#include "ac_binary.h"
#include "ac_llvm_build.h"
+#include "ac_llvm_util.h"
#include <stdio.h>
+// Use LDS symbols when supported by LLVM. Can be disabled for testing the old
+// path on newer LLVM for now. Should be removed in the long term.
+#define USE_LDS_SYMBOLS (true)
+
struct nir_shader;
struct si_shader;
struct si_context;
/* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
* index smaller than this.
*/
-#define SI_MAX_IO_GENERIC 46
+#define SI_MAX_IO_GENERIC 32
/* SGPR user data indices */
enum {
SI_SGPR_RW_BUFFERS, /* rings (& stream-out, VS only) */
-#if !HAVE_32BIT_POINTERS
- SI_SGPR_RW_BUFFERS_HI,
-#endif
SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
-#if !HAVE_32BIT_POINTERS
- SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES_HI,
-#endif
SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
-#if !HAVE_32BIT_POINTERS
- SI_SGPR_CONST_AND_SHADER_BUFFERS_HI,
-#endif
SI_SGPR_SAMPLERS_AND_IMAGES,
-#if !HAVE_32BIT_POINTERS
- SI_SGPR_SAMPLERS_AND_IMAGES_HI,
-#endif
SI_NUM_RESOURCE_SGPRS,
+ /* API VS, TES without GS, GS copy shader */
+ SI_SGPR_VS_STATE_BITS = SI_NUM_RESOURCE_SGPRS,
+ SI_NUM_VS_STATE_RESOURCE_SGPRS,
+
/* all VS variants */
- SI_SGPR_BASE_VERTEX = SI_NUM_RESOURCE_SGPRS,
+ SI_SGPR_BASE_VERTEX = SI_NUM_VS_STATE_RESOURCE_SGPRS,
SI_SGPR_START_INSTANCE,
SI_SGPR_DRAWID,
- SI_SGPR_VS_STATE_BITS,
SI_VS_NUM_USER_SGPR,
SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
/* TES */
- SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
+ SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_VS_STATE_RESOURCE_SGPRS,
SI_SGPR_TES_OFFCHIP_ADDR,
SI_TES_NUM_USER_SGPR,
GFX6_TCS_NUM_USER_SGPR,
/* GFX9: Merged shaders. */
-#if HAVE_32BIT_POINTERS
/* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
/* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
-#else
- /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO/HI (SGPR[0:1]). */
- GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES = SI_VS_NUM_USER_SGPR,
- GFX9_SGPR_2ND_SAMPLERS_AND_IMAGES_HI,
- GFX9_MERGED_NUM_USER_SGPR,
-#endif
/* GFX9: Merged LS-HS (VS-TCS) only. */
GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
GFX9_SGPR_TCS_OUT_OFFSETS,
GFX9_SGPR_TCS_OUT_LAYOUT,
-#if !HAVE_32BIT_POINTERS
- GFX9_SGPR_align_for_vb_pointer,
-#endif
GFX9_TCS_NUM_USER_SGPR,
/* GS limits */
GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
-#if HAVE_32BIT_POINTERS
GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
-#else
- GFX9_VSGS_NUM_USER_SGPR = GFX9_MERGED_NUM_USER_SGPR,
- GFX9_TESGS_NUM_USER_SGPR = GFX9_MERGED_NUM_USER_SGPR,
-#endif
- SI_GSCOPY_NUM_USER_SGPR = SI_SGPR_RW_BUFFERS + (HAVE_32BIT_POINTERS ? 1 : 2),
