#include "ac_surface.h"
#include "ac_llvm_build.h"
#include "ac_llvm_util.h"
+#include "radv_constants.h"
#include "radv_descriptor_set.h"
#include "radv_extensions.h"
-#include "radv_cs.h"
+#include "sid.h"
#include <llvm-c/TargetMachine.h>
#include "wsi_common.h"
#include "wsi_common_display.h"
-#define ATI_VENDOR_ID 0x1002
-
-#define MAX_VBS 32
-#define MAX_VERTEX_ATTRIBS 32
-#define MAX_RTS 8
-#define MAX_VIEWPORTS 16
-#define MAX_SCISSORS 16
-#define MAX_DISCARD_RECTANGLES 4
-#define MAX_SAMPLE_LOCATIONS 32
-#define MAX_PUSH_CONSTANTS_SIZE 128
-#define MAX_PUSH_DESCRIPTORS 32
-#define MAX_DYNAMIC_UNIFORM_BUFFERS 16
-#define MAX_DYNAMIC_STORAGE_BUFFERS 8
-#define MAX_DYNAMIC_BUFFERS (MAX_DYNAMIC_UNIFORM_BUFFERS + MAX_DYNAMIC_STORAGE_BUFFERS)
-#define MAX_SAMPLES_LOG2 4
-#define NUM_META_FS_KEYS 12
-#define RADV_MAX_DRM_DEVICES 8
-#define MAX_VIEWS 8
-#define MAX_SO_STREAMS 4
-#define MAX_SO_BUFFERS 4
-#define MAX_SO_OUTPUTS 64
-#define MAX_INLINE_UNIFORM_BLOCK_SIZE (4ull * 1024 * 1024)
-#define MAX_INLINE_UNIFORM_BLOCK_COUNT 64
-
-#define NUM_DEPTH_CLEAR_PIPELINES 3
+struct gfx10_format {
+ unsigned img_format:9;
-/*
- * This is the point we switch from using CP to compute shader
- * for certain buffer operations.
- */
-#define RADV_BUFFER_OPS_CS_THRESHOLD 4096
+ /* Various formats are only supported with workarounds for vertex fetch,
+ * and some 32_32_32 formats are supported natively, but only for buffers
+ * (possibly with some image support, actually, but no filtering). */
+ bool buffers_only:1;
+};
-#define RADV_BUFFER_UPDATE_THRESHOLD 1024
+#include "gfx10_format_table.h"
enum radv_mem_heap {
RADV_MEM_HEAP_VRAM,
* propagating errors. Might be useful to plug in a stack trace here.
*/
+struct radv_image_view;
struct radv_instance;
VkResult __vk_errorf(struct radv_instance *instance, VkResult error, const char *file, int line, const char *format, ...);
int master_fd;
struct wsi_device wsi_device;
- bool has_rbplus; /* if RB+ register exist */
- bool rbplus_allowed; /* if RB+ is allowed */
- bool has_clear_state;
- bool cpdma_prefetch_writes_memory;
- bool has_scissor_bug;
-
- bool has_out_of_order_rast;
bool out_of_order_rast_allowed;
/* Whether DCC should be enabled for MSAA textures. */
bool dcc_msaa_allowed;
- /* Whether LOAD_CONTEXT_REG packets are supported. */
- bool has_load_ctx_reg_pkt;
-
/* Whether to enable the AMD_shader_ballot extension */
bool use_shader_ballot;
+ /* Number of threads per wave. */
+ uint8_t ps_wave_size;
+ uint8_t cs_wave_size;
+ uint8_t ge_wave_size;
+
/* This is the drivers on-disk cache used as a fallback as opposed to
* the pipeline cache defined by apps.
