i915g: split up hw state emission into small atoms

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

2 files changed, 309 insertions, 275 deletions

diff --git a/src/gallium/drivers/i915/i915_state_emit.c b/src/gallium/drivers/i915/i915_state_emit.c
index 0323ad940f..504b60c64e 100644
--- a/src/gallium/drivers/i915/i915_state_emit.c
+++ b/src/gallium/drivers/i915/i915_state_emit.c
@@ -40,13 +40,19 @@
 
 struct i915_tracked_hw_state {
    const char *name;
-   void (*validate)(struct i915_context *);
+   void (*validate)(struct i915_context *, unsigned *batch_space);
    void (*emit)(struct i915_context *);
    unsigned dirty, batch_space;
 };
 
 
 static void
+validate_flush(struct i915_context *i915, unsigned *batch_space)
+{
+   *batch_space = i915->flush_dirty ? 1 : 0;
+}
+
+static void
 emit_flush(struct i915_context *i915)
 {
    /* Cache handling is very cheap atm. State handling can request to flushes:
@@ -61,82 +67,336 @@ emit_flush(struct i915_context *i915)
       OUT_BATCH(MI_FLUSH | INHIBIT_FLUSH_RENDER_CACHE);
 }
 
+uint32_t invariant_state[] = {
+   _3DSTATE_AA_CMD | AA_LINE_ECAAR_WIDTH_ENABLE | AA_LINE_ECAAR_WIDTH_1_0 |
+             AA_LINE_REGION_WIDTH_ENABLE | AA_LINE_REGION_WIDTH_1_0,
+
+   _3DSTATE_DFLT_DIFFUSE_CMD, 0,
+
+   _3DSTATE_DFLT_SPEC_CMD, 0,
+
+   _3DSTATE_DFLT_Z_CMD, 0,
+
+   _3DSTATE_COORD_SET_BINDINGS |
+             CSB_TCB(0, 0) |
+             CSB_TCB(1, 1) |
+             CSB_TCB(2, 2) |
+             CSB_TCB(3, 3) |
+             CSB_TCB(4, 4) |
+             CSB_TCB(5, 5) |
+             CSB_TCB(6, 6) |
+             CSB_TCB(7, 7),
+
+   _3DSTATE_RASTER_RULES_CMD |
+             ENABLE_POINT_RASTER_RULE |
+             OGL_POINT_RASTER_RULE |
+             ENABLE_LINE_STRIP_PROVOKE_VRTX |
+             ENABLE_TRI_FAN_PROVOKE_VRTX |
+             LINE_STRIP_PROVOKE_VRTX(1) |
+             TRI_FAN_PROVOKE_VRTX(2) |
+             ENABLE_TEXKILL_3D_4D |
+             TEXKILL_4D,
+
+   _3DSTATE_DEPTH_SUBRECT_DISABLE,
+
+   /* disable indirect state for now
+    */
+   _3DSTATE_LOAD_INDIRECT | 0, 0};
+
+static void
+emit_invariant(struct i915_context *i915)
+{
+   i915_winsys_batchbuffer_write(i915->batch, invariant_state,
+                                 Elements(invariant_state)*sizeof(uint32_t));
+}
+
 static void
-validate_immediate(struct i915_context *i915)
+validate_immediate(struct i915_context *i915, unsigned *batch_space)
 {
+   unsigned dirty = (1 << I915_IMMEDIATE_S0 | 1 << I915_IMMEDIATE_S1 |
+                     1 << I915_IMMEDIATE_S2 | 1 << I915_IMMEDIATE_S3 |
+                     1 << I915_IMMEDIATE_S3 | 1 << I915_IMMEDIATE_S4 |
+                     1 << I915_IMMEDIATE_S5 | 1 << I915_IMMEDIATE_S6) &
+                    i915->immediate_dirty;
+
    if (i915->immediate_dirty & (1 << I915_IMMEDIATE_S0))
       i915->validation_buffers[i915->num_validation_buffers++] = i915->vbo;
+
+   *batch_space = 1 + util_bitcount(dirty);
+}
+
+static void
+emit_immediate(struct i915_context *i915)
+{
+   /* remove unwatned bits and S7 */
+   unsigned dirty = (1 << I915_IMMEDIATE_S0 | 1 << I915_IMMEDIATE_S1 |
+                     1 << I915_IMMEDIATE_S2 | 1 << I915_IMMEDIATE_S3 |
+                     1 << I915_IMMEDIATE_S3 | 1 << I915_IMMEDIATE_S4 |
+                     1 << I915_IMMEDIATE_S5 | 1 << I915_IMMEDIATE_S6) &
+                    i915->immediate_dirty;
+   int i, num = util_bitcount(dirty);
+   assert(num && num <= I915_MAX_IMMEDIATE);
+
+   OUT_BATCH(_3DSTATE_LOAD_STATE_IMMEDIATE_1 |
+             dirty << 4 | (num - 1));
+
+   if (i915->immediate_dirty & (1 << I915_IMMEDIATE_S0)) {
+      if (i915->vbo)
+         OUT_RELOC(i915->vbo, I915_USAGE_VERTEX,
+                   i915->current.immediate[I915_IMMEDIATE_S0]);
+      else
+         OUT_BATCH(0);
+   }
+
+   for (i = 1; i < I915_MAX_IMMEDIATE; i++) {
+      if (dirty & (1 << i))
+         OUT_BATCH(i915->current.immediate[i]);
+   }
+}
+
+static void
+validate_dynamic(struct i915_context *i915, unsigned *batch_space)
+{
+   *batch_space = util_bitcount(i915->dynamic_dirty & ((1 << I915_MAX_DYNAMIC) - 1));
+}
+
+static void
+emit_dynamic(struct i915_context *i915)
+{
+   int i;
+   for (i = 0; i < I915_MAX_DYNAMIC; i++) {
+      if (i915->dynamic_dirty & (1 << i))
+         OUT_BATCH(i915->current.dynamic[i]);
+   }
 }
 
