2 * Copyright (c) 2014 Scott Mansell
3 * Copyright © 2014 Broadcom
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
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
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 NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 #include "util/u_format.h"
27 #include "util/u_hash.h"
28 #include "util/u_math.h"
29 #include "util/u_memory.h"
30 #include "util/ralloc.h"
31 #include "util/hash_table.h"
32 #include "tgsi/tgsi_dump.h"
33 #include "tgsi/tgsi_parse.h"
34 #include "compiler/nir/nir.h"
35 #include "compiler/nir/nir_builder.h"
36 #include "nir/tgsi_to_nir.h"
37 #include "vc4_context.h"
40 #ifdef USE_VC4_SIMULATOR
41 #include "simpenrose/simpenrose.h"
45 ntq_get_src(struct vc4_compile
*c
, nir_src src
, int i
);
48 resize_qreg_array(struct vc4_compile
*c
,
53 if (*size
>= decl_size
)
56 uint32_t old_size
= *size
;
57 *size
= MAX2(*size
* 2, decl_size
);
58 *regs
= reralloc(c
, *regs
, struct qreg
, *size
);
60 fprintf(stderr
, "Malloc failure\n");
64 for (uint32_t i
= old_size
; i
< *size
; i
++)
65 (*regs
)[i
] = c
->undef
;
69 indirect_uniform_load(struct vc4_compile
*c
, nir_intrinsic_instr
*intr
)
71 struct qreg indirect_offset
= ntq_get_src(c
, intr
->src
[0], 0);
72 uint32_t offset
= intr
->const_index
[0];
73 struct vc4_compiler_ubo_range
*range
= NULL
;
75 for (i
= 0; i
< c
->num_uniform_ranges
; i
++) {
76 range
= &c
->ubo_ranges
[i
];
77 if (offset
>= range
->src_offset
&&
78 offset
< range
->src_offset
+ range
->size
) {
82 /* The driver-location-based offset always has to be within a declared
88 range
->dst_offset
= c
->next_ubo_dst_offset
;
89 c
->next_ubo_dst_offset
+= range
->size
;
93 offset
-= range
->src_offset
;
95 /* Adjust for where we stored the TGSI register base. */
96 indirect_offset
= qir_ADD(c
, indirect_offset
,
97 qir_uniform_ui(c
, (range
->dst_offset
+
100 /* Clamp to [0, array size). Note that MIN/MAX are signed. */
101 indirect_offset
= qir_MAX(c
, indirect_offset
, qir_uniform_ui(c
, 0));
102 indirect_offset
= qir_MIN(c
, indirect_offset
,
103 qir_uniform_ui(c
, (range
->dst_offset
+
106 qir_TEX_DIRECT(c
, indirect_offset
, qir_uniform(c
, QUNIFORM_UBO_ADDR
, 0));
107 c
->num_texture_samples
++;
108 return qir_TEX_RESULT(c
);
111 nir_ssa_def
*vc4_nir_get_state_uniform(struct nir_builder
*b
,
112 enum quniform_contents contents
)
114 nir_intrinsic_instr
*intr
=
115 nir_intrinsic_instr_create(b
->shader
,
116 nir_intrinsic_load_uniform
);
117 intr
->const_index
[0] = (VC4_NIR_STATE_UNIFORM_OFFSET
+ contents
) * 4;
118 intr
->num_components
= 1;
119 intr
->src
[0] = nir_src_for_ssa(nir_imm_int(b
, 0));
120 nir_ssa_dest_init(&intr
->instr
, &intr
->dest
, 1, 32, NULL
);
121 nir_builder_instr_insert(b
, &intr
->instr
);
122 return &intr
->dest
.ssa
;
126 vc4_nir_get_swizzled_channel(nir_builder
*b
, nir_ssa_def
**srcs
, int swiz
)
130 case UTIL_FORMAT_SWIZZLE_NONE
:
131 fprintf(stderr
, "warning: unknown swizzle\n");
133 case UTIL_FORMAT_SWIZZLE_0
:
134 return nir_imm_float(b
, 0.0);
135 case UTIL_FORMAT_SWIZZLE_1
:
136 return nir_imm_float(b
, 1.0);
137 case UTIL_FORMAT_SWIZZLE_X
:
138 case UTIL_FORMAT_SWIZZLE_Y
:
139 case UTIL_FORMAT_SWIZZLE_Z
:
140 case UTIL_FORMAT_SWIZZLE_W
