radeonsi: always use Wave32 for GS fast launch, because Wave64 hangs

[mesa.git] / src / gallium / drivers / radeonsi / si_pipe.h

diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 746b39161ad1d839cf8fecb2459d54d490597943..9ad14cab96a28f44729f26dcfff71bd83400283e 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -709,6 +709,8 @@ struct si_framebuffer {
    bool CB_has_shader_readable_metadata;
    bool DB_has_shader_readable_metadata;
    bool all_DCC_pipe_aligned;
+   bool color_big_page;
+   bool zs_big_page;
 };
 
 enum si_quant_mode
@@ -1009,9 +1011,9 @@ struct si_context {
    struct pipe_scissor_state window_rectangles[4];
 
    /* Precomputed states. */
-   struct si_pm4_state *init_config;
-   struct si_pm4_state *init_config_gs_rings;
-   bool init_config_has_vgt_flush;
+   struct si_pm4_state *cs_preamble_state;
+   struct si_pm4_state *cs_preamble_gs_rings;
+   bool cs_preamble_has_vgt_flush;
    struct si_pm4_state *vgt_shader_config[SI_NUM_VGT_STAGES_STATES];
 
    /* shaders */
@@ -1030,6 +1032,8 @@ struct si_context {
    unsigned cs_max_waves_per_sh;
    bool flatshade;
    bool do_update_shaders;
+   bool compute_shaderbuf_sgprs_dirty;
+   bool compute_image_sgprs_dirty;
 
    /* shader descriptors */
    struct si_descriptors descriptors[SI_NUM_DESCS];
@@ -1407,6 +1411,7 @@ void si_init_screen_get_functions(struct si_screen *sscreen);
 /* si_gfx_cs.c */
 void si_flush_gfx_cs(struct si_context *ctx, unsigned flags, struct pipe_fence_handle **fence);
 void si_allocate_gds(struct si_context *ctx);
+void si_set_tracked_regs_to_clear_state(struct si_context *ctx);
 void si_begin_new_gfx_cs(struct si_context *ctx);
 void si_need_gfx_cs_space(struct si_context *ctx);
 void si_unref_sdma_uploads(struct si_context *sctx);
@@ -1875,12 +1880,14 @@ static inline bool si_compute_prim_discard_enabled(struct si_context *sctx)
 
 static inline unsigned si_get_wave_size(struct si_screen *sscreen,
                                         enum pipe_shader_type shader_type, bool ngg, bool es,
-                                        bool prim_discard_cs)
+                                        bool gs_fast_launch, bool prim_discard_cs)
 {
    if (shader_type == PIPE_SHADER_COMPUTE)
       return sscreen->compute_wave_size;
    else if (shader_type == PIPE_SHADER_FRAGMENT)
       return sscreen->ps_wave_size;
+   else if (gs_fast_launch)
+      return 32; /* GS fast launch hangs with Wave64, so always use Wave32. */
    else if ((shader_type == PIPE_SHADER_VERTEX && prim_discard_cs) || /* only Wave64 implemented */
             (shader_type == PIPE_SHADER_VERTEX && es && !ngg) ||
             (shader_type == PIPE_SHADER_TESS_EVAL && es && !ngg) ||
@@ -1893,7 +1900,9 @@ static inline unsigned si_get_wave_size(struct si_screen *sscreen,
 static inline unsigned si_get_shader_wave_size(struct si_shader *shader)
 {
    return si_get_wave_size(shader->selector->screen, shader->selector->type, shader->key.as_ngg,
-                           shader->key.as_es, shader->key.opt.vs_as_prim_discard_cs);
+                           shader->key.as_es,
+                           shader->key.opt.ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_ALL,
+                           shader->key.opt.vs_as_prim_discard_cs);
 }
 
 #define PRINT_ERR(fmt, args...)                                                                    \