i965/fs: Add support for translating ir_triop_fma into MAD.
[mesa.git] / src / mesa / drivers / dri / i965 / intel_batchbuffer.c
index 7a6293b557830aa1928f8656b6cbf1009f1edad1..56048293ade71225b9856adea734fe283767febb 100644 (file)
  * 
  **************************************************************************/
 
-#include "imports.h"
 #include "intel_batchbuffer.h"
-#include "intel_ioctl.h"
-#include "intel_decode.h"
-#include "bufmgr.h"
-
-
-static void intel_batchbuffer_reset( struct intel_batchbuffer *batch )
+#include "intel_buffer_objects.h"
+#include "intel_reg.h"
+#include "intel_bufmgr.h"
+#include "intel_buffers.h"
+#include "brw_context.h"
+
+static void
+intel_batchbuffer_reset(struct brw_context *brw);
+
+struct cached_batch_item {
+   struct cached_batch_item *next;
+   uint16_t header;
+   uint16_t size;
+};
+
+static void
+clear_cache(struct brw_context *brw)
 {
-   assert(batch->map == NULL);
+   struct cached_batch_item *item = brw->batch.cached_items;
 
-   batch->offset = (unsigned long)batch->ptr;
-   batch->offset = ALIGN(batch->offset, 64);
-   batch->ptr = (unsigned char *) batch->offset;
-
-   if (BATCH_SZ - batch->offset < BATCH_REFILL) {
-      bmBufferData(batch->intel, 
-                  batch->buffer,
-                  BATCH_SZ, 
-                  NULL, 
-                  0); 
-      batch->offset = 0;
-      batch->ptr = NULL;
+   while (item) {
+      struct cached_batch_item *next = item->next;
+      free(item);
+      item = next;
    }
-               
-   batch->flags = 0;
+
+   brw->batch.cached_items = NULL;
 }
 
-static void intel_batchbuffer_reset_cb( struct intel_context *intel,
-                                       void *ptr )
+void
+intel_batchbuffer_init(struct brw_context *brw)
 {
-   struct intel_batchbuffer *batch = (struct intel_batchbuffer *)ptr;
-   assert(batch->map == NULL);
-   batch->flags = 0;
-   batch->offset = 0;
-   batch->ptr = NULL;
+   intel_batchbuffer_reset(brw);
+
+   if (brw->gen >= 6) {
+      /* We can't just use brw_state_batch to get a chunk of space for
+       * the gen6 workaround because it involves actually writing to
+       * the buffer, and the kernel doesn't let us write to the batch.
+       */
+      brw->batch.workaround_bo = drm_intel_bo_alloc(brw->bufmgr,
+                                                     "pipe_control workaround",
+                                                     4096, 4096);
+   }
+
+   if (!brw->has_llc) {
+      brw->batch.cpu_map = malloc(BATCH_SZ);
+      brw->batch.map = brw->batch.cpu_map;
+   }
 }
 
-GLubyte *intel_batchbuffer_map( struct intel_batchbuffer *batch )
+static void
+intel_batchbuffer_reset(struct brw_context *brw)
 {
-   if (!batch->map) {
-      batch->map = bmMapBuffer(batch->intel, batch->buffer, 
-                              BM_MEM_AGP|BM_MEM_LOCAL|BM_CLIENT|BM_WRITE);
-      batch->ptr += (unsigned long)batch->map;
+   if (brw->batch.last_bo != NULL) {
+      drm_intel_bo_unreference(brw->batch.last_bo);
+      brw->batch.last_bo = NULL;
+   }
+   brw->batch.last_bo = brw->batch.bo;
+
+   clear_cache(brw);
+
+   brw->batch.bo = drm_intel_bo_alloc(brw->bufmgr, "batchbuffer",
+                                       BATCH_SZ, 4096);
+   if (brw->has_llc) {
+      drm_intel_bo_map(brw->batch.bo, true);
+      brw->batch.map = brw->batch.bo->virtual;
    }
 
-   return batch->map;
+   brw->batch.reserved_space = BATCH_RESERVED;
+   brw->batch.state_batch_offset = brw->batch.bo->size;
+   brw->batch.used = 0;
+   brw->batch.needs_sol_reset = false;
 }
 
