i965: Correct scratch space allocation.

One, it was allocating increments of 1kb, but per thread scratch space
is a power of two.  Two, the new FS wasn't getting total_scratch set
at all, so everyone thought they had 1kb and writes beyond 1kb would
go stomping on a neighbor thread.

With this plus the previous register spilling for the new FS,
glsl-fs-convolution-1 passes.

3 files changed, 15 insertions, 12 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index f38976aa8a..aa945c9771 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -3252,7 +3252,6 @@ brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c)
    }
 
    c->prog_data.total_grf = v.grf_used;
-   c->prog_data.total_scratch = 0;
 
    return GL_TRUE;
 }
diff --git a/src/mesa/drivers/dri/i965/brw_wm.c b/src/mesa/drivers/dri/i965/brw_wm.c
index f2ce756564..7f3ba5f058 100644
--- a/src/mesa/drivers/dri/i965/brw_wm.c
+++ b/src/mesa/drivers/dri/i965/brw_wm.c
@@ -114,14 +114,6 @@ brw_wm_non_glsl_emit(struct brw_context *brw, struct brw_wm_compile *c)
    /* how many general-purpose registers are used */
    c->prog_data.total_grf = c->max_wm_grf;
 
-   /* Scratch space is used for register spilling */
-   if (c->last_scratch) {
-      c->prog_data.total_scratch = c->last_scratch + 0x40;
-   }
-   else {
-      c->prog_data.total_scratch = 0;
-   }
-
    /* Emit GEN4 code.
     */
    brw_wm_emit(c);
@@ -193,6 +185,19 @@ static void do_wm_prog( struct brw_context *brw,
       }
    }
 
+   /* Scratch space is used for register spilling */
+   if (c->last_scratch) {
+      /* Per-thread scratch space is power-of-two sized. */
+      for (c->prog_data.total_scratch = 1024;
+	   c->prog_data.total_scratch <= c->last_scratch;
+	   c->prog_data.total_scratch *= 2) {
+	 /* empty */
+      }
+   }
+   else {
+      c->prog_data.total_scratch = 0;
+   }
+
    if (INTEL_DEBUG & DEBUG_WM)
       fprintf(stderr, "\n");
 
diff --git a/src/mesa/drivers/dri/i965/brw_wm_state.c b/src/mesa/drivers/dri/i965/brw_wm_state.c
index 411204e704..433ccc66f0 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_state.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_state.c
@@ -97,7 +97,7 @@ wm_unit_populate_key(struct brw_context *brw, struct brw_wm_unit_key *key)
    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);
+   key->total_scratch = brw->wm.prog_data->total_scratch;
 
    /* BRW_NEW_URB_FENCE */
    key->urb_size = brw->urb.vsize;
@@ -184,7 +184,7 @@ wm_unit_create_from_key(struct brw_context *brw, struct brw_wm_unit_key *key,
    if (key->total_scratch != 0) {
       wm.thread2.scratch_space_base_pointer =
 	 brw->wm.scratch_bo->offset >> 10; /* reloc */
-      wm.thread2.per_thread_scratch_space = key->total_scratch / 1024 - 1;
+      wm.thread2.per_thread_scratch_space = ffs(key->total_scratch) - 11;
    } else {
       wm.thread2.scratch_space_base_pointer = 0;
       wm.thread2.per_thread_scratch_space = 0;
@@ -293,7 +293,6 @@ static void upload_wm_unit( struct brw_context *brw )
     * 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 * brw->wm_max_threads;