vc4_qir.h revision 13ddd48b97474c261ef2d7412629748d6d91f2ad 
1/*
2 * Copyright © 2014 Broadcom
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24#ifndef VC4_QIR_H
25#define VC4_QIR_H
26
27#include <assert.h>
28#include <stdio.h>
29#include <stdlib.h>
30#include <stdbool.h>
31#include <stdint.h>
32#include <string.h>
33
34#include "util/macros.h"
35#include "glsl/nir/nir.h"
36#include "util/list.h"
37#include "util/u_math.h"
38
39#include "vc4_screen.h"
40#include "pipe/p_state.h"
41
42enum qfile {
43        QFILE_NULL,
44        QFILE_TEMP,
45        QFILE_VARY,
46        QFILE_UNIF,
47        QFILE_VPM,
48
49        /**
50         * Stores an immediate value in the index field that can be turned
51         * into a small immediate field by qpu_encode_small_immediate().
52         */
53        QFILE_SMALL_IMM,
54};
55
56struct qreg {
57        enum qfile file;
58        uint32_t index;
59};
60
61enum qop {
62        QOP_UNDEF,
63        QOP_MOV,
64        QOP_FADD,
65        QOP_FSUB,
66        QOP_FMUL,
67        QOP_MUL24,
68        QOP_FMIN,
69        QOP_FMAX,
70        QOP_FMINABS,
71        QOP_FMAXABS,
72        QOP_ADD,
73        QOP_SUB,
74        QOP_SHL,
75        QOP_SHR,
76        QOP_ASR,
77        QOP_MIN,
78        QOP_MAX,
79        QOP_AND,
80        QOP_OR,
81        QOP_XOR,
82        QOP_NOT,
83
84        /* Note: Orderings of these compares must be the same as in
85         * qpu_defines.h.  Selects the src[0] if the ns flag bit is set,
86         * otherwise 0. */
87        QOP_SEL_X_0_ZS,
88        QOP_SEL_X_0_ZC,
89        QOP_SEL_X_0_NS,
90        QOP_SEL_X_0_NC,
91        /* Selects the src[0] if the ns flag bit is set, otherwise src[1]. */
92        QOP_SEL_X_Y_ZS,
93        QOP_SEL_X_Y_ZC,
94        QOP_SEL_X_Y_NS,
95        QOP_SEL_X_Y_NC,
96
97        QOP_FTOI,
98        QOP_ITOF,
99        QOP_RCP,
100        QOP_RSQ,
101        QOP_EXP2,
102        QOP_LOG2,
103        QOP_VW_SETUP,
104        QOP_VR_SETUP,
105        QOP_PACK_SCALED,
106        QOP_PACK_8888_F,
107        QOP_PACK_8A_F,
108        QOP_PACK_8B_F,
109        QOP_PACK_8C_F,
110        QOP_PACK_8D_F,
111        QOP_TLB_DISCARD_SETUP,
112        QOP_TLB_STENCIL_SETUP,
113        QOP_TLB_Z_WRITE,
114        QOP_TLB_COLOR_WRITE,
115        QOP_TLB_COLOR_READ,
116        QOP_VARY_ADD_C,
117
118        QOP_FRAG_X,
119        QOP_FRAG_Y,
120        QOP_FRAG_Z,
121        QOP_FRAG_W,
122        QOP_FRAG_REV_FLAG,
123
124        QOP_UNPACK_8A_F,
125        QOP_UNPACK_8B_F,
126        QOP_UNPACK_8C_F,
127        QOP_UNPACK_8D_F,
128        QOP_UNPACK_16A_F,
129        QOP_UNPACK_16B_F,
130
131        QOP_UNPACK_8A_I,
132        QOP_UNPACK_8B_I,
133        QOP_UNPACK_8C_I,
134        QOP_UNPACK_8D_I,
135        QOP_UNPACK_16A_I,
136        QOP_UNPACK_16B_I,
137
138        /** Texture x coordinate parameter write */
139        QOP_TEX_S,
140        /** Texture y coordinate parameter write */
141        QOP_TEX_T,
142        /** Texture border color parameter or cube map z coordinate write */
143        QOP_TEX_R,
144        /** Texture LOD bias parameter write */
145        QOP_TEX_B,
146
147        /**
148         * Texture-unit 4-byte read with address provided direct in S
149         * cooordinate.
