#ifndef RADV_SHADER_H
#define RADV_SHADER_H
-#include "radv_debug.h"
-#include "radv_private.h"
+#include "ac_binary.h"
+#include "amd_family.h"
+#include "radv_constants.h"
#include "nir/nir.h"
+#include "vulkan/vulkan.h"
-/* descriptor index into scratch ring offsets */
-#define RING_SCRATCH 0
-#define RING_ESGS_VS 1
-#define RING_ESGS_GS 2
-#define RING_GSVS_VS 3
-#define RING_GSVS_GS 4
-#define RING_HS_TESS_FACTOR 5
-#define RING_HS_TESS_OFFCHIP 6
-#define RING_PS_SAMPLE_POSITIONS 7
-
-// Match MAX_SETS from radv_descriptor_set.h
-#define RADV_UD_MAX_SETS MAX_SETS
-
-#define RADV_NUM_PHYSICAL_VGPRS 256
+struct radv_device;
struct radv_shader_module {
struct nir_shader *nir;
RADV_ALPHA_ADJUST_SSCALED = 3,
};
+struct radv_vs_out_key {
+ uint32_t as_es:1;
+ uint32_t as_ls:1;
+ uint32_t as_ngg:1;
+ uint32_t export_prim_id:1;
+ uint32_t export_layer_id:1;
+ uint32_t export_clip_dists:1;
+};
+
struct radv_vs_variant_key {
+ struct radv_vs_out_key out;
+
uint32_t instance_rate_inputs;
uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
+ uint8_t vertex_attribute_formats[MAX_VERTEX_ATTRIBS];
+ uint32_t vertex_attribute_bindings[MAX_VERTEX_ATTRIBS];
+ uint32_t vertex_attribute_offsets[MAX_VERTEX_ATTRIBS];
+ uint32_t vertex_attribute_strides[MAX_VERTEX_ATTRIBS];
/* For 2_10_10_10 formats the alpha is handled as unsigned by pre-vega HW.
* so we may need to fix it up. */
uint64_t alpha_adjust;
- uint32_t as_es:1;
- uint32_t as_ls:1;
- uint32_t export_prim_id:1;
- uint32_t export_layer_id:1;
+ /* For some formats the channels have to be shuffled. */
+ uint32_t post_shuffle;
};
struct radv_tes_variant_key {
- uint32_t as_es:1;
- uint32_t export_prim_id:1;
- uint32_t export_layer_id:1;
+ struct radv_vs_out_key out;
+
uint8_t num_patches;
uint8_t tcs_num_outputs;
};
struct radv_fs_variant_key fs;
struct radv_tes_variant_key tes;
struct radv_tcs_variant_key tcs;
+
+ /* A common prefix of the vs and tes keys. */
+ struct radv_vs_out_key vs_common_out;
};
bool has_multiview_view_index;
};
bool unsafe_math;
bool supports_spill;
bool clamp_shadow_reference;
+ bool robust_buffer_access;
bool dump_shader;
bool dump_preoptir;
bool record_llvm_ir;
bool check_ir;
+ bool has_ls_vgpr_init_bug;
enum radeon_family family;
enum chip_class chip_class;
uint32_t tess_offchip_block_dw_size;
uint32_t address32_hi;
+ uint8_t wave_size;
};
enum radv_ud_index {
AC_UD_SCRATCH_RING_OFFSETS = 0,
AC_UD_PUSH_CONSTANTS = 1,
- AC_UD_INDIRECT_DESCRIPTOR_SETS = 2,
- AC_UD_VIEW_INDEX = 3,
- AC_UD_SHADER_START = 4,
+ AC_UD_INLINE_PUSH_CONSTANTS = 2,
+ AC_UD_INDIRECT_DESCRIPTOR_SETS = 3,
+ AC_UD_VIEW_INDEX = 4,
+ AC_UD_STREAMOUT_BUFFERS = 5,
+ AC_UD_SHADER_START = 6,
AC_UD_VS_VERTEX_BUFFERS = AC_UD_SHADER_START,
AC_UD_VS_BASE_VERTEX_START_INSTANCE,
AC_UD_VS_MAX_UD,
AC_UD_TES_MAX_UD,
AC_UD_MAX_UD = AC_UD_TCS_MAX_UD,
};
+
+struct radv_stream_output {
+ uint8_t location;
+ uint8_t buffer;
+ uint16_t offset;
+ uint8_t component_mask;
+ uint8_t stream;
+};
+
+struct radv_streamout_info {
+ uint16_t num_outputs;
+ struct radv_stream_output outputs[MAX_SO_OUTPUTS];
+ uint16_t strides[MAX_SO_BUFFERS];
+ uint32_t enabled_stream_buffers_mask;
+};
+
struct radv_shader_info {
bool loads_push_constants;
+ bool loads_dynamic_offsets;
+ uint8_t min_push_constant_used;
+ uint8_t max_push_constant_used;
+ bool has_only_32bit_push_constants;
+ bool has_indirect_push_constants;
+ uint8_t num_inline_push_consts;