:
146 ntq_init_ssa_def(struct vc4_compile
*c
, nir_ssa_def
*def
)
148 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
149 def
->num_components
);
150 _mesa_hash_table_insert(c
->def_ht
, def
, qregs
);
155 ntq_get_dest(struct vc4_compile
*c
, nir_dest
*dest
)
158 struct qreg
*qregs
= ntq_init_ssa_def(c
, &dest
->ssa
);
159 for (int i
= 0; i
< dest
->ssa
.num_components
; i
++)
163 nir_register
*reg
= dest
->reg
.reg
;
164 assert(dest
->reg
.base_offset
== 0);
165 assert(reg
->num_array_elems
== 0);
166 struct hash_entry
*entry
=
167 _mesa_hash_table_search(c
->def_ht
, reg
);
173 ntq_get_src(struct vc4_compile
*c
, nir_src src
, int i
)
175 struct hash_entry
*entry
;
177 entry
= _mesa_hash_table_search(c
->def_ht
, src
.ssa
);
178 assert(i
< src
.ssa
->num_components
);
180 nir_register
*reg
= src
.reg
.reg
;
181 entry
= _mesa_hash_table_search(c
->def_ht
, reg
);
182 assert(reg
->num_array_elems
== 0);
183 assert(src
.reg
.base_offset
== 0);
184 assert(i
< reg
->num_components
);
187 struct qreg
*qregs
= entry
->data
;
192 ntq_get_alu_src(struct vc4_compile
*c
, nir_alu_instr
*instr
,
195 assert(util_is_power_of_two(instr
->dest
.write_mask
));
196 unsigned chan
= ffs(instr
->dest
.write_mask
) - 1;
197 struct qreg r
= ntq_get_src(c
, instr
->src
[src
].src
,
198 instr
->src
[src
].swizzle
[chan
]);
200 assert(!instr
->src
[src
].abs
);
201 assert(!instr
->src
[src
].negate
);
206 static inline struct qreg
207 qir_SAT(struct vc4_compile
*c
, struct qreg val
)
210 qir_FMIN(c
, val
, qir_uniform_f(c
, 1.0)),
211 qir_uniform_f(c
, 0.0));
215 ntq_rcp(struct vc4_compile
*c
, struct qreg x
)
217 struct qreg r
= qir_RCP(c
, x
);
219 /* Apply a Newton-Raphson step to improve the accuracy. */
220 r
= qir_FMUL(c
, r
, qir_FSUB(c
,
221 qir_uniform_f(c
, 2.0),
228 ntq_rsq(struct vc4_compile
*c
, struct qreg x
)
230 struct qreg r
= qir_RSQ(c
, x
);
232 /* Apply a Newton-Raphson step to improve the accuracy. */
233 r
= qir_FMUL(c
, r
, qir_FSUB(c
,
234 qir_uniform_f(c
, 1.5),
236 qir_uniform_f(c
, 0.5),
238 qir_FMUL(c
, r
, r
)))));
244 qir_srgb_decode(struct vc4_compile
*c
, struct qreg srgb
)
246 struct qreg low
= qir_FMUL(c
, srgb
, qir_uniform_f(c
, 1.0 / 12.92));
247 struct qreg high
= qir_POW(c
,
251 qir_uniform_f(c
, 0.055)),
252 qir_uniform_f(c
, 1.0 / 1.055)),
253 qir_uniform_f(c
, 2.4));
255 qir_SF(c
, qir_FSUB(c
, srgb
, qir_uniform_f(c
, 0.04045)));
256 return qir_SEL(c
, QPU_COND_NS
, low
, high
);
260 ntq_umul(struct vc4_compile
*c
, struct qreg src0
, struct qreg src1
)
262 struct qreg src0_hi
= qir_SHR(c
, src0
,
263 qir_uniform_ui(c
, 24));
264 struct qreg src1_hi
= qir_SHR(c
, src1
,
265 qir_uniform_ui(c
, 24));
267 struct qreg hilo
= qir_MUL24(c
, src0_hi
, src1
);
268 struct qreg lohi
= qir_MUL24(c
, src0
, src1_hi
);
269 struct qreg lolo
= qir_MUL24(c
, src0
, src1
);
271 return qir_ADD(c
, lolo
, qir_SHL(c
,
272 qir_ADD(c
, hilo
, lohi
),
273 qir_uniform_ui(c
, 24)));
277 ntq_scale_depth_texture(struct vc4_compile
*c
, struct qreg src
)
279 struct qreg depthf
= qir_ITOF(c
, qir_SHR(c
, src
,
280 qir_uniform_ui(c
, 8)));
281 return qir_FMUL(c
, depthf
, qir_uniform_f(c
, 1.0f
/0xffffff));
285 * Emits a lowered TXF_MS from an MSAA texture.
