#include "nir/nir.h"
#include "vulkan/vulkan.h"
+#include "vulkan/util/vk_object.h"
+
+#define RADV_VERT_ATTRIB_MAX MAX2(VERT_ATTRIB_MAX, VERT_ATTRIB_GENERIC0 + MAX_VERTEX_ATTRIBS)
struct radv_device;
struct radv_shader_module {
+ struct vk_object_base base;
struct nir_shader *nir;
unsigned char sha1[20];
uint32_t size;
uint32_t export_prim_id:1;
uint32_t export_layer_id:1;
uint32_t export_clip_dists:1;
+ uint32_t export_viewport_index:1;
};
struct radv_vs_variant_key {
uint8_t num_samples;
uint32_t is_int8;
uint32_t is_int10;
+ bool is_dual_src;
};
struct radv_cs_variant_key {
bool has_multiview_view_index;
};
+enum radv_compiler_debug_level {
+ RADV_COMPILER_DEBUG_LEVEL_PERFWARN,
+ RADV_COMPILER_DEBUG_LEVEL_ERROR,
+};
+
struct radv_nir_compiler_options {
struct radv_pipeline_layout *layout;
struct radv_shader_variant_key key;
bool dump_shader;
bool dump_preoptir;
bool record_ir;
+ bool record_stats;
bool check_ir;
bool has_ls_vgpr_init_bug;
bool use_ngg_streamout;
+ bool enable_mrt_output_nan_fixup;
enum radeon_family family;
enum chip_class chip_class;
uint32_t tess_offchip_block_dw_size;
uint32_t address32_hi;
+
+ struct {
+ void (*func)(void *private_data,
+ enum radv_compiler_debug_level level,
+ const char *message);
+ void *private_data;
+ } debug;
};
enum radv_ud_index {
bool is_ngg_passthrough;
struct {
uint64_t ls_outputs_written;
- uint8_t input_usage_mask[VERT_ATTRIB_MAX];
+ uint8_t input_usage_mask[RADV_VERT_ATTRIB_MAX];
uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
bool has_vertex_buffers; /* needs vertex buffers and base/start */
bool needs_draw_id;
bool as_es;
bool as_ls;
bool export_prim_id;
+ uint8_t num_linked_outputs;
} vs;
struct {
uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
unsigned output_prim;
unsigned invocations;
unsigned es_type; /* GFX9: VS or TES */
+ uint8_t num_linked_inputs;
} gs;
struct {
uint8_t output_usage_mask[VARYING_SLOT_VAR31 + 1];
bool ccw;
bool point_mode;
bool export_prim_id;
+ uint8_t num_linked_inputs;
+ uint8_t num_linked_patch_inputs;
+ uint8_t num_linked_outputs;
} tes;
struct {
bool force_persample;
bool has_pcoord;
bool prim_id_input;
bool layer_input;
+ bool viewport_index_input;
uint8_t num_input_clips_culls;
uint32_t input_mask;
uint32_t flat_shaded_mask;
uint32_t explicit_shaded_mask;
uint32_t float16_shaded_mask;
uint32_t num_interp;
+ uint32_t cb_shader_mask;
bool can_discard;
bool early_fragment_test;
bool post_depth_coverage;
+ uint8_t depth_layout;
} ps;
struct {
bool uses_grid_size;
struct {
uint64_t outputs_written;
uint64_t patch_outputs_written;
+ uint64_t tes_inputs_read;
+ uint64_t tes_patch_inputs_read;
unsigned tcs_vertices_out;
uint32_t num_patches;
- uint32_t lds_size;
+ uint32_t num_lds_blocks;
+ uint8_t num_linked_inputs;
+ uint8_t num_linked_outputs;
+ uint8_t num_linked_patch_outputs;
} tcs;
struct radv_streamout_info so;
const struct radv_pipeline_layout *layout,
unsigned subgroup_size, unsigned ballot_bit_size);
-void *
-radv_alloc_shader_memory(struct radv_device *device,
- struct radv_shader_variant *shader);
-
void
radv_destroy_shader_slabs(struct radv_device *device);
-void
+VkResult
radv_create_shaders(struct radv_pipeline *pipeline,
struct radv_device *device,
struct radv_pipeline_cache *cache,
struct radv_pipeline_layout *layout,
const struct radv_shader_variant_key *key,
struct radv_shader_info *info,
- bool keep_shader_info,
+ bool keep_shader_info, bool keep_statistic_info,
struct radv_shader_binary **binary_out);
struct radv_shader_variant *
radv_create_gs_copy_shader(struct radv_device *device, struct nir_shader *nir,