 static void
-validate_static(struct i915_context *i915)
+validate_static(struct i915_context *i915, unsigned *batch_space)
 {
-   if (i915->current.cbuf_bo)
+   *batch_space = 2 + 5; /* including DRAW_RECT */
+
+   if (i915->current.cbuf_bo) {
       i915->validation_buffers[i915->num_validation_buffers++]
          = i915->current.cbuf_bo;
+      *batch_space += 3;
+   }
 
-   if (i915->current.depth_bo)
+   if (i915->current.depth_bo) {
       i915->validation_buffers[i915->num_validation_buffers++]
          = i915->current.depth_bo;
+      *batch_space += 3;
+   }
+}
+
+static void
+emit_static(struct i915_context *i915)
+{
+   if (i915->current.cbuf_bo) {
+      OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
+      OUT_BATCH(i915->current.cbuf_flags);
+      OUT_RELOC(i915->current.cbuf_bo,
+                I915_USAGE_RENDER,
+                0);
+   }
+
+   /* What happens if no zbuf??
+    */
+   if (i915->current.depth_bo) {
+      OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
+      OUT_BATCH(i915->current.depth_flags);
+      OUT_RELOC(i915->current.depth_bo,
+                I915_USAGE_RENDER,
+                0);
+   }
+
+   {
+      OUT_BATCH(_3DSTATE_DST_BUF_VARS_CMD);
+      OUT_BATCH(i915->current.dst_buf_vars);
+   }
 }
 
 static void
-validate_map(struct i915_context *i915)
+validate_map(struct i915_context *i915, unsigned *batch_space)
 {
    const uint enabled = i915->current.sampler_enable_flags;
    uint unit;
    struct i915_texture *tex;
 