287 * The addressing math has been lowered in NIR, and now we just need to read
291 ntq_emit_txf(struct vc4_compile
*c
, nir_tex_instr
*instr
)
293 uint32_t tile_width
= 32;
294 uint32_t tile_height
= 32;
295 uint32_t tile_size
= (tile_height
* tile_width
*
296 VC4_MAX_SAMPLES
* sizeof(uint32_t));
298 unsigned unit
= instr
->texture_index
;
299 uint32_t w
= align(c
->key
->tex
[unit
].msaa_width
, tile_width
);
300 uint32_t w_tiles
= w
/ tile_width
;
301 uint32_t h
= align(c
->key
->tex
[unit
].msaa_height
, tile_height
);
302 uint32_t h_tiles
= h
/ tile_height
;
303 uint32_t size
= w_tiles
* h_tiles
* tile_size
;
306 assert(instr
->num_srcs
== 1);
307 assert(instr
->src
[0].src_type
== nir_tex_src_coord
);
308 addr
= ntq_get_src(c
, instr
->src
[0].src
, 0);
310 /* Perform the clamping required by kernel validation. */
311 addr
= qir_MAX(c
, addr
, qir_uniform_ui(c
, 0));
312 addr
= qir_MIN(c
, addr
, qir_uniform_ui(c
, size
- 4));
314 qir_TEX_DIRECT(c
, addr
, qir_uniform(c
, QUNIFORM_TEXTURE_MSAA_ADDR
, unit
));
316 struct qreg tex
= qir_TEX_RESULT(c
);
317 c
->num_texture_samples
++;
319 struct qreg
*dest
= ntq_get_dest(c
, &instr
->dest
);
320 enum pipe_format format
= c
->key
->tex
[unit
].format
;
321 if (util_format_is_depth_or_stencil(format
)) {
322 struct qreg scaled
= ntq_scale_depth_texture(c
, tex
);
323 for (int i
= 0; i
< 4; i
++)
326 for (int i
= 0; i
< 4; i
++)
327 dest
[i
] = qir_UNPACK_8_F(c
, tex
, i
);
330 for (int i
= 0; i
< 4; i
++) {
331 if (c
->tex_srgb_decode
[unit
] & (1 << i
))
332 dest
[i
] = qir_srgb_decode(c
, dest
[i
]);
337 ntq_emit_tex(struct vc4_compile
*c
, nir_tex_instr
*instr
)
339 struct qreg s
, t
, r
, lod
, proj
, compare
;
340 bool is_txb
= false, is_txl
= false, has_proj
= false;
341 unsigned unit
= instr
->texture_index
;
343 if (instr
->op
== nir_texop_txf
) {
344 ntq_emit_txf(c
, instr
);
348 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
349 switch (instr
->src
[i
].src_type
) {
350 case nir_tex_src_coord
:
351 s
= ntq_get_src(c
, instr
->src
[i
].src
, 0);
352 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_1D
)
353 t
= qir_uniform_f(c
, 0.5);
355 t
= ntq_get_src(c
, instr
->src
[i
].src
, 1);
356 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
357 r
= ntq_get_src(c
, instr
->src
[i
].src
, 2);
359 case nir_tex_src_bias
:
360 lod
= ntq_get_src(c
, instr
->src
[i
].src
, 0);
363 case nir_tex_src_lod
:
364 lod
= ntq_get_src(c
, instr
->src
[i
].src
, 0);
367 case nir_tex_src_comparitor
:
368 compare
= ntq_get_src(c
, instr
->src
[i
].src
, 0);
370 case nir_tex_src_projector
:
371 proj
= qir_RCP(c
, ntq_get_src(c
, instr
->src
[i
].src
, 0));
372 s
= qir_FMUL(c
, s
, proj
);
373 t
= qir_FMUL(c
, t
, proj
);
377 unreachable("unknown texture source");
381 struct qreg texture_u
[] = {
382 qir_uniform(c
, QUNIFORM_TEXTURE_CONFIG_P0
, unit
),
383 qir_uniform(c
, QUNIFORM_TEXTURE_CONFIG_P1