*/
uint32_t optimisations_disabled : 1;
};
+struct radv_shader_binary;
+struct radv_shader_variant;
+
void
radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
struct radv_device *device);
radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
const void *data, size_t size);
-struct radv_shader_variant;
-
bool
radv_create_shader_variants_from_pipeline_cache(struct radv_device *device,
struct radv_pipeline_cache *cache,
struct radv_pipeline_cache *cache,
const unsigned char *sha1,
struct radv_shader_variant **variants,
- const void *const *codes,
- const unsigned *code_sizes);
+ struct radv_shader_binary *const *binaries);
enum radv_blit_ds_layout {
RADV_BLIT_DS_LAYOUT_TILE_ENABLE,
struct radeon_cmdbuf *empty_cs[RADV_MAX_QUEUE_FAMILIES];
bool always_use_syncobj;
- bool has_distributed_tess;
bool pbb_allowed;
bool dfsm_allowed;
uint32_t tess_offchip_block_dw_size;
struct radv_device_extension_table enabled_extensions;
+ /* Whether the app has enabled the robustBufferAccess feature. */
+ bool robust_buffer_access;
+
/* Whether the driver uses a global BO list. */
bool use_global_bo_list;
};
enum radv_cmd_flush_bits {
- RADV_CMD_FLAG_INV_ICACHE = 1 << 0,
- /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
- RADV_CMD_FLAG_INV_SMEM_L1 = 1 << 1,
- /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
- RADV_CMD_FLAG_INV_VMEM_L1 = 1 << 2,
- /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
- RADV_CMD_FLAG_INV_GLOBAL_L2 = 1 << 3,
- /* Same as above, but only writes back and doesn't invalidate */
- RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 = 1 << 4,
+ /* Instruction cache. */
+ RADV_CMD_FLAG_INV_ICACHE = 1 << 0,
+ /* Scalar L1 cache. */
+ RADV_CMD_FLAG_INV_SCACHE = 1 << 1,
+ /* Vector L1 cache. */
+ RADV_CMD_FLAG_INV_VCACHE = 1 << 2,
+ /* L2 cache + L2 metadata cache writeback & invalidate.
+ * GFX6-8: Used by shaders only. GFX9-10: Used by everything. */
+ RADV_CMD_FLAG_INV_L2 = 1 << 3,
+ /* L2 writeback (write dirty L2 lines to memory for non-L2 clients).
+ * Only used for coherency with non-L2 clients like CB, DB, CP on GFX6-8.
+ * GFX6-7 will do complete invalidation, because the writeback is unsupported. */
+ RADV_CMD_FLAG_WB_L2 = 1 << 4,
/* Framebuffer caches */
- RADV_CMD_FLAG_FLUSH_AND_INV_CB_META = 1 << 5,
- RADV_CMD_FLAG_FLUSH_AND_INV_DB_META = 1 << 6,
- RADV_CMD_FLAG_FLUSH_AND_INV_DB = 1 << 7,
- RADV_CMD_FLAG_FLUSH_AND_INV_CB = 1 << 8,
+ RADV_CMD_FLAG_FLUSH_AND_INV_CB_META = 1 << 5,
+ RADV_CMD_FLAG_FLUSH_AND_INV_DB_META = 1 << 6,
+ RADV_CMD_FLAG_FLUSH_AND_INV_DB = 1 << 7,
+ RADV_CMD_FLAG_FLUSH_AND_INV_CB = 1 << 8,
/* Engine synchronization. */
- RADV_CMD_FLAG_VS_PARTIAL_FLUSH = 1 << 9,
- RADV_CMD_FLAG_PS_PARTIAL_FLUSH = 1 << 10,
- RADV_CMD_FLAG_CS_PARTIAL_FLUSH = 1 << 11,
- RADV_CMD_FLAG_VGT_FLUSH = 1 << 12,
+ RADV_CMD_FLAG_VS_PARTIAL_FLUSH = 1 << 9,
+ RADV_CMD_FLAG_PS_PARTIAL_FLUSH = 1 << 10,
+ RADV_CMD_FLAG_CS_PARTIAL_FLUSH = 1 << 11,
+ RADV_CMD_FLAG_VGT_FLUSH = 1 << 12,
/* Pipeline query controls. */
- RADV_CMD_FLAG_START_PIPELINE_STATS = 1 << 13,
- RADV_CMD_FLAG_STOP_PIPELINE_STATS = 1 << 14,
- RADV_CMD_FLAG_VGT_STREAMOUT_SYNC = 1 << 15,
+ RADV_CMD_FLAG_START_PIPELINE_STATS = 1 << 13,
+ RADV_CMD_FLAG_STOP_PIPELINE_STATS = 1 << 14,
+ RADV_CMD_FLAG_VGT_STREAMOUT_SYNC = 1 << 15,
RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER = (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
const char *
radv_get_perftest_option_name(int id);
+struct radv_color_buffer_info {
+ uint64_t cb_color_base;
+ uint64_t cb_color_cmask;
+ uint64_t cb_color_fmask;
+ uint64_t cb_dcc_base;
+ uint32_t cb_color_slice;
+ uint32_t cb_color_view;
+ uint32_t cb_color_info;
+ uint32_t cb_color_attrib;
+ uint32_t cb_color_attrib2; /* GFX9 and later */
+ uint32_t cb_color_attrib3; /* GFX10 and later */
+ uint32_t cb_dcc_control;
+ uint32_t cb_color_cmask_slice;
+ uint32_t cb_color_fmask_slice;
+ union {
+ uint32_t cb_color_pitch; // GFX6-GFX8
+ uint32_t cb_mrt_epitch; // GFX9+
+ };
+};
+
+struct radv_ds_buffer_info {
+ uint64_t db_z_read_base;
+ uint64_t db_stencil_read_base;
+ uint64_t db_z_write_base;
+ uint64_t db_stencil_write_base;
+ uint64_t db_htile_data_base;
+ uint32_t db_depth_info;
+ uint32_t db_z_info;
+ uint32_t db_stencil_info;
+ uint32_t db_depth_view;
+ uint32_t db_depth_size;
+ uint32_t db_depth_slice;
+ uint32_t db_htile_surface;
+ uint32_t pa_su_poly_offset_db_fmt_cntl;
+ uint32_t db_z_info2; /* GFX9 only */
+ uint32_t db_stencil_info2; /* GFX9 only */
+ float offset_scale;
+};
+
+void
+radv_initialise_color_surface(struct radv_device *device,
+ struct radv_color_buffer_info *cb,
+ struct radv_image_view *iview);
+void
+radv_initialise_ds_surface(struct radv_device *device,
+ struct radv_ds_buffer_info *ds,
+ struct radv_image_view *iview);
+
/**
* Attachment state when recording a renderpass instance.
*
uint32_t cleared_views;
VkClearValue clear_value;
VkImageLayout current_layout;
+ bool current_in_render_loop;
struct radv_sample_locations_state sample_location;
+
+ union {
+ struct radv_color_buffer_info cb;
+ struct radv_ds_buffer_info ds;
+ };
+ struct radv_image_view *iview;
};
struct radv_descriptor_state {
enum chip_class chip_class,
bool is_mec,
unsigned event, unsigned event_flags,
- unsigned data_sel,
+ unsigned dst_sel, unsigned data_sel,
uint64_t va,
uint32_t new_fence,
uint64_t gfx9_eop_bug_va);
void radv_device_init_msaa(struct radv_device *device);
void radv_update_ds_clear_metadata(struct radv_cmd_buffer *cmd_buffer,
- struct radv_image *image,
+ const struct radv_image_view *iview,
VkClearDepthStencilValue ds_clear_value,
VkImageAspectFlags aspects);
#define RADV_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
#define RADV_HASH_SHADER_SISCHED (1 << 1)
#define RADV_HASH_SHADER_UNSAFE_MATH (1 << 2)
+#define RADV_HASH_SHADER_NO_NGG (1 << 3)
+#define RADV_HASH_SHADER_CS_WAVE32 (1 << 4)
+#define RADV_HASH_SHADER_PS_WAVE32 (1 << 5)
+#define RADV_HASH_SHADER_GE_WAVE32 (1 << 6)
+
void
radv_hash_shaders(unsigned char *hash,
const VkPipelineShaderStageCreateInfo **stages,
bool partial_vs_wave;
};
+struct radv_binning_state {
+ uint32_t pa_sc_binner_cntl_0;
+ uint32_t db_dfsm_control;
+};
+
#define SI_GS_PER_ES 128
struct radv_pipeline {
union {
struct {
struct radv_multisample_state ms;
+ struct radv_binning_state binning;
uint32_t spi_baryc_cntl;
bool prim_restart_enable;
unsigned esgs_ring_size;
return pipeline->shaders[MESA_SHADER_TESS_CTRL] ? true : false;
}
+bool radv_pipeline_has_ngg(const struct radv_pipeline *pipeline);
+
