[Mesa-dev] [PATCH][RFC] mesa/main: Clamp rgba with streamed sse

[Roland Scheidegger]

Am 31.10.2014 um 18:17 schrieb Matt Turner:
> On Fri, Oct 31, 2014 at 3:13 AM, Juha-Pekka Heikkila
> <juhapekka.heikkila at gmail.com> wrote:
>> Signed-off-by: Juha-Pekka Heikkila <juhapekka.heikkila at gmail.com>
>> ---
>>  src/mesa/main/colormac.h      | 20 +++++++++++++++
>>  src/mesa/main/pixeltransfer.c | 59 ++++++++++++++++++++++++++++++++-----------
>>  2 files changed, 64 insertions(+), 15 deletions(-)
>>
>> diff --git a/src/mesa/main/colormac.h b/src/mesa/main/colormac.h
>> index c8adca6..da5e094 100644
>> --- a/src/mesa/main/colormac.h
>> +++ b/src/mesa/main/colormac.h
>> @@ -51,6 +51,26 @@ _mesa_unclamped_float_rgba_to_ubyte(GLubyte dst[4], const GLfloat src[4])
>>
>>
>>  /**
>> + * Clamp four float values to [min,max]
>> + */
>> +#if defined(__SSE2__) && defined(__GNUC__)
>> +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);
> 
> Surely 128-bit pixels will be 128-bit aligned? I think we can use an
> aligned load here.
I can't see why that would be the case (sure in some cases it might be
but I don't see anything which would actually guarantee that). temp
images themselves don't seem to have any forced alignment (that could be
fixed though I didn't check fi there's other paths into it which can't
be easily fixed).

> 
>> +    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);
> 
> And an aligned store here.
> 
>> +}
>> +#endif
>> +
>> +
>> +/**
>>   * \name Generic color packing macros.  All inputs should be GLubytes.
>>   *
>>   * \todo We may move these into texstore.h at some point.
>> diff --git a/src/mesa/main/pixeltransfer.c b/src/mesa/main/pixeltransfer.c
>> index 8bbeeb8..e16eb59 100644
>> --- a/src/mesa/main/pixeltransfer.c
>> +++ b/src/mesa/main/pixeltransfer.c
>> @@ -35,7 +35,7 @@
>>  #include "pixeltransfer.h"
>>  #include "imports.h"
>>  #include "mtypes.h"
>> -
>> +#include "x86/common_x86_asm.h"
>>
>>  /*
>>   * Apply scale and bias factors to an array of RGBA pixels.
>> @@ -89,16 +89,34 @@ _mesa_map_rgba( const struct gl_context *ctx, GLuint n, GLfloat rgba[][4] )
>>     const GLfloat *bMap = ctx->PixelMaps.BtoB.Map;
>>     const GLfloat *aMap = ctx->PixelMaps.AtoA.Map;
>>     GLuint i;
>> -   for (i=0;i<n;i++) {
>> -      GLfloat r = CLAMP(rgba[i][RCOMP], 0.0F, 1.0F);
>> -      GLfloat g = CLAMP(rgba[i][GCOMP], 0.0F, 1.0F);
>> -      GLfloat b = CLAMP(rgba[i][BCOMP], 0.0F, 1.0F);
>> -      GLfloat a = CLAMP(rgba[i][ACOMP], 0.0F, 1.0F);
>> -      rgba[i][RCOMP] = rMap[F_TO_I(r * rscale)];
>> -      rgba[i][GCOMP] = gMap[F_TO_I(g * gscale)];
>> -      rgba[i][BCOMP] = bMap[F_TO_I(b * bscale)];
>> -      rgba[i][ACOMP] = aMap[F_TO_I(a * ascale)];
>> +
>> +#if defined(__SSE2__) && defined(__GNUC__)
>> +   if (cpu_has_xmm2) {
> 
> #ifdef __SSE2__ means the compiler is free to use SSE2 instructions
> whenever it pleases. That's not what you want here, if you're also
> doing runtime checking (cpu_has_xmm2).
> 
> The typical way to do this is to put the function containing SSE
> instructions in a separate file that is compiled with -msse2. See
> streaming-load-memcpy.c for example. gcc has a way to mark specific
> functions, but since we have to compile with MSVC...
> 
> I think you just want copy-and-paste the SSE4.1 testing code in
> configure.ac for SSE2 and then wrap these uses in #ifdef USE_SSE2.
> 
>> +      for (i=0;i<n;i++) {
>> +         GLfloat rgba_temp[4];
>> +         _mesa_clamp_float_rgba(rgba[i], rgba_temp, 0.0F, 1.0F);
>> +         rgba[i][RCOMP] = rMap[F_TO_I(rgba_temp[RCOMP] * rscale)];
>> +         rgba[i][GCOMP] = gMap[F_TO_I(rgba_temp[GCOMP] * gscale)];
>> +         rgba[i][BCOMP] = bMap[F_TO_I(rgba_temp[BCOMP] * bscale)];
>> +         rgba[i][ACOMP] = aMap[F_TO_I(rgba_temp[ACOMP] * ascale)];
> 
> Oh, but we shouldn't be bothering to store the floats back to memory
> anyway. We should just do this part with SSE as well.

Indeed the mul scale and F_TO_I look like obvious candidates for
vectorization (though the table part unfortunately not).

Roland