+ uint8_t base_inline_push_consts;
uint32_t desc_set_used_mask;
bool needs_multiview_view_index;
bool uses_invocation_id;
bool uses_prim_id;
+ uint8_t wave_size;
struct {
uint64_t ls_outputs_written;
uint8_t input_usage_mask[VERT_ATTRIB_MAX];
} vs;
struct {
uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
+ uint8_t num_stream_output_components[4];
+ uint8_t output_streams[VARYING_SLOT_VAR31 + 1];
+ uint8_t max_stream;
} gs;
struct {
uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
struct {
bool force_persample;
bool needs_sample_positions;
- bool uses_input_attachments;
bool writes_memory;
bool writes_z;
bool writes_stencil;
bool has_pcoord;
bool prim_id_input;
bool layer_input;
+ uint8_t num_input_clips_culls;
+ uint32_t input_mask;
+ uint32_t flat_shaded_mask;
+ uint32_t float16_shaded_mask;
+ uint32_t num_interp;
} ps;
struct {
bool uses_grid_size;
uint64_t outputs_written;
uint64_t patch_outputs_written;
} tcs;
+
+ struct radv_streamout_info so;
};
struct radv_userdata_info {
int8_t sgpr_idx;
uint8_t num_sgprs;
- bool indirect;
- uint32_t indirect_offset;
};
struct radv_userdata_locations {
- struct radv_userdata_info descriptor_sets[RADV_UD_MAX_SETS];
+ struct radv_userdata_info descriptor_sets[MAX_SETS];
struct radv_userdata_info shader_data[AC_UD_MAX_UD];
uint32_t descriptor_sets_enabled;
};
unsigned num_input_vgprs;
unsigned private_mem_vgprs;
bool need_indirect_descriptor_sets;
+ bool is_ngg;
struct {
struct {
struct radv_vs_output_info outinfo;
struct radv_es_output_info es_info;
- unsigned vgpr_comp_cnt;
bool as_es;
bool as_ls;
+ bool export_prim_id;
} vs;
struct {
- unsigned num_interp;
- uint32_t input_mask;
- uint32_t flat_shaded_mask;
bool can_discard;
bool early_fragment_test;
+ bool post_depth_coverage;
} fs;
struct {
unsigned block_size[3];
enum gl_tess_spacing spacing;
bool ccw;
bool point_mode;
+ bool export_prim_id;
} tes;
};
};
+enum radv_shader_binary_type {
+ RADV_BINARY_TYPE_LEGACY,
+ RADV_BINARY_TYPE_RTLD
+};
+
+struct radv_shader_binary {
+ enum radv_shader_binary_type type;
+ gl_shader_stage stage;
+ bool is_gs_copy_shader;
+
+ struct radv_shader_variant_info variant_info;
+
+ /* Self-referential size so we avoid consistency issues. */
+ uint32_t total_size;
+};
+
+struct radv_shader_binary_legacy {
+ struct radv_shader_binary base;
+ struct ac_shader_config config;
+ unsigned code_size;
+ unsigned llvm_ir_size;
+ unsigned disasm_size;
+
+ /* data has size of code_size + llvm_ir_size + disasm_size + 2, where
+ * the +2 is for 0 of the ir strings. */
+ uint8_t data[0];
+};
+
+struct radv_shader_binary_rtld {
+ struct radv_shader_binary base;
+ unsigned elf_size;
+ unsigned llvm_ir_size;
+ uint8_t data[0];
+};
+
struct radv_shader_variant {
uint32_t ref_count;
struct ac_shader_config config;
uint32_t code_size;
struct radv_shader_variant_info info;
- unsigned rsrc1;
- unsigned rsrc2;
/* debug only */
uint32_t *spirv;
uint32_t spirv_size;
- struct nir_shader *nir;
+ char *nir_string;
char *disasm_string;
char *llvm_ir_string;
};
void
-radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively);
+radv_optimize_nir(struct nir_shader *shader, bool optimize_conservatively,
+ bool allow_copies);
+bool
+radv_nir_lower_ycbcr_textures(nir_shader *shader,
+ const struct radv_pipeline_layout *layout);
nir_shader *
radv_shader_compile_to_nir(struct radv_device *device,
const char *entrypoint_name,
gl_shader_stage stage,
