radeonsi: update hack for HTILE corruption in ARK: Survival Evolved
[mesa.git] / src / gallium / drivers / radeonsi / si_pipe.h
index 93f1d653e13b02b2f9a466411fd074f49dc14613..8d7703b0e4b0f150d4dd1badd202b20c84724374 100644 (file)
@@ -29,6 +29,7 @@
 #include "si_shader.h"
 
 #include "util/u_dynarray.h"
+#include "util/u_idalloc.h"
 
 #ifdef PIPE_ARCH_BIG_ENDIAN
 #define SI_BIG_ENDIAN 1
 /* Write dirty L2 lines back to memory (shader and CP DMA stores), but don't
  * invalidate L2. SI-CIK can't do it, so they will do complete invalidation. */
 #define SI_CONTEXT_WRITEBACK_GLOBAL_L2 (R600_CONTEXT_PRIVATE_FLAG << 4)
-/* gaps */
+/* Writeback & invalidate the L2 metadata cache. It can only be coupled with
+ * a CB or DB flush. */
+#define SI_CONTEXT_INV_L2_METADATA     (R600_CONTEXT_PRIVATE_FLAG << 5)
 /* Framebuffer caches. */
-#define SI_CONTEXT_FLUSH_AND_INV_DB    (R600_CONTEXT_PRIVATE_FLAG << 7)
+#define SI_CONTEXT_FLUSH_AND_INV_DB    (R600_CONTEXT_PRIVATE_FLAG << 6)
+#define SI_CONTEXT_FLUSH_AND_INV_DB_META (R600_CONTEXT_PRIVATE_FLAG << 7)
 #define SI_CONTEXT_FLUSH_AND_INV_CB    (R600_CONTEXT_PRIVATE_FLAG << 8)
 /* Engine synchronization. */
 #define SI_CONTEXT_VS_PARTIAL_FLUSH    (R600_CONTEXT_PRIVATE_FLAG << 9)
 #define SI_CONTEXT_VGT_FLUSH           (R600_CONTEXT_PRIVATE_FLAG << 12)
 #define SI_CONTEXT_VGT_STREAMOUT_SYNC  (R600_CONTEXT_PRIVATE_FLAG << 13)
 
+#define SI_PREFETCH_VBO_DESCRIPTORS    (1 << 0)
+#define SI_PREFETCH_LS                 (1 << 1)
+#define SI_PREFETCH_HS                 (1 << 2)
+#define SI_PREFETCH_ES                 (1 << 3)
+#define SI_PREFETCH_GS                 (1 << 4)
+#define SI_PREFETCH_VS                 (1 << 5)
+#define SI_PREFETCH_PS                 (1 << 6)
+
 #define SI_MAX_BORDER_COLORS   4096
+#define SI_MAX_VIEWPORTS        16
 #define SIX_BITS               0x3F
 
 struct si_compute;
@@ -79,10 +92,16 @@ struct si_screen {
        struct r600_common_screen       b;
        unsigned                        gs_table_depth;
        unsigned                        tess_offchip_block_dw_size;
+       bool                            has_clear_state;
        bool                            has_distributed_tess;
        bool                            has_draw_indirect_multi;
-       bool                            has_ds_bpermute;
+       bool                            has_out_of_order_rast;
+       bool                            assume_no_z_fights;
+       bool                            commutative_blend_add;
+       bool                            clear_db_cache_before_clear;
        bool                            has_msaa_sample_loc_bug;
+       bool                            dpbb_allowed;
+       bool                            dfsm_allowed;
        bool                            llvm_has_working_vgpr_indexing;
 
        /* Whether shaders are monolithic (1-part) or separate (3-part). */
@@ -140,6 +159,7 @@ struct si_sampler_view {
        ubyte                           base_level;
        ubyte                           block_width;
        bool is_stencil_sampler;
+       bool is_integer;
        bool dcc_incompatible;
 };
 
