• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2012 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * on the rights to use, copy, modify, merge, publish, distribute, sub
9  * license, and/or sell copies of the Software, and to permit persons to whom
10  * the Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
23  */
24 
25 /* The compiler middle-end architecture: Explaining (non-)monolithic shaders
26  * -------------------------------------------------------------------------
27  *
28  * Typically, there is one-to-one correspondence between API and HW shaders,
29  * that is, for every API shader, there is exactly one shader binary in
30  * the driver.
31  *
32  * The problem with that is that we also have to emulate some API states
33  * (e.g. alpha-test, and many others) in shaders too. The two obvious ways
34  * to deal with it are:
35  * - each shader has multiple variants for each combination of emulated states,
36  *   and the variants are compiled on demand, possibly relying on a shader
37  *   cache for good performance
38  * - patch shaders at the binary level
39  *
40  * This driver uses something completely different. The emulated states are
41  * usually implemented at the beginning or end of shaders. Therefore, we can
42  * split the shader into 3 parts:
43  * - prolog part (shader code dependent on states)
44  * - main part (the API shader)
45  * - epilog part (shader code dependent on states)
46  *
47  * Each part is compiled as a separate shader and the final binaries are
48  * concatenated. This type of shader is called non-monolithic, because it
49  * consists of multiple independent binaries. Creating a new shader variant
50  * is therefore only a concatenation of shader parts (binaries) and doesn't
51  * involve any compilation. The main shader parts are the only parts that are
52  * compiled when applications create shader objects. The prolog and epilog
53  * parts are compiled on the first use and saved, so that their binaries can
54  * be reused by many other shaders.
55  *
56  * One of the roles of the prolog part is to compute vertex buffer addresses
57  * for vertex shaders. A few of the roles of the epilog part are color buffer
58  * format conversions in pixel shaders that we have to do manually, and write
59  * tessellation factors in tessellation control shaders. The prolog and epilog
60  * have many other important responsibilities in various shader stages.
61  * They don't just "emulate legacy stuff".
62  *
63  * Monolithic shaders are shaders where the parts are combined before LLVM
64  * compilation, and the whole thing is compiled and optimized as one unit with
65  * one binary on the output. The result is the same as the non-monolithic
66  * shader, but the final code can be better, because LLVM can optimize across
67  * all shader parts. Monolithic shaders aren't usually used except for these
68  * special cases:
69  *
70  * 1) Some rarely-used states require modification of the main shader part
71  *    itself, and in such cases, only the monolithic shader variant is
72  *    compiled, and that's always done on the first use.
73  *
74  * 2) When we do cross-stage optimizations for separate shader objects and
75  *    e.g. eliminate unused shader varyings, the resulting optimized shader
76  *    variants are always compiled as monolithic shaders, and always
77  *    asynchronously (i.e. not stalling ongoing rendering). We call them
78  *    "optimized monolithic" shaders. The important property here is that
79  *    the non-monolithic unoptimized shader variant is always available for use
80  *    when the asynchronous compilation of the optimized shader is not done
81  *    yet.
82  *
83  * Starting with GFX9 chips, some shader stages are merged, and the number of
84  * shader parts per shader increased. The complete new list of shader parts is:
85  * - 1st shader: prolog part
86  * - 1st shader: main part
87  * - 2nd shader: prolog part
88  * - 2nd shader: main part
89  * - 2nd shader: epilog part
90  */
91 
92 /* How linking shader inputs and outputs between vertex, tessellation, and
93  * geometry shaders works.
94  *
95  * Inputs and outputs between shaders are stored in a buffer. This buffer
96  * lives in LDS (typical case for tessellation), but it can also live
97  * in memory (ESGS). Each input or output has a fixed location within a vertex.
98  * The highest used input or output determines the stride between vertices.
99  *
100  * Since GS and tessellation are only possible in the OpenGL core profile,
101  * only these semantics are valid for per-vertex data:
102  *
103  *   Name             Location
104  *
105  *   POSITION         0
106  *   PSIZE            1
107  *   CLIPDIST0..1     2..3
108  *   CULLDIST0..1     (not implemented)
109  *   GENERIC0..31     4..35
110  *
111  * For example, a shader only writing GENERIC0 has the output stride of 5.
112  *
113  * Only these semantics are valid for per-patch data:
114  *
115  *   Name             Location
116  *
117  *   TESSOUTER        0
118  *   TESSINNER        1
119  *   PATCH0..29       2..31
120  *
121  * That's how independent shaders agree on input and output locations.
122  * The si_shader_io_get_unique_index function assigns the locations.
123  *
124  * For tessellation, other required information for calculating the input and
125  * output addresses like the vertex stride, the patch stride, and the offsets
126  * where per-vertex and per-patch data start, is passed to the shader via
127  * user data SGPRs. The offsets and strides are calculated at draw time and
128  * aren't available at compile time.
