#define SI_CONTEXT_FLUSH_FOR_RENDER_COND (1 << 2)
/* Instruction cache. */
#define SI_CONTEXT_INV_ICACHE (1 << 3)
-/* Scalar L1 cache. */
+/* Scalar cache. (GFX6-9: scalar L1; GFX10: scalar L0)
+ * GFX10: This also invalidates the L1 shader array cache. */
#define SI_CONTEXT_INV_SCACHE (1 << 4)
-/* Vector L1 cache. */
+/* Vector cache. (GFX6-9: vector L1; GFX10: vector L0)
+ * GFX10: This also invalidates the L1 shader array cache. */
#define SI_CONTEXT_INV_VCACHE (1 << 5)
/* L2 cache + L2 metadata cache writeback & invalidate.
* GFX6-8: Used by shaders only. GFX9-10: Used by everything. */
/* Shader compiler options the shader cache should be aware of: */
DBG_FS_CORRECT_DERIVS_AFTER_KILL,
- DBG_UNSAFE_MATH,
DBG_SI_SCHED,
DBG_GISEL,
+ DBG_W32_GE,
+ DBG_W32_PS,
+ DBG_W32_CS,
+ DBG_W64_GE,
+ DBG_W64_PS,
+ DBG_W64_CS,
/* Shader compiler options (with no effect on the shader cache): */
DBG_CHECK_IR,
DBG_ZERO_VRAM,
/* 3D engine options: */
+ DBG_NO_GFX,
+ DBG_NO_NGG,
DBG_ALWAYS_PD,
DBG_PD,
DBG_NO_PD,
struct si_resource buffer;
struct radeon_surf surface;
- uint64_t size;
struct si_texture *flushed_depth_texture;
/* One texture allocation can contain these buffers:
* - displayable DCC buffer (if the DCC buffer is not displayable)
* - DCC retile mapping buffer (if the DCC buffer is not displayable)
*/
- uint64_t fmask_offset;
- uint64_t cmask_offset;
uint64_t cmask_base_address_reg;
struct si_resource *cmask_buffer;
- uint64_t dcc_offset; /* 0 = disabled */
- uint64_t display_dcc_offset;
- uint64_t dcc_retile_map_offset;
unsigned cb_color_info; /* fast clear enable bit */
unsigned color_clear_value[2];
unsigned last_msaa_resolve_target_micro_mode;
unsigned num_level0_transfers;
+ unsigned plane_index; /* other planes are different pipe_resources */
+ unsigned num_planes;
/* Depth buffer compression and fast clear. */
- uint64_t htile_offset;
float depth_clear_value;
uint16_t dirty_level_mask; /* each bit says if that mipmap is compressed */
uint16_t stencil_dirty_level_mask; /* each bit says if that mipmap is compressed */
enum pipe_format db_render_format:16;
uint8_t stencil_clear_value;
+ bool fmask_is_not_identity:1;
bool tc_compatible_htile: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 */
bool upgraded_depth:1; /* upgraded from unorm to Z32_FLOAT */
unsigned eqaa_force_coverage_samples;
unsigned eqaa_force_z_samples;
unsigned eqaa_force_color_samples;
- bool has_clear_state;
- bool has_distributed_tess;
bool has_draw_indirect_multi;
bool has_out_of_order_rast;
bool assume_no_z_fights;
bool commutative_blend_add;
- bool has_gfx9_scissor_bug;
- bool has_msaa_sample_loc_bug;
- bool has_ls_vgpr_init_bug;
- bool has_dcc_constant_encode;
bool dpbb_allowed;
bool dfsm_allowed;
bool llvm_has_working_vgpr_indexing;
+ bool use_ngg;
+ bool use_ngg_streamout;
struct {
#define OPT_BOOL(name, dflt, description) bool name:1;
/* Whether shaders are monolithic (1-part) or separate (3-part). */
bool use_monolithic_shaders;
bool record_llvm_ir;
- bool has_rbplus; /* if RB+ registers exist */
- bool rbplus_allowed; /* if RB+ is allowed */
bool dcc_msaa_allowed;
- bool cpdma_prefetch_writes_memory;
struct slab_parent_pool pool_transfers;
/* Auxiliary context. Mainly used to initialize resources.