@@ -150,6 +170,8 @@ struct si_sampler_state {
        unsigned                        magic;
 #endif
        uint32_t                        val[4];
+       uint32_t                        integer_val[4];
+       uint32_t                        upgraded_depth_val[4];
 };
 
 struct si_cs_shader_state {
@@ -160,13 +182,17 @@ struct si_cs_shader_state {
        bool                            uses_scratch;
 };
 
-struct si_textures_info {
-       struct si_sampler_views         views;
+struct si_samplers {
+       struct pipe_sampler_view        *views[SI_NUM_SAMPLERS];
+       struct si_sampler_state         *sampler_states[SI_NUM_SAMPLERS];
+
+       /* The i-th bit is set if that element is enabled (non-NULL resource). */
+       unsigned                        enabled_mask;
        uint32_t                        needs_depth_decompress_mask;
        uint32_t                        needs_color_decompress_mask;
 };
 
-struct si_images_info {
+struct si_images {
        struct pipe_image_view          views[SI_NUM_IMAGES];
        uint32_t                        needs_color_decompress_mask;
        unsigned                        enabled_mask;
@@ -188,7 +214,29 @@ struct si_framebuffer {
        ubyte                           dirty_cbufs;
        bool                            dirty_zsbuf;
        bool                            any_dst_linear;
-       bool                            do_update_surf_dirtiness;
+       bool                            CB_has_shader_readable_metadata;
+       bool                            DB_has_shader_readable_metadata;
+};
+
+struct si_signed_scissor {
+       int minx;
+       int miny;
+       int maxx;
+       int maxy;
+};
+
+struct si_scissors {
+       struct r600_atom                atom;
+       unsigned                        dirty_mask;
+       struct pipe_scissor_state       states[SI_MAX_VIEWPORTS];
+};
+
+struct si_viewports {
+       struct r600_atom                atom;
+       unsigned                        dirty_mask;
+       unsigned                        depth_range_dirty_mask;
+       struct pipe_viewport_state      states[SI_MAX_VIEWPORTS];
+       struct si_signed_scissor        as_scissor[SI_MAX_VIEWPORTS];
 };
 
 struct si_clip_state {
@@ -207,6 +255,43 @@ struct si_sample_mask {
        uint16_t                sample_mask;
 };
 
+struct si_streamout_target {
+       struct pipe_stream_output_target b;
+
+       /* The buffer where BUFFER_FILLED_SIZE is stored. */
+       struct r600_resource    *buf_filled_size;
+       unsigned                buf_filled_size_offset;
+       bool                    buf_filled_size_valid;
+
+       unsigned                stride_in_dw;
+};
+
+struct si_streamout {
+       struct r600_atom                begin_atom;
+       bool                            begin_emitted;
+
+       unsigned                        enabled_mask;
+       unsigned                        num_targets;
+       struct si_streamout_target      *targets[PIPE_MAX_SO_BUFFERS];
+
+       unsigned                        append_bitmask;
+       bool                            suspended;
+
+       /* External state which comes from the vertex shader,
+        * it must be set explicitly when binding a shader. */
+       uint16_t                        *stride_in_dw;
+       unsigned                        enabled_stream_buffers_mask; /* stream0 buffers0-3 in 4 LSB */
+
+       /* The state of VGT_STRMOUT_BUFFER_(CONFIG|EN). */
+       unsigned                        hw_enabled_mask;
+
+       /* The state of VGT_STRMOUT_(CONFIG|EN). */
+       struct r600_atom                enable_atom;
+       bool                            streamout_enabled;
+       bool                            prims_gen_query_enabled;
+       int                             num_prims_gen_queries;
+};
+
 /* A shader state consists of the shader selector, which is a constant state
  * object shared by multiple contexts and shouldn't be modified, and
  * the current shader variant selected for this context.
@@ -238,28 +323,32 @@ union si_vgt_param_key {
        uint32_t index;
 };
 
