Gfxbench Manhattan3.0 optimization

Eliminate draw call for Manhattan robot shadow when conditions
apply. The performance of Gfxbench 1080pManhattan3.0 offscreen can
be improved 2.4%.

Tracked-On: OAM-122053
Signed-off-by: Huang Rui <rui1.huang@intel.com>

Author: huangrui666

src/gallium/drivers/iris/driinfo_iris.h

@@ -6,6 +6,7 @@ DRI_CONF_SECTION_DEBUG
    DRI_CONF_ALWAYS_FLUSH_CACHE(false)
    DRI_CONF_OPT_B(sync_compile, false, "Always compile synchronously (will cause stalls)")
    DRI_CONF_LIMIT_TRIG_INPUT_RANGE(false)
+   DRI_CONF_MHT_ROBOT_SHADOW_SHADER_ELIMINATION(false)
 DRI_CONF_SECTION_END
 
 DRI_CONF_SECTION_PERFORMANCE

src/gallium/drivers/iris/iris_clear.c

@@ -561,6 +561,8 @@ fast_clear_depth(struct iris_context *ice,
                                ISL_AUX_STATE_CLEAR);
    ice->state.dirty |= IRIS_DIRTY_DEPTH_BUFFER;
    ice->state.stage_dirty |= IRIS_ALL_STAGE_DIRTY_BINDINGS;
+   struct pipe_resource *p_res = (void *) res;
+   p_res->depth_cleared = true;
 }
 
 static void

src/gallium/drivers/iris/iris_context.h

@@ -866,6 +866,8 @@ struct iris_context {
 
       /** Resource holding the pixel pipe hashing tables. */
       struct pipe_resource *pixel_hashing_tables;
+
+      bool skipManhattanRobotShadowShader;
    } state;
 };
 

src/gallium/drivers/iris/iris_draw.c

@@ -296,6 +296,10 @@ iris_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info,
                                         stage, true);
       }
       iris_predraw_resolve_framebuffer(ice, batch, draw_aux_buffer_disabled);
+      if (screen->driconf.mht_robot_shadow_shader_elimination && ice->state.skipManhattanRobotShadowShader) {
+         ice->state.skipManhattanRobotShadowShader = false;
+         return;
+      };
    }
 
    if (ice->state.dirty & IRIS_DIRTY_RENDER_MISC_BUFFER_FLUSHES) {

src/gallium/drivers/iris/iris_resolve.c

@@ -97,7 +97,15 @@ resolve_sampler_views(struct iris_context *ice,
          continue;
 
       struct iris_sampler_view *isv = shs->textures[i];
-
+      if (isv->res->base.b.depth_cleared) {
+         isv->res->base.b.depth_cleared = false;
+         struct iris_screen *screen = (struct iris_screen*)ice->ctx.screen;
+         if ( screen->driconf.mht_robot_shadow_shader_elimination &&
+            (*(uint32_t*)ice->shaders.uncompiled[MESA_SHADER_FRAGMENT]->nir->info.source_sha1 == 880113298 ||
+             *(uint32_t*)ice->shaders.uncompiled[MESA_SHADER_FRAGMENT]->nir->info.source_sha1 == 2932317437)) {
+            ice->state.skipManhattanRobotShadowShader = true;
+         }
+      }
       if (isv->res->base.b.target != PIPE_BUFFER) {
          if (consider_framebuffer) {
             disable_rb_aux_buffer(ice, draw_aux_buffer_disabled, isv->res,
@@ -215,6 +223,9 @@ iris_predraw_resolve_framebuffer(struct iris_context *ice,
       if (zs_surf) {
          struct iris_resource *z_res, *s_res;
          iris_get_depth_stencil_resources(zs_surf->texture, &z_res, &s_res);
+         if (z_res->base.b.depth_cleared) {
+            z_res->base.b.depth_cleared = false;
+         }
          unsigned num_layers =
             zs_surf->u.tex.last_layer - zs_surf->u.tex.first_layer + 1;
 

src/gallium/drivers/iris/iris_screen.c

@@ -845,6 +845,8 @@ iris_screen_create(int fd, const struct pipe_screen_config *config)
       driQueryOptionb(config->options, "limit_trig_input_range");
    screen->driconf.lower_depth_range_rate =
       driQueryOptionf(config->options, "lower_depth_range_rate");
+   screen->driconf.mht_robot_shadow_shader_elimination =
+      driQueryOptionf(config->options, "mht_robot_shadow_shader_elimination");
 
    screen->precompile = debug_get_bool_option("shader_precompile", true);
 

src/gallium/drivers/iris/iris_screen.h

@@ -183,6 +183,7 @@ struct iris_screen {
       bool sync_compile;
       bool limit_trig_input_range;
       float lower_depth_range_rate;
+      bool mht_robot_shadow_shader_elimination;
    } driconf;
 
    /** Does the kernel support various features (KERNEL_HAS_* bitfield)? */

src/gallium/include/pipe/p_state.h

@@ -565,6 +565,7 @@ struct pipe_resource
    uint16_t height0; /* Textures: The maximum height/depth/array_size is 16k. */
    uint16_t depth0;
    uint16_t array_size;
+   bool depth_cleared;
 
    enum pipe_format format:16;         /**< PIPE_FORMAT_x */
    enum pipe_texture_target target:8; /**< PIPE_TEXTURE_x */

src/util/00-mesa-defaults.conf

@@ -980,6 +980,12 @@ TODO: document the other workarounds.
         <application name="Insurgency" executable="insurgency_linux">
             <option name="force_gl_vendor" value="X.Org" />
         </application>
+        <application name="Manhattan3.0" executable="testfw_app">
+            <option name="mht_robot_shadow_shader_elimination" value="true" />
+        </application>
+        <application name="Manhattan3.0" executable="net.kishonti.gfxbench.gl.v50000.corporate">
+            <option name="mht_robot_shadow_shader_elimination" value="true" />
+        </application>
     </device>
     <device driver="crocus">
         <application name="glmark2" executable="glmark2">

src/util/driconf.h

@@ -302,6 +302,10 @@
    DRI_CONF_OPT_B(limit_trig_input_range, def, \
                   "Limit trig input range to [-2p : 2p] to improve sin/cos calculation precision on Intel")
 
+#define DRI_CONF_MHT_ROBOT_SHADOW_SHADER_ELIMINATION(def) \
+   DRI_CONF_OPT_B(mht_robot_shadow_shader_elimination, def, \
+                  "Eliminate draw call for Gfxbench Manhattan3.0 robot shadow when conditions apply.")
+
 /**
  * \brief Image quality-related options
  */