/*
* Copyright 2012 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+/* The compiler middle-end architecture: Explaining (non-)monolithic shaders
+ * -------------------------------------------------------------------------
+ *
+ * Typically, there is one-to-one correspondence between API and HW shaders,
+ * that is, for every API shader, there is exactly one shader binary in
+ * the driver.
+ *
+ * The problem with that is that we also have to emulate some API states
+ * (e.g. alpha-test, and many others) in shaders too. The two obvious ways
+ * to deal with it are:
+ * - each shader has multiple variants for each combination of emulated states,
+ * and the variants are compiled on demand, possibly relying on a shader
+ * cache for good performance
+ * - patch shaders at the binary level
+ *
+ * This driver uses something completely different. The emulated states are
+ * usually implemented at the beginning or end of shaders. Therefore, we can
+ * split the shader into 3 parts:
+ * - prolog part (shader code dependent on states)
+ * - main part (the API shader)
+ * - epilog part (shader code dependent on states)
+ *
+ * Each part is compiled as a separate shader and the final binaries are
+ * concatenated. This type of shader is called non-monolithic, because it
+ * consists of multiple independent binaries. Creating a new shader variant
+ * is therefore only a concatenation of shader parts (binaries) and doesn't
+ * involve any compilation. The main shader parts are the only parts that are
+ * compiled when applications create shader objects. The prolog and epilog
+ * parts are compiled on the first use and saved, so that their binaries can
+ * be reused by many other shaders.
+ *
+ * One of the roles of the prolog part is to compute vertex buffer addresses
+ * for vertex shaders. A few of the roles of the epilog part are color buffer
+ * format conversions in pixel shaders that we have to do manually, and write
+ * tessellation factors in tessellation control shaders. The prolog and epilog
+ * have many other important responsibilities in various shader stages.
+ * They don't just "emulate legacy stuff".
+ *
+ * Monolithic shaders are shaders where the parts are combined before LLVM
+ * compilation, and the whole thing is compiled and optimized as one unit with
+ * one binary on the output. The result is the same as the non-monolithic
+ * shader, but the final code can be better, because LLVM can optimize across
+ * all shader parts. Monolithic shaders aren't usually used except for these
+ * special cases:
*
- * Authors:
- * Tom Stellard <thomas.stellard@amd.com>
- * Michel Dänzer <michel.daenzer@amd.com>
- * Christian König <christian.koenig@amd.com>
+ * 1) Some rarely-used states require modification of the main shader part
+ * itself, and in such cases, only the monolithic shader variant is
+ * compiled, and that's always done on the first use.
+ *
+ * 2) When we do cross-stage optimizations for separate shader objects and
+ * e.g. eliminate unused shader varyings, the resulting optimized shader
+ * variants are always compiled as monolithic shaders, and always
+ * asynchronously (i.e. not stalling ongoing rendering). We call them
+ * "optimized monolithic" shaders. The important property here is that
+ * the non-monolithic unoptimized shader variant is always available for use
+ * when the asynchronous compilation of the optimized shader is not done
+ * yet.
+ *
+ * Starting with GFX9 chips, some shader stages are merged, and the number of
+ * shader parts per shader increased. The complete new list of shader parts is:
+ * - 1st shader: prolog part
+ * - 1st shader: main part
+ * - 2nd shader: prolog part
+ * - 2nd shader: main part
+ * - 2nd shader: epilog part
*/
/* How linking shader inputs and outputs between vertex, tessellation, and
#include <llvm-c/Core.h> /* LLVMModuleRef */
#include <llvm-c/TargetMachine.h>
#include "tgsi/tgsi_scan.h"
+#include "util/u_inlines.h"
#include "util/u_queue.h"
-#include "si_state.h"
-struct ac_shader_binary;
+#include "ac_binary.h"
+#include "ac_llvm_build.h"
+
+#include <stdio.h>
+
+struct nir_shader;
+struct si_shader;
+struct si_context;
+#define SI_MAX_ATTRIBS 16
#define SI_MAX_VS_OUTPUTS 40
+/* Shader IO unique indices are supported for TGSI_SEMANTIC_GENERIC with an
+ * index smaller than this.
+ */
+#define SI_MAX_IO_GENERIC 46
+
/* SGPR user data indices */
enum {
- /* GFX9 merged shaders have RW_BUFFERS among the first 8 system SGPRs,
- * and these two are used for other purposes.
