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 6b934c1dff47d0aa22d871d2b58095755c34b97e..9ad14cab96a28f44729f26dcfff71bd83400283e 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -105,7 +105,7 @@
 #define SI_MAP_BUFFER_ALIGNMENT           64
 #define SI_MAX_VARIABLE_THREADS_PER_BLOCK 1024
 
-#define SI_RESOURCE_FLAG_TRANSFER          (PIPE_RESOURCE_FLAG_DRV_PRIV << 0)
+#define SI_RESOURCE_FLAG_FORCE_LINEAR      (PIPE_RESOURCE_FLAG_DRV_PRIV << 0)
 #define SI_RESOURCE_FLAG_FLUSHED_DEPTH     (PIPE_RESOURCE_FLAG_DRV_PRIV << 1)
 #define SI_RESOURCE_FLAG_FORCE_MSAA_TILING (PIPE_RESOURCE_FLAG_DRV_PRIV << 2)
 #define SI_RESOURCE_FLAG_DISABLE_DCC       (PIPE_RESOURCE_FLAG_DRV_PRIV << 3)
@@ -122,6 +122,7 @@
    (((x)&0x3) << SI_RESOURCE_FLAG_MICRO_TILE_MODE_SHIFT)
 #define SI_RESOURCE_FLAG_MICRO_TILE_MODE_GET(x)                                                    \
    (((x) >> SI_RESOURCE_FLAG_MICRO_TILE_MODE_SHIFT) & 0x3)
+#define SI_RESOURCE_FLAG_UNCACHED          (PIPE_RESOURCE_FLAG_DRV_PRIV << 12)
 
 enum si_clear_code
 {
@@ -239,6 +240,7 @@ enum si_coherency
    SI_COHERENCY_NONE, /* no cache flushes needed */
    SI_COHERENCY_SHADER,
    SI_COHERENCY_CB_META,
+   SI_COHERENCY_DB_META,
    SI_COHERENCY_CP,
 };
 
@@ -336,6 +338,7 @@ struct si_texture {
    uint8_t stencil_clear_value;
    bool fmask_is_identity : 1;
    bool tc_compatible_htile : 1;
+   bool enable_tc_compatible_htile_next_clear : 1;
    bool htile_stencil_disabled : 1;
    bool depth_cleared : 1;   /* if it was cleared at least once */
    bool stencil_cleared : 1; /* if it was cleared at least once */
@@ -706,6 +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
@@ -1006,9 +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 */
@@ -1027,6 +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];
@@ -1334,9 +1341,10 @@ void si_init_compute_blit_functions(struct si_context *sctx);
 #define SI_CPDMA_SKIP_SYNC_BEFORE    (1 << 2) /* don't wait for DMA before the copy (RAW hazards) */
 #define SI_CPDMA_SKIP_GFX_SYNC       (1 << 3) /* don't flush caches and don't wait for PS/CS */
 #define SI_CPDMA_SKIP_BO_LIST_UPDATE (1 << 4) /* don't update the BO list */
+#define SI_CPDMA_SKIP_TMZ            (1 << 5) /* don't update tmz state */
 #define SI_CPDMA_SKIP_ALL                                                                          \
    (SI_CPDMA_SKIP_CHECK_CS_SPACE | SI_CPDMA_SKIP_SYNC_AFTER | SI_CPDMA_SKIP_SYNC_BEFORE |          \
-    SI_CPDMA_SKIP_GFX_SYNC | SI_CPDMA_SKIP_BO_LIST_UPDATE)
+    SI_CPDMA_SKIP_GFX_SYNC | SI_CPDMA_SKIP_BO_LIST_UPDATE | SI_CPDMA_SKIP_TMZ)
 
 void si_cp_dma_wait_for_idle(struct si_context *sctx);
 void si_cp_dma_clear_buffer(struct si_context *sctx, struct radeon_cmdbuf *cs,
@@ -1403,6 +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);
@@ -1494,7 +1503,8 @@ void si_init_viewport_functions(struct si_context *ctx);
 bool si_prepare_for_dma_blit(struct si_context *sctx, struct si_texture *dst, unsigned dst_level,
                              unsigned dstx, unsigned dsty, unsigned dstz, struct si_texture *src,
                              unsigned src_level, const struct pipe_box *src_box);
-void si_eliminate_fast_color_clear(struct si_context *sctx, struct si_texture *tex);
+void si_eliminate_fast_color_clear(struct si_context *sctx, struct si_texture *tex,
+                                   bool *ctx_flushed);
 void si_texture_discard_cmask(struct si_screen *sscreen, struct si_texture *tex);
 bool si_init_flushed_depth_texture(struct pipe_context *ctx, struct pipe_resource *texture);
 void si_print_texture_info(struct si_screen *sscreen, struct si_texture *tex,
@@ -1870,12 +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) ||
@@ -1888,7 +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...)                                                                    \