129  */
130 
131 #ifndef SI_SHADER_H
132 #define SI_SHADER_H
133 
134 #include "ac_binary.h"
135 #include "ac_llvm_build.h"
136 #include "ac_llvm_util.h"
137 #include "util/simple_mtx.h"
138 #include "util/u_inlines.h"
139 #include "util/u_live_shader_cache.h"
140 #include "util/u_queue.h"
141 
142 #include <stdio.h>
143 
144 // Use LDS symbols when supported by LLVM. Can be disabled for testing the old
145 // path on newer LLVM for now. Should be removed in the long term.
146 #define USE_LDS_SYMBOLS (true)
147 
148 struct nir_shader;
149 struct si_shader;
150 struct si_context;
151 
152 #define SI_MAX_ATTRIBS    16
153 #define SI_MAX_VS_OUTPUTS 40
154 
155 /* Shader IO unique indices are supported for VARYING_SLOT_VARn with an
156  * index smaller than this.
157  */
158 #define SI_MAX_IO_GENERIC 32
159 
160 #define SI_NGG_PRIM_EDGE_FLAG_BITS ((1 << 9) | (1 << 19) | (1 << 29))
161 
162 /* SGPR user data indices */
163 enum
164 {
165    SI_SGPR_RW_BUFFERS, /* rings (& stream-out, VS only) */
166    SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES,
167    SI_SGPR_CONST_AND_SHADER_BUFFERS, /* or just a constant buffer 0 pointer */
168    SI_SGPR_SAMPLERS_AND_IMAGES,
169    SI_NUM_RESOURCE_SGPRS,
170 
171    /* API VS, TES without GS, GS copy shader */
172    SI_SGPR_VS_STATE_BITS = SI_NUM_RESOURCE_SGPRS,
173    SI_NUM_VS_STATE_RESOURCE_SGPRS,
174 
175    /* all VS variants */
176    SI_SGPR_BASE_VERTEX = SI_NUM_VS_STATE_RESOURCE_SGPRS,
177    SI_SGPR_START_INSTANCE,
178    SI_SGPR_DRAWID,
179    SI_VS_NUM_USER_SGPR,
180 
181    SI_SGPR_VS_BLIT_DATA = SI_SGPR_CONST_AND_SHADER_BUFFERS,
182 
183    /* TES */
184    SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_VS_STATE_RESOURCE_SGPRS,
185    SI_SGPR_TES_OFFCHIP_ADDR,
186    SI_TES_NUM_USER_SGPR,
187 
188    /* GFX6-8: TCS only */
189    GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
190    GFX6_SGPR_TCS_OUT_OFFSETS,
191    GFX6_SGPR_TCS_OUT_LAYOUT,
192    GFX6_SGPR_TCS_IN_LAYOUT,
193    GFX6_TCS_NUM_USER_SGPR,
194 
195    /* GFX9: Merged shaders. */
196    /* 2ND_CONST_AND_SHADER_BUFFERS is set in USER_DATA_ADDR_LO (SGPR0). */
197    /* 2ND_SAMPLERS_AND_IMAGES is set in USER_DATA_ADDR_HI (SGPR1). */
198    GFX9_MERGED_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
199 
200    /* GFX9: Merged LS-HS (VS-TCS) only. */
201    GFX9_SGPR_TCS_OFFCHIP_LAYOUT = GFX9_MERGED_NUM_USER_SGPR,
202    GFX9_SGPR_TCS_OUT_OFFSETS,
203    GFX9_SGPR_TCS_OUT_LAYOUT,
204    GFX9_TCS_NUM_USER_SGPR,
205 
206    /* GS limits */
207    GFX6_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
208    GFX9_VSGS_NUM_USER_SGPR = SI_VS_NUM_USER_SGPR,
209    GFX9_TESGS_NUM_USER_SGPR = SI_TES_NUM_USER_SGPR,
210    SI_GSCOPY_NUM_USER_SGPR = SI_NUM_VS_STATE_RESOURCE_SGPRS,
211 
212    /* PS only */
213    SI_SGPR_ALPHA_REF = SI_NUM_RESOURCE_SGPRS,
214    SI_PS_NUM_USER_SGPR,
215 
216    /* The value has to be 12, because the hw requires that descriptors
217     * are aligned to 4 SGPRs.
