anv/gpu_memcpy: CS Stall before a MI memcpy on gen7

[mesa.git] / src / intel / vulkan / genX_blorp_exec.c

diff --git a/src/intel/vulkan/genX_blorp_exec.c b/src/intel/vulkan/genX_blorp_exec.c
index 7f22b677ea69a38ce79df625ef8292d5023fce49..f956715228604bf46753108291e190111727a37a 100644 (file)
--- a/src/intel/vulkan/genX_blorp_exec.c
+++ b/src/intel/vulkan/genX_blorp_exec.c
@@ -64,6 +64,16 @@ blorp_surface_reloc(struct blorp_batch *batch, uint32_t ss_offset,
       anv_batch_set_error(&cmd_buffer->batch, result);
 }
 
+static struct blorp_address
+blorp_get_surface_base_address(struct blorp_batch *batch)
+{
+   struct anv_cmd_buffer *cmd_buffer = batch->driver_batch;
+   return (struct blorp_address) {
+      .buffer = &cmd_buffer->device->surface_state_pool.block_pool.bo,
+      .offset = 0,
+   };
+}
+
 static void *
 blorp_alloc_dynamic_state(struct blorp_batch *batch,
                           uint32_t size,
@@ -132,27 +142,43 @@ blorp_alloc_vertex_buffer(struct blorp_batch *batch, uint32_t size,
       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 64);
 
    *addr = (struct blorp_address) {
-      .buffer = &cmd_buffer->device->dynamic_state_block_pool.bo,
+      .buffer = &cmd_buffer->device->dynamic_state_pool.block_pool.bo,
       .offset = vb_state.offset,
+      .mocs = cmd_buffer->device->default_mocs,
    };
 
    return vb_state.map;
 }
 
+#if GEN_GEN >= 8
+static struct blorp_address
+blorp_get_workaround_page(struct blorp_batch *batch)
+{
+   struct anv_cmd_buffer *cmd_buffer = batch->driver_batch;
+
+   return (struct blorp_address) {
+      .buffer = &cmd_buffer->device->workaround_bo,
+   };
+}
+#endif
+
 static void
 blorp_flush_range(struct blorp_batch *batch, void *start, size_t size)
 {
    struct anv_device *device = batch->blorp->driver_ctx;
    if (!device->info.has_llc)
-      anv_flush_range(start, size);
+      gen_flush_range(start, size);
 }
 
 static void
-blorp_emit_urb_config(struct blorp_batch *batch, unsigned vs_entry_size)
+blorp_emit_urb_config(struct blorp_batch *batch,
+                      unsigned vs_entry_size, unsigned sf_entry_size)
 {
    struct anv_device *device = batch->blorp->driver_ctx;
    struct anv_cmd_buffer *cmd_buffer = batch->driver_batch;
 
+   assert(sf_entry_size == 0);
+
    const unsigned entry_size[4] = { vs_entry_size, 1, 1, 1 };
 
    genX(emit_urb_setup)(device, &cmd_buffer->batch,
@@ -174,6 +200,16 @@ genX(blorp_exec)(struct blorp_batch *batch,
       genX(cmd_buffer_config_l3)(cmd_buffer, cfg);
    }
 
+#if GEN_GEN == 7
+   /* The MI_LOAD/STORE_REGISTER_MEM commands which BLORP uses to implement
+    * indirect fast-clear colors can cause GPU hangs if we don't stall first.
+    * See genX(cmd_buffer_mi_memcpy) for more details.
+    */
+   assert(params->src.clear_color_addr.buffer == NULL);
+   if (params->dst.clear_color_addr.buffer)
+      cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
+#endif
+
    genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
 
    genX(flush_pipeline_select_3d)(cmd_buffer);
@@ -192,7 +228,7 @@ genX(blorp_exec)(struct blorp_batch *batch,
 
    blorp_exec(batch, params);
 
-   cmd_buffer->state.vb_dirty = ~0;
-   cmd_buffer->state.dirty = ~0;
+   cmd_buffer->state.gfx.vb_dirty = ~0;
+   cmd_buffer->state.gfx.dirty = ~0;
    cmd_buffer->state.push_constants_dirty = ~0;
 }