* It must be locked prior to using and flushed before unlocking. */
struct pipe_context *aux_context;
- mtx_t aux_context_lock;
+ simple_mtx_t aux_context_lock;
/* This must be in the screen, because UE4 uses one context for
* compilation and another one for rendering.
unsigned num_shader_cache_hits;
/* GPU load thread. */
- mtx_t gpu_load_mutex;
+ simple_mtx_t gpu_load_mutex;
thrd_t gpu_load_thread;
union si_mmio_counters mmio_counters;
volatile unsigned gpu_load_stop_thread; /* bool */
unsigned L2_to_cp;
} barrier_flags;
- mtx_t shader_parts_mutex;
+ simple_mtx_t shader_parts_mutex;
struct si_shader_part *vs_prologs;
struct si_shader_part *tcs_epilogs;
struct si_shader_part *gs_prologs;
* - GS and CS aren't cached, but it's certainly possible to cache
* those as well.
*/
- mtx_t shader_cache_mutex;
+ simple_mtx_t shader_cache_mutex;
struct hash_table *shader_cache;
/* Shader compiler queue for multithreaded compilation. */
/* Use at most 2 low priority threads on quadcore and better.
* We want to minimize the impact on multithreaded Mesa. */
struct ac_llvm_compiler compiler_lowp[10];
+
+ unsigned compute_wave_size;
+ unsigned ps_wave_size;
+ unsigned ge_wave_size;
};
struct si_blend_color {
ubyte color_is_int10;
ubyte dirty_cbufs;
ubyte dcc_overwrite_combiner_watermark;
+ ubyte min_bytes_per_pixel;
bool dirty_zsbuf;
bool any_dst_linear;
bool CB_has_shader_readable_metadata;
uint32_t index;
};
-#define SI_NUM_VGT_STAGES_KEY_BITS 2
+#define SI_NUM_VGT_STAGES_KEY_BITS 4
#define SI_NUM_VGT_STAGES_STATES (1 << SI_NUM_VGT_STAGES_KEY_BITS)
/* The VGT_SHADER_STAGES key used to index the table of precomputed values.
#ifdef PIPE_ARCH_LITTLE_ENDIAN
unsigned tess:1;
unsigned gs:1;
+ unsigned ngg:1; /* gfx10+ */
+ unsigned streamout:1; /* only used with NGG */
unsigned _pad:32 - SI_NUM_VGT_STAGES_KEY_BITS;
#else /* PIPE_ARCH_BIG_ENDIAN */
unsigned _pad:32 - SI_NUM_VGT_STAGES_KEY_BITS;
+ unsigned streamout:1;
+ unsigned ngg:1;
unsigned gs:1;
unsigned tess:1;
#endif
struct pipe_device_reset_callback device_reset_callback;
struct u_log_context *log;
void *query_result_shader;
+ void *sh_query_result_shader;
+
+ void (*emit_cache_flush)(struct si_context *ctx);
+
struct blitter_context *blitter;
+ void *noop_blend;
+ void *noop_dsa;
+ void *discard_rasterizer_state;
void *custom_dsa_flush;
void *custom_blend_resolve;
void *custom_blend_fmask_decompress;
void *cs_clear_render_target;
void *cs_clear_render_target_1d_array;
void *cs_dcc_retile;
+ void *cs_fmask_expand[3][2]; /* [log2(samples)-1][is_array] */
struct si_screen *screen;
struct pipe_debug_callback debug;
struct ac_llvm_compiler compiler; /* only non-threaded compilation */
bool gs_tri_strip_adj_fix:1;
bool ls_vgpr_fix:1;
bool prim_discard_cs_instancing:1;
+ bool ngg:1;
int last_index_size;
int last_base_vertex;
int last_start_instance;
int last_prim;
int last_multi_vgt_param;
int last_rast_prim;
+ int last_flatshade_first;
+ int last_binning_enabled;
unsigned last_sc_line_stipple;
unsigned current_vs_state;
unsigned last_vs_state;
struct si_resource *scratch_buffer;
unsigned scratch_waves;
unsigned spi_tmpring_size;
+ unsigned max_seen_scratch_bytes_per_wave;
+ unsigned max_seen_compute_scratch_bytes_per_wave;
struct si_resource *compute_scratch_buffer;
unsigned num_sdma_uploads;
unsigned max_sdma_uploads;
+ /* Shader-based queries. */
+ struct list_head shader_query_buffers;
+ unsigned num_active_shader_queries;
+
/* Statistics gathering for the DCC enablement heuristic. It can't be
* in si_texture because si_texture can be shared by multiple
