1 files changed, 86 insertions, 99 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_curbe.c b/src/mesa/drivers/dri/i965/brw_curbe.c
index 3f0aaa1f86..4ec0cfb816 100644
--- a/src/mesa/drivers/dri/i965/brw_curbe.c
+++ b/src/mesa/drivers/dri/i965/brw_curbe.c
@@ -31,10 +31,10 @@
 
 
 
-#include "glheader.h"
-#include "context.h"
-#include "macros.h"
-#include "enums.h"
+#include "main/glheader.h"
+#include "main/context.h"
+#include "main/macros.h"
+#include "main/enums.h"
 #include "shader/prog_parameter.h"
 #include "shader/prog_statevars.h"
 #include "intel_batchbuffer.h"
@@ -42,13 +42,13 @@
 #include "brw_defines.h"
 #include "brw_state.h"
 #include "brw_util.h"
-#include "brw_aub.h"
 
 
 /* Partition the CURBE between the various users of constant values:
  */
 static void calculate_curbe_offsets( struct brw_context *brw )
 {
+   GLcontext *ctx = &brw->intel.ctx;
    /* CACHE_NEW_WM_PROG */
    GLuint nr_fp_regs = (brw->wm.prog_data->nr_params + 15) / 16;
    
@@ -59,8 +59,8 @@ static void calculate_curbe_offsets( struct brw_context *brw )
    GLuint total_regs;
 
    /* _NEW_TRANSFORM */
-   if (brw->attribs.Transform->ClipPlanesEnabled) {
-      GLuint nr_planes = 6 + brw_count_bits(brw->attribs.Transform->ClipPlanesEnabled);
+   if (ctx->Transform.ClipPlanesEnabled) {
+      GLuint nr_planes = 6 + brw_count_bits(ctx->Transform.ClipPlanesEnabled);
       nr_clip_regs = (nr_planes * 4 + 15) / 16;
    }
 
@@ -90,7 +90,7 @@ static void calculate_curbe_offsets( struct brw_context *brw )
     */
    if (nr_fp_regs > brw->curbe.wm_size ||
        nr_vp_regs > brw->curbe.vs_size ||
-       nr_clip_regs > brw->curbe.clip_size ||
+       nr_clip_regs != brw->curbe.clip_size ||
        (total_regs < brw->curbe.total_size / 4 &&
 	brw->curbe.total_size > 16)) {
 
@@ -127,7 +127,7 @@ const struct brw_tracked_state brw_curbe_offsets = {
       .brw  = BRW_NEW_VERTEX_PROGRAM,
       .cache = CACHE_NEW_WM_PROG
    },
-   .update = calculate_curbe_offsets
+   .prepare = calculate_curbe_offsets
 };
 
 
@@ -138,37 +138,25 @@ const struct brw_tracked_state brw_curbe_offsets = {
  * fixed-function hardware in a double-buffering scheme to avoid a
  * pipeline stall each time the contents of the curbe is changed.
  */
-void brw_upload_constant_buffer_state(struct brw_context *brw)
+void brw_upload_cs_urb_state(struct brw_context *brw)
 {
-   struct brw_constant_buffer_state cbs; 
-   memset(&cbs, 0, sizeof(cbs));
+   struct brw_cs_urb_state cs_urb;
+   memset(&cs_urb, 0, sizeof(cs_urb));
 
    /* It appears that this is the state packet for the CS unit, ie. the
     * urb entries detailed here are housed in the CS range from the
     * URB_FENCE command.
     */
-   cbs.header.opcode = CMD_CONST_BUFFER_STATE;
-   cbs.header.length = sizeof(cbs)/4 - 2;
+   cs_urb.header.opcode = CMD_CS_URB_STATE;
+   cs_urb.header.length = sizeof(cs_urb)/4 - 2;
 
    /* BRW_NEW_URB_FENCE */
-   cbs.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
-   cbs.bits0.urb_entry_size = brw->urb.csize - 1;
+   cs_urb.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
+   cs_urb.bits0.urb_entry_size = brw->urb.csize - 1;
 
    assert(brw->urb.nr_cs_entries);
-   BRW_CACHED_BATCH_STRUCT(brw, &cbs);
-}      
-
-#if 0
-const struct brw_tracked_state brw_constant_buffer_state = {
-   .dirty = {
-      .mesa = 0,
-      .brw = BRW_NEW_URB_FENCE,
-      .cache = 0
-   },
-   .update = brw_upload_constant_buffer_state
-};
-#endif
-
+   BRW_CACHED_BATCH_STRUCT(brw, &cs_urb);
+}
 