+ SI_GSCOPY_NUM_USER_SGPR = SI_NUM_VS_STATE_RESOURCE_SGPRS,
/* PS only */
SI_SGPR_ALPHA_REF = SI_NUM_RESOURCE_SGPRS,
};
/* Fields of driver-defined VS state SGPR. */
-/* Clamp vertex color output (only used in VS as VS). */
#define S_VS_STATE_CLAMP_VERTEX_COLOR(x) (((unsigned)(x) & 0x1) << 0)
#define C_VS_STATE_CLAMP_VERTEX_COLOR 0xFFFFFFFE
#define S_VS_STATE_INDEXED(x) (((unsigned)(x) & 0x1) << 1)
#define C_VS_STATE_INDEXED 0xFFFFFFFD
+#define S_VS_STATE_OUTPRIM(x) (((unsigned)(x) & 0x3) << 2)
+#define C_VS_STATE_OUTPRIM 0xFFFFFFF3
+#define S_VS_STATE_PROVOKING_VTX_INDEX(x) (((unsigned)(x) & 0x3) << 4)
+#define C_VS_STATE_PROVOKING_VTX_INDEX 0xFFFFFFCF
+#define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x) & 0x1) << 6)
+#define C_VS_STATE_STREAMOUT_QUERY_ENABLED 0xFFFFFFBF
#define S_VS_STATE_LS_OUT_PATCH_SIZE(x) (((unsigned)(x) & 0x1FFF) << 8)
#define C_VS_STATE_LS_OUT_PATCH_SIZE 0xFFE000FF
#define S_VS_STATE_LS_OUT_VERTEX_SIZE(x) (((unsigned)(x) & 0xFF) << 24)
#define C_VS_STATE_LS_OUT_VERTEX_SIZE 0x00FFFFFF
-/* SI-specific system values. */
+/* Driver-specific system values. */
enum {
+ /* Values from set_tess_state. */
TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI = TGSI_SEMANTIC_COUNT,
TGSI_SEMANTIC_DEFAULT_TESSINNER_SI,
};
enum {
- /* Use a property enum that VS wouldn't use. */
- TGSI_PROPERTY_VS_BLIT_SGPRS = TGSI_PROPERTY_FS_COORD_ORIGIN,
+ /* Use a property enum that CS wouldn't use. */
+ TGSI_PROPERTY_CS_LOCAL_SIZE = TGSI_PROPERTY_FS_COORD_ORIGIN,
/* These represent the number of SGPRs the shader uses. */
SI_VS_BLIT_SGPRS_POS = 3,
SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
};
-/* For VS shader key fix_fetch. */
-enum {
- SI_FIX_FETCH_NONE = 0,
- SI_FIX_FETCH_A2_SNORM,
- SI_FIX_FETCH_A2_SSCALED,
- SI_FIX_FETCH_A2_SINT,
- SI_FIX_FETCH_RGBA_32_UNORM,
- SI_FIX_FETCH_RGBX_32_UNORM,
- SI_FIX_FETCH_RGBA_32_SNORM,
- SI_FIX_FETCH_RGBX_32_SNORM,
- SI_FIX_FETCH_RGBA_32_USCALED,
- SI_FIX_FETCH_RGBA_32_SSCALED,
- SI_FIX_FETCH_RGBA_32_FIXED,
- SI_FIX_FETCH_RGBX_32_FIXED,
- SI_FIX_FETCH_RG_64_FLOAT,
- SI_FIX_FETCH_RGB_64_FLOAT,
- SI_FIX_FETCH_RGBA_64_FLOAT,
- SI_FIX_FETCH_RGB_8, /* A = 1.0 */
- SI_FIX_FETCH_RGB_8_INT, /* A = 1 */
- SI_FIX_FETCH_RGB_16,
- SI_FIX_FETCH_RGB_16_INT,
+/**
+ * For VS shader keys, describe any fixups required for vertex fetch.
+ *
+ * \ref log_size, \ref format, and the number of channels are interpreted as
+ * by \ref ac_build_opencoded_load_format.
+ *
+ * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
+ * impossible format and indicates that no fixup is needed (just use
+ * buffer_load_format_xyzw).