150         *
151         * The first operand is the offset from the start of the UBO, and the
152         * second is the uniform that has the UBO's base pointer.
153         */
154        QOP_TEX_DIRECT,
155
156        /**
157         * Signal of texture read being necessary and then reading r4 into
158         * the destination
159         */
160        QOP_TEX_RESULT,
161        QOP_R4_UNPACK_A,
162        QOP_R4_UNPACK_B,
163        QOP_R4_UNPACK_C,
164        QOP_R4_UNPACK_D
165};
166
167struct queued_qpu_inst {
168        struct list_head link;
169        uint64_t inst;
170};
171
172struct qinst {
173        struct list_head link;
174
175        enum qop op;
176        struct qreg dst;
177        struct qreg *src;
178        bool sf;
179};
180
181enum qstage {
182        /**
183         * Coordinate shader, runs during binning, before the VS, and just
184         * outputs position.
185         */
186        QSTAGE_COORD,
187        QSTAGE_VERT,
188        QSTAGE_FRAG,
189};
190
191enum quniform_contents {
192        /**
193         * Indicates that a constant 32-bit value is copied from the program's
194         * uniform contents.
195         */
196        QUNIFORM_CONSTANT,
197        /**
198         * Indicates that the program's uniform contents are used as an index
199         * into the GL uniform storage.
200         */
201        QUNIFORM_UNIFORM,
202
203        /** @{
204         * Scaling factors from clip coordinates to relative to the viewport
205         * center.
206         *
207         * This is used by the coordinate and vertex shaders to produce the
208         * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
209         * point offsets from the viewport ccenter.
210         */
211        QUNIFORM_VIEWPORT_X_SCALE,
212        QUNIFORM_VIEWPORT_Y_SCALE,
213        /** @} */
214
215        QUNIFORM_VIEWPORT_Z_OFFSET,
216        QUNIFORM_VIEWPORT_Z_SCALE,
217
218        QUNIFORM_USER_CLIP_PLANE,
219
220        /**
221         * A reference to a texture config parameter 0 uniform.
222         *
223         * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
224         * defines texture type, miplevels, and such.  It will be found as a
225         * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
226         */
227        QUNIFORM_TEXTURE_CONFIG_P0,
228
229        /**
230         * A reference to a texture config parameter 1 uniform.
231         *
232         * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
233         * defines texture width, height, filters, and wrap modes.  It will be
234         * found as a parameter to the second QOP_TEX_[STRB] instruction in a
235         * sequence.
236         */
237        QUNIFORM_TEXTURE_CONFIG_P1,
238
239        /** A reference to a texture config parameter 2 cubemap stride uniform */
240        QUNIFORM_TEXTURE_CONFIG_P2,
241
242        QUNIFORM_UBO_ADDR,
243
244        QUNIFORM_TEXRECT_SCALE_X,
245        QUNIFORM_TEXRECT_SCALE_Y,
246
247        QUNIFORM_TEXTURE_BORDER_COLOR,
248
249        QUNIFORM_BLEND_CONST_COLOR,
250        QUNIFORM_STENCIL,
251
252        QUNIFORM_ALPHA_REF,
253};
254
255struct vc4_varying_semantic {
256        uint8_t semantic;
257        uint8_t index;
258        uint8_t swizzle;
259};
260
261struct vc4_compiler_ubo_range {
262        /**
263         * offset in bytes from the start of the ubo where this range is
264         * uploaded.
265         *
266         * Only set once used is set.
267         */
268        uint32_t dst_offset;
269
270        /**
271         * offset in bytes from the start of the gallium uniforms where the
272         * data comes from.
273         */
274        uint32_t src_offset;
275
276        /** size in bytes of this ubo range */
277        uint32_t size;
278
279        /**
280         * Set if this range is used by the shader for indirect uniforms
281         * access.
