OSMesa for gallium would be really helpful. Are there plans to implement some sort of antialising (GL_POLYGON_SMOOTH, fsaa) in softpipe/llvmpipe?
Andreas -----Ursprüngliche Nachricht----- Von: Brian Paul [mailto:bri...@vmware.com] Gesendet: Donnerstag, 11. August 2011 16:31 An: Andreas Fänger Cc: mesa-dev@lists.freedesktop.org Betreff: Re: [Mesa-dev] [PATCH] swrast: initial multi-threaded span rendering These patches don't look too intrusive so I'm OK with them. I'll apply them, test and push. I've been meaning to write a new OSMesa interface for gallium for some time now. That would probably be useful to a few people. I'm tempted to redesign the OSMesa API a bit, but it's probably more important to keep it unchanged for the sake of existing apps. -Brian On 08/10/2011 03:49 AM, Andreas Fänger wrote: > Hi Keith, > > you are right. The main purpose of this patch is to speedup osmesa rendering > as there is no llvmpipe target at the moment. Also llvmpipe is currently > missing some important features like aa/fsaa and anisotropic filtering, which > is available in swrast now. > So I need to stick with the old rasterizer at the moment, with some > improvements. > > Andreas > > -----Ursprüngliche Nachricht----- > Von: Keith Whitwell [mailto:kei...@vmware.com] > Gesendet: Mittwoch, 10. August 2011 11:17 > An: Andreas Fänger > Cc: mesa-dev@lists.freedesktop.org > Betreff: Re: [Mesa-dev] [PATCH] swrast: initial multi-threaded span rendering > > I'm not sure it makes a lot of sense to be optimizing swrast at this > stage. Take a look at llvmpipe and perhaps consider improving the > multithreading already in place in that rasterizer, which is far better > optimized than swrast already. > > Keith > > On Wed, 2011-08-10 at 08:07 +0000, Andreas Fänger wrote: >> Optional parallel rendering of spans using OpenMP. >> Initial implementation for aa triangles. A new option for scons is >> also provided to activate the openmp support (off by default). >> --- >> common.py | 1 + >> scons/gallium.py | 12 +++++++ >> src/mesa/swrast/s_aatritemp.h | 68 >> ++++++++++++++++++++++----------------- >> src/mesa/swrast/s_context.c | 26 ++++++++++++--- >> src/mesa/swrast/s_texcombine.c | 4 ++ >> src/mesa/tnl/t_pipeline.c | 12 +++++++ >> 6 files changed, 87 insertions(+), 36 deletions(-) >> >> diff --git a/common.py b/common.py >> index 8657030..cfee1b5 100644 >> --- a/common.py >> +++ b/common.py >> @@ -88,6 +88,7 @@ def AddOptions(opts): >> opts.Add('toolchain', 'compiler toolchain', default_toolchain) >> opts.Add(BoolOption('gles', 'EXPERIMENTAL: enable OpenGL ES support', >> 'no')) >> opts.Add(BoolOption('llvm', 'use LLVM', default_llvm)) >> + opts.Add(BoolOption('openmp', 'EXPERIMENTAL: compile with openmp >> (swrast)', 'no')) >> opts.Add(BoolOption('debug', 'DEPRECATED: debug build', 'yes')) >> opts.Add(BoolOption('profile', 'DEPRECATED: profile build', 'no')) >> opts.Add(BoolOption('quiet', 'DEPRECATED: profile build', 'yes')) >> diff --git a/scons/gallium.py b/scons/gallium.py >> index 8cd3bc7..7135251 100755 >> --- a/scons/gallium.py >> +++ b/scons/gallium.py >> @@ -596,6 +596,18 @@ def generate(env): >> libs += ['m', 'pthread', 'dl'] >> env.Append(LIBS = libs) >> >> + # OpenMP >> + if env['openmp']: >> + if env['msvc']: >> + env.Append(CCFLAGS = ['/openmp']) >> + # When building openmp release VS2008 link.exe crashes with >> LNK1103 error. >> + # Workaround: overwrite PDB flags with empty value as it isn't >> required anyways >> + if env['build'] == 'release': >> + env['PDB'] = '' >> + if env['gcc']: >> + env.Append(CCFLAGS = ['-fopenmp']) >> + env.Append(LIBS = ['gomp']) >> + >> # Load tools >> env.Tool('lex') >> env.Tool('yacc') >> diff --git a/src/mesa/swrast/s_aatritemp.h b/src/mesa/swrast/s_aatritemp.h >> index 91d4f7a..005d12c 100644 >> --- a/src/mesa/swrast/s_aatritemp.h >> +++ b/src/mesa/swrast/s_aatritemp.h >> @@ -181,13 +181,18 @@ >> const GLfloat *pMax = vMax->attrib[FRAG_ATTRIB_WPOS]; >> const GLfloat dxdy = majDx / majDy; >> const GLfloat xAdj = dxdy< 0.0F ? -dxdy : 0.0F; >> - GLfloat x = pMin[0] - (yMin - iyMin) * dxdy; >> GLint iy; >> - for (iy = iyMin; iy< iyMax; iy++, x += dxdy) { >> + #pragma omp parallel for schedule(dynamic) private(iy) >> firstprivate(span) >> + for (iy = iyMin; iy< iyMax; iy++) { >> + GLfloat x = pMin[0] - (yMin - iy) * dxdy; >> GLint ix, startX = (GLint) (x - xAdj); >> GLuint count; >> GLfloat coverage = 0.0F; >> >> +#ifdef _OPENMP >> + /* each thread needs to use a different (global) SpanArrays >> variable */ >> + span.array = SWRAST_CONTEXT(ctx)->SpanArrays + >> omp_get_thread_num(); >> +#endif >> /* skip over fragments with zero coverage */ >> while (startX< MAX_WIDTH) { >> coverage = compute_coveragef(pMin, pMid, pMax, startX, iy); >> @@ -228,13 +233,12 @@ >> coverage = compute_coveragef(pMin, pMid, pMax, ix, iy); >> } >> >> - if (ix<= startX) >> - continue; >> - >> - span.x = startX; >> - span.y = iy; >> - span.end = (GLuint) ix - (GLuint) startX; >> - _swrast_write_rgba_span(ctx,&span); >> + if (ix> startX) { >> + span.x = startX; >> + span.y = iy; >> + span.end = (GLuint) ix - (GLuint) startX; >> + _swrast_write_rgba_span(ctx,&span); >> + } >> } >> } >> else { >> @@ -244,13 +248,18 @@ >> const GLfloat *pMax = vMax->attrib[FRAG_ATTRIB_WPOS]; >> const GLfloat dxdy = majDx / majDy; >> const GLfloat xAdj = dxdy> 0 ? dxdy : 0.0F; >> - GLfloat x = pMin[0] - (yMin - iyMin) * dxdy; >> GLint iy; >> - for (iy = iyMin; iy< iyMax; iy++, x += dxdy) { >> + #pragma omp parallel for schedule(dynamic) private(iy) >> firstprivate(span) >> + for (iy = iyMin; iy< iyMax; iy++) { >> + GLfloat x = pMin[0] - (yMin - iy) * dxdy; >> GLint ix, left, startX = (GLint) (x + xAdj); >> GLuint count, n; >> GLfloat coverage = 0.0F; >> >> +#ifdef _OPENMP >> + /* each thread needs to use a different (global) SpanArrays >> variable */ >> + span.array = SWRAST_CONTEXT(ctx)->SpanArrays + >> omp_get_thread_num(); >> +#endif >> /* make sure we're not past the window edge */ >> if (startX>= ctx->DrawBuffer->_Xmax) { >> startX = ctx->DrawBuffer->_Xmax - 1; >> @@ -296,31 +305,30 @@ >> ATTRIB_LOOP_END >> #endif >> >> - if (startX<= ix) >> - continue; >> + if (startX> ix) { >> + n = (GLuint) startX - (GLuint) ix; >> >> - n = (GLuint) startX - (GLuint) ix; >> + left = ix + 1; >> >> - left = ix + 1; >> - >> - /* shift all values to the left */ >> - /* XXX this is temporary */ >> - { >> - SWspanarrays *array = span.array; >> - GLint j; >> - for (j = 0; j< (GLint) n; j++) { >> - array->coverage[j] = array->coverage[j + left]; >> - COPY_CHAN4(array->rgba[j], array->rgba[j + left]); >> + /* shift all values to the left */ >> + /* XXX this is temporary */ >> + { >> + SWspanarrays *array = span.array; >> + GLint j; >> + for (j = 0; j< (GLint) n; j++) { >> + array->coverage[j] = array->coverage[j + left]; >> + COPY_CHAN4(array->rgba[j], array->rgba[j + left]); >> #ifdef DO_Z >> - array->z[j] = array->z[j + left]; >> + array->z[j] = array->z[j + left]; >> #endif >> + } >> } >> - } >> >> - span.x = left; >> - span.y = iy; >> - span.end = n; >> - _swrast_write_rgba_span(ctx,&span); >> + span.x = left; >> + span.y = iy; >> + span.end = n; >> + _swrast_write_rgba_span(ctx,&span); >> + } >> } >> } >> } >> diff --git a/src/mesa/swrast/s_context.c b/src/mesa/swrast/s_context.c >> index def1531..4434f11 100644 >> --- a/src/mesa/swrast/s_context.c >> +++ b/src/mesa/swrast/s_context.c >> @@ -772,6 +772,11 @@ _swrast_CreateContext( struct gl_context *ctx ) >> { >> GLuint i; >> SWcontext *swrast = (SWcontext *)CALLOC(sizeof(SWcontext)); >> +#ifdef _OPENMP >> + const GLint maxThreads = omp_get_max_threads(); >> +#else >> + const GLint maxThreads = 1; >> +#endif >> >> if (SWRAST_DEBUG) { >> _mesa_debug(ctx, "_swrast_CreateContext\n"); >> @@ -806,19 +811,25 @@ _swrast_CreateContext( struct gl_context *ctx ) >> for (i = 0; i< MAX_TEXTURE_IMAGE_UNITS; i++) >> swrast->TextureSample[i] = NULL; >> >> - swrast->SpanArrays = MALLOC_STRUCT(sw_span_arrays); >> + /* SpanArrays is global and shared by all SWspan instances. However, when >> + * using multiple threads, it is necessary to have one SpanArrays >> instance >> + * per thread. >> + */ >> + swrast->SpanArrays = (SWspanarrays *) MALLOC(maxThreads * >> sizeof(SWspanarrays)); >> if (!swrast->SpanArrays) { >> FREE(swrast); >> return GL_FALSE; >> } >> - swrast->SpanArrays->ChanType = CHAN_TYPE; >> + for(i = 0; i< maxThreads; i++) { >> + swrast->SpanArrays[i].ChanType = CHAN_TYPE; >> #if CHAN_TYPE == GL_UNSIGNED_BYTE >> - swrast->SpanArrays->rgba = swrast->SpanArrays->rgba8; >> + swrast->SpanArrays[i].rgba = swrast->SpanArrays[i].rgba8; >> #elif CHAN_TYPE == GL_UNSIGNED_SHORT >> - swrast->SpanArrays->rgba = swrast->SpanArrays->rgba16; >> + swrast->SpanArrays[i].rgba = swrast->SpanArrays[i].rgba16; >> #else >> - swrast->SpanArrays->rgba = swrast->SpanArrays->attribs[FRAG_ATTRIB_COL0]; >> + swrast->SpanArrays[i].rgba = >> swrast->SpanArrays[i].attribs[FRAG_ATTRIB_COL0]; >> #endif >> + } >> >> /* init point span buffer */ >> swrast->PointSpan.primitive = GL_POINT; >> @@ -826,7 +837,10 @@ _swrast_CreateContext( struct gl_context *ctx ) >> swrast->PointSpan.facing = 0; >> swrast->PointSpan.array = swrast->SpanArrays; >> >> - swrast->TexelBuffer = (GLfloat *) MALLOC(ctx->Const.MaxTextureImageUnits >> * >> + /* TexelBuffer is also global and normally shared by all SWspan >> instances; >> + * when running with multiple threads, create one per thread. >> + */ >> + swrast->TexelBuffer = (GLfloat *) MALLOC(ctx->Const.MaxTextureImageUnits >> * maxThreads * >> MAX_WIDTH * 4 * >> sizeof(GLfloat)); >> if (!swrast->TexelBuffer) { >> FREE(swrast->SpanArrays); >> diff --git a/src/mesa/swrast/s_texcombine.c b/src/mesa/swrast/s_texcombine.c >> index 086ed0b..80b9dff 100644 >> --- a/src/mesa/swrast/s_texcombine.c >> +++ b/src/mesa/swrast/s_texcombine.c >> @@ -48,7 +48,11 @@ typedef float (*float4_array)[4]; >> static INLINE float4_array >> get_texel_array(SWcontext *swrast, GLuint unit) >> { >> +#ifdef _OPENMP >> + return (float4_array) (swrast->TexelBuffer + unit * MAX_WIDTH * 4 * >> omp_get_num_threads() + (MAX_WIDTH * 4 * omp_get_thread_num())); >> +#else >> return (float4_array) (swrast->TexelBuffer + unit * MAX_WIDTH * 4); >> +#endif >> } >> >> >> diff --git a/src/mesa/tnl/t_pipeline.c b/src/mesa/tnl/t_pipeline.c >> index 18f095f..881d5d5 100644 >> --- a/src/mesa/tnl/t_pipeline.c >> +++ b/src/mesa/tnl/t_pipeline.c >> @@ -146,7 +146,17 @@ void _tnl_run_pipeline( struct gl_context *ctx ) >> _tnl_notify_pipeline_output_change( ctx ); >> } >> >> +#ifndef _OPENMP >> + /* Don't adjust FPU precision mode in case multiple threads are to be >> used. >> + * This would require that the additional threads also changed the FPU >> mode >> + * which is quite a mess as this had to be done in all parallelized >> sections; >> + * otherwise the master thread and all other threads are running in >> different >> + * modes, producing inconsistent results. >> + * Note that all x64 implementations don't define/use START_FAST_MATH, so >> + * this is "hack" is only used in i386 mode >> + */ >> START_FAST_MATH(__tmp); >> +#endif >> >> for (i = 0; i< tnl->pipeline.nr_stages ; i++) { >> struct tnl_pipeline_stage *s =&tnl->pipeline.stages[i]; >> @@ -154,7 +164,9 @@ void _tnl_run_pipeline( struct gl_context *ctx ) >> break; >> } >> >> +#ifndef _OPENMP >> END_FAST_MATH(__tmp); >> +#endif >> } >> >> > > > > _______________________________________________ > mesa-dev mailing list > mesa-dev@lists.freedesktop.org > http://lists.freedesktop.org/mailman/listinfo/mesa-dev _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