+   *batch_space = i915->current.sampler_enable_nr ?
+     2 + 3*i915->current.sampler_enable_nr : 0;
 
    for (unit = 0; unit < I915_TEX_UNITS; unit++) {
       if (enabled & (1 << unit)) {
-	 tex = i915_texture(i915->fragment_sampler_views[unit]->texture);
-	 i915->validation_buffers[i915->num_validation_buffers++] = tex->buffer;
+         tex = i915_texture(i915->fragment_sampler_views[unit]->texture);
+         i915->validation_buffers[i915->num_validation_buffers++] = tex->buffer;
       }
    }
 }
 
-const static struct i915_tracked_hw_state hw_atoms[] = {
-   { "flush", NULL, emit_flush, I915_HW_FLUSH, 1 },
-   { "immediate", validate_immediate, NULL, I915_HW_IMMEDIATE },
-   { "static", validate_static, NULL, I915_HW_STATIC },
-   { "map", validate_map, NULL, I915_HW_MAP }
-};
+static void
+emit_map(struct i915_context *i915)
+{
+   const uint nr = i915->current.sampler_enable_nr;
+   if (nr) {
+      const uint enabled = i915->current.sampler_enable_flags;
+      uint unit;
+      uint count = 0;
+      OUT_BATCH(_3DSTATE_MAP_STATE | (3 * nr));
+      OUT_BATCH(enabled);
+      for (unit = 0; unit < I915_TEX_UNITS; unit++) {
+         if (enabled & (1 << unit)) {
+            struct i915_texture *texture = i915_texture(i915->fragment_sampler_views[unit]->texture);
+            struct i915_winsys_buffer *buf = texture->buffer;
+            assert(buf);
+
+            count++;
+
+            OUT_RELOC(buf, I915_USAGE_SAMPLER, 0);
+            OUT_BATCH(i915->current.texbuffer[unit][0]); /* MS3 */
+            OUT_BATCH(i915->current.texbuffer[unit][1]); /* MS4 */
+         }
+      }
+      assert(count == nr);
+   }
+}
+
+static void
+validate_sampler(struct i915_context *i915, unsigned *batch_space)
+{
+   *batch_space = i915->current.sampler_enable_nr ?
+     2 + 3*i915->current.sampler_enable_nr : 0;
+}
+
+static void
+emit_sampler(struct i915_context *i915)
+{
+   if (i915->current.sampler_enable_nr) {
+      int i;
+
+      OUT_BATCH( _3DSTATE_SAMPLER_STATE | 
+                 (3 * i915->current.sampler_enable_nr) );
+
+      OUT_BATCH( i915->current.sampler_enable_flags );
+
+      for (i = 0; i < I915_TEX_UNITS; i++) {
+         if (i915->current.sampler_enable_flags & (1<<i)) {
+            OUT_BATCH( i915->current.sampler[i][0] );
+            OUT_BATCH( i915->current.sampler[i][1] );
+            OUT_BATCH( i915->current.sampler[i][2] );
+         }
+      }
+   }
+}
+
+static void
+validate_constants(struct i915_context *i915, unsigned *batch_space)
+{
+   *batch_space = i915->fs->num_constants ?
+      2 + 4*i915->fs->num_constants : 0;
+}
+
+static void
+emit_constants(struct i915_context *i915)
+{
+   /* Collate the user-defined constants with the fragment shader's
+    * immediates according to the constant_flags[] array.
+    */
+   const uint nr = i915->fs->num_constants;
+   if (nr) {
+      uint i;
+
+      OUT_BATCH( _3DSTATE_PIXEL_SHADER_CONSTANTS | (nr * 4) );
+      OUT_BATCH((1 << nr) - 1);
+
+      for (i = 0; i < nr; i++) {
+         const uint *c;
+         if (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER) {
+            /* grab user-defined constant */
+            c = (uint *) i915_buffer(i915->constants[PIPE_SHADER_FRAGMENT])->data;
+            c += 4 * i;
+         }
+         else {
+            /* emit program constant */
+            c = (uint *) i915->fs->constants[i];
+         }
+#if 0 /* debug */
+         {
+            float *f = (float *) c;
+            printf("Const %2d: %f %f %f %f %s\n", i, f[0], f[1], f[2], f[3],
+                   (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER
+                    ? "user" : "immediate"));
+         }
+#endif
+         OUT_BATCH(*c++);
+         OUT_BATCH(*c++);
+         OUT_BATCH(*c++);
+         OUT_BATCH(*c++);
+      }
+   }
+}
+
+static void
+validate_program(struct i915_context *i915, unsigned *batch_space)
+{
+   *batch_space = i915->fs->program_len;
+}
+
+static void
+emit_program(struct i915_context *i915)
+{
+      uint i;
+      /* we should always have, at least, a pass-through program */
+      assert(i915->fs->program_len > 0);
+      for (i = 0; i < i915->fs->program_len; i++) {
+         OUT_BATCH(i915->fs->program[i]);
+      }
+}
+
+static void
+emit_draw_rect(struct i915_context *i915)
+{
+   OUT_BATCH(_3DSTATE_DRAW_RECT_CMD);
+   OUT_BATCH(DRAW_RECT_DIS_DEPTH_OFS);
+   OUT_BATCH(i915->current.draw_offset);
+   OUT_BATCH(i915->current.draw_size);
+   OUT_BATCH(i915->current.draw_offset);
+}
 