-void intel_batchbuffer_unmap( struct intel_batchbuffer *batch )
+void
+intel_batchbuffer_save_state(struct brw_context *brw)
 {
-   if (batch->map) {
-      batch->ptr -= (unsigned long)batch->map;
-      batch->map = NULL;
-      bmUnmapBuffer(batch->intel, batch->buffer);
-   }
+   brw->batch.saved.used = brw->batch.used;
+   brw->batch.saved.reloc_count =
+      drm_intel_gem_bo_get_reloc_count(brw->batch.bo);
 }
 
+void
+intel_batchbuffer_reset_to_saved(struct brw_context *brw)
+{
+   drm_intel_gem_bo_clear_relocs(brw->batch.bo, brw->batch.saved.reloc_count);
 
+   brw->batch.used = brw->batch.saved.used;
 
-/*======================================================================
- * Public functions
- */
-struct intel_batchbuffer *intel_batchbuffer_alloc( struct intel_context *intel )
-{
-   struct intel_batchbuffer *batch = calloc(sizeof(*batch), 1);
+   /* Cached batch state is dead, since we just cleared some unknown part of the
+    * batchbuffer.  Assume that the caller resets any other state necessary.
+    */
+   clear_cache(brw);
+}
 
-   batch->intel = intel;
+void
+intel_batchbuffer_free(struct brw_context *brw)
+{
+   free(brw->batch.cpu_map);
+   drm_intel_bo_unreference(brw->batch.last_bo);
+   drm_intel_bo_unreference(brw->batch.bo);
+   drm_intel_bo_unreference(brw->batch.workaround_bo);
+   clear_cache(brw);
+}
 
-   bmGenBuffers(intel, "batch", 1, &batch->buffer, 12);
+static void
+do_batch_dump(struct brw_context *brw)
+{
+   struct drm_intel_decode *decode;
+   struct intel_batchbuffer *batch = &brw->batch;
+   int ret;
+
+   decode = drm_intel_decode_context_alloc(brw->intelScreen->deviceID);
+   if (!decode)
+      return;
+
+   ret = drm_intel_bo_map(batch->bo, false);
+   if (ret == 0) {
+      drm_intel_decode_set_batch_pointer(decode,
+                                        batch->bo->virtual,
+                                        batch->bo->offset,
+                                        batch->used);
+   } else {
+      fprintf(stderr,
+             "WARNING: failed to map batchbuffer (%s), "
+             "dumping uploaded data instead.\n", strerror(ret));
+
+      drm_intel_decode_set_batch_pointer(decode,
+                                        batch->map,
+                                        batch->bo->offset,
+                                        batch->used);
+   }
 
-   bmBufferSetInvalidateCB(intel, batch->buffer,
-                          intel_batchbuffer_reset_cb,
-                          batch,
-                          GL_TRUE);
+   drm_intel_decode(decode);
 
-   bmBufferData(batch->intel,
-               batch->buffer,
-               BATCH_SZ,
-               NULL,
-               0);
+   drm_intel_decode_context_free(decode);
 
+   if (ret == 0) {
+      drm_intel_bo_unmap(batch->bo);
 
-   return batch;
+      brw_debug_batch(brw);
+   }
 }
 
-void intel_batchbuffer_free( struct intel_batchbuffer *batch )
+/* TODO: Push this whole function into bufmgr.
+ */
+static int
+do_flush_locked(struct brw_context *brw)
 {
-   if (batch->map) 
-      bmUnmapBuffer(batch->intel, batch->buffer);
-   
-   bmDeleteBuffers(batch->intel, 1, &batch->buffer);
-   free(batch);
-}
+   struct intel_batchbuffer *batch = &brw->batch;
+   int ret = 0;
+
+   if (brw->has_llc) {
+      drm_intel_bo_unmap(batch->bo);
+   } else {
+      ret = drm_intel_bo_subdata(batch->bo, 0, 4*batch->used, batch->map);
+      if (ret == 0 && batch->state_batch_offset != batch->bo->size) {
+        ret = drm_intel_bo_subdata(batch->bo,
+                                   batch->state_batch_offset,
+                                   batch->bo->size - batch->state_batch_offset,
+                                   (char *)batch->map + batch->state_batch_offset);
+      }
+   }
 