struct radv_shader_info *info,
struct radv_shader_binary **binary_out,
- bool multiview, bool keep_shader_info);
+ bool multiview, bool keep_shader_info,
+ bool keep_statistic_info);
+
+struct radv_shader_variant *
+radv_create_trap_handler_shader(struct radv_device *device);
void
radv_shader_variant_destroy(struct radv_device *device,
radv_get_shader_name(struct radv_shader_info *info,
gl_shader_stage stage);
-void
-radv_shader_dump_stats(struct radv_device *device,
- struct radv_shader_variant *variant,
- gl_shader_stage stage,
- FILE *file);
-
bool
radv_can_dump_shader(struct radv_device *device,
struct radv_shader_module *module,
radv_can_dump_shader_stats(struct radv_device *device,
struct radv_shader_module *module);
+VkResult
+radv_dump_shader_stats(struct radv_device *device,
+ struct radv_pipeline *pipeline,
+ gl_shader_stage stage, FILE *output);
+
static inline unsigned
shader_io_get_unique_index(gl_varying_slot slot)
{
}
static inline unsigned
-calculate_tess_lds_size(unsigned tcs_num_input_vertices,
+calculate_tess_lds_size(enum chip_class chip_class,
+ unsigned tcs_num_input_vertices,
unsigned tcs_num_output_vertices,
unsigned tcs_num_inputs,
unsigned tcs_num_patches,
- unsigned tcs_outputs_written,
- unsigned tcs_per_patch_outputs_written)
+ unsigned tcs_num_outputs,
+ unsigned tcs_num_patch_outputs)
{
- unsigned num_tcs_outputs = util_last_bit64(tcs_outputs_written);
- unsigned num_tcs_patch_outputs = util_last_bit64(tcs_per_patch_outputs_written);
-
unsigned input_vertex_size = tcs_num_inputs * 16;
- unsigned output_vertex_size = num_tcs_outputs * 16;
+ unsigned output_vertex_size = tcs_num_outputs * 16;
unsigned input_patch_size = tcs_num_input_vertices * input_vertex_size;
unsigned pervertex_output_patch_size = tcs_num_output_vertices * output_vertex_size;
- unsigned output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
+ unsigned output_patch_size = pervertex_output_patch_size + tcs_num_patch_outputs * 16;
unsigned output_patch0_offset = input_patch_size * tcs_num_patches;
- return output_patch0_offset + output_patch_size * tcs_num_patches;
+ unsigned lds_size = output_patch0_offset + output_patch_size * tcs_num_patches;
+
+ if (chip_class >= GFX7) {
+ assert(lds_size <= 65536);
+ lds_size = align(lds_size, 512) / 512;
+ } else {
+ assert(lds_size <= 32768);
+ lds_size = align(lds_size, 256) / 256;
+ }
+
+ return lds_size;
}
static inline unsigned
get_tcs_num_patches(unsigned tcs_num_input_vertices,
unsigned tcs_num_output_vertices,
unsigned tcs_num_inputs,
- unsigned tcs_outputs_written,
- unsigned tcs_per_patch_outputs_written,
+ unsigned tcs_num_outputs,
+ unsigned tcs_num_patch_outputs,
unsigned tess_offchip_block_dw_size,
enum chip_class chip_class,
enum radeon_family family)
{
uint32_t input_vertex_size = tcs_num_inputs * 16;
uint32_t input_patch_size = tcs_num_input_vertices * input_vertex_size;
- uint32_t num_tcs_outputs = util_last_bit64(tcs_outputs_written);
- uint32_t num_tcs_patch_outputs = util_last_bit64(tcs_per_patch_outputs_written);
- uint32_t output_vertex_size = num_tcs_outputs * 16;
+ uint32_t output_vertex_size = tcs_num_outputs * 16;
uint32_t pervertex_output_patch_size = tcs_num_output_vertices * output_vertex_size;
- uint32_t output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
+ uint32_t output_patch_size = pervertex_output_patch_size + tcs_num_patch_outputs * 16;
/* Ensure that we only need one wave per SIMD so we don't need to check
* resource usage. Also ensures that the number of tcs in and out