Import i915 and i965 dri drivers from mesa 7.1.0.7.1.0

1 files changed, 192 insertions, 91 deletions

diff --git a/i965/brw_wm_state.c b/i965/brw_wm_state.c
index 5b4f2ab..f4da0f2 100644
--- a/i965/brw_wm_state.c
+++ b/i965/brw_wm_state.c
@@ -34,109 +34,136 @@
 #include "brw_context.h"
 #include "brw_state.h"
 #include "brw_defines.h"
-#include "bufmgr.h"
+#include "dri_bufmgr.h"
+#include "brw_wm.h"
 
 /***********************************************************************
  * WM unit - fragment programs and rasterization
  */
 
-static void invalidate_scratch_cb( struct intel_context *intel,
-				   void *unused )
-{
-   /* nothing */
-}
+struct brw_wm_unit_key {
+   unsigned int total_grf, total_scratch;
+   unsigned int urb_entry_read_length;
+   unsigned int curb_entry_read_length;
+   unsigned int dispatch_grf_start_reg;
+
+   unsigned int curbe_offset;
+   unsigned int urb_size;
+
+   unsigned int max_threads;
 
+   unsigned int nr_surfaces, sampler_count;
+   GLboolean uses_depth, computes_depth, uses_kill, is_glsl;
+   GLboolean polygon_stipple, stats_wm, line_stipple, offset_enable;
+   GLfloat offset_units, offset_factor;
+};
 
-static void upload_wm_unit(struct brw_context *brw )
+static void
+wm_unit_populate_key(struct brw_context *brw, struct brw_wm_unit_key *key)
 {
+   const struct gl_fragment_program *fp = brw->fragment_program;
    struct intel_context *intel = &brw->intel;
-   struct brw_wm_unit_state wm;
-   GLuint max_threads;
+
+   memset(key, 0, sizeof(*key));
 
    if (INTEL_DEBUG & DEBUG_SINGLE_THREAD)
-      max_threads = 0; 
+      key->max_threads = 1;
    else
-      max_threads = 31;
-
-
-   memset(&wm, 0, sizeof(wm));
+      key->max_threads = 32;
 
    /* CACHE_NEW_WM_PROG */
-   wm.thread0.grf_reg_count = ((brw->wm.prog_data->total_grf-1) & ~15) / 16;
-   wm.thread0.kernel_start_pointer = brw->wm.prog_gs_offset >> 6;
-   wm.thread3.dispatch_grf_start_reg = brw->wm.prog_data->first_curbe_grf;
-   wm.thread3.urb_entry_read_length = brw->wm.prog_data->urb_read_length;
-   wm.thread3.const_urb_entry_read_length = brw->wm.prog_data->curb_read_length;
-
-   wm.wm5.max_threads = max_threads;      
-
-   if (brw->wm.prog_data->total_scratch) {
-      GLuint per_thread = (brw->wm.prog_data->total_scratch + 1023) / 1024;
-      GLuint total = per_thread * (max_threads + 1);
-
-      /* Scratch space -- just have to make sure there is sufficient
-       * allocated for the active program and current number of threads.
-       */      
-
-      if (!brw->wm.scratch_buffer) {
-	 bmGenBuffers(intel, "wm scratch", 1, &brw->wm.scratch_buffer, 12);
-	 bmBufferSetInvalidateCB(intel,
-				 brw->wm.scratch_buffer,
-				 invalidate_scratch_cb,
-				 NULL,
-				 GL_FALSE);
-      }
+   key->total_grf = brw->wm.prog_data->total_grf;
+   key->urb_entry_read_length = brw->wm.prog_data->urb_read_length;
+   key->curb_entry_read_length = brw->wm.prog_data->curb_read_length;
+   key->dispatch_grf_start_reg = brw->wm.prog_data->first_curbe_grf;
+   key->total_scratch = ALIGN(brw->wm.prog_data->total_scratch, 1024);
 