+   if (!brw->intelScreen->no_hw) {
+      int flags;
+
+      if (brw->gen < 6 || !batch->is_blit) {
+        flags = I915_EXEC_RENDER;
+      } else {
+        flags = I915_EXEC_BLT;
+      }
+
+      if (batch->needs_sol_reset)
+        flags |= I915_EXEC_GEN7_SOL_RESET;
+
+      if (ret == 0) {
+         if (unlikely(INTEL_DEBUG & DEBUG_AUB))
+            brw_annotate_aub(brw);
+        if (brw->hw_ctx == NULL || batch->is_blit) {
+           ret = drm_intel_bo_mrb_exec(batch->bo, 4 * batch->used, NULL, 0, 0,
+                                       flags);
+        } else {
+           ret = drm_intel_gem_bo_context_exec(batch->bo, brw->hw_ctx,
+                                               4 * batch->used, flags);
+        }
+      }
+   }
 
-#define MI_BATCH_BUFFER_END    (0xA<<23)
+   if (unlikely(INTEL_DEBUG & DEBUG_BATCH))
+      do_batch_dump(brw);
 
+   if (ret != 0) {
+      fprintf(stderr, "intel_do_flush_locked failed: %s\n", strerror(-ret));
+      exit(1);
+   }
+   brw->vtbl.new_batch(brw);
 
-GLboolean intel_batchbuffer_flush( struct intel_batchbuffer *batch )
-{
-   struct intel_context *intel = batch->intel;
-   GLuint used = batch->ptr - (batch->map + batch->offset);
-   GLuint offset;
-   GLint retval = GL_TRUE;
+   return ret;
+}
 
-   assert(intel->locked);
+int
+_intel_batchbuffer_flush(struct brw_context *brw,
+                        const char *file, int line)
+{
+   int ret;
 
-   if (used == 0) {
-      bmReleaseBuffers( batch->intel );
-      return GL_TRUE;
-   }
+   if (brw->batch.used == 0)
+      return 0;
 
-   /* Add the MI_BATCH_BUFFER_END.  Always add an MI_FLUSH - this is a
-    * performance drain that we would like to avoid.
-    */
-   if (used & 4) {
-      ((int *)batch->ptr)[0] = MI_BATCH_BUFFER_END;
-      batch->ptr += 4;
-      used += 4;
+   if (brw->first_post_swapbuffers_batch == NULL) {
+      brw->first_post_swapbuffers_batch = brw->batch.bo;
+      drm_intel_bo_reference(brw->first_post_swapbuffers_batch);
    }
-   else {
-      ((int *)batch->ptr)[0] = 0;
-      ((int *)batch->ptr)[1] = MI_BATCH_BUFFER_END;
 
-      batch->ptr += 8;
-      used += 8;
+   if (unlikely(INTEL_DEBUG & DEBUG_BATCH)) {
+      int bytes_for_commands = 4 * brw->batch.used;
+      int bytes_for_state = brw->batch.bo->size - brw->batch.state_batch_offset;
+      int total_bytes = bytes_for_commands + bytes_for_state;
+      fprintf(stderr, "%s:%d: Batchbuffer flush with %4db (pkt) + "
+              "%4db (state) = %4db (%0.1f%%)\n", file, line,
+              bytes_for_commands, bytes_for_state,
+              total_bytes,
+              100.0f * total_bytes / BATCH_SZ);
    }
 
-   intel_batchbuffer_unmap(batch);
+   brw->batch.reserved_space = 0;
 
-   /* Get the batch buffer offset: Must call bmBufferOffset() before
-    * bmValidateBuffers(), otherwise the buffer won't be on the inuse
-    * list.
-    */
-   offset = bmBufferOffset(batch->intel, batch->buffer);
+   if (brw->vtbl.finish_batch)
+      brw->vtbl.finish_batch(brw);
 
-   if (bmValidateBuffers( batch->intel ) != 0) {
-      assert(intel->locked);
-      bmReleaseBuffers( batch->intel );
-      retval = GL_FALSE;
-      goto out;
+   /* Mark the end of the buffer. */
+   intel_batchbuffer_emit_dword(brw, MI_BATCH_BUFFER_END);
+   if (brw->batch.used & 1) {
+      /* Round batchbuffer usage to 2 DWORDs. */
+      intel_batchbuffer_emit_dword(brw, MI_NOOP);
    }
 