218     */
219    SI_SGPR_VS_VB_DESCRIPTOR_FIRST = 12,
220 };
221 
222 /* LLVM function parameter indices */
223 enum
224 {
225    SI_NUM_RESOURCE_PARAMS = 4,
226 
227    /* PS only parameters */
228    SI_PARAM_ALPHA_REF = SI_NUM_RESOURCE_PARAMS,
229    SI_PARAM_PRIM_MASK,
230    SI_PARAM_PERSP_SAMPLE,
231    SI_PARAM_PERSP_CENTER,
232    SI_PARAM_PERSP_CENTROID,
233    SI_PARAM_PERSP_PULL_MODEL,
234    SI_PARAM_LINEAR_SAMPLE,
235    SI_PARAM_LINEAR_CENTER,
236    SI_PARAM_LINEAR_CENTROID,
237    SI_PARAM_LINE_STIPPLE_TEX,
238    SI_PARAM_POS_X_FLOAT,
239    SI_PARAM_POS_Y_FLOAT,
240    SI_PARAM_POS_Z_FLOAT,
241    SI_PARAM_POS_W_FLOAT,
242    SI_PARAM_FRONT_FACE,
243    SI_PARAM_ANCILLARY,
244    SI_PARAM_SAMPLE_COVERAGE,
245    SI_PARAM_POS_FIXED_PT,
246 
247    SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
248 };
249 
250 /* Fields of driver-defined VS state SGPR. */
251 #define S_VS_STATE_CLAMP_VERTEX_COLOR(x)      (((unsigned)(x)&0x1) << 0)
252 #define C_VS_STATE_CLAMP_VERTEX_COLOR         0xFFFFFFFE
253 #define S_VS_STATE_INDEXED(x)                 (((unsigned)(x)&0x1) << 1)
254 #define C_VS_STATE_INDEXED                    0xFFFFFFFD
255 #define S_VS_STATE_OUTPRIM(x)                 (((unsigned)(x)&0x3) << 2)
256 #define C_VS_STATE_OUTPRIM                    0xFFFFFFF3
257 #define S_VS_STATE_PROVOKING_VTX_INDEX(x)     (((unsigned)(x)&0x3) << 4)
258 #define C_VS_STATE_PROVOKING_VTX_INDEX        0xFFFFFFCF
259 #define S_VS_STATE_STREAMOUT_QUERY_ENABLED(x) (((unsigned)(x)&0x1) << 6)
260 #define C_VS_STATE_STREAMOUT_QUERY_ENABLED    0xFFFFFFBF
261 #define S_VS_STATE_SMALL_PRIM_PRECISION(x)    (((unsigned)(x)&0xF) << 7)
262 #define C_VS_STATE_SMALL_PRIM_PRECISION       0xFFFFF87F
263 #define S_VS_STATE_LS_OUT_PATCH_SIZE(x)       (((unsigned)(x)&0x1FFF) << 11)
264 #define C_VS_STATE_LS_OUT_PATCH_SIZE          0xFF0007FF
265 #define S_VS_STATE_LS_OUT_VERTEX_SIZE(x)      (((unsigned)(x)&0xFF) << 24)
266 #define C_VS_STATE_LS_OUT_VERTEX_SIZE         0x00FFFFFF
267 
268 enum
269 {
270    /* These represent the number of SGPRs the shader uses. */
271    SI_VS_BLIT_SGPRS_POS = 3,
272    SI_VS_BLIT_SGPRS_POS_COLOR = 7,
273    SI_VS_BLIT_SGPRS_POS_TEXCOORD = 9,
274 };
275 
276 #define SI_NGG_CULL_VIEW_SMALLPRIMS          (1 << 0)   /* view.xy + small prims */
277 #define SI_NGG_CULL_BACK_FACE                (1 << 1)   /* back faces */
278 #define SI_NGG_CULL_FRONT_FACE               (1 << 2)   /* front faces */
279 #define SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST  (1 << 3)   /* GS fast launch: triangles */
280 #define SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP (1 << 4)   /* GS fast launch: triangle strip */
281 #define SI_NGG_CULL_GS_FAST_LAUNCH_ALL       (0x3 << 3) /* GS fast launch (both prim types) */
282 
283 /**
284  * For VS shader keys, describe any fixups required for vertex fetch.
285  *
286  * \ref log_size, \ref format, and the number of channels are interpreted as
287  * by \ref ac_build_opencoded_load_format.
288  *
289  * Note: all bits 0 (size = 1 byte, num channels = 1, format = float) is an
290  * impossible format and indicates that no fixup is needed (just use
291  * buffer_load_format_xyzw).