 static boolean
 i915_validate_state(struct i915_context *i915, unsigned *batch_space)
 {
-   int i;
+   unsigned tmp;
 
    i915->num_validation_buffers = 0;
-   *batch_space = 0;
-
-   for (i = 0; i < Elements(hw_atoms); i++)
-      if ((i915->hardware_dirty & hw_atoms[i].dirty) && hw_atoms[i].validate) {
-	 hw_atoms[i].validate(i915);
-	 *batch_space += hw_atoms[i].batch_space;
-      }
+   if (i915->hardware_dirty & I915_HW_INVARIANT)
+      *batch_space = Elements(invariant_state);
+   else
+      *batch_space = 0;
+
+#define VALIDATE_ATOM(atom, hw_dirty) \
+   if (i915->hardware_dirty & hw_dirty) { \
+      validate_##atom(i915, &tmp); \
+      *batch_space += tmp; }
+   VALIDATE_ATOM(flush, I915_HW_FLUSH);
+   VALIDATE_ATOM(immediate, I915_HW_IMMEDIATE);
+   VALIDATE_ATOM(dynamic, I915_HW_DYNAMIC);
+   VALIDATE_ATOM(static, I915_HW_STATIC);
+   VALIDATE_ATOM(map, I915_HW_MAP);
+   VALIDATE_ATOM(sampler, I915_HW_SAMPLER);
+   VALIDATE_ATOM(constants, I915_HW_CONSTANTS);
+   VALIDATE_ATOM(program, I915_HW_PROGRAM);
+#undef VALIDATE_ATOM
 
    if (i915->num_validation_buffers == 0)
       return TRUE;
 
    if (!i915_winsys_validate_buffers(i915->batch, i915->validation_buffers,
-				     i915->num_validation_buffers))
+                                     i915->num_validation_buffers))
       return FALSE;
 
    return TRUE;
 }
 
-static void
-emit_state(struct i915_context *i915)
-{
-   int i;
-
-   for (i = 0; i < Elements(hw_atoms); i++)
-      if ((i915->hardware_dirty & hw_atoms[i].dirty) && hw_atoms[i].emit)
-	 hw_atoms[i].emit(i915);
-}
-
 /* Push the state into the sarea and/or texture memory.
  */
 void
@@ -182,247 +442,20 @@ i915_emit_hardware_state(struct i915_context *i915 )
    save_ptr = (uintptr_t)i915->batch->ptr;
    save_relocs = i915->batch->relocs;
 