- */
SI_SGPR_RW_BUFFERS, /* rings (& stream-out, VS only) */
+#if !HAVE_32BIT_POINTERS
SI_SGPR_RW_BUFFERS_HI,
- SI_SGPR_CONST_BUFFERS,
- SI_SGPR_CONST_BUFFERS_HI,
- SI_SGPR_SAMPLERS, /* images & sampler states interleaved */
- SI_SGPR_SAMPLERS_HI,
- SI_SGPR_IMAGES,
- SI_SGPR_IMAGES_HI,
- SI_SGPR_SHADER_BUFFERS,
- SI_SGPR_SHADER_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,
/* all VS variants */
- SI_SGPR_VERTEX_BUFFERS = SI_NUM_RESOURCE_SGPRS,
- SI_SGPR_VERTEX_BUFFERS_HI,
- SI_SGPR_BASE_VERTEX,
+ SI_SGPR_BASE_VERTEX = SI_NUM_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_ADDR,
SI_TES_NUM_USER_SGPR,
/* GFX6-8: TCS only */
GFX6_SGPR_TCS_IN_LAYOUT,
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 = SI_VS_NUM_USER_SGPR,
+ GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
GFX9_SGPR_TCS_OUT_OFFSETS,
GFX9_SGPR_TCS_OUT_LAYOUT,
- GFX9_SGPR_unused_to_align_the_next_pointer,
- GFX9_SGPR_TCS_CONST_BUFFERS,
- GFX9_SGPR_TCS_CONST_BUFFERS_HI,
- GFX9_SGPR_TCS_SAMPLERS, /* images & sampler states interleaved */
- GFX9_SGPR_TCS_SAMPLERS_HI,
- GFX9_SGPR_TCS_IMAGES,
- GFX9_SGPR_TCS_IMAGES_HI,
- GFX9_SGPR_TCS_SHADER_BUFFERS,
- GFX9_SGPR_TCS_SHADER_BUFFERS_HI,
+#if !HAVE_32BIT_POINTERS
+ GFX9_SGPR_align_for_vb_pointer,
+#endif
GFX9_TCS_NUM_USER_SGPR,
- /* GFX9: Merged ES-GS (VS-GS or TES-GS). */
- GFX9_SGPR_GS_CONST_BUFFERS = SI_VS_NUM_USER_SGPR,
- GFX9_SGPR_GS_CONST_BUFFERS_HI,
- GFX9_SGPR_GS_SAMPLERS,
- GFX9_SGPR_GS_SAMPLERS_HI,
- GFX9_SGPR_GS_IMAGES,
- GFX9_SGPR_GS_IMAGES_HI,
- GFX9_SGPR_GS_SHADER_BUFFERS,
- GFX9_SGPR_GS_SHADER_BUFFERS_HI,
- GFX9_GS_NUM_USER_SGPR,
-
/* GS limits */
GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
- SI_GSCOPY_NUM_USER_SGPR = SI_SGPR_RW_BUFFERS_HI + 1,
+#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),
/* PS only */
SI_SGPR_ALPHA_REF = SI_NUM_RESOURCE_SGPRS,
SI_PS_NUM_USER_SGPR,
-
- /* CS only */
- SI_SGPR_GRID_SIZE = SI_NUM_RESOURCE_SGPRS,
- SI_SGPR_BLOCK_SIZE = SI_SGPR_GRID_SIZE + 3,
- SI_CS_NUM_USER_SGPR = SI_SGPR_BLOCK_SIZE + 3
};
/* LLVM function parameter indices */
enum {
- SI_NUM_RESOURCE_PARAMS = 5,
+ SI_NUM_RESOURCE_PARAMS = 4,
/* PS only parameters */
SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
SI_PARAM_SAMPLE_COVERAGE,
SI_PARAM_POS_FIXED_PT,
- /* CS only parameters */
- SI_PARAM_GRID_SIZE = SI_NUM_RESOURCE_PARAMS,
- SI_PARAM_BLOCK_SIZE,
- SI_PARAM_BLOCK_ID,
- SI_PARAM_THREAD_ID,
-
SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
};
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,
+
+ /* These represent the number of SGPRs the shader uses. */
+ SI_VS_BLIT_SGPRS_POS = 3,
+ SI_VS_BLIT_SGPRS_POS_COLOR = 7,
+ SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
+};
+
/* For VS shader key fix_fetch. */
enum {
SI_FIX_FETCH_NONE = 0,
struct si_shader;
+/* Per-thread persistent LLVM objects. */
+struct si_compiler {
+ LLVMTargetMachineRef tm;
+ const char *triple;
+};
+
/* State of the context creating the shader object. */
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 si_compiler *compiler;
/* Used if thread_index == -1 or if debug.async is true. */
struct pipe_debug_callback debug;
struct si_shader *gs_copy_shader;
struct tgsi_token *tokens;
+ struct nir_shader *nir;
struct pipe_stream_output_info so;
struct tgsi_shader_info info;
+ struct tgsi_tessctrl_info tcs_info;
/* PIPE_SHADER_[VERTEX|FRAGMENT|...] */
unsigned type;
bool vs_needs_prolog;
+ bool force_correct_derivs_after_kill;
+ unsigned pa_cl_vs_out_cntl;
+ ubyte clipdist_mask;
+ ubyte culldist_mask;
- /* GS parameters. */
+ /* ES parameters. */
unsigned esgs_itemsize;
+
+ /* GS parameters. */
unsigned gs_input_verts_per_prim;
unsigned gs_output_prim;