282         */
283        bool used;
284};
285
286struct vc4_key {
287        struct vc4_uncompiled_shader *shader_state;
288        struct {
289                enum pipe_format format;
290                unsigned compare_mode:1;
291                unsigned compare_func:3;
292                unsigned wrap_s:3;
293                unsigned wrap_t:3;
294                uint8_t swizzle[4];
295        } tex[VC4_MAX_TEXTURE_SAMPLERS];
296        uint8_t ucp_enables;
297};
298
299struct vc4_fs_key {
300        struct vc4_key base;
301        enum pipe_format color_format;
302        bool depth_enabled;
303        bool stencil_enabled;
304        bool stencil_twoside;
305        bool stencil_full_writemasks;
306        bool is_points;
307        bool is_lines;
308        bool alpha_test;
309        bool point_coord_upper_left;
310        bool light_twoside;
311        uint8_t alpha_test_func;
312        uint8_t logicop_func;
313        uint32_t point_sprite_mask;
314
315        struct pipe_rt_blend_state blend;
316};
317
318struct vc4_vs_key {
319        struct vc4_key base;
320
321        /**
322         * This is a proxy for the array of FS input semantics, which is
323         * larger than we would want to put in the key.
324         */
325        uint64_t compiled_fs_id;
326
327        enum pipe_format attr_formats[8];
328        bool is_coord;
329        bool per_vertex_point_size;
330};
331
332struct vc4_compile {
333        struct vc4_context *vc4;
334        nir_shader *s;
335        nir_function_impl *impl;
336        struct exec_list *cf_node_list;
337
338        /**
339         * Mapping from nir_register * or nir_ssa_def * to array of struct
340         * qreg for the values.
341         */
342        struct hash_table *def_ht;
343
344        /* For each temp, the instruction generating its value. */
345        struct qinst **defs;
346        uint32_t defs_array_size;
347
348        /**
349         * Inputs to the shader, arranged by TGSI declaration order.
350         *
351         * Not all fragment shader QFILE_VARY reads are present in this array.
352         */
353        struct qreg *inputs;
354        struct qreg *outputs;
355        uint32_t inputs_array_size;
356        uint32_t outputs_array_size;
357        uint32_t uniforms_array_size;
358
359        struct vc4_compiler_ubo_range *ubo_ranges;
360        uint32_t ubo_ranges_array_size;
361        /** Number of uniform areas declared in ubo_ranges. */
362        uint32_t num_uniform_ranges;
363        /** Number of uniform areas used for indirect addressed loads. */
364        uint32_t num_ubo_ranges;
365        uint32_t next_ubo_dst_offset;
366
367        struct qreg line_x, point_x, point_y;
368        struct qreg discard;
369
370        uint8_t vattr_sizes[8];
371
372        /**
373         * Array of the TGSI semantics of all FS QFILE_VARY reads.
374         *
375         * This includes those that aren't part of the VPM varyings, like
376         * point/line coordinates.
377         */
378        struct vc4_varying_semantic *input_semantics;
379        uint32_t num_input_semantics;
380        uint32_t input_semantics_array_size;
381
382        /**
383         * An entry per outputs[] in the VS indicating what the semantic of
384         * the output is.  Used to emit from the VS in the order that the FS
385         * needs.