-struct si_bindless_descriptor
-{
-       struct pb_slab_entry            entry;
-       struct r600_resource            *buffer;
-       unsigned                        offset;
-       uint32_t                        desc_list[16];
-       bool                            dirty;
-};
-
 struct si_texture_handle
 {
-       struct si_bindless_descriptor   *desc;
+       unsigned                        desc_slot;
+       bool                            desc_dirty;
        struct pipe_sampler_view        *view;
        struct si_sampler_state         sstate;
 };
 
 struct si_image_handle
 {
-       struct si_bindless_descriptor   *desc;
+       unsigned                        desc_slot;
+       bool                            desc_dirty;
        struct pipe_image_view          view;
 };
 
+struct si_saved_cs {
+       struct pipe_reference   reference;
+       struct si_context       *ctx;
+       struct radeon_saved_cs  gfx;
+       struct r600_resource    *trace_buf;
+       unsigned                trace_id;
+
+       unsigned                gfx_last_dw;
+       bool                    flushed;
+};
+
 struct si_context {
        struct r600_common_context      b;
        struct blitter_context          *blitter;
@@ -268,19 +357,17 @@ struct si_context {
        void                            *custom_blend_fmask_decompress;
        void                            *custom_blend_eliminate_fastclear;
        void                            *custom_blend_dcc_decompress;
+       void                            *vs_blit_pos;
+       void                            *vs_blit_pos_layered;
+       void                            *vs_blit_color;
+       void                            *vs_blit_color_layered;
+       void                            *vs_blit_texcoord;
        struct si_screen                *screen;
        LLVMTargetMachineRef            tm; /* only non-threaded compilation */
        struct si_shader_ctx_state      fixed_func_tcs_shader;
        struct r600_resource            *wait_mem_scratch;
        unsigned                        wait_mem_number;
-
-       struct radeon_winsys_cs         *ce_ib;
-       struct radeon_winsys_cs         *ce_preamble_ib;
-       struct r600_resource            *ce_ram_saved_buffer;
-       struct u_suballocator           *ce_suballocator;
-       unsigned                        ce_ram_saved_offset;
-       uint16_t                        total_ce_ram_allocated;
-       bool                            ce_need_synchronization:1;
+       uint16_t                        prefetch_L2_mask;
 
        bool                            gfx_flush_in_progress:1;
        bool                            compute_is_busy:1;
@@ -294,10 +381,10 @@ struct si_context {
        union si_state                  emitted;
 
        /* Atom declarations. */
-       struct r600_atom                prefetch_L2;
        struct si_framebuffer           framebuffer;
        struct si_sample_locs           msaa_sample_locs;
        struct r600_atom                db_render_state;
+       struct r600_atom                dpbb_state;
        struct r600_atom                msaa_config;
        struct si_sample_mask           sample_mask;
        struct r600_atom                cb_render_state;
@@ -305,9 +392,12 @@ struct si_context {
        struct si_blend_color           blend_color;
        struct r600_atom                clip_regs;
        struct si_clip_state            clip_state;
-       struct si_shader_data           shader_userdata;
+       struct si_shader_data           shader_pointers;
        struct si_stencil_ref           stencil_ref;
        struct r600_atom                spi_map;
+       struct si_scissors              scissors;
+       struct si_streamout             streamout;
+       struct si_viewports             viewports;
 
        /* Precomputed states. */
        struct si_pm4_state             *init_config;
@@ -337,8 +427,8 @@ struct si_context {
        unsigned                        shader_needs_decompress_mask;
        struct si_buffer_resources      rw_buffers;
        struct si_buffer_resources      const_and_shader_buffers[SI_NUM_SHADERS];
-       struct si_textures_info         samplers[SI_NUM_SHADERS];
-       struct si_images_info           images[SI_NUM_SHADERS];
+       struct si_samplers              samplers[SI_NUM_SHADERS];
+       struct si_images                images[SI_NUM_SHADERS];
 
