[Mesa-dev] [PATCH 2/3] mesa/main: Add sse2 streaming clamping
Juha-Pekka Heikkila
juhapekka.heikkila at gmail.com
Wed Nov 12 11:47:58 PST 2014
On 12.11.2014 19:36, Bruno Jimenez wrote:
> On Wed, 2014-11-12 at 14:50 +0200, Juha-Pekka Heikkila wrote:
>> Signed-off-by: Juha-Pekka Heikkila <juhapekka.heikkila at gmail.com>
>> ---
>> src/mesa/Makefile.am | 8 +++
>> src/mesa/main/sse2_clamping.c | 138 ++++++++++++++++++++++++++++++++++++++++++
>> src/mesa/main/sse2_clamping.h | 49 +++++++++++++++
>> 3 files changed, 195 insertions(+)
>> create mode 100644 src/mesa/main/sse2_clamping.c
>> create mode 100644 src/mesa/main/sse2_clamping.h
>>
>> diff --git a/src/mesa/Makefile.am b/src/mesa/Makefile.am
>> index 932db4f..43dbe87 100644
>> --- a/src/mesa/Makefile.am
>> +++ b/src/mesa/Makefile.am
>> @@ -111,6 +111,10 @@ if SSE41_SUPPORTED
>> ARCH_LIBS += libmesa_sse41.la
>> endif
>>
>> +if SSE2_SUPPORTED
>> +ARCH_LIBS += libmesa_sse2.la
>> +endif
>> +
>> MESA_ASM_FILES_FOR_ARCH =
>>
>> if HAVE_X86_ASM
>> @@ -155,6 +159,10 @@ libmesa_sse41_la_SOURCES = \
>> main/sse_minmax.c
>> libmesa_sse41_la_CFLAGS = $(AM_CFLAGS) -msse4.1
>>
>> +libmesa_sse2_la_SOURCES = \
>> + main/sse2_clamping.c
>> +libmesa_sse2_la_CFLAGS = $(AM_CFLAGS) -msse2
>> +
>> pkgconfigdir = $(libdir)/pkgconfig
>> pkgconfig_DATA = gl.pc
>>
>> diff --git a/src/mesa/main/sse2_clamping.c b/src/mesa/main/sse2_clamping.c
>> new file mode 100644
>> index 0000000..66c7dc7
>> --- /dev/null
>> +++ b/src/mesa/main/sse2_clamping.c
>> @@ -0,0 +1,138 @@
>> +/*
>> + * 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>
>> + *
>> + */
>> +
>> +#ifdef __SSE2__
>> +#include "main/macros.h"
>> +#include "main/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
>> + */
>> +__attribute__((optimize("unroll-loops")))
>> +void
>> +_mesa_streaming_clamp_float_rgba(const GLuint n, GLfloat rgba_src[][4],
>> + GLfloat rgba_dst[][4], const GLfloat min,
>> + const GLfloat max)
>> +{
>> + int c, prefetch_c;
>> + float* worker = &rgba_src[0][0];
>> + __m128 operand[2], minval, maxval;
>> +
>> + _mm_prefetch((char*) (((unsigned long)worker)|0x1f) + 65, _MM_HINT_T0);
> ^^^^ ^^^
>
> Hi,
>
> May I ask why precisely this numbers?
0x1f as you note below is a typo, should be 0x0f. 65 is cache line
length added with one to even the |0x1f operation.
>
>> +
>> + minval = _mm_set1_ps(min);
>> + maxval = _mm_set1_ps(max);
>> +
>> + for (c = n*4; c > 0 && (((unsigned long)worker)&0x1f) != 0; c--, worker++) {
> ^^^^^
>
> I guess that this is for alignment, but you only need to align to a 16
> bytes boundary, not 32. Or maybe I am missing something obvious.
>
You are correct, 0x1f is typo. should be 0x0f
>> + operand[0] = _mm_load_ss(worker);
>> + operand[0] = _mm_max_ss(operand[0], minval);
>> + operand[0] = _mm_min_ss(operand[0], maxval);
>> + _mm_store_ss(worker, operand[0]);
>> + }
>> +
>> + while (c >= 8) {
>> + _mm_prefetch((char*) worker + 64, _MM_HINT_T0);
> ^^^
>> +
>> + for (prefetch_c = 64/8; prefetch_c > 0 && c >= 8; prefetch_c--, c-=8,
> ^^^^
>
> May I ask also why this numbers?
>
64 is cache line length in bytes, 8 mean this loop handle 8 floats in
one go, operand[0] get 4 floats and same goes for operand[1]. I found
interleaving this way to give more performance, adding more operands did
not give any more performance thus 2 operands to work with which turn
into the number 8. Now that you asked I think 64 has to be divided by 32
and not 8 (8 elements, each 4 bytes -> 32)
Idea here is we send prefetch for next cache line before we start to
work with current line, arriving to next loop iteration we already have
new cache line ready.
>
>> + worker += 8) {
>> +
>> + operand[0] = _mm_load_ps(worker);
>> + operand[1] = _mm_load_ps(worker+4);
>> + operand[0] = _mm_max_ps(operand[0], minval);
>> + operand[1] = _mm_max_ps(operand[1], minval);
>> + operand[0] = _mm_min_ps(operand[0], maxval);
>> + operand[1] = _mm_min_ps(operand[1], maxval);
>> +
>> + _mm_store_ps(worker, operand[0]);
>> + _mm_store_ps(worker+4, operand[1]);
>> + }
>> + }
>> +
>> + for (; c > 0; c--, worker++) {
>> + operand[0] = _mm_load_ss(worker);
>> + operand[0] = _mm_max_ss(operand[0], minval);
>> + operand[0] = _mm_min_ss(operand[0], maxval);
>> + _mm_store_ss(worker, operand[0]);
>> + }
>> +}
>> +
>> +
>> +/* 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/sse2_clamping.h b/src/mesa/main/sse2_clamping.h
>> new file mode 100644
>> index 0000000..688fab7
>> --- /dev/null
>> +++ b/src/mesa/main/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>
>> + *
>> + */
>> +
>> +#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__ */
>
>
More information about the mesa-dev
mailing list