-   if (INTEL_DEBUG & DEBUG_BATCH) {
-      char *map;
+   intel_upload_finish(brw);
 
-      map = bmMapBuffer(batch->intel, batch->buffer,
-                       BM_MEM_AGP|BM_MEM_LOCAL|BM_CLIENT);
-      intel_decode((uint32_t *)(map + batch->offset), used / 4,
-                  offset + batch->offset, intel->intelScreen->deviceID);
-      bmUnmapBuffer(batch->intel, batch->buffer);
-   }
+   /* Check that we didn't just wrap our batchbuffer at a bad time. */
+   assert(!brw->no_batch_wrap);
 
-   /* Fire the batch buffer, which was uploaded above:
-    */
-   intel_batch_ioctl(batch->intel, 
-                    offset + batch->offset,
-                    used);
+   ret = do_flush_locked(brw);
+
+   if (unlikely(INTEL_DEBUG & DEBUG_SYNC)) {
+      fprintf(stderr, "waiting for idle\n");
+      drm_intel_bo_wait_rendering(brw->batch.bo);
+   }
 
    /* Reset the buffer:
     */
- out:
-   intel_batchbuffer_reset( batch );
-   intel_batchbuffer_map( batch );
+   intel_batchbuffer_reset(brw);
 
-   if (!retval)
-      DBG("%s failed\n", __FUNCTION__);
-
-   return retval;
+   return ret;
 }
 
 
+/*  This is the only way buffers get added to the validate list.
+ */
+bool
+intel_batchbuffer_emit_reloc(struct brw_context *brw,
+                             drm_intel_bo *buffer,
+                             uint32_t read_domains, uint32_t write_domain,
+                            uint32_t delta)
+{
+   int ret;
+
+   ret = drm_intel_bo_emit_reloc(brw->batch.bo, 4*brw->batch.used,
+                                buffer, delta,
+                                read_domains, write_domain);
+   assert(ret == 0);
+   (void)ret;
+
+   /*
+    * Using the old buffer offset, write in what the right data would be, in case
+    * the buffer doesn't move and we can short-circuit the relocation processing
+    * in the kernel
+    */
+   intel_batchbuffer_emit_dword(brw, buffer->offset + delta);
 
+   return true;
+}
 
+bool
+intel_batchbuffer_emit_reloc_fenced(struct brw_context *brw,
+                                   drm_intel_bo *buffer,
+                                   uint32_t read_domains,
+                                   uint32_t write_domain,
+                                   uint32_t delta)
+{
+   int ret;
+
+   ret = drm_intel_bo_emit_reloc_fence(brw->batch.bo, 4*brw->batch.used,
+                                      buffer, delta,
+                                      read_domains, write_domain);
+   assert(ret == 0);
+   (void)ret;
+
+   /*
+    * Using the old buffer offset, write in what the right data would
+    * be, in case the buffer doesn't move and we can short-circuit the
+    * relocation processing in the kernel
+    */
+   intel_batchbuffer_emit_dword(brw, buffer->offset + delta);
 
+   return true;
+}
 
-
-void intel_batchbuffer_align( struct intel_batchbuffer *batch,
-                             GLuint align,
-                             GLuint sz )
+void
+intel_batchbuffer_data(struct brw_context *brw,
+                       const void *data, GLuint bytes, bool is_blit)
 {
-   unsigned long ptr = (unsigned long) batch->ptr;
-   unsigned long aptr = ALIGN(ptr, align);
-   GLuint fixup = aptr - ptr;
+   assert((bytes & 3) == 0);
+   intel_batchbuffer_require_space(brw, bytes, is_blit);
+   __memcpy(brw->batch.map + brw->batch.used, data, bytes);
+   brw->batch.used += bytes >> 2;
+}
 
