anv: further optimize dirty state after secondary emission

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41581>

src/intel/vulkan/anv_private.h

@@ -1984,6 +1984,7 @@ enum anv_gfx_state_bits {
    ANV_GFX_STATE_WM,
    ANV_GFX_STATE_WM_DEPTH_STENCIL,
    ANV_GFX_STATE_PS_EXTRA,
+
    ANV_GFX_STATE_PMA_FIX, /* Fake state to implement workaround */
    ANV_GFX_STATE_WA_18019816803, /* Fake state to implement workaround */
    ANV_GFX_STATE_WA_14018283232, /* Fake state to implement workaround */
@@ -2492,6 +2493,9 @@ struct anv_gfx_dynamic_state {
 
    /** Dirty bits of what needs to be reemitted */
    BITSET_DECLARE(emit_dirty, ANV_GFX_STATE_MAX);
+
+   /** Emitted bits */
+   BITSET_DECLARE(emitted, ANV_GFX_STATE_MAX);
 };
 
 enum anv_internal_kernel_name {

src/intel/vulkan/genX_blorp_exec.c

@@ -432,6 +432,9 @@ blorp_exec_on_render(struct blorp_batch *batch,
       BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_PS_BLEND);
    }
 
+   /* Add the flagged instructions as emitted */
+   BITSET_OR(hw_state->emitted, hw_state->emitted, hw_state->emit_dirty);
+
    anv_cmd_dirty_mask_t dirty = ~(ANV_CMD_DIRTY_INDEX_BUFFER |
                                   ANV_CMD_DIRTY_XFB_ENABLE |
                                   ANV_CMD_DIRTY_OCCLUSION_QUERY_ACTIVE |

src/intel/vulkan/genX_cmd_buffer.c

@@ -4415,6 +4415,20 @@ genX(CmdExecuteCommands)(
 
       container->state.compute.trace_rays_active |=
          secondary->state.compute.trace_rays_active;
+
+      /* For each GFX instruction emitted in the secondary, mark it dirty in
+       * the container, so it's reemited. Even though Vulkan spec says that
+       * after a secondary command buffer is executed the state in the primary
+       * is undefined, our emission optimization code will avoid dirtying an
+       * instruction if the values inside an instruction haven't changed, but
+       * it doesn't see that this was potentially changed by the secondary.
+       *
+       * TODO: do an ultimate version of this by diffing secondary/container
+       *       emitted instructions
+       */
+      BITSET_OR(container->state.gfx.dyn_state.emit_dirty,
+                container->state.gfx.dyn_state.emit_dirty,
+                secondary->state.gfx.dyn_state.emitted);
    }
 
    /* The secondary isn't counted in our VF cache tracking so we need to
@@ -4446,10 +4460,6 @@ genX(CmdExecuteCommands)(
 
    memset(&container->state.gfx.urb_cfg, 0, sizeof(struct intel_urb_config));
 
-   /* Reemit all GFX instructions in container */
-   BITSET_OR(container->state.gfx.dyn_state.emit_dirty,
-             container->state.gfx.dyn_state.emit_dirty,
-             device->gfx_dirty_state);
    if (container->device->vk.enabled_extensions.KHR_fragment_shading_rate) {
       /* Also recompute the CPS_STATE offset */
       struct vk_dynamic_graphics_state *dyn =

src/intel/vulkan/genX_gfx_state.c

@@ -3646,6 +3646,9 @@ cmd_buffer_gfx_state_emission(struct anv_cmd_buffer *cmd_buffer)
       mi_builder_set_mocs(&b, isl_mocs(&device->isl_dev, 0, false));
    }
 
+   /* Save all the instructions we're about to emit */
+   BITSET_OR(hw_state->emitted, hw_state->emitted, hw_state->emit_dirty);
+
 #if INTEL_WA_16011107343_GFX_VER
    /* Will be emitted in front of every draw instead */
    if (intel_needs_workaround(device->info, 16011107343) &&

src/intel/vulkan/genX_gpu_memcpy.c

@@ -373,6 +373,9 @@ genX(emit_so_memcpy_fini)(struct anv_memcpy_state *state,
          BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TASK_CONTROL);
       }
 
+      /* Add the flagged instructions as emitted */
+      BITSET_OR(hw_state->emitted, hw_state->emitted, hw_state->emit_dirty);
+
       state->cmd_buffer->state.gfx.dirty |=
          ~(ANV_CMD_DIRTY_ALL_SHADERS(state->device) |
            ANV_CMD_DIRTY_INDEX_BUFFER |

src/intel/vulkan/genX_simple_shader.c

@@ -358,6 +358,9 @@ genX(emit_simpler_shader_init_fragment)(struct anv_simple_shader *state)
       BITSET_SET(hw_state->emit_dirty, ANV_GFX_STATE_TASK_CONTROL);
    }
 
+   /* Add the flagged instructions as emitted */
+   BITSET_OR(hw_state->emitted, hw_state->emitted, hw_state->emit_dirty);
+
    /* Update urb config after simple shader. */
    memcpy(&state->cmd_buffer->state.gfx.urb_cfg, &urb_cfg,
           sizeof(urb_cfg));