radeonsi: overhaul the vertex fetch fixup mechanism
[mesa.git] / src / gallium / drivers / radeonsi / si_state.h
index fffc63680d048b9a5cc6bf85e308eeb723c4de11..e4b1cf791326f3e99cf9e5335760cac29a18775f 100644 (file)
@@ -71,6 +71,7 @@ struct si_state_rasterizer {
        float                   max_point_size;
        unsigned                sprite_coord_enable:8;
        unsigned                clip_plane_enable:8;
+       unsigned                half_pixel_center:1;
        unsigned                flatshade:1;
        unsigned                two_side:1;
        unsigned                multisample_enable:1;
@@ -131,13 +132,32 @@ struct si_stencil_ref {
 
 struct si_vertex_elements
 {
-       uint32_t                        instance_divisors[SI_MAX_ATTRIBS];
+       struct si_resource              *instance_divisor_factor_buffer;
        uint32_t                        rsrc_word3[SI_MAX_ATTRIBS];
        uint16_t                        src_offset[SI_MAX_ATTRIBS];
        uint8_t                         fix_fetch[SI_MAX_ATTRIBS];
        uint8_t                         format_size[SI_MAX_ATTRIBS];
        uint8_t                         vertex_buffer_index[SI_MAX_ATTRIBS];
 
+       /* Bitmask of elements that always need a fixup to be applied. */
+       uint16_t                        fix_fetch_always;
+
+       /* Bitmask of elements whose fetch should always be opencoded. */
+       uint16_t                        fix_fetch_opencode;
+
+       /* Bitmask of elements which need to be opencoded if the vertex buffer
+        * is unaligned. */
+       uint16_t                        fix_fetch_unaligned;
+
+       /* For elements in fix_fetch_unaligned: whether the effective
+        * element load size as seen by the hardware is a dword (as opposed
+        * to a short).
+        */
+       uint16_t                        hw_load_is_dword;
+
+       /* Bitmask of vertex buffers requiring alignment check */
+       uint16_t                        vb_alignment_check_mask;
+
        uint8_t                         count;
        bool                            uses_instance_divisors;
 
@@ -170,17 +190,13 @@ union si_state {
 #define SI_STATE_BIT(name) (1 << SI_STATE_IDX(name))
 #define SI_NUM_STATES (sizeof(union si_state) / sizeof(struct si_pm4_state *))
 
-static inline unsigned si_states_that_roll_context(void)
+static inline unsigned si_states_that_always_roll_context(void)
 {
        return (SI_STATE_BIT(blend) |
                SI_STATE_BIT(rasterizer) |
                SI_STATE_BIT(dsa) |
                SI_STATE_BIT(poly_offset) |
-               SI_STATE_BIT(es) |
-               SI_STATE_BIT(gs) |
-               SI_STATE_BIT(vgt_shader_config) |
-               SI_STATE_BIT(vs) |
-               SI_STATE_BIT(ps));
+               SI_STATE_BIT(vgt_shader_config));
 }
 
 union si_state_atoms {
@@ -215,26 +231,20 @@ union si_state_atoms {
                                 sizeof(struct si_atom)))
 #define SI_NUM_ATOMS (sizeof(union si_state_atoms)/sizeof(struct si_atom*))
 
-static inline unsigned si_atoms_that_roll_context(void)
+static inline unsigned si_atoms_that_always_roll_context(void)
 {
        return (SI_ATOM_BIT(streamout_begin) |
                SI_ATOM_BIT(streamout_enable) |
                SI_ATOM_BIT(framebuffer) |
                SI_ATOM_BIT(msaa_sample_locs) |
-               SI_ATOM_BIT(db_render_state) |
-               SI_ATOM_BIT(dpbb_state) |
-               SI_ATOM_BIT(msaa_config) |
                SI_ATOM_BIT(sample_mask) |
-               SI_ATOM_BIT(cb_render_state) |
                SI_ATOM_BIT(blend_color) |
-               SI_ATOM_BIT(clip_regs) |
                SI_ATOM_BIT(clip_state) |
-               SI_ATOM_BIT(guardband) |
                SI_ATOM_BIT(scissors) |
                SI_ATOM_BIT(viewports) |
                SI_ATOM_BIT(stencil_ref) |
-               SI_ATOM_BIT(spi_map) |
-               SI_ATOM_BIT(scratch_state));
+               SI_ATOM_BIT(scratch_state) |
+               SI_ATOM_BIT(window_rectangles));
 }
 
 struct si_shader_data {
@@ -262,6 +272,7 @@ enum si_tracked_reg {
        SI_TRACKED_DB_EQAA,
        SI_TRACKED_PA_SC_MODE_CNTL_1,
 