292  */
293 union si_vs_fix_fetch {
294    struct {
295       uint8_t log_size : 2;        /* 1, 2, 4, 8 or bytes per channel */
296       uint8_t num_channels_m1 : 2; /* number of channels minus 1 */
297       uint8_t format : 3;          /* AC_FETCH_FORMAT_xxx */
298       uint8_t reverse : 1;         /* reverse XYZ channels */
299    } u;
300    uint8_t bits;
301 };
302 
303 struct si_shader;
304 
305 /* State of the context creating the shader object. */
306 struct si_compiler_ctx_state {
307    /* Should only be used by si_init_shader_selector_async and
308     * si_build_shader_variant if thread_index == -1 (non-threaded). */
309    struct ac_llvm_compiler *compiler;
310 
311    /* Used if thread_index == -1 or if debug.async is true. */
312    struct pipe_debug_callback debug;
313 
314    /* Used for creating the log string for gallium/ddebug. */
315    bool is_debug_context;
316 };
317 
318 enum si_color_output_type {
319    SI_TYPE_ANY32,
320    SI_TYPE_FLOAT16,
321    SI_TYPE_INT16,
322    SI_TYPE_UINT16,
323 };
324 
325 struct si_shader_info {
326    shader_info base;
327 
328    gl_shader_stage stage;
329 
330    ubyte num_inputs;
331    ubyte num_outputs;
332    ubyte input_semantic[PIPE_MAX_SHADER_INPUTS];
333    ubyte input_interpolate[PIPE_MAX_SHADER_INPUTS];
334    ubyte input_usage_mask[PIPE_MAX_SHADER_INPUTS];
335    ubyte output_semantic[PIPE_MAX_SHADER_OUTPUTS];
336    char output_semantic_to_slot[VARYING_SLOT_TESS_MAX];
337    ubyte output_usagemask[PIPE_MAX_SHADER_OUTPUTS];
338    ubyte output_readmask[PIPE_MAX_SHADER_OUTPUTS];
339    ubyte output_streams[PIPE_MAX_SHADER_OUTPUTS];
340    ubyte output_type[PIPE_MAX_SHADER_OUTPUTS]; /* enum nir_alu_type */
341 
342    ubyte color_interpolate[2];
343    ubyte color_interpolate_loc[2];
344 
345    int constbuf0_num_slots;
346    ubyte num_stream_output_components[4];
347 
348    uint num_memory_stores;
349 
350    ubyte colors_read; /**< which color components are read by the FS */
351    ubyte colors_written;
352    uint16_t output_color_types; /**< Each bit pair is enum si_color_output_type */
353    bool color0_writes_all_cbufs; /**< gl_FragColor */
354    bool reads_samplemask;   /**< does fragment shader read sample mask? */
355    bool reads_tess_factors; /**< If TES reads TESSINNER or TESSOUTER */
356    bool writes_z;           /**< does fragment shader write Z value? */
357    bool writes_stencil;     /**< does fragment shader write stencil value? */
358    bool writes_samplemask;  /**< does fragment shader write sample mask? */
359    bool writes_edgeflag;    /**< vertex shader outputs edgeflag */
360    bool uses_persp_center;
361    bool uses_persp_centroid;
362    bool uses_persp_sample;
363    bool uses_linear_center;
364    bool uses_linear_centroid;
365    bool uses_linear_sample;
366    bool uses_interp_at_sample;
367    bool uses_instanceid;
368    bool uses_drawid;
369    bool uses_primid;
370    bool uses_frontface;
371    bool uses_invocationid;
372    bool uses_thread_id[3];
373    bool uses_block_id[3];
374    bool uses_variable_block_size;
375    bool uses_grid_size;
376    bool uses_subgroup_info;
377    bool writes_position;
378    bool writes_psize;
379    bool writes_clipvertex;
380    bool writes_primid;
381    bool writes_viewport_index;
382    bool writes_layer;
383    bool uses_bindless_samplers;
384    bool uses_bindless_images;
385 
386    /** Whether all codepaths write tess factors in all invocations. */
387    bool tessfactors_are_def_in_all_invocs;
388 };
389 
390 /* A shader selector is a gallium CSO and contains shader variants and
391  * binaries for one NIR program. This can be shared by multiple contexts.
392  */
393 struct si_shader_selector {
394    struct util_live_shader base;
395    struct si_screen *screen;
396    struct util_queue_fence ready;
397    struct si_compiler_ctx_state compiler_ctx_state;
398 
399    simple_mtx_t mutex;
400    struct si_shader *first_variant; /* immutable after the first variant */
401    struct si_shader *last_variant;  /* mutable */
402 
403    /* The compiled NIR shader without a prolog and/or epilog (not
404     * uploaded to a buffer object).
405     */
406    struct si_shader *main_shader_part;
407    struct si_shader *main_shader_part_ls;     /* as_ls is set in the key */
408    struct si_shader *main_shader_part_es;     /* as_es is set in the key */
409    struct si_shader *main_shader_part_ngg;    /* as_ngg is set in the key */
410    struct si_shader *main_shader_part_ngg_es; /* for Wave32 TES before legacy GS */
411 
412    struct si_shader *gs_copy_shader;
413 
414    struct nir_shader *nir;
415    void *nir_binary;
416    unsigned nir_size;
417 
418    struct pipe_stream_output_info so;
419    struct si_shader_info info;
420 
421    enum pipe_shader_type pipe_shader_type;
422    ubyte const_and_shader_buf_descriptors_index;
423    ubyte sampler_and_images_descriptors_index;
424    bool vs_needs_prolog;
425    bool prim_discard_cs_allowed;
426    ubyte cs_shaderbufs_sgpr_index;
427    ubyte cs_num_shaderbufs_in_user_sgprs;
428    ubyte cs_images_sgpr_index;
429    ubyte cs_images_num_sgprs;
430    ubyte cs_num_images_in_user_sgprs;
431    ubyte num_vs_inputs;
432    ubyte num_vbos_in_user_sgprs;
433    unsigned pa_cl_vs_out_cntl;
434    unsigned ngg_cull_vert_threshold; /* 0 = disabled */
435    unsigned ngg_cull_nonindexed_fast_launch_vert_threshold; /* 0 = disabled */
436    ubyte clipdist_mask;
437    ubyte culldist_mask;
438    ubyte rast_prim;
439 
440    /* ES parameters. */
441    uint16_t esgs_itemsize; /* vertex stride */
442    uint16_t lshs_vertex_stride;
443 
444    /* GS parameters. */
445    uint16_t gsvs_vertex_size;
446    ubyte gs_input_verts_per_prim;
447    unsigned max_gsvs_emit_size;
448    uint16_t enabled_streamout_buffer_mask;
449    bool tess_turns_off_ngg;
450 
451    /* PS parameters. */
452    ubyte color_attr_index[2];
453    unsigned db_shader_control;
454    /* Set 0xf or 0x0 (4 bits) per each written output.