, unit
),
384 qir_uniform(c
, QUNIFORM_CONSTANT
, 0),
385 qir_uniform(c
, QUNIFORM_CONSTANT
, 0),
387 uint32_t next_texture_u
= 0;
389 /* There is no native support for GL texture rectangle coordinates, so
390 * we have to rescale from ([0, width], [0, height]) to ([0, 1], [0,
393 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) {
395 qir_uniform(c
, QUNIFORM_TEXRECT_SCALE_X
, unit
));
397 qir_uniform(c
, QUNIFORM_TEXRECT_SCALE_Y
, unit
));
400 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
|| is_txl
) {
401 texture_u
[2] = qir_uniform(c
, QUNIFORM_TEXTURE_CONFIG_P2
,
402 unit
| (is_txl
<< 16));
405 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
406 struct qreg ma
= qir_FMAXABS(c
, qir_FMAXABS(c
, s
, t
), r
);
407 struct qreg rcp_ma
= qir_RCP(c
, ma
);
408 s
= qir_FMUL(c
, s
, rcp_ma
);
409 t
= qir_FMUL(c
, t
, rcp_ma
);
410 r
= qir_FMUL(c
, r
, rcp_ma
);
412 qir_TEX_R(c
, r
, texture_u
[next_texture_u
++]);
413 } else if (c
->key
->tex
[unit
].wrap_s
== PIPE_TEX_WRAP_CLAMP_TO_BORDER
||
414 c
->key
->tex
[unit
].wrap_s
== PIPE_TEX_WRAP_CLAMP
||
415 c
->key
->tex
[unit
].wrap_t
== PIPE_TEX_WRAP_CLAMP_TO_BORDER
||
416 c
->key
->tex
[unit
].wrap_t
== PIPE_TEX_WRAP_CLAMP
) {
417 qir_TEX_R(c
, qir_uniform(c
, QUNIFORM_TEXTURE_BORDER_COLOR
, unit
),
418 texture_u
[next_texture_u
++]);
421 if (c
->key
->tex
[unit
].wrap_s
== PIPE_TEX_WRAP_CLAMP
) {
425 if (c
->key
->tex
[unit
].wrap_t
== PIPE_TEX_WRAP_CLAMP
) {
429 qir_TEX_T(c
, t
, texture_u
[next_texture_u
++]);
431 if (is_txl
|| is_txb
)
432 qir_TEX_B(c
, lod
, texture_u
[next_texture_u
++]);
434 qir_TEX_S(c
, s
, texture_u
[next_texture_u
++]);
436 c
->num_texture_samples
++;
437 struct qreg tex
= qir_TEX_RESULT(c
);
439 enum pipe_format format
= c
->key
->tex
[unit
].format
;
441 struct qreg
*dest
= ntq_get_dest(c
, &instr
->dest
);
442 if (util_format_is_depth_or_stencil(format
)) {
443 struct qreg normalized
= ntq_scale_depth_texture(c
, tex
);
444 struct qreg depth_output
;
446 struct qreg u0
= qir_uniform_f(c
, 0.0f
);
447 struct qreg u1
= qir_uniform_f(c
, 1.0f
);
448 if (c
->key
->tex
[unit
].compare_mode
) {
450 compare
= qir_FMUL(c
, compare
, proj
);
452 switch (c
->key
->tex
[unit
].compare_func
) {
453 case PIPE_FUNC_NEVER
:
454 depth_output
= qir_uniform_f(c
, 0.0f
);
456 case PIPE_FUNC_ALWAYS
:
459 case PIPE_FUNC_EQUAL
:
460 qir_SF(c
, qir_FSUB(c
, compare
, normalized
));
461 depth_output
= qir_SEL(c
, QPU_COND_ZS
, u1
, u0
);
463 case PIPE_FUNC_NOTEQUAL
:
464 qir_SF(c
, qir_FSUB(c
, compare
, normalized
));
465 depth_output
= qir_SEL(c
, QPU_COND_ZC
, u1
, u0
);
467 case PIPE_FUNC_GREATER
:
468 qir_SF(c
, qir_FSUB(c
, compare
, normalized
));
469 depth_output
= qir_SEL(c
, QPU_COND_NC
, u1
, u0
);
471 case PIPE_FUNC_GEQUAL
:
472 qir_SF(c
, qir_FSUB(c
, normalized
, compare
));
473 depth_output
= qir_SEL(c
, QPU_COND_NS
, u1
, u0
);
476 qir_SF(c