unsigned gs_max_out_vertices;
unsigned max_gs_stream; /* count - 1 */
unsigned gsvs_vertex_size;
unsigned max_gsvs_emit_size;
+ unsigned enabled_streamout_buffer_mask;
/* PS parameters. */
unsigned color_attr_index[2];
unsigned local_size;
uint64_t outputs_written; /* "get_unique_index" bits */
- uint32_t patch_outputs_written; /* "get_unique_index" bits */
- uint32_t outputs_written2; /* "get_unique_index2" bits */
+ uint32_t patch_outputs_written; /* "get_unique_index_patch" bits */
uint64_t inputs_read; /* "get_unique_index" bits */
- uint32_t inputs_read2; /* "get_unique_index2" bits */
+
+ /* bitmasks of used descriptor slots */
+ uint32_t active_const_and_shader_buffers;
+ uint64_t active_samplers_and_images;
};
/* Valid shader configurations:
* -> = merged with the next stage
*/
+/* Use the byte alignment for all following structure members for optimal
+ * shader key memory footprint.
+ */
+#pragma pack(push, 1)
+
/* Common VS bits between the shader key and the prolog key. */
struct si_vs_prolog_bits {
- unsigned instance_divisors[SI_MAX_ATTRIBS];
-};
-
-/* Common VS bits between the shader key and the epilog key. */
-struct si_vs_epilog_bits {
- unsigned export_prim_id:1; /* when PS needs it and GS is disabled */
+ /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
+ * divisor is 0.
+ * - If "is_one" has a bit set, the instance divisor is 1.
+ * - If "is_fetched" has a bit set, the instance divisor will be loaded
+ * from the constant buffer.
+ */
+ uint16_t instance_divisor_is_one; /* bitmask of inputs */
+ uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
+ unsigned ls_vgpr_fix:1;
};
/* Common TCS bits between the shader key and the epilog key. */
struct si_tcs_epilog_bits {
unsigned prim_mode:3;
+ unsigned invoc0_tess_factors_are_def:1;
unsigned tes_reads_tess_factors:1;
};
struct si_gs_prolog_bits {
unsigned tri_strip_adj_fix:1;
+ unsigned gfx9_prev_is_vs:1;
};
/* Common PS bits between the shader key and the prolog key. */
unsigned force_linear_center_interp:1;
unsigned bc_optimize_for_persp:1;
unsigned bc_optimize_for_linear:1;
+ unsigned samplemask_log_ps_iter:3;
};
/* Common PS bits between the shader key and the epilog key. */
/* For merged stages such as LS-HS, HS input VGPRs are first. */
unsigned num_merged_next_stage_vgprs:3;
unsigned last_input:4;
+ unsigned as_ls:1;
+ unsigned as_es:1;
/* Prologs for monolithic shaders shouldn't set EXEC. */
unsigned is_monolithic:1;
} vs_prolog;
- struct {
- struct si_vs_epilog_bits states;
- unsigned prim_id_param_offset:5;
- } vs_epilog;
struct {
struct si_tcs_epilog_bits states;
} tcs_epilog;
unsigned colors_read:8; /* color input components read */
unsigned num_interp_inputs:5; /* BCOLOR is at this location */
unsigned face_vgpr_index:5;
+ unsigned ancillary_vgpr_index:5;
unsigned wqm:1;
char color_attr_index[2];
char color_interp_vgpr_index[2]; /* -1 == constant */
union {
struct {
struct si_vs_prolog_bits prolog;
- struct si_vs_epilog_bits epilog;
} vs;
struct {
struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
struct si_shader_selector *ls; /* for merged LS-HS */
struct si_tcs_epilog_bits epilog;
} tcs; /* tessellation control shader */
- struct {
- struct si_vs_epilog_bits epilog; /* same as VS */
- } tes; /* tessellation evaluation shader */
struct {
struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
struct si_shader_selector *es; /* for merged ES-GS */
struct {
/* One byte for every input: SI_FIX_FETCH_* enums. */
uint8_t vs_fix_fetch[SI_MAX_ATTRIBS];
- uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
+
+ union {
+ uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
+ /* When PS needs PrimID and GS is disabled. */