386         */
387        struct vc4_varying_semantic *output_semantics;
388
389        struct pipe_shader_state *shader_state;
390        struct vc4_key *key;
391        struct vc4_fs_key *fs_key;
392        struct vc4_vs_key *vs_key;
393
394        uint32_t *uniform_data;
395        enum quniform_contents *uniform_contents;
396        uint32_t uniform_array_size;
397        uint32_t num_uniforms;
398        uint32_t num_outputs;
399        uint32_t num_texture_samples;
400        uint32_t output_position_index;
401        uint32_t output_clipvertex_index;
402        uint32_t output_color_index;
403        uint32_t output_point_size_index;
404
405        struct qreg undef;
406        enum qstage stage;
407        uint32_t num_temps;
408        struct list_head instructions;
409        uint32_t immediates[1024];
410
411        struct list_head qpu_inst_list;
412        uint64_t *qpu_insts;
413        uint32_t qpu_inst_count;
414        uint32_t qpu_inst_size;
415        uint32_t num_inputs;
416
417        uint32_t program_id;
418        uint32_t variant_id;
419};
420
421struct vc4_compile *qir_compile_init(void);
422void qir_compile_destroy(struct vc4_compile *c);
423struct qinst *qir_inst(enum qop op, struct qreg dst,
424                       struct qreg src0, struct qreg src1);
425struct qinst *qir_inst4(enum qop op, struct qreg dst,
426                        struct qreg a,
427                        struct qreg b,
428                        struct qreg c,
429                        struct qreg d);
430void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
431struct qreg qir_uniform(struct vc4_compile *c,
432                        enum quniform_contents contents,
433                        uint32_t data);
434void qir_reorder_uniforms(struct vc4_compile *c);
435void qir_emit(struct vc4_compile *c, struct qinst *inst);
436struct qreg qir_get_temp(struct vc4_compile *c);
437int qir_get_op_nsrc(enum qop qop);
438bool qir_reg_equals(struct qreg a, struct qreg b);
439bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
440bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
441bool qir_is_multi_instruction(struct qinst *inst);
442bool qir_is_tex(struct qinst *inst);
443bool qir_depends_on_flags(struct qinst *inst);
444bool qir_writes_r4(struct qinst *inst);
445bool qir_reads_r4(struct qinst *inst);
446bool qir_src_needs_a_file(struct qinst *inst);
447struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
448
449void qir_dump(struct vc4_compile *c);
450void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
451const char *qir_get_stage_name(enum qstage stage);
452
453void qir_optimize(struct vc4_compile *c);
454bool qir_opt_algebraic(struct vc4_compile *c);
455bool qir_opt_constant_folding(struct vc4_compile *c);
456bool qir_opt_copy_propagation(struct vc4_compile *c);
457bool qir_opt_cse(struct vc4_compile *c);
458bool qir_opt_dead_code(struct vc4_compile *c);
459bool qir_opt_small_immediates(struct vc4_compile *c);
460bool qir_opt_vpm_writes(struct vc4_compile *c);
461void vc4_nir_lower_io(struct vc4_compile *c);
462void qir_lower_uniforms(struct vc4_compile *c);
463
464void qpu_schedule_instructions(struct vc4_compile *c);
465
466void qir_SF(struct vc4_compile *c, struct qreg src);
467
468static inline struct qreg
469qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
470{
471        return qir_uniform(c, QUNIFORM_CONSTANT, ui);
472}
473
474static inline struct qreg
475qir_uniform_f(struct vc4_compile *c, float f)
476{
477        return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
478}
479
480#define QIR_ALU0(name)                                                   \
481static inline struct qreg                                                \
482qir_##name(struct vc4_compile *c)                                        \
483{                                                                        \
484        struct qreg t = qir_get_temp(c);                                 \
485        qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
486        return t;                                                        \
487}
488
489#define QIR_ALU1(name)                                                   \
490static inline struct qreg                                                \
491qir_##name(struct vc4_compile *c, struct qreg a)                         \
492{                                                                        \
493        struct qreg t = qir_get_temp(c);                                 \
494        qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
495        return t;                                                        \
496}
497