-   emit_state(i915);
-   /* 14 dwords, 0 relocs */
-   if (i915->hardware_dirty & I915_HW_INVARIANT)
-   {
-      OUT_BATCH(_3DSTATE_AA_CMD |
-                AA_LINE_ECAAR_WIDTH_ENABLE |
-                AA_LINE_ECAAR_WIDTH_1_0 |
-                AA_LINE_REGION_WIDTH_ENABLE | AA_LINE_REGION_WIDTH_1_0);
-
-      OUT_BATCH(_3DSTATE_DFLT_DIFFUSE_CMD);
-      OUT_BATCH(0);
-
-      OUT_BATCH(_3DSTATE_DFLT_SPEC_CMD);
-      OUT_BATCH(0);
-      
-      OUT_BATCH(_3DSTATE_DFLT_Z_CMD);
-      OUT_BATCH(0);
-
-      OUT_BATCH(_3DSTATE_COORD_SET_BINDINGS |
-                CSB_TCB(0, 0) |
-                CSB_TCB(1, 1) |
-                CSB_TCB(2, 2) |
-                CSB_TCB(3, 3) |
-                CSB_TCB(4, 4) | 
-                CSB_TCB(5, 5) | 
-                CSB_TCB(6, 6) | 
-                CSB_TCB(7, 7));
-
-      OUT_BATCH(_3DSTATE_RASTER_RULES_CMD |
-                ENABLE_POINT_RASTER_RULE |
-                OGL_POINT_RASTER_RULE |
-                ENABLE_LINE_STRIP_PROVOKE_VRTX |
-                ENABLE_TRI_FAN_PROVOKE_VRTX |
-                LINE_STRIP_PROVOKE_VRTX(1) |
-                TRI_FAN_PROVOKE_VRTX(2) | 
-                ENABLE_TEXKILL_3D_4D | 
-                TEXKILL_4D);
-
-      OUT_BATCH(_3DSTATE_DEPTH_SUBRECT_DISABLE);
-
-      /* disable indirect state for now
-       */
-      OUT_BATCH(_3DSTATE_LOAD_INDIRECT | 0);
-      OUT_BATCH(0);
-   }
-
-   /* 7 dwords, 1 relocs */
-   if (i915->hardware_dirty & I915_HW_IMMEDIATE)
-   {
-      /* remove unwatned bits and S7 */
-      unsigned dirty = (1 << I915_IMMEDIATE_S0 | 1 << I915_IMMEDIATE_S1 |
-                        1 << I915_IMMEDIATE_S2 | 1 << I915_IMMEDIATE_S3 |
-                        1 << I915_IMMEDIATE_S3 | 1 << I915_IMMEDIATE_S4 |
-                        1 << I915_IMMEDIATE_S5 | 1 << I915_IMMEDIATE_S6) &
-                       i915->immediate_dirty;
-      int i, num = util_bitcount(dirty);
-      assert(num && num <= I915_MAX_IMMEDIATE);
-
-      OUT_BATCH(_3DSTATE_LOAD_STATE_IMMEDIATE_1 |
-                dirty << 4 | (num - 1));
-
-      if (i915->immediate_dirty & (1 << I915_IMMEDIATE_S0)) {
-         if (i915->vbo)
-            OUT_RELOC(i915->vbo, I915_USAGE_VERTEX,
-                      i915->current.immediate[I915_IMMEDIATE_S0]);
-         else
-            OUT_BATCH(0);
-      }
-
-      for (i = 1; i < I915_MAX_IMMEDIATE; i++) {
-         if (dirty & (1 << i))
-            OUT_BATCH(i915->current.immediate[i]);
-      }
-   } 
-
-#if 01
-   /* I915_MAX_DYNAMIC dwords, 0 relocs */
-   if (i915->hardware_dirty & I915_HW_DYNAMIC) 
-   {
-      int i;
-      for (i = 0; i < I915_MAX_DYNAMIC; i++) {
-         if (i915->dynamic_dirty & (1 << i))
-            OUT_BATCH(i915->current.dynamic[i]);
-      }
-   }
-#endif
-
-#if 01
-   /* 8 dwords, 2 relocs */
-   if (i915->hardware_dirty & I915_HW_STATIC)
-   {
-      if (i915->current.cbuf_bo) {
-         OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
-         OUT_BATCH(i915->current.cbuf_flags);
-         OUT_RELOC(i915->current.cbuf_bo,
