radv: allow to disable sinking of load inputs for FS via drirc

To workaround game bugs where partial derivatives are used in
non-uniform control flow. A proper solution needs to be implemented,
but as a quick fix disabling nir_opt_sink() works.

Cc: mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16165>

src/amd/vulkan/radv_device.c

@@ -959,6 +959,7 @@ static const driOptionDescription radv_dri_options[] = {
       DRI_CONF_RADV_REQUIRE_ETC2(false)
       DRI_CONF_RADV_DISABLE_HTILE_LAYERS(false)
       DRI_CONF_RADV_DISABLE_ANISO_SINGLE_LEVEL(false)
+      DRI_CONF_RADV_DISABLE_SINKING_LOAD_INPUT_FS(false)
    DRI_CONF_SECTION_END
 };
 // clang-format on
@@ -1010,6 +1011,9 @@ radv_init_dri_options(struct radv_instance *instance)
 
    instance->disable_aniso_single_level =
       driQueryOptionb(&instance->dri_options, "radv_disable_aniso_single_level");
+
+   instance->disable_sinking_load_input_fs =
+      driQueryOptionb(&instance->dri_options, "radv_disable_sinking_load_input_fs");
 }
 
 VKAPI_ATTR VkResult VKAPI_CALL

src/amd/vulkan/radv_pipeline.c

@@ -3151,6 +3151,9 @@ radv_generate_graphics_pipeline_key(const struct radv_pipeline *pipeline,
         device->physical_device->rad_info.family == CHIP_VANGOGH))
       key.adjust_frag_coord_z = true;
 
+   if (device->instance->disable_sinking_load_input_fs)
+      key.disable_sinking_load_input_fs = true;
+
    return key;
 }
 
@@ -4470,7 +4473,11 @@ radv_create_shaders(struct radv_pipeline *pipeline, struct radv_pipeline_layout
                            .allow_fp16 = device->physical_device->rad_info.chip_class >= GFX9,
                         });
 
-         nir_opt_sink(stages[i].nir, nir_move_load_input | nir_move_const_undef | nir_move_copies);
+         nir_move_options sink_opts = nir_move_const_undef | nir_move_copies;
+         if (i != MESA_SHADER_FRAGMENT || !pipeline_key->disable_sinking_load_input_fs)
+            sink_opts |= nir_move_load_input;
+
+         nir_opt_sink(stages[i].nir, sink_opts);
          nir_opt_move(stages[i].nir, nir_move_load_input | nir_move_const_undef | nir_move_copies);
 
          /* Lower I/O intrinsics to memory instructions. */
@@ -4513,9 +4520,12 @@ radv_create_shaders(struct radv_pipeline *pipeline, struct radv_pipeline_layout
 
          /* cleanup passes */
          nir_lower_load_const_to_scalar(stages[i].nir);
+
+         sink_opts |= nir_move_comparisons | nir_move_load_ubo | nir_move_load_ssbo;
+         nir_opt_sink(stages[i].nir, sink_opts);
+
          nir_move_options move_opts = nir_move_const_undef | nir_move_load_ubo |
                                       nir_move_load_input | nir_move_comparisons | nir_move_copies;
-         nir_opt_sink(stages[i].nir, move_opts | nir_move_load_ssbo);
          nir_opt_move(stages[i].nir, move_opts);
 
          stages[i].feedback.duration += os_time_get_nano() - stage_start;

src/amd/vulkan/radv_private.h

@@ -354,6 +354,7 @@ struct radv_instance {
    bool disable_htile_layers;
    bool disable_aniso_single_level;
    bool zero_vram;
+   bool disable_sinking_load_input_fs;
 };
 
 VkResult radv_init_wsi(struct radv_physical_device *physical_device);

src/amd/vulkan/radv_shader.h

@@ -63,6 +63,7 @@ struct radv_pipeline_key {
    uint32_t use_ngg : 1;
    uint32_t adjust_frag_coord_z : 1;
    uint32_t disable_aniso_single_level : 1;
+   uint32_t disable_sinking_load_input_fs : 1;
 
    struct {
       uint32_t instance_rate_inputs;

src/util/driconf.h

@@ -576,6 +576,10 @@
   DRI_CONF_OPT_B(radv_disable_aniso_single_level, def, \
                  "Disable anisotropic filtering for single level images")
 
+#define DRI_CONF_RADV_DISABLE_SINKING_LOAD_INPUT_FS(def) \
+   DRI_CONF_OPT_B(radv_disable_sinking_load_input_fs, def, \
+                  "Disable sinking load inputs for fragment shaders")
+
 /**
  * \brief ANV specific configuration options
  */