+ unsigned vs_export_prim_id:1;
+ struct {
+ unsigned interpolate_at_sample_force_center:1;
+ unsigned fbfetch_msaa;
+ unsigned fbfetch_is_1D;
+ unsigned fbfetch_layered;
+ } ps;
+ } u;
} mono;
/* Optimization flags for asynchronous compilation only. */
struct {
- struct {
- uint64_t kill_outputs; /* "get_unique_index" bits */
- uint32_t kill_outputs2; /* "get_unique_index2" bits */
- unsigned clip_disable:1;
- } hw_vs; /* HW VS (it can be VS, TES, GS) */
+ /* For HW VS (it can be VS, TES, GS) */
+ uint64_t kill_outputs; /* "get_unique_index" bits */
+ unsigned clip_disable:1;
/* For shaders where monolithic variants have better code.
*
} opt;
};
+/* Restore the pack alignment to default. */
+#pragma pack(pop)
+
struct si_shader_config {
unsigned num_sgprs;
unsigned num_vgprs;
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;
ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
ubyte num_input_sgprs;
ubyte num_input_vgprs;
- char face_vgpr_index;
+ signed char face_vgpr_index;
+ signed char ancillary_vgpr_index;
bool uses_instanceid;
ubyte nr_pos_exports;
ubyte nr_param_exports;
struct r600_resource *bo;
struct r600_resource *scratch_bo;
struct si_shader_key key;
- struct util_queue_fence optimized_ready;
+ struct util_queue_fence ready;
bool compilation_failed;
bool is_monolithic;
bool is_optimized;
/* si_shader.c */
struct si_shader *
si_generate_gs_copy_shader(struct si_screen *sscreen,
- LLVMTargetMachineRef tm,
+ struct si_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 si_compiler *compiler,
struct si_shader *shader,
bool is_monolithic,
struct pipe_debug_callback *debug);
-int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
+int si_shader_create(struct si_screen *sscreen, struct si_compiler *compiler,
struct si_shader *shader,
struct pipe_debug_callback *debug);
-int si_compile_llvm(struct si_screen *sscreen,
- struct ac_shader_binary *binary,
- struct si_shader_config *conf,
- LLVMTargetMachineRef tm,
- LLVMModuleRef mod,
- struct pipe_debug_callback *debug,
- unsigned processor,
- const char *name);
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);
-unsigned si_shader_io_get_unique_index2(unsigned name, unsigned index);
int si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader);
-void si_shader_dump(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,
FILE *f, bool check_debug_option);
+void si_shader_dump_stats_for_shader_db(const 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_context *sctx,
- struct si_shader *shader,
- struct si_shader_config *config,
- uint64_t scratch_va);
+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);
-unsigned si_get_spi_shader_z_format(bool writes_z, bool writes_stencil,
- bool writes_samplemask);
-const char *si_get_shader_name(struct si_shader *shader, unsigned processor);
+const char *si_get_shader_name(const struct si_shader *shader, unsigned processor);
+
+/* 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);
/* Inline helpers. */
return &sel->main_shader_part;
}
+static inline bool
+si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
+{
+ return selector ? selector->info.uses_bindless_samplers : false;
+}
+
+static inline bool
+si_shader_uses_bindless_images(struct si_shader_selector *selector)
+{
+ return selector ? selector->info.uses_bindless_images : false;
+}
+
+void si_destroy_shader_selector(struct si_context *sctx,
+ struct si_shader_selector *sel);
+
+static inline void
+si_shader_selector_reference(struct si_context *sctx,
+ struct si_shader_selector **dst,
+ struct si_shader_selector *src)
+{
+ if (pipe_reference(&(*dst)->reference, &src->reference))
+ si_destroy_shader_selector(sctx, *dst);
+
+ *dst = src;
+}
+
#endif