-                   I915_USAGE_RENDER,
-                   0);
-      }
-
-      /* What happens if no zbuf??
-       */
-      if (i915->current.depth_bo) {
-         OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
-         OUT_BATCH(i915->current.depth_flags);
-         OUT_RELOC(i915->current.depth_bo,
-                   I915_USAGE_RENDER,
-                   0);
-      }
-
-      {
-         OUT_BATCH(_3DSTATE_DST_BUF_VARS_CMD);
-         OUT_BATCH(i915->current.dst_buf_vars);
-      }
-   }
-#endif
-
-#if 01
-      /* texture images */
-      /* 2 + I915_TEX_UNITS*3 dwords, I915_TEX_UNITS relocs */
-      if (i915->hardware_dirty & (I915_HW_MAP | I915_HW_SAMPLER))
-      {
-         const uint nr = i915->current.sampler_enable_nr;
-         if (nr) {
-            const uint enabled = i915->current.sampler_enable_flags;
-            uint unit;
-            uint count = 0;
-            OUT_BATCH(_3DSTATE_MAP_STATE | (3 * nr));
-            OUT_BATCH(enabled);
-            for (unit = 0; unit < I915_TEX_UNITS; unit++) {
-               if (enabled & (1 << unit)) {
-                  struct i915_texture *texture = i915_texture(i915->fragment_sampler_views[unit]->texture);
-                  struct i915_winsys_buffer *buf = texture->buffer;
-                  assert(buf);
-
-                  count++;
-
-                  OUT_RELOC(buf, I915_USAGE_SAMPLER, 0);
-                  OUT_BATCH(i915->current.texbuffer[unit][0]); /* MS3 */
-                  OUT_BATCH(i915->current.texbuffer[unit][1]); /* MS4 */
-               }
-            }
-            assert(count == nr);
-         }
-      }
-#endif
-
-#if 01
-   /* samplers */
-   /* 2 + I915_TEX_UNITS*3 dwords, 0 relocs */
-   if (i915->hardware_dirty & I915_HW_SAMPLER) 
-   {
-      if (i915->current.sampler_enable_nr) {
-         int i;
-
-         OUT_BATCH( _3DSTATE_SAMPLER_STATE | 
-                    (3 * i915->current.sampler_enable_nr) );
-
-         OUT_BATCH( i915->current.sampler_enable_flags );
-
-         for (i = 0; i < I915_TEX_UNITS; i++) {
-            if (i915->current.sampler_enable_flags & (1<<i)) {
-               OUT_BATCH( i915->current.sampler[i][0] );
-               OUT_BATCH( i915->current.sampler[i][1] );
-               OUT_BATCH( i915->current.sampler[i][2] );
-            }
-         }
-      }
-   }
-#endif
-
-#if 01
-   /* constants */
-   /* 2 + I915_MAX_CONSTANT*4 dwords, 0 relocs */
-   if (i915->hardware_dirty & I915_HW_CONSTANTS)
-   {
-      /* Collate the user-defined constants with the fragment shader's
-       * immediates according to the constant_flags[] array.
-       */
-      const uint nr = i915->fs->num_constants;
-      if (nr) {
-         uint i;
-
-         OUT_BATCH( _3DSTATE_PIXEL_SHADER_CONSTANTS | (nr * 4) );
-	 OUT_BATCH((1 << nr) - 1);
-
-         for (i = 0; i < nr; i++) {
-            const uint *c;
-            if (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER) {
-               /* grab user-defined constant */