+bool radv_pipeline_has_gs_copy_shader(const struct radv_pipeline *pipeline);
+
struct radv_userdata_info *radv_lookup_user_sgpr(struct radv_pipeline *pipeline,
gl_shader_stage stage,
int idx);
VkFormat format2);
bool radv_device_supports_etc(struct radv_physical_device *physical_device);
-struct radv_fmask_info {
- uint64_t offset;
- uint64_t size;
- unsigned alignment;
- unsigned pitch_in_pixels;
- unsigned bank_height;
- unsigned slice_tile_max;
- unsigned tile_mode_index;
- unsigned tile_swizzle;
- uint64_t slice_size;
-};
-
-struct radv_cmask_info {
- uint64_t offset;
- uint64_t size;
- unsigned alignment;
- unsigned slice_tile_max;
- unsigned slice_size;
-};
-
-
struct radv_image_plane {
VkFormat format;
struct radeon_surf surface;
bool tc_compatible_htile;
bool tc_compatible_cmask;
- struct radv_fmask_info fmask;
- struct radv_cmask_info cmask;
+ uint64_t cmask_offset;
+ uint64_t fmask_offset;
uint64_t clear_value_offset;
uint64_t fce_pred_offset;
uint64_t dcc_pred_offset;
* the image. */
bool radv_layout_has_htile(const struct radv_image *image,
VkImageLayout layout,
+ bool in_render_loop,
unsigned queue_mask);
/* Whether the image has a htile that is known consistent with the contents of
*/
bool radv_layout_is_htile_compressed(const struct radv_image *image,
VkImageLayout layout,
+ bool in_render_loop,
unsigned queue_mask);
bool radv_layout_can_fast_clear(const struct radv_image *image,
VkImageLayout layout,
+ bool in_render_loop,
unsigned queue_mask);
-bool radv_layout_dcc_compressed(const struct radv_image *image,
+bool radv_layout_dcc_compressed(const struct radv_device *device,
+ const struct radv_image *image,
VkImageLayout layout,
+ bool in_render_loop,
unsigned queue_mask);
/**
static inline bool
radv_image_has_cmask(const struct radv_image *image)
{
- return image->cmask.size;
+ return image->cmask_offset;
}
/**
static inline bool
radv_image_has_fmask(const struct radv_image *image)
{
- return image->fmask.size;
+ return image->fmask_offset;
}
/**
return va;
}
+static inline uint64_t
+radv_get_tc_compat_zrange_va(const struct radv_image *image,
+ uint32_t base_level)
+{
+ uint64_t va = radv_buffer_get_va(image->bo);
+ va += image->offset + image->tc_compat_zrange_offset + base_level * 4;
+ return va;
+}
+
+static inline uint64_t
+radv_get_ds_clear_value_va(const struct radv_image *image,
+ uint32_t base_level)
+{
+ uint64_t va = radv_buffer_get_va(image->bo);
+ va += image->offset + image->clear_value_offset + base_level * 8;
+ return va;
+}
+
unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family);
static inline uint32_t
const VkAllocationCallbacks *alloc,
VkImage *out_image_h);
+struct radv_image_view_extra_create_info {
+ bool disable_compression;
+};
+
void radv_image_view_init(struct radv_image_view *view,
struct radv_device *device,
- const VkImageViewCreateInfo* pCreateInfo);
+ const VkImageViewCreateInfo *pCreateInfo,
+ const struct radv_image_view_extra_create_info* extra_create_info);
VkFormat radv_get_aspect_format(struct radv_image *image, VkImageAspectFlags mask);
struct radv_sampler_ycbcr_conversion *ycbcr_sampler;
};
-struct radv_color_buffer_info {
- uint64_t cb_color_base;
- uint64_t cb_color_cmask;
- uint64_t cb_color_fmask;
- uint64_t cb_dcc_base;
- uint32_t cb_color_slice;
- uint32_t cb_color_view;