498#define QIR_ALU2(name)                                                   \
499static inline struct qreg                                                \
500qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
501{                                                                        \
502        struct qreg t = qir_get_temp(c);                                 \
503        qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
504        return t;                                                        \
505}
506
507#define QIR_NODST_1(name)                                               \
508static inline void                                                      \
509qir_##name(struct vc4_compile *c, struct qreg a)                        \
510{                                                                       \
511        qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef));       \
512}
513
514#define QIR_NODST_2(name)                                               \
515static inline void                                                      \
516qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
517{                                                                       \
518        qir_emit(c, qir_inst(QOP_##name, c->undef, a, b));       \
519}
520
521QIR_ALU1(MOV)
522QIR_ALU2(FADD)
523QIR_ALU2(FSUB)
524QIR_ALU2(FMUL)
525QIR_ALU2(MUL24)
526QIR_ALU1(SEL_X_0_ZS)
527QIR_ALU1(SEL_X_0_ZC)
528QIR_ALU1(SEL_X_0_NS)
529QIR_ALU1(SEL_X_0_NC)
530QIR_ALU2(SEL_X_Y_ZS)
531QIR_ALU2(SEL_X_Y_ZC)
532QIR_ALU2(SEL_X_Y_NS)
533QIR_ALU2(SEL_X_Y_NC)
534QIR_ALU2(FMIN)
535QIR_ALU2(FMAX)
536QIR_ALU2(FMINABS)
537QIR_ALU2(FMAXABS)
538QIR_ALU1(FTOI)
539QIR_ALU1(ITOF)
540
541QIR_ALU2(ADD)
542QIR_ALU2(SUB)
543QIR_ALU2(SHL)
544QIR_ALU2(SHR)
545QIR_ALU2(ASR)
546QIR_ALU2(MIN)
547QIR_ALU2(MAX)
548QIR_ALU2(AND)
549QIR_ALU2(OR)
550QIR_ALU2(XOR)
551QIR_ALU1(NOT)
552
553QIR_ALU1(RCP)
554QIR_ALU1(RSQ)
555QIR_ALU1(EXP2)
556QIR_ALU1(LOG2)
557QIR_ALU2(PACK_SCALED)
558QIR_ALU1(PACK_8888_F)
559QIR_ALU2(PACK_8A_F)
560QIR_ALU2(PACK_8B_F)
561QIR_ALU2(PACK_8C_F)
562QIR_ALU2(PACK_8D_F)
563QIR_ALU1(VARY_ADD_C)
564QIR_NODST_2(TEX_S)
565QIR_NODST_2(TEX_T)
566QIR_NODST_2(TEX_R)
567QIR_NODST_2(TEX_B)
568QIR_NODST_2(TEX_DIRECT)
569QIR_ALU0(FRAG_X)
570QIR_ALU0(FRAG_Y)
571QIR_ALU0(FRAG_Z)
572QIR_ALU0(FRAG_W)
573QIR_ALU0(FRAG_REV_FLAG)
574QIR_ALU0(TEX_RESULT)
575QIR_ALU0(TLB_COLOR_READ)
576QIR_NODST_1(TLB_COLOR_WRITE)
577QIR_NODST_1(TLB_Z_WRITE)
578QIR_NODST_1(TLB_DISCARD_SETUP)
579QIR_NODST_1(TLB_STENCIL_SETUP)
580
581static inline struct qreg
582qir_R4_UNPACK(struct vc4_compile *c, struct qreg r4, int i)
583{
584        struct qreg t = qir_get_temp(c);
585        qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, r4, c->undef));
586        return t;
587}
588
589static inline struct qreg
590qir_SEL_X_0_COND(struct vc4_compile *c, int i)
591{
592        struct qreg t = qir_get_temp(c);
593        qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, c->undef, c->undef));
594        return t;
595}
596
597static inline struct qreg
598qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
599{
600        struct qreg t = qir_get_temp(c);
601        qir_emit(c, qir_inst(QOP_UNPACK_8A_F + i, t, src, c->undef));
602        return t;
603}
604
605static inline struct qreg
606qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
607{
608        struct qreg t = qir_get_temp(c);
609        qir_emit(c, qir_inst(QOP_UNPACK_8A_I + i, t, src, c->undef));
610        return t;
611}
612
613static inline struct qreg
614qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
615{
616        struct qreg t = qir_get_temp(c);
617        qir_emit(c, qir_inst(QOP_UNPACK_16A_F + i, t, src, c->undef));
618        return t;
619}
620
621static inline struct qreg
622qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
623{
624        struct qreg t = qir_get_temp(c);
625        qir_emit(c, qir_inst(QOP_UNPACK_16A_I + i, t, src, c->undef));
626        return t;
627}
628
629static inline struct qreg
630qir_PACK_8_F(struct vc4_compile *c, struct qreg rest, struct qreg val, int chan)
631{
632        struct qreg t = qir_get_temp(c);
633        qir_emit(c, qir_inst(QOP_PACK_8A_F + chan, t, rest, val));
634        return t;
635}
636
637static inline struct qreg
638qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
639{
640        return qir_EXP2(c, qir_FMUL(c,
641                                    y,
642                                    qir_LOG2(c, x)));
643}
644
645static inline void
646qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
647{
648        static const struct qreg vpm = { QFILE_VPM, 0 };
649        qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
650}
651
652#endif /* VC4_QIR_H */
653