455     * ANDed with spi_shader_col_format.
456     */
457    unsigned colors_written_4bit;
458 
459    uint64_t outputs_written_before_ps; /* "get_unique_index" bits */
460    uint64_t outputs_written;           /* "get_unique_index" bits */
461    uint32_t patch_outputs_written;     /* "get_unique_index_patch" bits */
462 
463    uint64_t inputs_read; /* "get_unique_index" bits */
464 
465    /* bitmasks of used descriptor slots */
466    uint64_t active_const_and_shader_buffers;
467    uint64_t active_samplers_and_images;
468 };
469 
470 /* Valid shader configurations:
471  *
472  * API shaders           VS | TCS | TES | GS |pass| PS
473  * are compiled as:         |     |     |    |thru|
474  *                          |     |     |    |    |
475  * Only VS & PS:         VS |     |     |    |    | PS
476  * GFX6     - with GS:   ES |     |     | GS | VS | PS
477  *          - with tess: LS | HS  | VS  |    |    | PS
478  *          - with both: LS | HS  | ES  | GS | VS | PS
479  * GFX9     - with GS:   -> |     |     | GS | VS | PS
480  *          - with tess: -> | HS  | VS  |    |    | PS
481  *          - with both: -> | HS  | ->  | GS | VS | PS
482  *                          |     |     |    |    |
483  * NGG      - VS & PS:   GS |     |     |    |    | PS
484  * (GFX10+) - with GS:   -> |     |     | GS |    | PS
485  *          - with tess: -> | HS  | GS  |    |    | PS
486  *          - with both: -> | HS  | ->  | GS |    | PS
487  *
488  * -> = merged with the next stage
489  */
490 
491 /* Use the byte alignment for all following structure members for optimal
492  * shader key memory footprint.
493  */
494 #pragma pack(push, 1)
495 
496 /* Common VS bits between the shader key and the prolog key. */
497 struct si_vs_prolog_bits {
498    /* - If neither "is_one" nor "is_fetched" has a bit set, the instance
499     *   divisor is 0.
500     * - If "is_one" has a bit set, the instance divisor is 1.
501     * - If "is_fetched" has a bit set, the instance divisor will be loaded
502     *   from the constant buffer.
503     */
504    uint16_t instance_divisor_is_one;     /* bitmask of inputs */
505    uint16_t instance_divisor_is_fetched; /* bitmask of inputs */
506    unsigned ls_vgpr_fix : 1;
507    unsigned unpack_instance_id_from_vertex_id : 1;
508 };
509 
510 /* Common TCS bits between the shader key and the epilog key. */
511 struct si_tcs_epilog_bits {
512    unsigned prim_mode : 3;
513    unsigned invoc0_tess_factors_are_def : 1;
514    unsigned tes_reads_tess_factors : 1;
515 };
516 
517 struct si_gs_prolog_bits {
518    unsigned tri_strip_adj_fix : 1;
519    unsigned gfx9_prev_is_vs : 1;
520 };
521 
522 /* Common PS bits between the shader key and the prolog key. */
523 struct si_ps_prolog_bits {
524    unsigned color_two_side : 1;
525    unsigned flatshade_colors : 1;
526    unsigned poly_stipple : 1;
527    unsigned force_persp_sample_interp : 1;
528    unsigned force_linear_sample_interp : 1;
529    unsigned force_persp_center_interp : 1;
530    unsigned force_linear_center_interp : 1;
531    unsigned bc_optimize_for_persp : 1;
532    unsigned bc_optimize_for_linear : 1;
533    unsigned samplemask_log_ps_iter : 3;
534 };
535 
536 /* Common PS bits between the shader key and the epilog key. */
537 struct si_ps_epilog_bits {
538    unsigned spi_shader_col_format;
539    unsigned color_is_int8 : 8;
540    unsigned color_is_int10 : 8;
541    unsigned last_cbuf : 3;
542    unsigned alpha_func : 3;
543    unsigned alpha_to_one : 1;
544    unsigned poly_line_smoothing : 1;
545    unsigned clamp_color : 1;
546 };
547 
548 union si_shader_part_key {
549    struct {
550       struct si_vs_prolog_bits states;
551       unsigned num_input_sgprs : 6;
552       /* For merged stages such as LS-HS, HS input VGPRs are first. */
553       unsigned num_merged_next_stage_vgprs : 3;
554       unsigned num_inputs : 5;
555       unsigned as_ls : 1;
556       unsigned as_es : 1;