* contexts. This is for back buffers only. We shouldn't get too many
/* si_clear.c */
enum pipe_format si_simplify_cb_format(enum pipe_format format);
-bool vi_alpha_is_on_msb(enum pipe_format format);
-void vi_dcc_clear_level(struct si_context *sctx,
+bool vi_alpha_is_on_msb(struct si_screen *sscreen, enum pipe_format format);
+bool vi_dcc_clear_level(struct si_context *sctx,
struct si_texture *tex,
unsigned level, unsigned clear_value);
void si_init_clear_functions(struct si_context *sctx);
unsigned width, unsigned height,
bool render_condition_enabled);
void si_retile_dcc(struct si_context *sctx, struct si_texture *tex);
+void si_compute_expand_fmask(struct pipe_context *ctx, struct pipe_resource *tex);
void si_init_compute_blit_functions(struct si_context *sctx);
/* si_cp_dma.c */
/* 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_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);
/* si_compute.c */
void si_emit_initial_compute_regs(struct si_context *sctx, struct radeon_cmdbuf *cs);
-unsigned si_get_compute_resource_limits(struct si_screen *sscreen,
- unsigned waves_per_threadgroup,
- unsigned max_waves_per_sh,
- unsigned threadgroups_per_cu);
void si_init_compute_functions(struct si_context *sctx);
/* si_compute_prim_discard.c */
void si_init_perfcounters(struct si_screen *screen);
void si_destroy_perfcounters(struct si_screen *screen);
-/* si_pipe.c */
-bool si_check_device_reset(struct si_context *sctx);
-
/* si_query.c */
void si_init_screen_query_functions(struct si_screen *sscreen);
void si_init_query_functions(struct si_context *sctx);
void *si_clear_render_target_shader(struct pipe_context *ctx);
void *si_clear_render_target_shader_1d_array(struct pipe_context *ctx);
void *si_create_dcc_retile_cs(struct pipe_context *ctx);
+void *si_create_fmask_expand_cs(struct pipe_context *ctx, unsigned num_samples,
+ bool is_array);
void *si_create_query_result_cs(struct si_context *sctx);
+void *gfx10_create_sh_query_result_cs(struct si_context *sctx);
+
+/* gfx10_query.c */
+void gfx10_init_query(struct si_context *sctx);
+void gfx10_destroy_query(struct si_context *sctx);
/* si_test_dma.c */
void si_test_dma(struct si_screen *sscreen);
struct si_texture *tex, struct u_log_context *log);
struct pipe_resource *si_texture_create(struct pipe_screen *screen,
const struct pipe_resource *templ);
-bool vi_dcc_formats_compatible(enum pipe_format format1,
+bool vi_dcc_formats_compatible(struct si_screen *sscreen,
+ enum pipe_format format1,
enum pipe_format format2);
bool vi_dcc_formats_are_incompatible(struct pipe_resource *tex,
unsigned level,
static inline bool
vi_dcc_enabled(struct si_texture *tex, unsigned level)
{
- return tex->dcc_offset && level < tex->surface.num_dcc_levels;
+ return tex->surface.dcc_offset && level < tex->surface.num_dcc_levels;
}
static inline unsigned
static inline struct si_shader* si_get_vs_state(struct si_context *sctx)
{
- if (sctx->gs_shader.cso)
+ if (sctx->gs_shader.cso &&
+ sctx->gs_shader.current &&
+ !sctx->gs_shader.current->key.as_ngg)
return sctx->gs_shader.cso->gs_copy_shader;
struct si_shader_ctx_state *vs = si_get_vs(sctx);
sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
SI_CONTEXT_INV_VCACHE;
- if (sctx->chip_class >= GFX9) {
+ if (sctx->chip_class >= GFX10) {
+ if (sctx->screen->info.tcc_harvested)
+ sctx->flags |= SI_CONTEXT_INV_L2;
+ else if (shaders_read_metadata)
+ sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
+ } else if (sctx->chip_class == GFX9) {
/* Single-sample color is coherent with shaders on GFX9, but
* L2 metadata must be flushed if shaders read metadata.