, qir_FSUB(c
, compare
, normalized
));
477 depth_output
= qir_SEL(c
, QPU_COND_NS
, u1
, u0
);
479 case PIPE_FUNC_LEQUAL
:
480 qir_SF(c
, qir_FSUB(c
, normalized
, compare
));
481 depth_output
= qir_SEL(c
, QPU_COND_NC
, u1
, u0
);
485 depth_output
= normalized
;
488 for (int i
= 0; i
< 4; i
++)
489 dest
[i
] = depth_output
;
491 for (int i
= 0; i
< 4; i
++)
492 dest
[i
] = qir_UNPACK_8_F(c
, tex
, i
);
495 for (int i
= 0; i
< 4; i
++) {
496 if (c
->tex_srgb_decode
[unit
] & (1 << i
))
497 dest
[i
] = qir_srgb_decode(c
, dest
[i
]);
502 * Computes x - floor(x), which is tricky because our FTOI truncates (rounds
506 ntq_ffract(struct vc4_compile
*c
, struct qreg src
)
508 struct qreg trunc
= qir_ITOF(c
, qir_FTOI(c
, src
));
509 struct qreg diff
= qir_FSUB(c
, src
, trunc
);
511 return qir_SEL(c
, QPU_COND_NS
,
512 qir_FADD(c
, diff
, qir_uniform_f(c
, 1.0)), diff
);
516 * Computes floor(x), which is tricky because our FTOI truncates (rounds to
520 ntq_ffloor(struct vc4_compile
*c
, struct qreg src
)
522 struct qreg trunc
= qir_ITOF(c
, qir_FTOI(c
, src
));
524 /* This will be < 0 if we truncated and the truncation was of a value
525 * that was < 0 in the first place.
527 qir_SF(c
, qir_FSUB(c
, src
, trunc
));
529 return qir_SEL(c
, QPU_COND_NS
,
530 qir_FSUB(c
, trunc
, qir_uniform_f(c
, 1.0)), trunc
);
534 * Computes ceil(x), which is tricky because our FTOI truncates (rounds to
538 ntq_fceil(struct vc4_compile
*c
, struct qreg src
)
540 struct qreg trunc
= qir_ITOF(c
, qir_FTOI(c
, src
));
542 /* This will be < 0 if we truncated and the truncation was of a value
543 * that was > 0 in the first place.
545 qir_SF(c
, qir_FSUB(c
, trunc
, src
));
547 return qir_SEL(c
, QPU_COND_NS
,
548 qir_FADD(c
, trunc
, qir_uniform_f(c
, 1.0)), trunc
);
552 ntq_fsin(struct vc4_compile
*c
, struct qreg src
)
556 pow(2.0 * M_PI
, 3) / (3 * 2 * 1),
557 -pow(2.0 * M_PI
, 5) / (5 * 4 * 3 * 2 * 1),
558 pow(2.0 * M_PI
, 7) / (7 * 6 * 5 * 4 * 3 * 2 * 1),
559 -pow(2.0 * M_PI
, 9) / (9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
562 struct qreg scaled_x
=
565 qir_uniform_f(c
, 1.0 / (M_PI
* 2.0)));
567 struct qreg x
= qir_FADD(c
,
568 ntq_ffract(c
, scaled_x
),
569 qir_uniform_f(c
, -0.5));
570 struct qreg x2
= qir_FMUL(c
, x
, x
);
571 struct qreg sum
= qir_FMUL(c
, x
, qir_uniform_f(c
, coeff
[0]));
572 for (int i
= 1; i
< ARRAY_SIZE(coeff
); i
++) {
573 x
= qir_FMUL(c
, x
, x2
);
578 qir_uniform_f(c
, coeff
[i
])));
584 ntq_fcos(struct vc4_compile
*c
, struct qreg src
)
588 pow(2.0 * M_PI
, 2) / (2 * 1),
589 -pow(2.0 * M_PI
, 4) / (4 * 3 * 2 * 1),
590 pow(2.0 * M_PI
, 6) / (6 * 5 * 4 * 3 * 2 * 1),
591 -pow(2.0 * M_PI
, 8) / (8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
592 pow(2.0 * M_PI
, 10) / (10 * 9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
595 struct qreg scaled_x
=
597 qir_uniform_f(c
, 1.0f
/ (M_PI
* 2.0f
)));