557       unsigned as_ngg : 1;
558       unsigned as_prim_discard_cs : 1;
559       unsigned gs_fast_launch_tri_list : 1;  /* for NGG culling */
560       unsigned gs_fast_launch_tri_strip : 1; /* for NGG culling */
561       /* Prologs for monolithic shaders shouldn't set EXEC. */
562       unsigned is_monolithic : 1;
563    } vs_prolog;
564    struct {
565       struct si_tcs_epilog_bits states;
566    } tcs_epilog;
567    struct {
568       struct si_gs_prolog_bits states;
569       /* Prologs of monolithic shaders shouldn't set EXEC. */
570       unsigned is_monolithic : 1;
571       unsigned as_ngg : 1;
572    } gs_prolog;
573    struct {
574       struct si_ps_prolog_bits states;
575       unsigned num_input_sgprs : 6;
576       unsigned num_input_vgprs : 5;
577       /* Color interpolation and two-side color selection. */
578       unsigned colors_read : 8;       /* color input components read */
579       unsigned num_interp_inputs : 5; /* BCOLOR is at this location */
580       unsigned face_vgpr_index : 5;
581       unsigned ancillary_vgpr_index : 5;
582       unsigned wqm : 1;
583       char color_attr_index[2];
584       signed char color_interp_vgpr_index[2]; /* -1 == constant */
585    } ps_prolog;
586    struct {
587       struct si_ps_epilog_bits states;
588       unsigned colors_written : 8;
589       unsigned color_types : 16;
590       unsigned writes_z : 1;
591       unsigned writes_stencil : 1;
592       unsigned writes_samplemask : 1;
593    } ps_epilog;
594 };
595 
596 struct si_shader_key {
597    /* Prolog and epilog flags. */
598    union {
599       struct {
600          struct si_vs_prolog_bits prolog;
601       } vs;
602       struct {
603          struct si_vs_prolog_bits ls_prolog; /* for merged LS-HS */
604          struct si_shader_selector *ls;      /* for merged LS-HS */
605          struct si_tcs_epilog_bits epilog;
606       } tcs; /* tessellation control shader */
607       struct {
608          struct si_vs_prolog_bits vs_prolog; /* for merged ES-GS */
609          struct si_shader_selector *es;      /* for merged ES-GS */
610          struct si_gs_prolog_bits prolog;
611       } gs;
612       struct {
613          struct si_ps_prolog_bits prolog;
614          struct si_ps_epilog_bits epilog;
615       } ps;
616    } part;
617 
618    /* These three are initially set according to the NEXT_SHADER property,
619     * or guessed if the property doesn't seem correct.
620     */
621    unsigned as_es : 1;  /* export shader, which precedes GS */
622    unsigned as_ls : 1;  /* local shader, which precedes TCS */
623    unsigned as_ngg : 1; /* VS, TES, or GS compiled as NGG primitive shader */
624 
625    /* Flags for monolithic compilation only. */
626    struct {
627       /* Whether fetch should be opencoded according to vs_fix_fetch.
628        * Otherwise, if vs_fix_fetch is non-zero, buffer_load_format_xyzw
629        * with minimal fixups is used. */
630       uint16_t vs_fetch_opencode;
631       union si_vs_fix_fetch vs_fix_fetch[SI_MAX_ATTRIBS];
632 
633       union {
634          uint64_t ff_tcs_inputs_to_copy; /* for fixed-func TCS */
635          /* When PS needs PrimID and GS is disabled. */
636          unsigned vs_export_prim_id : 1;
637          struct {
638             unsigned interpolate_at_sample_force_center : 1;
639             unsigned fbfetch_msaa : 1;
640             unsigned fbfetch_is_1D : 1;
641             unsigned fbfetch_layered : 1;
642          } ps;
643       } u;
644    } mono;
645 
646    /* Optimization flags for asynchronous compilation only. */
647    struct {
648       /* For HW VS (it can be VS, TES, GS) */
649       uint64_t kill_outputs; /* "get_unique_index" bits */
650       unsigned kill_clip_distances : 8;
651       unsigned kill_pointsize : 1;
652 
653       /* For NGG VS and TES. */
654       unsigned ngg_culling : 5; /* SI_NGG_CULL_* */
655 
656       /* For shaders where monolithic variants have better code.