* (DCC, CMASK).
sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB |
SI_CONTEXT_INV_VCACHE;
- if (sctx->chip_class >= GFX9) {
+ if (sctx->chip_class >= GFX10) {
+ if (sctx->screen->info.tcc_harvested)
+ sctx->flags |= SI_CONTEXT_INV_L2;
+ else if (shaders_read_metadata)
+ sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
+ } else if (sctx->chip_class == GFX9) {
/* Single-sample depth (not stencil) is coherent with shaders
* on GFX9, but L2 metadata must be flushed if shaders read
* metadata.
}
static inline bool
-si_htile_enabled(struct si_texture *tex, unsigned level)
+si_htile_enabled(struct si_texture *tex, unsigned level, unsigned zs_mask)
{
- return tex->htile_offset && level == 0;
+ if (zs_mask == PIPE_MASK_S && tex->htile_stencil_disabled)
+ return false;
+
+ return tex->surface.htile_offset && level == 0;
}
static inline bool
-vi_tc_compat_htile_enabled(struct si_texture *tex, unsigned level)
+vi_tc_compat_htile_enabled(struct si_texture *tex, unsigned level, unsigned zs_mask)
{
- assert(!tex->tc_compatible_htile || tex->htile_offset);
- return tex->tc_compatible_htile && level == 0;
+ assert(!tex->tc_compatible_htile || tex->surface.htile_offset);
+ return tex->tc_compatible_htile && si_htile_enabled(tex, level, zs_mask);
}
static inline unsigned si_get_ps_iter_samples(struct si_context *sctx)
(1 << PIPE_PRIM_POINTS))) != 0;
}
+static inline bool util_rast_prim_is_triangles(unsigned prim)
+{
+ return ((1 << prim) & ((1 << PIPE_PRIM_TRIANGLES) |
+ (1 << PIPE_PRIM_TRIANGLE_STRIP) |
+ (1 << PIPE_PRIM_TRIANGLE_FAN) |
+ (1 << PIPE_PRIM_QUADS) |
+ (1 << PIPE_PRIM_QUAD_STRIP) |
+ (1 << PIPE_PRIM_POLYGON) |
+ (1 << PIPE_PRIM_TRIANGLES_ADJACENCY) |
+ (1 << PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY)));
+}
+
/**
* Return true if there is enough memory in VRAM and GTT for the buffers
* added so far.
return sctx->prim_discard_vertex_count_threshold != UINT_MAX;
}
+static inline unsigned si_get_wave_size(struct si_screen *sscreen,
+ enum pipe_shader_type shader_type,
+ bool ngg, bool es)
+{
+ 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 ((shader_type == PIPE_SHADER_VERTEX && es && !ngg) ||
+ (shader_type == PIPE_SHADER_TESS_EVAL && es && !ngg) ||
+ (shader_type == PIPE_SHADER_GEOMETRY && !ngg)) /* legacy GS only supports Wave64 */
+ return 64;
+ else
+ return sscreen->ge_wave_size;
+}
+
+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);
+}
+
#define PRINT_ERR(fmt, args...) \
fprintf(stderr, "EE %s:%d %s - " fmt, __FILE__, __LINE__, __func__, ##args)