const VkSpecializationInfo *spec_info,
- const VkPipelineCreateFlags flags);
+ const VkPipelineCreateFlags flags,
+ const struct radv_pipeline_layout *layout);
void *
radv_alloc_shader_memory(struct radv_device *device,
struct radv_shader_variant *
radv_shader_variant_create(struct radv_device *device,
- struct radv_shader_module *module,
- struct nir_shader *const *shaders,
- int shader_count,
- struct radv_pipeline_layout *layout,
- const struct radv_shader_variant_key *key,
- void **code_out,
- unsigned *code_size_out);
+ const struct radv_shader_binary *binary,
+ bool keep_shader_info);
+struct radv_shader_variant *
+radv_shader_variant_compile(struct radv_device *device,
+ struct radv_shader_module *module,
+ struct nir_shader *const *shaders,
+ int shader_count,
+ struct radv_pipeline_layout *layout,
+ const struct radv_shader_variant_key *key,
+ bool keep_shader_info,
+ struct radv_shader_binary **binary_out);
struct radv_shader_variant *
radv_create_gs_copy_shader(struct radv_device *device, struct nir_shader *nir,
- void **code_out, unsigned *code_size_out,
- bool multiview);
+ struct radv_shader_binary **binary_out,
+ bool multiview, bool keep_shader_info);
void
radv_shader_variant_destroy(struct radv_device *device,
struct radv_shader_variant *variant);
+
+unsigned
+radv_get_max_waves(struct radv_device *device,
+ struct radv_shader_variant *variant,
+ gl_shader_stage stage);
+
+unsigned
+radv_get_max_workgroup_size(enum chip_class chip_class,
+ gl_shader_stage stage,
+ const unsigned *sizes);
+
const char *
-radv_get_shader_name(struct radv_shader_variant *var, gl_shader_stage stage);
+radv_get_shader_name(struct radv_shader_variant_info *info,
+ gl_shader_stage stage);
void
radv_shader_dump_stats(struct radv_device *device,
gl_shader_stage stage,
FILE *file);
-static inline bool
+bool
radv_can_dump_shader(struct radv_device *device,
struct radv_shader_module *module,
- bool is_gs_copy_shader)
-{
- if (!(device->instance->debug_flags & RADV_DEBUG_DUMP_SHADERS))
- return false;
+ bool is_gs_copy_shader);
- /* Only dump non-meta shaders, useful for debugging purposes. */
- return (module && !module->nir) || is_gs_copy_shader;
-}
-
-static inline bool
+bool
radv_can_dump_shader_stats(struct radv_device *device,
- struct radv_shader_module *module)
-{
- /* Only dump non-meta shader stats. */
- return device->instance->debug_flags & RADV_DEBUG_DUMP_SHADER_STATS &&
- module && !module->nir;
-}
-
-static inline unsigned shader_io_get_unique_index(gl_varying_slot slot)
-{
- /* handle patch indices separate */
- if (slot == VARYING_SLOT_TESS_LEVEL_OUTER)
- return 0;
- if (slot == VARYING_SLOT_TESS_LEVEL_INNER)
- return 1;
- if (slot >= VARYING_SLOT_PATCH0 && slot <= VARYING_SLOT_TESS_MAX)
- return 2 + (slot - VARYING_SLOT_PATCH0);
- if (slot == VARYING_SLOT_POS)
- return 0;
- if (slot == VARYING_SLOT_PSIZ)
- return 1;
- if (slot == VARYING_SLOT_CLIP_DIST0)
- return 2;
- /* 3 is reserved for clip dist as well */
- if (slot >= VARYING_SLOT_VAR0 && slot <= VARYING_SLOT_VAR31)
- return 4 + (slot - VARYING_SLOT_VAR0);
- unreachable("illegal slot in get unique index\n");
-}
-
-static inline uint32_t
-radv_get_num_physical_sgprs(struct radv_physical_device *physical_device)
-{
- return physical_device->rad_info.chip_class >= VI ? 800 : 512;
-}
+ struct radv_shader_module *module);
+
+unsigned
+shader_io_get_unique_index(gl_varying_slot slot);
#endif