 static GLfloat fixed_plane[6][4] = {
    { 0,    0,   -1, 1 },
@@ -183,12 +171,11 @@ static GLfloat fixed_plane[6][4] = {
  * cache mechanism, but maybe would benefit from a comparison against
  * the current uploaded set of constants.
  */
-static void upload_constant_buffer(struct brw_context *brw)
+static void prepare_constant_buffer(struct brw_context *brw)
 {
    GLcontext *ctx = &brw->intel.ctx;
    struct brw_vertex_program *vp = (struct brw_vertex_program *)brw->vertex_program;
    struct brw_fragment_program *fp = (struct brw_fragment_program *)brw->fragment_program;
-   struct brw_mem_pool *pool = &brw->pool[BRW_GS_POOL];
    GLuint sz = brw->curbe.total_size;
    GLuint bufsz = sz * 16 * sizeof(GLfloat);
    GLfloat *buf;
@@ -198,24 +185,17 @@ static void upload_constant_buffer(struct brw_context *brw)
     * function will also be called whenever fp or vp changes.
     */
    brw->curbe.tracked_state.dirty.mesa = (_NEW_TRANSFORM|_NEW_PROJECTION);
-   brw->curbe.tracked_state.dirty.mesa |= vp->param_state;
-   brw->curbe.tracked_state.dirty.mesa |= fp->param_state;
+   brw->curbe.tracked_state.dirty.mesa |= vp->program.Base.Parameters->StateFlags;
+   brw->curbe.tracked_state.dirty.mesa |= fp->program.Base.Parameters->StateFlags;
 
    if (sz == 0) {
-      struct brw_constant_buffer cb;
-      cb.header.opcode = CMD_CONST_BUFFER;
-      cb.header.length = sizeof(cb)/4 - 2;
-      cb.header.valid = 0;
-      cb.bits0.buffer_length = 0;
-      cb.bits0.buffer_address = 0;
-      BRW_BATCH_STRUCT(brw, &cb);
 
       if (brw->curbe.last_buf) {
 	 free(brw->curbe.last_buf);
 	 brw->curbe.last_buf = NULL;
 	 brw->curbe.last_bufsz  = 0;
       }
-       
+
       return;
    }
 
@@ -254,11 +234,11 @@ static void upload_constant_buffer(struct brw_context *brw)
        */
       assert(MAX_CLIP_PLANES == 6);
       for (j = 0; j < MAX_CLIP_PLANES; j++) {
-	 if (brw->attribs.Transform->ClipPlanesEnabled & (1<<j)) {
-	    buf[offset + i * 4 + 0] = brw->attribs.Transform->_ClipUserPlane[j][0];
-	    buf[offset + i * 4 + 1] = brw->attribs.Transform->_ClipUserPlane[j][1];
-	    buf[offset + i * 4 + 2] = brw->attribs.Transform->_ClipUserPlane[j][2];
-	    buf[offset + i * 4 + 3] = brw->attribs.Transform->_ClipUserPlane[j][3];
+	 if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
+	    buf[offset + i * 4 + 0] = ctx->Transform._ClipUserPlane[j][0];
+	    buf[offset + i * 4 + 1] = ctx->Transform._ClipUserPlane[j][1];
+	    buf[offset + i * 4 + 2] = ctx->Transform._ClipUserPlane[j][2];
+	    buf[offset + i * 4 + 3] = ctx->Transform._ClipUserPlane[j][3];
 	    i++;
 	 }
       }
@@ -290,11 +270,11 @@ static void upload_constant_buffer(struct brw_context *brw)
 		   brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
    }
 
-   if (brw->curbe.last_buf &&
+   if (brw->curbe.curbe_bo != NULL &&
+       brw->curbe.last_buf &&
        bufsz == brw->curbe.last_bufsz &&
        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
       free(buf);
-/*       return; */
    } 
    else {
       if (brw->curbe.last_buf)
@@ -302,61 +282,66 @@ static void upload_constant_buffer(struct brw_context *brw)
       brw->curbe.last_buf = buf;
       brw->curbe.last_bufsz = bufsz;
 