        /* other shader resources */
        struct pipe_constant_buffer     null_const_buf; /* used for set_constant_buffer(NULL) on CIK */
@@ -350,6 +440,8 @@ struct si_context {
        struct r600_resource            *border_color_buffer;
        union pipe_color_union          *border_color_map; /* in VRAM (slow access), little endian */
        unsigned                        border_color_count;
+       unsigned                        num_vs_blit_sgprs;
+       uint32_t                        vs_blit_sh_data[SI_VS_BLIT_SGPRS_POS_TEXCOORD];
 
        /* Vertex and index buffers. */
        bool                            vertex_buffers_dirty;
@@ -376,6 +468,7 @@ struct si_context {
 
        /* Emitted draw state. */
        bool                    gs_tri_strip_adj_fix:1;
+       bool                    ls_vgpr_fix:1;
        int                     last_index_size;
        int                     last_base_vertex;
        int                     last_start_instance;
@@ -411,10 +504,7 @@ struct si_context {
 
        /* Debug state. */
        bool                    is_debug;
-       struct radeon_saved_cs  last_gfx;
-       struct r600_resource    *last_trace_buf;
-       struct r600_resource    *trace_buf;
-       unsigned                trace_id;
+       struct si_saved_cs      *current_saved_cs;
        uint64_t                dmesg_timestamp;
        unsigned                apitrace_call_number;
 
@@ -422,16 +512,20 @@ struct si_context {
        bool need_check_render_feedback;
        bool                    decompression_enabled;
 
+       bool                    vs_writes_viewport_index;
+       bool                    vs_disables_clipping_viewport;
+
        /* Precomputed IA_MULTI_VGT_PARAM */
        union si_vgt_param_key  ia_multi_vgt_param_key;
        unsigned                ia_multi_vgt_param[SI_NUM_VGT_PARAM_STATES];
 
-       /* Slab allocator for bindless descriptors. */
-       struct pb_slabs         bindless_descriptor_slabs;
-
        /* Bindless descriptors. */
-       struct util_dynarray    bindless_descriptors;
+       struct si_descriptors   bindless_descriptors;
+       struct util_idalloc     bindless_used_slots;
+       unsigned                num_bindless_descriptors;
        bool                    bindless_descriptors_dirty;
+       bool                    graphics_bindless_pointer_dirty;
+       bool                    compute_bindless_pointer_dirty;
 
        /* Allocated bindless handles */
        struct hash_table       *tex_handles;
@@ -449,6 +543,15 @@ struct si_context {
        /* Bindless state */
        bool                    uses_bindless_samplers;
        bool                    uses_bindless_images;
+
+       /* MSAA sample locations.
+        * The first index is the sample index.
+        * The second index is the coordinate: X, Y. */
+       float                   sample_locations_1x[1][2];
+       float                   sample_locations_2x[2][2];
+       float                   sample_locations_4x[4][2];
+       float                   sample_locations_8x[8][2];
+       float                   sample_locations_16x[16][2];
 };
 
 /* cik_sdma.c */
@@ -456,8 +559,7 @@ void cik_init_sdma_functions(struct si_context *sctx);
 
 /* si_blit.c */
 void si_init_blit_functions(struct si_context *sctx);
-void si_decompress_graphics_textures(struct si_context *sctx);
-void si_decompress_compute_textures(struct si_context *sctx);
+void si_decompress_textures(struct si_context *sctx, unsigned shader_mask);
 void si_resource_copy_region(struct pipe_context *ctx,
                             struct pipe_resource *dst,
                             unsigned dst_level,
@@ -484,9 +586,14 @@ void si_copy_buffer(struct si_context *sctx,
                    unsigned user_flags);
 void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
                              uint64_t offset, unsigned size);
+void cik_emit_prefetch_L2(struct si_context *sctx);
 void si_init_cp_dma_functions(struct si_context *sctx);
 
