMap)
> +{
> + int i;
> + GLfloat __attribute__((aligned(16))) temp[4];
> + __m128 *operand = (__m128*) &temp, multiplier, mmove;
> + __m128i truncated_integers;
> +
> + const unsigned int* map_p = (const unsigned int*)
> &truncated_integers;
> +
> + multiplier = _mm_loadu_ps(scale);
> +
> + for(i = 0; i < n; i++) {
> + _mesa_clamp_float_rgba(rgba_src[i], temp, min, max);
> +
> + *operand = _mm_mul_ps(multiplier, *operand);
> + truncated_integers = _mm_cvttps_epi32(*operand);
> + mmove = _mm_set_ps(aMap[map_p[ACOMP]], bMap[map_p[BCOMP]],
> + gMap[map_p[GCOMP]], rMap[map_p[RCOMP]] );
> +
> + _mm_storeu_ps(rgba_dst[i], mmove);
> + }
> +}
> +
> +
> +#endif /* __SSE2__ */
> diff --git a/src/mesa/main/x86/sse2_clamping.h
> b/src/mesa/main/x86/sse2_clamping.h
> new file mode 100644
> index 0000000..688fab7
> --- /dev/null
> +++ b/src/mesa/main/x86/sse2_clamping.h
> @@ -0,0 +1,49 @@
> +/*
> + * Copyright © 2014 Intel Corporation
> + *
> + * 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 (including
> the next
> + * paragraph) 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.
> + *
> + * Authors:
> + * Juha-Pekka Heikkila <juhapekka.heikkila at gmail.com
> <mailto:juhapekka.heikkila at gmail.com>>
> + *
> + */
> +
> +#ifdef __SSE2__
> +
> +/* Clamp n amount float rgba pixels to [min,max] using SSE2
> + */
> +void
> +_mesa_streaming_clamp_float_rgba(const GLuint n, GLfloat rgba_src[][4],
> + GLfloat rgba_dst[][4], const
> GLfloat min,
> + const GLfloat max);
> +
> +
> +/* Clamp n amount float rgba pixels to [min,max] using SSE2 and apply
> + * scaling and mapping to components.
> + */
> +void
> +_mesa_clamp_float_rgba_scale_and_map(const GLuint n, GLfloat
> rgba_src[][4],
> + GLfloat rgba_dst[][4], const
> GLfloat min,
> + const GLfloat max,
> + const GLfloat scale[4],
> + const GLfloat* rMap, const
> GLfloat* gMap,
> + const GLfloat* bMap, const
> GLfloat* aMap);
> +
> +#endif /* __SSE2__ */
> --
> 1.8.5.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org <mailto:mesa-dev at lists.freedesktop.org>
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
>
More information about the mesa-dev
mailing list