-      if (total > brw->wm.scratch_buffer_size) {
-	 brw->wm.scratch_buffer_size = total;
-	 bmBufferData(intel,
-		      brw->wm.scratch_buffer,
-		      brw->wm.scratch_buffer_size,
-		      NULL,
-		      0);
-      }
-		   
-      assert(per_thread <= 12 * 1024);
-      wm.thread2.per_thread_scratch_space = (per_thread / 1024) - 1;
+   /* BRW_NEW_URB_FENCE */
+   key->urb_size = brw->urb.vsize;
 
-      /* XXX: could make this dynamic as this is so rarely active:
-       */
-      /* BRW_NEW_LOCK */
-      wm.thread2.scratch_space_base_pointer = 
-	 bmBufferOffset(intel, brw->wm.scratch_buffer) >> 10;
-   }
+   /* BRW_NEW_CURBE_OFFSETS */
+   key->curbe_offset = brw->curbe.wm_start;
 
    /* CACHE_NEW_SURFACE */
-   wm.thread1.binding_table_entry_count = brw->wm.nr_surfaces;
+   key->nr_surfaces = brw->wm.nr_surfaces;
 
-   /* BRW_NEW_CURBE_OFFSETS */
-   wm.thread3.const_urb_entry_read_offset = brw->curbe.wm_start * 2;
+   /* CACHE_NEW_SAMPLER */
+   key->sampler_count = brw->wm.sampler_count;
 
-   wm.thread3.urb_entry_read_offset = 0;
+   /* _NEW_POLYGONSTIPPLE */
+   key->polygon_stipple = brw->attribs.Polygon->StippleFlag;
+
+   /* BRW_NEW_FRAGMENT_PROGRAM */
+   key->uses_depth = (fp->Base.InputsRead & (1 << FRAG_ATTRIB_WPOS)) != 0;
+
+   /* as far as we can tell */
+   key->computes_depth =
+      (fp->Base.OutputsWritten & (1 << FRAG_RESULT_DEPR)) != 0;
+
+   /* _NEW_COLOR */
+   key->uses_kill = fp->UsesKill || brw->attribs.Color->AlphaEnabled;
+   key->is_glsl = brw_wm_is_glsl(fp);
+
+   /* XXX: This needs a flag to indicate when it changes. */
+   key->stats_wm = intel->stats_wm;
+
+   /* _NEW_LINE */
+   key->line_stipple = brw->attribs.Line->StippleFlag;
+
+   /* _NEW_POLYGON */
+   key->offset_enable = brw->attribs.Polygon->OffsetFill;
+   key->offset_units = brw->attribs.Polygon->OffsetUnits;
+   key->offset_factor = brw->attribs.Polygon->OffsetFactor;
+}
+
+static dri_bo *
+wm_unit_create_from_key(struct brw_context *brw, struct brw_wm_unit_key *key,
+			dri_bo **reloc_bufs)
+{
+   struct brw_wm_unit_state wm;
+   dri_bo *bo;
+
+   memset(&wm, 0, sizeof(wm));
+
+   wm.thread0.grf_reg_count = ALIGN(key->total_grf, 16) / 16 - 1;
+   wm.thread0.kernel_start_pointer = brw->wm.prog_bo->offset >> 6; /* reloc */
    wm.thread1.depth_coef_urb_read_offset = 1;
    wm.thread1.floating_point_mode = BRW_FLOATING_POINT_NON_IEEE_754;
+   wm.thread1.binding_table_entry_count = key->nr_surfaces;
+
+   if (key->total_scratch != 0) {
+      wm.thread2.scratch_space_base_pointer =
+	 brw->wm.scratch_buffer->offset >> 10; /* reloc */
+      wm.thread2.per_thread_scratch_space = key->total_scratch / 1024 - 1;
+   } else {
+      wm.thread2.scratch_space_base_pointer = 0;
+      wm.thread2.per_thread_scratch_space = 0;
+   }
 