 /* si_debug.c */
+void si_auto_log_cs(void *data, struct u_log_context *log);
+void si_log_hw_flush(struct si_context *sctx);
+void si_log_draw_state(struct si_context *sctx, struct u_log_context *log);
+void si_log_compute_state(struct si_context *sctx, struct u_log_context *log);
 void si_init_debug_functions(struct si_context *sctx);
 void si_check_vm_faults(struct r600_common_context *ctx,
                        struct radeon_saved_cs *saved, enum ring_type ring);
@@ -496,6 +603,7 @@ bool si_replace_shader(unsigned num, struct ac_shader_binary *binary);
 void si_init_dma_functions(struct si_context *sctx);
 
 /* si_hw_context.c */
+void si_destroy_saved_cs(struct si_saved_cs *scs);
 void si_context_gfx_flush(void *context, unsigned flags,
                          struct pipe_fence_handle **fence);
 void si_begin_new_cs(struct si_context *ctx);
@@ -514,6 +622,11 @@ struct pipe_video_codec *si_uvd_create_decoder(struct pipe_context *context,
 struct pipe_video_buffer *si_video_buffer_create(struct pipe_context *pipe,
                                                 const struct pipe_video_buffer *tmpl);
 
+/* si_viewport.c */
+void si_update_vs_viewport_state(struct si_context *ctx);
+void si_init_viewport_functions(struct si_context *ctx);
+
+
 /*
  * common helpers
  */
@@ -578,6 +691,12 @@ static inline struct si_shader* si_get_vs_state(struct si_context *sctx)
        return vs->current ? vs->current : NULL;
 }
 
+static inline bool si_get_strmout_en(struct si_context *sctx)
+{
+       return sctx->streamout.streamout_enabled ||
+              sctx->streamout.prims_gen_query_enabled;
+}
+
 static inline unsigned
 si_optimal_tcc_alignment(struct si_context *sctx, unsigned upload_size)
 {
@@ -593,4 +712,57 @@ si_optimal_tcc_alignment(struct si_context *sctx, unsigned upload_size)
        return MIN2(alignment, tcc_cache_line_size);
 }
 
+static inline void
+si_saved_cs_reference(struct si_saved_cs **dst, struct si_saved_cs *src)
+{
+       if (pipe_reference(&(*dst)->reference, &src->reference))
+               si_destroy_saved_cs(*dst);
+
+       *dst = src;
+}
+
+static inline void
+si_make_CB_shader_coherent(struct si_context *sctx, unsigned num_samples,
+                          bool shaders_read_metadata)
+{
+       sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
+                        SI_CONTEXT_INV_VMEM_L1;
+
+       if (sctx->b.chip_class >= GFX9) {
+               /* Single-sample color is coherent with shaders on GFX9, but
+                * L2 metadata must be flushed if shaders read metadata.
+                * (DCC, CMASK).
+                */
+               if (num_samples >= 2)
+                       sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
+               else if (shaders_read_metadata)
+                       sctx->b.flags |= SI_CONTEXT_INV_L2_METADATA;
+       } else {
+               /* SI-CI-VI */
+               sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
+       }
+}
+
+static inline void
+si_make_DB_shader_coherent(struct si_context *sctx, unsigned num_samples,
+                          bool include_stencil, bool shaders_read_metadata)
+{
+       sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_DB |
+                        SI_CONTEXT_INV_VMEM_L1;
+
+       if (sctx->b.chip_class >= GFX9) {
+               /* Single-sample depth (not stencil) is coherent with shaders
+                * on GFX9, but L2 metadata must be flushed if shaders read
+                * metadata.
+                */
+               if (num_samples >= 2 || include_stencil)
+                       sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
+               else if (shaders_read_metadata)
+                       sctx->b.flags |= SI_CONTEXT_INV_L2_METADATA;
+       } else {
+               /* SI-CI-VI */
+               sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
+       }
+}
+
 #endif