- uint32_t cb_color_info;
- uint32_t cb_color_attrib;
- uint32_t cb_color_attrib2;
- uint32_t cb_dcc_control;
- uint32_t cb_color_cmask_slice;
- uint32_t cb_color_fmask_slice;
- union {
- uint32_t cb_color_pitch; // GFX6-GFX8
- uint32_t cb_mrt_epitch; // GFX9+
- };
-};
-
-struct radv_ds_buffer_info {
- uint64_t db_z_read_base;
- uint64_t db_stencil_read_base;
- uint64_t db_z_write_base;
- uint64_t db_stencil_write_base;
- uint64_t db_htile_data_base;
- uint32_t db_depth_info;
- uint32_t db_z_info;
- uint32_t db_stencil_info;
- uint32_t db_depth_view;
- uint32_t db_depth_size;
- uint32_t db_depth_slice;
- uint32_t db_htile_surface;
- uint32_t pa_su_poly_offset_db_fmt_cntl;
- uint32_t db_z_info2;
- uint32_t db_stencil_info2;
- float offset_scale;
-};
-
-struct radv_attachment_info {
- union {
- struct radv_color_buffer_info cb;
- struct radv_ds_buffer_info ds;
- };
- struct radv_image_view *attachment;
-};
-
struct radv_framebuffer {
uint32_t width;
uint32_t height;
uint32_t layers;
uint32_t attachment_count;
- struct radv_attachment_info attachments[0];
+ struct radv_image_view *attachments[0];
};
struct radv_subpass_barrier {
struct radv_subpass_attachment {
uint32_t attachment;
VkImageLayout layout;
+ bool in_render_loop;
};
struct radv_subpass {
};
/* radv_nir_to_llvm.c */
-struct radv_shader_variant_info;
+struct radv_shader_info;
struct radv_nir_compiler_options;
void radv_compile_gs_copy_shader(struct ac_llvm_compiler *ac_llvm,
struct nir_shader *geom_shader,
- struct ac_shader_binary *binary,
- struct ac_shader_config *config,
- struct radv_shader_variant_info *shader_info,
+ struct radv_shader_binary **rbinary,
+ struct radv_shader_info *info,
const struct radv_nir_compiler_options *option);
void radv_compile_nir_shader(struct ac_llvm_compiler *ac_llvm,
- struct ac_shader_binary *binary,
- struct ac_shader_config *config,
- struct radv_shader_variant_info *shader_info,
+ struct radv_shader_binary **rbinary,
+ struct radv_shader_info *info,
struct nir_shader *const *nir,
int nir_count,
const struct radv_nir_compiler_options *options);
unsigned radv_nir_get_max_workgroup_size(enum chip_class chip_class,
+ gl_shader_stage stage,
const struct nir_shader *nir);
/* radv_shader_info.h */
struct radv_shader_info;
+struct radv_shader_variant_key;
void radv_nir_shader_info_pass(const struct nir_shader *nir,
- const struct radv_nir_compiler_options *options,
+ const struct radv_pipeline_layout *layout,
+ const struct radv_shader_variant_key *key,
struct radv_shader_info *info);
void radv_nir_shader_info_init(struct radv_shader_info *info);
uint64_t radv_get_current_time(void);
+static inline uint32_t
+si_conv_gl_prim_to_vertices(unsigned gl_prim)
+{
+ switch (gl_prim) {
+ case 0: /* GL_POINTS */
+ return 1;
+ case 1: /* GL_LINES */
+ case 3: /* GL_LINE_STRIP */
+ return 2;
+ case 4: /* GL_TRIANGLES */
+ case 5: /* GL_TRIANGLE_STRIP */
+ return 3;
+ case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
+ return 4;
+ case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
+ return 6;
+ case 7: /* GL_QUADS */
+ return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
+ default:
+ assert(0);
+ return 0;
+ }
+}
+
#define RADV_DEFINE_HANDLE_CASTS(__radv_type, __VkType) \
\
static inline struct __radv_type * \