-               c = (uint *) i915_buffer(i915->constants[PIPE_SHADER_FRAGMENT])->data;
-               c += 4 * i;
-            }
-            else {
-               /* emit program constant */
-               c = (uint *) i915->fs->constants[i];
-            }
-#if 0 /* debug */
-            {
-               float *f = (float *) c;
-               printf("Const %2d: %f %f %f %f %s\n", i, f[0], f[1], f[2], f[3],
-                      (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER
-                       ? "user" : "immediate"));
-            }
-#endif
-            OUT_BATCH(*c++);
-            OUT_BATCH(*c++);
-            OUT_BATCH(*c++);
-            OUT_BATCH(*c++);
-         }
-      }
-   }
-#endif
-
-#if 01
-   /* Fragment program */
-   /* i915->current.program_len dwords, 0 relocs */
-   if (i915->hardware_dirty & I915_HW_PROGRAM)
-   {
-      uint i;
-      /* we should always have, at least, a pass-through program */
-      assert(i915->fs->program_len > 0);
-      for (i = 0; i < i915->fs->program_len; i++) {
-         OUT_BATCH(i915->fs->program[i]);
-      }
-   }
-#endif
-
-#if 01
-   /* drawing surface size */
-   /* 6 dwords, 0 relocs */
-   if (i915->hardware_dirty & I915_HW_STATIC)
-   {
-      /* XXX flush only required when the draw_offset changes! */
-      OUT_BATCH(MI_FLUSH | INHIBIT_FLUSH_RENDER_CACHE);
-      OUT_BATCH(_3DSTATE_DRAW_RECT_CMD);
-      OUT_BATCH(DRAW_RECT_DIS_DEPTH_OFS);
-      OUT_BATCH(i915->current.draw_offset);
-      OUT_BATCH(i915->current.draw_size);
-      OUT_BATCH(i915->current.draw_offset);
-   }
-#endif
+#define EMIT_ATOM(atom, hw_dirty) \
+   if (i915->hardware_dirty & hw_dirty) \
+      emit_##atom(i915);
+   EMIT_ATOM(flush, I915_HW_FLUSH);
+   EMIT_ATOM(invariant, I915_HW_INVARIANT);
+   EMIT_ATOM(immediate, I915_HW_IMMEDIATE);
+   EMIT_ATOM(dynamic, I915_HW_DYNAMIC);
+   EMIT_ATOM(static, I915_HW_STATIC);
+   EMIT_ATOM(map, I915_HW_MAP);
+   EMIT_ATOM(sampler, I915_HW_SAMPLER);
+   EMIT_ATOM(constants, I915_HW_CONSTANTS);
+   EMIT_ATOM(program, I915_HW_PROGRAM);
+   EMIT_ATOM(draw_rect, I915_HW_STATIC);
+#undef EMIT_ATOM
 
    I915_DBG(DBG_EMIT, "%s: used %d dwords, %d relocs\n", __FUNCTION__,
             ((uintptr_t)i915->batch->ptr - save_ptr) / 4,
@@ -431,4 +464,5 @@ i915_emit_hardware_state(struct i915_context *i915 )
    i915->hardware_dirty = 0;
    i915->immediate_dirty = 0;
    i915->dynamic_dirty = 0;
+   i915->flush_dirty = 0;
 }
diff --git a/src/gallium/drivers/i915/i915_state_static.c b/src/gallium/drivers/i915/i915_state_static.c
index 9704449999..20cd23f8f7 100644
--- a/src/gallium/drivers/i915/i915_state_static.c
+++ b/src/gallium/drivers/i915/i915_state_static.c
@@ -164,7 +164,7 @@ static void update_framebuffer(struct i915_context *i915)
    assert(ret);
    if (i915->current.draw_offset != draw_offset) {
       i915->current.draw_offset = draw_offset;
-      /* XXX: only emit flush on change and not always in emit */
+      i915_set_flush_dirty(i915, I915_PIPELINE_FLUSH);
    }
    i915->current.draw_size = (w - 1 + x) | ((h - 1 + y) << 16);