657        *
658        * This is a flag that has no effect on code generation,
659        * but forces monolithic shaders to be used as soon as
660        * possible, because it's in the "opt" group.
661        */
662       unsigned prefer_mono : 1;
663 
664       /* Primitive discard compute shader. */
665       unsigned vs_as_prim_discard_cs : 1;
666       unsigned cs_prim_type : 4;
667       unsigned cs_indexed : 1;
668       unsigned cs_instancing : 1;
669       unsigned cs_primitive_restart : 1;
670       unsigned cs_provoking_vertex_first : 1;
671       unsigned cs_need_correct_orientation : 1;
672       unsigned cs_cull_front : 1;
673       unsigned cs_cull_back : 1;
674       unsigned cs_cull_z : 1;
675       unsigned cs_halfz_clip_space : 1;
676       unsigned inline_uniforms:1;
677 
678       uint32_t inlined_uniform_values[MAX_INLINABLE_UNIFORMS];
679    } opt;
680 };
681 
682 /* Restore the pack alignment to default. */
683 #pragma pack(pop)
684 
685 /* GCN-specific shader info. */
686 struct si_shader_binary_info {
687    ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
688    ubyte num_input_sgprs;
689    ubyte num_input_vgprs;
690    signed char face_vgpr_index;
691    signed char ancillary_vgpr_index;
692    bool uses_instanceid;
693    ubyte nr_pos_exports;
694    ubyte nr_param_exports;
695    unsigned private_mem_vgprs;
696    unsigned max_simd_waves;
697 };
698 
699 struct si_shader_binary {
700    const char *elf_buffer;
701    size_t elf_size;
702 
703    char *llvm_ir_string;
704 };
705 
706 struct gfx9_gs_info {
707    unsigned es_verts_per_subgroup;
708    unsigned gs_prims_per_subgroup;
709    unsigned gs_inst_prims_in_subgroup;
710    unsigned max_prims_per_subgroup;
711    unsigned esgs_ring_size; /* in bytes */
712 };
713 
714 struct si_shader {
715    struct si_compiler_ctx_state compiler_ctx_state;
716 
717    struct si_shader_selector *selector;
718    struct si_shader_selector *previous_stage_sel; /* for refcounting */
719    struct si_shader *next_variant;
720 
721    struct si_shader_part *prolog;
722    struct si_shader *previous_stage; /* for GFX9 */
723    struct si_shader_part *prolog2;
724    struct si_shader_part *epilog;
725 
726    struct si_pm4_state *pm4;
727    struct si_resource *bo;
728    struct si_resource *scratch_bo;
729    struct si_shader_key key;
730    struct util_queue_fence ready;
731    bool compilation_failed;
732    bool is_monolithic;
733    bool is_optimized;
734    bool is_binary_shared;
735    bool is_gs_copy_shader;
736 
737    /* The following data is all that's needed for binary shaders. */
738    struct si_shader_binary binary;
739    struct ac_shader_config config;
740    struct si_shader_binary_info info;
741 
742    struct {
743       uint16_t ngg_emit_size; /* in dwords */
744       uint16_t hw_max_esverts;
745       uint16_t max_gsprims;
746       uint16_t max_out_verts;
747       uint16_t prim_amp_factor;
748       bool max_vert_out_per_gs_instance;
749    } ngg;
750 
751    /* Shader key + LLVM IR + disassembly + statistics.
752     * Generated for debug contexts only.
753     */
754    char *shader_log;
755    size_t shader_log_size;
756 
757    struct gfx9_gs_info gs_info;
758 
759    /* For save precompute context registers values. */
760    union {
761       struct {
762          unsigned vgt_gsvs_ring_offset_1;
763          unsigned vgt_gsvs_ring_offset_2;
764          unsigned vgt_gsvs_ring_offset_3;
765          unsigned vgt_gsvs_ring_itemsize;
766          unsigned vgt_gs_max_vert_out;
767          unsigned vgt_gs_vert_itemsize;
768          unsigned vgt_gs_vert_itemsize_1;
769          unsigned vgt_gs_vert_itemsize_2;
770          unsigned vgt_gs_vert_itemsize_3;
771          unsigned vgt_gs_instance_cnt;
772          unsigned vgt_gs_onchip_cntl;
773          unsigned vgt_gs_max_prims_per_subgroup;
774          unsigned vgt_esgs_ring_itemsize;
775       } gs;
776 
777       struct {
778          unsigned ge_max_output_per_subgroup;
779          unsigned ge_ngg_subgrp_cntl;
780          unsigned vgt_primitiveid_en;
781          unsigned vgt_gs_onchip_cntl;
782          unsigned vgt_gs_instance_cnt;
783          unsigned vgt_esgs_ring_itemsize;
784          unsigned spi_vs_out_config;
785          unsigned spi_shader_idx_format;
786          unsigned spi_shader_pos_format;
787          unsigned pa_cl_vte_cntl;
788          unsigned pa_cl_ngg_cntl;
789          unsigned vgt_gs_max_vert_out; /* for API GS */
790          unsigned ge_pc_alloc;         /* uconfig register */
791       } ngg;
792 
793       struct {
794          unsigned vgt_gs_mode;
795          unsigned vgt_primitiveid_en;
796          unsigned vgt_reuse_off;