-      
-      if (!brw_pool_alloc(pool, 
-			  bufsz,
-			  6,
-			  &brw->curbe.gs_offset)) {
-	 _mesa_printf("out of GS memory for curbe\n");
-	 assert(0);
-	 return;
+      if (brw->curbe.curbe_bo != NULL &&
+	  (brw->curbe.need_new_bo ||
+	   brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size))
+      {
+	 dri_bo_unreference(brw->curbe.curbe_bo);
+	 brw->curbe.curbe_bo = NULL;
+      }
+
+      if (brw->curbe.curbe_bo == NULL) {
+	 /* Allocate a single page for CURBE entries for this batchbuffer.
+	  * They're generally around 64b.
+	  */
+	 brw->curbe.curbe_bo = dri_bo_alloc(brw->intel.bufmgr, "CURBE",
+					    4096, 1 << 6);
+	 brw->curbe.curbe_next_offset = 0;
       }
-            
+
+      brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
+      brw->curbe.curbe_next_offset += bufsz;
+      brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 
       /* Copy data to the buffer:
        */
-      bmBufferSubDataAUB(&brw->intel,
-			 pool->buffer,
-			 brw->curbe.gs_offset, 
-			 bufsz, 
-			 buf,
-			 DW_CONSTANT_BUFFER,
-			 0);
+      dri_bo_subdata(brw->curbe.curbe_bo, brw->curbe.curbe_offset, bufsz, buf);
    }
 
-   /* TODO: only emit the constant_buffer packet when necessary, ie:
-      - contents have changed
-      - offset has changed
-      - hw requirements due to other packets emitted.
-   */
-   {
-      struct brw_constant_buffer cb;
-      
-      memset(&cb, 0, sizeof(cb));
-
-      cb.header.opcode = CMD_CONST_BUFFER;
-      cb.header.length = sizeof(cb)/4 - 2;
-      cb.header.valid = 1;
-      cb.bits0.buffer_length = sz - 1;
-      cb.bits0.buffer_address = brw->curbe.gs_offset >> 6;
-      
-      /* Because this provokes an action (ie copy the constants into the
-       * URB), it shouldn't be shortcircuited if identical to the
-       * previous time - because eg. the urb destination may have
-       * changed, or the urb contents different to last time.  
-       *
-       * Note that the data referred to is actually copied internally,
-       * not just used in place according to passed pointer.
-       *
-       * It appears that the CS unit takes care of using each available
-       * URB entry (Const URB Entry == CURBE) in turn, and issuing
-       * flushes as necessary when doublebuffering of CURBEs isn't
-       * possible.
-       */
-/*       intel_batchbuffer_align(brw->intel.batch, 64, sizeof(cb)); */
-      BRW_BATCH_STRUCT(brw, &cb);
-/*       intel_batchbuffer_align(brw->intel.batch, 64, 0); */
+   brw_add_validated_bo(brw, brw->curbe.curbe_bo);
+
+   /* Because this provokes an action (ie copy the constants into the
+    * URB), it shouldn't be shortcircuited if identical to the
+    * previous time - because eg. the urb destination may have
+    * changed, or the urb contents different to last time.
+    *
+    * Note that the data referred to is actually copied internally,
+    * not just used in place according to passed pointer.
+    *
+    * It appears that the CS unit takes care of using each available
+    * URB entry (Const URB Entry == CURBE) in turn, and issuing
+    * flushes as necessary when doublebuffering of CURBEs isn't
+    * possible.
+    */
+}
+
+
+static void emit_constant_buffer(struct brw_context *brw)
+{
+   struct intel_context *intel = &brw->intel;
+   GLuint sz = brw->curbe.total_size;
+
+   BEGIN_BATCH(2, IGNORE_CLIPRECTS);
+   if (sz == 0) {
+      OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
+      OUT_BATCH(0);
+   } else {
+      OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
+      OUT_RELOC(brw->curbe.curbe_bo,
+		I915_GEM_DOMAIN_INSTRUCTION, 0,
+		(sz - 1) + brw->curbe.curbe_offset);
    }
+   ADVANCE_BATCH();
 }
 
 /* This tracked state is unique in that the state it monitors varies
@@ -372,9 +357,11 @@ const struct brw_tracked_state brw_constant_buffer = {
 	       BRW_NEW_VERTEX_PROGRAM |
 	       BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 	       BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
-	       BRW_NEW_CURBE_OFFSETS),
+	       BRW_NEW_CURBE_OFFSETS |
+	       BRW_NEW_BATCH),
       .cache = (CACHE_NEW_WM_PROG) 
    },
-   .update = upload_constant_buffer
+   .prepare = prepare_constant_buffer,
+   .emit = emit_constant_buffer,
 };