-   if (intel_batchbuffer_space(batch) < fixup + sz)
-      intel_batchbuffer_flush(batch);
-   else {
-      memset(batch->ptr, 0, fixup);      
-      batch->ptr += fixup;
+void
+intel_batchbuffer_cached_advance(struct brw_context *brw)
+{
+   struct cached_batch_item **prev = &brw->batch.cached_items, *item;
+   uint32_t sz = (brw->batch.used - brw->batch.emit) * sizeof(uint32_t);
+   uint32_t *start = brw->batch.map + brw->batch.emit;
+   uint16_t op = *start >> 16;
+
+   while (*prev) {
+      uint32_t *old;
+
+      item = *prev;
+      old = brw->batch.map + item->header;
+      if (op == *old >> 16) {
+        if (item->size == sz && memcmp(old, start, sz) == 0) {
+           if (prev != &brw->batch.cached_items) {
+              *prev = item->next;
+              item->next = brw->batch.cached_items;
+              brw->batch.cached_items = item;
+           }
+           brw->batch.used = brw->batch.emit;
+           return;
+        }
+
+        goto emit;
+      }
+      prev = &item->next;
    }
-}
 
+   item = malloc(sizeof(struct cached_batch_item));
+   if (item == NULL)
+      return;
 
+   item->next = brw->batch.cached_items;
+   brw->batch.cached_items = item;
 
+emit:
+   item->size = sz;
+   item->header = brw->batch.emit;
+}
 
-void intel_batchbuffer_data(struct intel_batchbuffer *batch,
-                           const void *data,
-                           GLuint bytes,
-                           GLuint flags)
+/**
+ * Restriction [DevSNB, DevIVB]:
+ *
+ * Prior to changing Depth/Stencil Buffer state (i.e. any combination of
+ * 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS, 3DSTATE_STENCIL_BUFFER,
+ * 3DSTATE_HIER_DEPTH_BUFFER) SW must first issue a pipelined depth stall
+ * (PIPE_CONTROL with Depth Stall bit set), followed by a pipelined depth
+ * cache flush (PIPE_CONTROL with Depth Flush Bit set), followed by
+ * another pipelined depth stall (PIPE_CONTROL with Depth Stall bit set),
+ * unless SW can otherwise guarantee that the pipeline from WM onwards is
+ * already flushed (e.g., via a preceding MI_FLUSH).
+ */
+void
+intel_emit_depth_stall_flushes(struct brw_context *brw)
 {
-   assert((bytes & 3) == 0);
-   intel_batchbuffer_require_space(batch, bytes, flags);
-   __memcpy(batch->ptr, data, bytes);
-   batch->ptr += bytes;
+   assert(brw->gen >= 6 && brw->gen <= 7);
+
+   BEGIN_BATCH(4);
+   OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
+   OUT_BATCH(PIPE_CONTROL_DEPTH_STALL);
+   OUT_BATCH(0); /* address */
+   OUT_BATCH(0); /* write data */
+   ADVANCE_BATCH()
+
+   BEGIN_BATCH(4);
+   OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
+   OUT_BATCH(PIPE_CONTROL_DEPTH_CACHE_FLUSH);
+   OUT_BATCH(0); /* address */
+   OUT_BATCH(0); /* write data */
+   ADVANCE_BATCH();
+
+   BEGIN_BATCH(4);
+   OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
+   OUT_BATCH(PIPE_CONTROL_DEPTH_STALL);
+   OUT_BATCH(0); /* address */
+   OUT_BATCH(0); /* write data */
+   ADVANCE_BATCH();
+}
+
+/**
+ * From the Ivybridge PRM, Volume 2 Part 1, Section 3.2 (VS Stage Input):
+ * "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth
+ *  stall needs to be sent just prior to any 3DSTATE_VS, 3DSTATE_URB_VS,
+ *  3DSTATE_CONSTANT_VS, 3DSTATE_BINDING_TABLE_POINTER_VS,
+ *  3DSTATE_SAMPLER_STATE_POINTER_VS command.  Only one PIPE_CONTROL needs
+ *  to be sent before any combination of VS associated 3DSTATE."
+ */
+void
+gen7_emit_vs_workaround_flush(struct brw_context *brw)
+{
+   assert(brw->gen == 7);
+
+   BEGIN_BATCH(4);
+   OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
+   OUT_BATCH(PIPE_CONTROL_DEPTH_STALL | PIPE_CONTROL_WRITE_IMMEDIATE);
+   OUT_RELOC(brw->batch.workaround_bo,
+            I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, 0);
+   OUT_BATCH(0); /* write data */
+   ADVANCE_BATCH();
 }
 