+ */
+union si_vs_fix_fetch {
+ struct {
+ uint8_t log_size : 2; /* 1, 2, 4, 8 or bytes per channel */
+ uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
+ uint8_t format : 3; /* AC_FETCH_FORMAT_xxx */
+ uint8_t reverse : 1; /* reverse XYZ channels */
+ } u;
+ uint8_t bits;
};
struct si_shader;
struct si_compiler_ctx_state {
/* Should only be used by si_init_shader_selector_async and
* si_build_shader_variant if thread_index == -1 (non-threaded). */
- LLVMTargetMachineRef tm;
+ struct ac_llvm_compiler *compiler;
/* Used if thread_index == -1 or if debug.async is true. */
struct pipe_debug_callback debug;
struct si_shader *main_shader_part;
struct si_shader *main_shader_part_ls; /* as_ls is set in the key */
struct si_shader *main_shader_part_es; /* as_es is set in the key */
+ struct si_shader *main_shader_part_ngg; /* as_ngg is set in the key */
+ struct si_shader *main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
struct si_shader *gs_copy_shader;
struct tgsi_tessctrl_info tcs_info;
/* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
- unsigned type;
+ enum pipe_shader_type type;
bool vs_needs_prolog;
bool force_correct_derivs_after_kill;
+ bool prim_discard_cs_allowed;
+ bool ngg_writes_edgeflag;
+ bool pos_writes_edgeflag;
unsigned pa_cl_vs_out_cntl;
ubyte clipdist_mask;
ubyte culldist_mask;
+ unsigned rast_prim;
/* ES parameters. */
- unsigned esgs_itemsize;
+ unsigned esgs_itemsize; /* vertex stride */
+ unsigned lshs_vertex_stride;
/* GS parameters. */
unsigned gs_input_verts_per_prim;
unsigned gsvs_vertex_size;
unsigned max_gsvs_emit_size;
unsigned enabled_streamout_buffer_mask;
+ bool tess_turns_off_ngg;
/* PS parameters. */
unsigned color_attr_index[2];
*/
unsigned colors_written_4bit;
- /* CS parameters */
- unsigned local_size;
-
+ uint64_t outputs_written_before_ps; /* "get_unique_index" bits */
uint64_t outputs_written; /* "get_unique_index" bits */
uint32_t patch_outputs_written; /* "get_unique_index_patch" bits */
/* Valid shader configurations:
*
- * API shaders VS | TCS | TES | GS |pass| PS
- * are compiled as: | | | |thru|
- * | | | | |
- * Only VS & PS: VS | | | | | PS
- * GFX6 - with GS: ES | | | GS | VS | PS
- * - with tess: LS | HS | VS | | | PS
- * - with both: LS | HS | ES | GS | VS | PS
- * GFX9 - with GS: -> | | | GS | VS | PS
- * - with tess: -> | HS | VS | | | PS
- * - with both: -> | HS | -> | GS | VS | PS
+ * API shaders VS | TCS | TES | GS |pass| PS
+ * are compiled as: | | | |thru|
+ * | | | | |
+ * Only VS & PS: VS | | | | | PS
+ * GFX6 - with GS: ES | | | GS | VS | PS
+ * - with tess: LS | HS | VS | | | PS
+ * - with both: LS | HS | ES | GS | VS | PS
+ * GFX9 - with GS: -> | | | GS | VS | PS
+ * - with tess: -> | HS | VS | | | PS
+ * - with both: -> | HS | -> | GS | VS | PS
+ * | | | | |
+ * NGG - VS & PS: GS | | | | | PS
+ * (GFX10+) - with GS: -> | | | GS | | PS
+ * - with tess: -> | HS | GS | | | PS
+ * - with both: -> | HS | -> | GS | | PS
*
* -> = merged with the next stage
*/
uint16_t instance_divisor_is_one; /* bitmask of inputs */
uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
unsigned ls_vgpr_fix:1;
+ unsigned unpack_instance_id_from_vertex_id:1;
};
/* Common TCS bits between the shader key and the epilog key. */
unsigned last_input:4;
unsigned as_ls:1;
unsigned as_es:1;
+ unsigned as_ngg:1;
/* Prologs for monolithic shaders shouldn't set EXEC. */
unsigned is_monolithic:1;
} vs_prolog;
struct si_gs_prolog_bits states;
/* Prologs of monolithic shaders shouldn't set EXEC. */
unsigned is_monolithic:1;
+ unsigned as_ngg:1;
} gs_prolog;
struct {
struct si_ps_prolog_bits states;
unsigned ancillary_vgpr_index:5;
unsigned wqm:1;
char color_attr_index[2];
- char color_interp_vgpr_index[2]; /* -1 == constant */
+ signed char color_interp_vgpr_index[2]; /* -1 == constant */
} ps_prolog;
struct {
struct si_ps_epilog_bits states;
} ps;
} part;
- /* These two are initially set according to the NEXT_SHADER property,
+ /* These three are initially set according to the NEXT_SHADER property,
* or guessed if the property doesn't seem correct.
*/
unsigned as_es:1; /* export shader, which precedes GS */
unsigned as_ls:1; /* local shader, which precedes TCS */
+ unsigned as_ngg:1; /* VS, TES, or GS compiled as NGG primitive shader */
/* Flags for monolithic compilation only. */
struct {
- /* One byte for every input: SI_FIX_FETCH_* enums. */
- uint8_t vs_fix_fetch[SI_MAX_ATTRIBS];
+ /* Whether fetch should be opencoded according to vs_fix_fetch.
+ * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
+ * with minimal fixups is used. */
+ uint16_t vs_fetch_opencode;
+ union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
union {
uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
unsigned vs_export_prim_id:1;
struct {
unsigned interpolate_at_sample_force_center:1;
- unsigned fbfetch_msaa;
- unsigned fbfetch_is_1D;
- unsigned fbfetch_layered;
+ unsigned fbfetch_msaa:1;
+ unsigned fbfetch_is_1D:1;
+ unsigned fbfetch_layered:1;
} ps;
} u;
} mono;
* possible, because it's in the "opt" group.
*/
unsigned prefer_mono:1;
+
+ /* Primitive discard compute shader. */
+ unsigned vs_as_prim_discard_cs:1;
+ unsigned cs_prim_type:4;
+ unsigned cs_indexed:1;
+ unsigned cs_instancing:1;
+ unsigned cs_primitive_restart:1;
+ unsigned cs_provoking_vertex_first:1;
+ unsigned cs_need_correct_orientation:1;
+ unsigned cs_cull_front:1;
+ unsigned cs_cull_back:1;
+ unsigned cs_cull_z:1;
+ unsigned cs_halfz_clip_space:1;
} opt;
};
/* Restore the pack alignment to default. */
#pragma pack(pop)
-struct si_shader_config {
- unsigned num_sgprs;
- unsigned num_vgprs;
- unsigned spilled_sgprs;
- unsigned spilled_vgprs;
- unsigned private_mem_vgprs;
- unsigned lds_size;
- unsigned max_simd_waves;
- unsigned spi_ps_input_ena;
- unsigned spi_ps_input_addr;
- unsigned float_mode;
- unsigned scratch_bytes_per_wave;
- unsigned rsrc1;
- unsigned rsrc2;
-};
-
/* GCN-specific shader info. */
struct si_shader_info {
ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
bool uses_instanceid;
ubyte nr_pos_exports;
ubyte nr_param_exports;
+ unsigned private_mem_vgprs;
+ unsigned max_simd_waves;
+};
+
+struct si_shader_binary {
+ const char *elf_buffer;
+ size_t elf_size;
+
+ char *llvm_ir_string;
+};
+
+struct gfx9_gs_info {
+ unsigned es_verts_per_subgroup;
+ unsigned gs_prims_per_subgroup;
+ unsigned gs_inst_prims_in_subgroup;
+ unsigned max_prims_per_subgroup;
+ unsigned esgs_ring_size; /* in bytes */
};
struct si_shader {
struct si_shader_part *epilog;
struct si_pm4_state *pm4;
- struct r600_resource *bo;
- struct r600_resource *scratch_bo;
+ struct si_resource *bo;
+ struct si_resource *scratch_bo;
struct si_shader_key key;
struct util_queue_fence ready;
bool compilation_failed;
bool is_gs_copy_shader;
/* The following data is all that's needed for binary shaders. */
- struct ac_shader_binary binary;
- struct si_shader_config config;
+ struct si_shader_binary binary;
+ struct ac_shader_config config;
struct si_shader_info info;
+ struct {
+ uint16_t ngg_emit_size; /* in dwords */
+ uint16_t hw_max_esverts;
+ uint16_t max_gsprims;
+ uint16_t max_out_verts;
+ uint16_t prim_amp_factor;
+ bool max_vert_out_per_gs_instance;
+ } ngg;
+
/* Shader key + LLVM IR + disassembly + statistics.
* Generated for debug contexts only.