-   /* CACHE_NEW_SAMPLER */
-   wm.wm4.sampler_count = (brw->wm.sampler_count + 1) / 4;
-   wm.wm4.sampler_state_pointer = brw->wm.sampler_gs_offset >> 5;
+   wm.thread3.dispatch_grf_start_reg = key->dispatch_grf_start_reg;
+   wm.thread3.urb_entry_read_length = key->urb_entry_read_length;
+   wm.thread3.const_urb_entry_read_length = key->curb_entry_read_length;
+   wm.thread3.const_urb_entry_read_offset = key->curbe_offset * 2;
+   wm.thread3.urb_entry_read_offset = 0;
 
-   /* BRW_NEW_FRAGMENT_PROGRAM */
-   {
-      const struct gl_fragment_program *fp = brw->fragment_program; 
-
-      if (fp->Base.InputsRead & (1<<FRAG_ATTRIB_WPOS)) 
-	 wm.wm5.program_uses_depth = 1; /* as far as we can tell */
-   
-      if (fp->Base.OutputsWritten & (1<<FRAG_RESULT_DEPR)) 
-	 wm.wm5.program_computes_depth = 1;
-   
-      /* _NEW_COLOR */
-      if (fp->UsesKill || 
-	  brw->attribs.Color->AlphaEnabled) 
-	 wm.wm5.program_uses_killpixel = 1; 
+   wm.wm4.sampler_count = (key->sampler_count + 1) / 4;
+   if (brw->wm.sampler_bo != NULL) {
+      /* reloc */
+      wm.wm4.sampler_state_pointer = brw->wm.sampler_bo->offset >> 5;
+   } else {
+      wm.wm4.sampler_state_pointer = 0;
    }
 
-   wm.wm5.enable_16_pix = 1;
+   wm.wm5.program_uses_depth = key->uses_depth;
+   wm.wm5.program_computes_depth = key->computes_depth;
+   wm.wm5.program_uses_killpixel = key->uses_kill;
+
+   if (key->is_glsl)
+      wm.wm5.enable_8_pix = 1;
+   else
+      wm.wm5.enable_16_pix = 1;
+
+   wm.wm5.max_threads = key->max_threads - 1;
    wm.wm5.thread_dispatch_enable = 1;	/* AKA: color_write */
    wm.wm5.legacy_line_rast = 0;
    wm.wm5.legacy_global_depth_bias = 0;
@@ -144,34 +171,108 @@ static void upload_wm_unit(struct brw_context *brw )
    wm.wm5.line_aa_region_width = 0;
    wm.wm5.line_endcap_aa_region_width = 1;
 
-   /* _NEW_POLYGONSTIPPLE */
-   if (brw->attribs.Polygon->StippleFlag) 
-      wm.wm5.polygon_stipple = 1;
+   wm.wm5.polygon_stipple = key->polygon_stipple;
 
-   /* _NEW_POLYGON */
-   if (brw->attribs.Polygon->OffsetFill) {
+   if (key->offset_enable) {
       wm.wm5.depth_offset = 1;
       /* Something wierd going on with legacy_global_depth_bias,
        * offset_constant, scaling and MRD.  This value passes glean
        * but gives some odd results elsewere (eg. the
        * quad-offset-units test).
        */
-      wm.global_depth_offset_constant = brw->attribs.Polygon->OffsetUnits * 2;
+      wm.global_depth_offset_constant = key->offset_units * 2;
 
       /* This is the only value that passes glean:
        */
-      wm.global_depth_offset_scale = brw->attribs.Polygon->OffsetFactor;
+      wm.global_depth_offset_scale = key->offset_factor;
    }
 
-   /* _NEW_LINE */
-   if (brw->attribs.Line->StippleFlag) {
-      wm.wm5.line_stipple = 1;
-   }
+   wm.wm5.line_stipple = key->line_stipple;
 
