swrast: initial multi-threaded span rendering

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).

Signed-off-by: Brian Paul <brianp@vmware.com>

6 files changed, 91 insertions, 36 deletions

diff --git a/common.py b/common.py
index 8657030ea3f..cfee1b5dc2e 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 8cd3bc7f6e0..7135251d7a3 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 91d4f7a10ab..77b3ae6ec7a 100644
--- a/src/mesa/swrast/s_aatritemp.h
+++ b/src/mesa/swrast/s_aatritemp.h
@@ -181,13 +181,20 @@
       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) {
+#ifdef _OPENMP
+#pragma omp parallel for schedule(dynamic) private(iy) firstprivate(span)
+#endif
+      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 +235,12 @@
             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 +250,20 @@
       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) {
+#ifdef _OPENMP
+#pragma omp parallel for schedule(dynamic) private(iy) firstprivate(span)
+#endif
+      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 +309,30 @@
          ATTRIB_LOOP_END
 #endif
 
-         if (startX <= ix)
-            continue;
-
-         n = (GLuint) startX - (GLuint) ix;
+         if (startX > 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 def1531d7ff..4434f11b990 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 086ed0b33d7..80b9dff3cc2 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 18f095f0d4b..881d5d5f535 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
 }