*/
char *shader_log;
size_t shader_log_size;
+
+ struct gfx9_gs_info gs_info;
+
+ /* For save precompute context registers values. */
+ union {
+ struct {
+ unsigned vgt_gsvs_ring_offset_1;
+ unsigned vgt_gsvs_ring_offset_2;
+ unsigned vgt_gsvs_ring_offset_3;
+ unsigned vgt_gsvs_ring_itemsize;
+ unsigned vgt_gs_max_vert_out;
+ unsigned vgt_gs_vert_itemsize;
+ unsigned vgt_gs_vert_itemsize_1;
+ unsigned vgt_gs_vert_itemsize_2;
+ unsigned vgt_gs_vert_itemsize_3;
+ unsigned vgt_gs_instance_cnt;
+ unsigned vgt_gs_onchip_cntl;
+ unsigned vgt_gs_max_prims_per_subgroup;
+ unsigned vgt_esgs_ring_itemsize;
+ } gs;
+
+ struct {
+ unsigned ge_max_output_per_subgroup;
+ unsigned ge_ngg_subgrp_cntl;
+ unsigned vgt_primitiveid_en;
+ unsigned vgt_gs_onchip_cntl;
+ unsigned vgt_gs_instance_cnt;
+ unsigned vgt_esgs_ring_itemsize;
+ unsigned spi_vs_out_config;
+ unsigned spi_shader_idx_format;
+ unsigned spi_shader_pos_format;
+ unsigned pa_cl_vte_cntl;
+ unsigned pa_cl_ngg_cntl;
+ unsigned vgt_gs_max_vert_out; /* for API GS */
+ } ngg;
+
+ struct {
+ unsigned vgt_gs_mode;
+ unsigned vgt_primitiveid_en;
+ unsigned vgt_reuse_off;
+ unsigned spi_vs_out_config;
+ unsigned spi_shader_pos_format;
+ unsigned pa_cl_vte_cntl;
+ } vs;
+
+ struct {
+ unsigned spi_ps_input_ena;
+ unsigned spi_ps_input_addr;
+ unsigned spi_baryc_cntl;
+ unsigned spi_ps_in_control;
+ unsigned spi_shader_z_format;
+ unsigned spi_shader_col_format;
+ unsigned cb_shader_mask;
+ } ps;
+ } ctx_reg;
+
+ /*For save precompute registers value */
+ unsigned vgt_tf_param; /* VGT_TF_PARAM */
+ unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
+ unsigned ge_cntl;
};
struct si_shader_part {
struct si_shader_part *next;
union si_shader_part_key key;
- struct ac_shader_binary binary;
- struct si_shader_config config;
+ struct si_shader_binary binary;
+ struct ac_shader_config config;
};
/* si_shader.c */
struct si_shader *
si_generate_gs_copy_shader(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
+ struct ac_llvm_compiler *compiler,
struct si_shader_selector *gs_selector,
struct pipe_debug_callback *debug);
int si_compile_tgsi_shader(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
+ struct ac_llvm_compiler *compiler,
struct si_shader *shader,
- bool is_monolithic,
struct pipe_debug_callback *debug);
-int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
+bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
struct si_shader *shader,
struct pipe_debug_callback *debug);
void si_shader_destroy(struct si_shader *shader);
unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index);
-unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index);
-int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader);
-void si_shader_dump(struct si_screen *sscreen, const struct si_shader *shader,
- struct pipe_debug_callback *debug, unsigned processor,
+unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
+ unsigned is_varying);
+bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
+ uint64_t scratch_va);
+void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
+ struct pipe_debug_callback *debug,
FILE *f, bool check_debug_option);
-void si_shader_dump_stats_for_shader_db(const struct si_shader *shader,
+void si_shader_dump_stats_for_shader_db(struct si_screen *screen,
+ struct si_shader *shader,
struct pipe_debug_callback *debug);
void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
unsigned *lds_size);
-void si_shader_apply_scratch_relocs(struct si_shader *shader,
- uint64_t scratch_va);
-void si_shader_binary_read_config(struct ac_shader_binary *binary,
- struct si_shader_config *conf,
- unsigned symbol_offset);
-const char *si_get_shader_name(const struct si_shader *shader, unsigned processor);
+const char *si_get_shader_name(const struct si_shader *shader);
+void si_shader_binary_clean(struct si_shader_binary *binary);
/* si_shader_nir.c */
void si_nir_scan_shader(const struct nir_shader *nir,
struct tgsi_shader_info *info);
void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
- const struct tgsi_shader_info *info,
struct tgsi_tessctrl_info *out);
-void si_lower_nir(struct si_shader_selector *sel);
+void si_nir_lower_ps_inputs(struct nir_shader *nir);
+void si_lower_nir(struct si_shader_selector *sel, unsigned wave_size);
+void si_nir_opts(struct nir_shader *nir);
+
+/* si_state_shaders.c */
+void gfx9_get_gs_info(struct si_shader_selector *es,
+ struct si_shader_selector *gs,
+ struct gfx9_gs_info *out);
/* Inline helpers. */
{
if (key->as_ls)
return &sel->main_shader_part_ls;
+ if (key->as_es && key->as_ngg)
+ return &sel->main_shader_part_ngg_es;
if (key->as_es)
return &sel->main_shader_part_es;
+ if (key->as_ngg)
+ return &sel->main_shader_part_ngg;
return &sel->main_shader_part;
}
projects / mesa.git / blobdiff