+/**
+ * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
+ * implementing two workarounds on gen6.  From section 1.4.7.1
+ * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
+ *
+ * [DevSNB-C+{W/A}] Before any depth stall flush (including those
+ * produced by non-pipelined state commands), software needs to first
+ * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
+ * 0.
+ *
+ * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
+ * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
+ *
+ * And the workaround for these two requires this workaround first:
+ *
+ * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
+ * BEFORE the pipe-control with a post-sync op and no write-cache
+ * flushes.
+ *
+ * And this last workaround is tricky because of the requirements on
+ * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
+ * volume 2 part 1:
+ *
+ *     "1 of the following must also be set:
+ *      - Render Target Cache Flush Enable ([12] of DW1)
+ *      - Depth Cache Flush Enable ([0] of DW1)
+ *      - Stall at Pixel Scoreboard ([1] of DW1)
+ *      - Depth Stall ([13] of DW1)
+ *      - Post-Sync Operation ([13] of DW1)
+ *      - Notify Enable ([8] of DW1)"
+ *
+ * The cache flushes require the workaround flush that triggered this
+ * one, so we can't use it.  Depth stall would trigger the same.
+ * Post-sync nonzero is what triggered this second workaround, so we
+ * can't use that one either.  Notify enable is IRQs, which aren't
+ * really our business.  That leaves only stall at scoreboard.
+ */
+void
+intel_emit_post_sync_nonzero_flush(struct brw_context *brw)
+{
+   if (!brw->batch.need_workaround_flush)
+      return;
+
+   BEGIN_BATCH(4);
+   OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
+   OUT_BATCH(PIPE_CONTROL_CS_STALL |
+            PIPE_CONTROL_STALL_AT_SCOREBOARD);
+   OUT_BATCH(0); /* address */
+   OUT_BATCH(0); /* write data */
+   ADVANCE_BATCH();
+
+   BEGIN_BATCH(4);
+   OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
+   OUT_BATCH(PIPE_CONTROL_WRITE_IMMEDIATE);
+   OUT_RELOC(brw->batch.workaround_bo,
+            I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION, 0);
+   OUT_BATCH(0); /* write data */
+   ADVANCE_BATCH();
+
+   brw->batch.need_workaround_flush = false;
+}
+
+/* Emit a pipelined flush to either flush render and texture cache for
+ * reading from a FBO-drawn texture, or flush so that frontbuffer
+ * render appears on the screen in DRI1.
+ *
+ * This is also used for the always_flush_cache driconf debug option.
+ */
+void
+intel_batchbuffer_emit_mi_flush(struct brw_context *brw)
+{
+   if (brw->gen >= 6) {
+      if (brw->batch.is_blit) {
+        BEGIN_BATCH_BLT(4);
+        OUT_BATCH(MI_FLUSH_DW);
+        OUT_BATCH(0);
+        OUT_BATCH(0);
+        OUT_BATCH(0);
+        ADVANCE_BATCH();
+      } else {
+        if (brw->gen == 6) {
+           /* Hardware workaround: SNB B-Spec says:
+            *
+            * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache
+            * Flush Enable =1, a PIPE_CONTROL with any non-zero
+            * post-sync-op is required.
+            */
+           intel_emit_post_sync_nonzero_flush(brw);
+        }
+
+        BEGIN_BATCH(4);
+        OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
+        OUT_BATCH(PIPE_CONTROL_INSTRUCTION_FLUSH |
+                  PIPE_CONTROL_WRITE_FLUSH |
+                  PIPE_CONTROL_DEPTH_CACHE_FLUSH |
+                   PIPE_CONTROL_VF_CACHE_INVALIDATE |
+                  PIPE_CONTROL_TC_FLUSH |
+                  PIPE_CONTROL_NO_WRITE |
+                   PIPE_CONTROL_CS_STALL);
+        OUT_BATCH(0); /* write address */
+        OUT_BATCH(0); /* write data */
+        ADVANCE_BATCH();
+      }
+   } else {
+      BEGIN_BATCH(4);
+      OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
+               PIPE_CONTROL_WRITE_FLUSH |
+               PIPE_CONTROL_NO_WRITE);
+      OUT_BATCH(0); /* write address */
+      OUT_BATCH(0); /* write data */
+      OUT_BATCH(0); /* write data */
+      ADVANCE_BATCH();
+   }
+}