-   if (INTEL_DEBUG & DEBUG_STATS || intel->stats_wm)
+   if (INTEL_DEBUG & DEBUG_STATS || key->stats_wm)
       wm.wm4.stats_enable = 1;
 
-   brw->wm.state_gs_offset = brw_cache_data( &brw->cache[BRW_WM_UNIT], &wm );
+   bo = brw_upload_cache(&brw->cache, BRW_WM_UNIT,
+			 key, sizeof(*key),
+			 reloc_bufs, 3,
+			 &wm, sizeof(wm),
+			 NULL, NULL);
+
+   /* Emit WM program relocation */
+   dri_emit_reloc(bo,
+		  DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_READ,
+		  wm.thread0.grf_reg_count << 1,
+		  offsetof(struct brw_wm_unit_state, thread0),
+		  brw->wm.prog_bo);
+
+   /* Emit scratch space relocation */
+   if (key->total_scratch != 0) {
+      dri_emit_reloc(bo,
+		     DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_READ | DRM_BO_FLAG_WRITE,
+		     wm.thread2.per_thread_scratch_space,
+		     offsetof(struct brw_wm_unit_state, thread2),
+		     brw->wm.scratch_buffer);
+   }
+
+   /* Emit sampler state relocation */
+   if (key->sampler_count != 0) {
+      dri_emit_reloc(bo,
+		     DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_READ,
+		     wm.wm4.stats_enable | (wm.wm4.sampler_count << 2),
+		     offsetof(struct brw_wm_unit_state, wm4),
+		     brw->wm.sampler_bo);
+   }
+
+   return bo;
+}
+
+
+static int upload_wm_unit( struct brw_context *brw )
+{
+   struct intel_context *intel = &brw->intel;
+   struct brw_wm_unit_key key;
+   dri_bo *reloc_bufs[3];
+   int ret = 0, i;
+   wm_unit_populate_key(brw, &key);
+
+   /* Allocate the necessary scratch space if we haven't already.  Don't
+    * bother reducing the allocation later, since we use scratch so
+    * rarely.
+    */
+   assert(key.total_scratch <= 12 * 1024);
+   if (key.total_scratch) {
+      GLuint total = key.total_scratch * key.max_threads;
+
+      if (brw->wm.scratch_buffer && total > brw->wm.scratch_buffer->size) {
+	 dri_bo_unreference(brw->wm.scratch_buffer);
+	 brw->wm.scratch_buffer = NULL;
+      }
+      if (brw->wm.scratch_buffer == NULL) {
+	 brw->wm.scratch_buffer = dri_bo_alloc(intel->bufmgr,
+					       "wm scratch",
+					       total,
+					       4096, DRM_BO_FLAG_MEM_TT);
+      }
+   }
+
+   reloc_bufs[0] = brw->wm.prog_bo;
+   reloc_bufs[1] = brw->wm.scratch_buffer;
+   reloc_bufs[2] = brw->wm.sampler_bo;
+
+   dri_bo_unreference(brw->wm.state_bo);
+   brw->wm.state_bo = brw_search_cache(&brw->cache, BRW_WM_UNIT,
+				       &key, sizeof(key),
+				       reloc_bufs, 3,
+				       NULL);
+   if (brw->wm.state_bo == NULL) {
+      brw->wm.state_bo = wm_unit_create_from_key(brw, &key, reloc_bufs);
+   }
+
+   for (i = 0; i < 3; i++)
+     if (reloc_bufs[i])
+       ret |= dri_bufmgr_check_aperture_space(reloc_bufs[i]);
+   ret |= dri_bufmgr_check_aperture_space(brw->wm.state_bo);
+   return ret;
 }
 
 const struct brw_tracked_state brw_wm_unit = {
@@ -189,6 +290,6 @@ const struct brw_tracked_state brw_wm_unit = {
 		CACHE_NEW_WM_PROG | 
 		CACHE_NEW_SAMPLER)
    },
-   .update = upload_wm_unit
+   .prepare = upload_wm_unit,
 };