+       SI_TRACKED_PA_SU_PRIM_FILTER_CNTL,
        SI_TRACKED_PA_SU_SMALL_PRIM_FILTER_CNTL,
 
        SI_TRACKED_PA_CL_VS_OUT_CNTL,
@@ -275,6 +286,9 @@ enum si_tracked_reg {
        SI_TRACKED_PA_CL_GB_HORZ_CLIP_ADJ,
        SI_TRACKED_PA_CL_GB_HORZ_DISC_ADJ,
 
+       SI_TRACKED_PA_SU_HARDWARE_SCREEN_OFFSET,
+       SI_TRACKED_PA_SU_VTX_CNTL,
+
        SI_TRACKED_PA_SC_CLIPRECT_RULE,
 
        SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
@@ -390,7 +404,7 @@ struct si_descriptors {
        uint32_t *gpu_list;
 
        /* The buffer where the descriptors have been uploaded. */
-       struct r600_resource *buffer;
+       struct si_resource *buffer;
        uint64_t gpu_address;
 
        /* The maximum number of descriptors. */
@@ -415,13 +429,12 @@ struct si_descriptors {
 struct si_buffer_resources {
        struct pipe_resource            **buffers; /* this has num_buffers elements */
 
-       enum radeon_bo_usage            shader_usage:4; /* READ, WRITE, or READWRITE */
-       enum radeon_bo_usage            shader_usage_constbuf:4;
        enum radeon_bo_priority         priority:6;
        enum radeon_bo_priority         priority_constbuf:6;
 
        /* The i-th bit is set if that element is enabled (non-NULL resource). */
        unsigned                        enabled_mask;
+       unsigned                        writable_mask;
 };
 
 #define si_pm4_state_changed(sctx, member) \
@@ -469,9 +482,10 @@ bool si_upload_vertex_buffer_descriptors(struct si_context *sctx);
 bool si_upload_graphics_shader_descriptors(struct si_context *sctx);
 bool si_upload_compute_shader_descriptors(struct si_context *sctx);
 void si_release_all_descriptors(struct si_context *sctx);
+void si_gfx_resources_add_all_to_bo_list(struct si_context *sctx);
+void si_compute_resources_add_all_to_bo_list(struct si_context *sctx);
 void si_all_descriptors_begin_new_cs(struct si_context *sctx);
-void si_all_resident_buffers_begin_new_cs(struct si_context *sctx);
-void si_upload_const_buffer(struct si_context *sctx, struct r600_resource **rbuffer,
+void si_upload_const_buffer(struct si_context *sctx, struct si_resource **buf,
                            const uint8_t *ptr, unsigned size, uint32_t *const_offset);
 void si_update_all_texture_descriptors(struct si_context *sctx);
 void si_shader_change_notify(struct si_context *sctx);
@@ -480,6 +494,8 @@ void si_emit_graphics_shader_pointers(struct si_context *sctx);
 void si_emit_compute_shader_pointers(struct si_context *sctx);
 void si_set_rw_buffer(struct si_context *sctx,
                      uint slot, const struct pipe_constant_buffer *input);
+void si_set_rw_shader_buffer(struct si_context *sctx, uint slot,
+                            const struct pipe_shader_buffer *sbuffer);
 void si_set_active_descriptors(struct si_context *sctx, unsigned desc_idx,
                               uint64_t new_active_mask);
 void si_set_active_descriptors_for_shader(struct si_context *sctx,
@@ -493,10 +509,11 @@ void si_bindless_descriptor_slab_free(void *priv, struct pb_slab *pslab);
 void si_rebind_buffer(struct si_context *sctx, struct pipe_resource *buf,
                      uint64_t old_va);
 /* si_state.c */
+void si_init_state_compute_functions(struct si_context *sctx);
 void si_init_state_functions(struct si_context *sctx);
 void si_init_screen_state_functions(struct si_screen *sscreen);
 void
-si_make_buffer_descriptor(struct si_screen *screen, struct r600_resource *buf,
+si_make_buffer_descriptor(struct si_screen *screen, struct si_resource *buf,
                          enum pipe_format format,
                          unsigned offset, unsigned size,
                          uint32_t *state);
@@ -547,17 +564,9 @@ void si_get_active_slot_masks(const struct tgsi_shader_info *info,
                              uint64_t *samplers_and_images);
 
 /* si_state_draw.c */
-void si_init_ia_multi_vgt_param_table(struct si_context *sctx);
 void si_emit_cache_flush(struct si_context *sctx);
-void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *dinfo);
-void si_draw_rectangle(struct blitter_context *blitter,
-                      void *vertex_elements_cso,
-                      blitter_get_vs_func get_vs,
-                      int x1, int y1, int x2, int y2,
-                      float depth, unsigned num_instances,
-                      enum blitter_attrib_type type,
-                      const union blitter_attrib *attrib);
 void si_trace_emit(struct si_context *sctx);
+void si_init_draw_functions(struct si_context *sctx);
 
 /* si_state_msaa.c */
 void si_init_msaa_functions(struct si_context *sctx);