797          unsigned spi_vs_out_config;
798          unsigned spi_shader_pos_format;
799          unsigned pa_cl_vte_cntl;
800          unsigned ge_pc_alloc; /* uconfig register */
801       } vs;
802 
803       struct {
804          unsigned spi_ps_input_ena;
805          unsigned spi_ps_input_addr;
806          unsigned spi_baryc_cntl;
807          unsigned spi_ps_in_control;
808          unsigned spi_shader_z_format;
809          unsigned spi_shader_col_format;
810          unsigned cb_shader_mask;
811       } ps;
812    } ctx_reg;
813 
814    /*For save precompute registers value */
815    unsigned vgt_tf_param;                /* VGT_TF_PARAM */
816    unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
817    unsigned pa_cl_vs_out_cntl;
818    unsigned ge_cntl;
819 };
820 
821 struct si_shader_part {
822    struct si_shader_part *next;
823    union si_shader_part_key key;
824    struct si_shader_binary binary;
825    struct ac_shader_config config;
826 };
827 
828 /* si_shader.c */
829 bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
830                        struct si_shader *shader, struct pipe_debug_callback *debug);
831 bool si_create_shader_variant(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
832                               struct si_shader *shader, struct pipe_debug_callback *debug);
833 void si_shader_destroy(struct si_shader *shader);
834 unsigned si_shader_io_get_unique_index_patch(unsigned semantic);
835 unsigned si_shader_io_get_unique_index(unsigned semantic, bool is_varying);
836 bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
837                              uint64_t scratch_va);
838 void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
839                     struct pipe_debug_callback *debug, FILE *f, bool check_debug_option);
840 void si_shader_dump_stats_for_shader_db(struct si_screen *screen, struct si_shader *shader,
841                                         struct pipe_debug_callback *debug);
842 void si_multiwave_lds_size_workaround(struct si_screen *sscreen, unsigned *lds_size);
843 const char *si_get_shader_name(const struct si_shader *shader);
844 void si_shader_binary_clean(struct si_shader_binary *binary);
845 
846 /* si_shader_llvm_gs.c */
847 struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
848                                              struct ac_llvm_compiler *compiler,
849                                              struct si_shader_selector *gs_selector,
850                                              struct pipe_debug_callback *debug);
851 
852 /* si_shader_nir.c */
853 void si_nir_scan_shader(const struct nir_shader *nir, struct si_shader_info *info);
854 void si_nir_opts(struct si_screen *sscreen, struct nir_shader *nir, bool first);
855 void si_finalize_nir(struct pipe_screen *screen, void *nirptr, bool optimize);
856 
857 /* si_state_shaders.c */
858 void gfx9_get_gs_info(struct si_shader_selector *es, struct si_shader_selector *gs,
859                       struct gfx9_gs_info *out);
860 
861 /* Inline helpers. */
862 
863 /* Return the pointer to the main shader part's pointer. */
si_get_main_shader_part(struct si_shader_selector * sel,struct si_shader_key * key)864 static inline struct si_shader **si_get_main_shader_part(struct si_shader_selector *sel,
865                                                          struct si_shader_key *key)
866 {
867    if (key->as_ls)
868       return &sel->main_shader_part_ls;
869    if (key->as_es && key->as_ngg)
870       return &sel->main_shader_part_ngg_es;
871    if (key->as_es)
872       return &sel->main_shader_part_es;
873    if (key->as_ngg)
874       return &sel->main_shader_part_ngg;
875    return &sel->main_shader_part;
876 }
877 
gfx10_is_ngg_passthrough(struct si_shader * shader)878 static inline bool gfx10_is_ngg_passthrough(struct si_shader *shader)
879 {
880    struct si_shader_selector *sel = shader->selector;
881 
882    return sel->info.stage != MESA_SHADER_GEOMETRY && !sel->so.num_outputs && !sel->info.writes_edgeflag &&
883           !shader->key.opt.ngg_culling &&
884           (sel->info.stage != MESA_SHADER_VERTEX || !shader->key.mono.u.vs_export_prim_id);
885 }
886 
si_shader_uses_bindless_samplers(struct si_shader_selector * selector)887 static inline bool si_shader_uses_bindless_samplers(struct si_shader_selector *selector)
888 {
889    return selector ? selector->info.uses_bindless_samplers : false;
890 }
891 
si_shader_uses_bindless_images(struct si_shader_selector * selector)892 static inline bool si_shader_uses_bindless_images(struct si_shader_selector *selector)
893 {
894    return selector ? selector->info.uses_bindless_images : false;
895 }
896 
897 #endif
898