598 struct qreg x_frac
= qir_FADD(c
,
599 ntq_ffract(c
, scaled_x
),
600 qir_uniform_f(c
, -0.5));
602 struct qreg sum
= qir_uniform_f(c
, coeff
[0]);
603 struct qreg x2
= qir_FMUL(c
, x_frac
, x_frac
);
604 struct qreg x
= x2
; /* Current x^2, x^4, or x^6 */
605 for (int i
= 1; i
< ARRAY_SIZE(coeff
); i
++) {
607 x
= qir_FMUL(c
, x
, x2
);
609 struct qreg mul
= qir_FMUL(c
,
611 qir_uniform_f(c
, coeff
[i
]));
615 sum
= qir_FADD(c
, sum
, mul
);
621 ntq_fsign(struct vc4_compile
*c
, struct qreg src
)
623 struct qreg t
= qir_get_temp(c
);
626 qir_MOV_dest(c
, t
, qir_uniform_f(c
, 0.0));
627 qir_MOV_dest(c
, t
, qir_uniform_f(c
, 1.0))->cond
= QPU_COND_ZC
;
628 qir_MOV_dest(c
, t
, qir_uniform_f(c
, -1.0))->cond
= QPU_COND_NS
;
633 emit_vertex_input(struct vc4_compile
*c
, int attr
)
635 enum pipe_format format
= c
->vs_key
->attr_formats
[attr
];
636 uint32_t attr_size
= util_format_get_blocksize(format
);
638 c
->vattr_sizes
[attr
] = align(attr_size
, 4);
639 for (int i
= 0; i
< align(attr_size
, 4) / 4; i
++) {
640 struct qreg vpm
= { QFILE_VPM
, attr
* 4 + i
};
641 c
->inputs
[attr
* 4 + i
] = qir_MOV(c
, vpm
);
647 emit_fragcoord_input(struct vc4_compile
*c
, int attr
)
649 c
->inputs
[attr
* 4 + 0] = qir_FRAG_X(c
);
650 c
->inputs
[attr
* 4 + 1] = qir_FRAG_Y(c
);
651 c
->inputs
[attr
* 4 + 2] =
653 qir_ITOF(c
, qir_FRAG_Z(c
)),
654 qir_uniform_f(c
, 1.0 / 0xffffff));
655 c
->inputs
[attr
* 4 + 3] = qir_RCP(c
, qir_FRAG_W(c
));
659 emit_fragment_varying(struct vc4_compile
*c
, gl_varying_slot slot
,
662 uint32_t i
= c
->num_input_slots
++;
668 if (c
->num_input_slots
>= c
->input_slots_array_size
) {
669 c
->input_slots_array_size
=
670 MAX2(4, c
->input_slots_array_size
* 2);
672 c
->input_slots
= reralloc(c
, c
->input_slots
,
673 struct vc4_varying_slot
,
674 c
->input_slots_array_size
);
677 c
->input_slots
[i
].slot
= slot
;
678 c
->input_slots
[i
].swizzle
= swizzle
;
680 return qir_VARY_ADD_C(c
, qir_FMUL(c
, vary
, qir_FRAG_W(c
)));
684 emit_fragment_input(struct vc4_compile
*c
, int attr
, gl_varying_slot slot
)
686 for (int i
= 0; i
< 4; i
++) {
687 c
->inputs
[attr
* 4 + i
] =
688 emit_fragment_varying(c
, slot
, i
);
694 add_output(struct vc4_compile
*c
,
695 uint32_t decl_offset
,
699 uint32_t old_array_size
= c
->outputs_array_size
;
700 resize_qreg_array(c
, &c
->outputs
, &c
->outputs_array_size
,
703 if (old_array_size
!= c
->outputs_array_size
) {
704 c
->output_slots
= reralloc(c
,
706 struct vc4_varying_slot
,
707 c
->outputs_array_size
);
710 c
->output_slots
[decl_offset
].slot
= slot
;
711 c
->output_slots
[decl_offset
].swizzle
= swizzle
;
715 declare_uniform_range(struct vc4_compile
*c
, uint32_t start
, uint32_t size
)
717 unsigned array_id
= c
->num_uniform_ranges
++;
718 if (array_id
>= c
->ubo_ranges_array_size
) {
719 c
->ubo_ranges_array_size
= MAX2(c
->ubo_ranges_array_size
* 2,
721 c
->ubo_ranges
= reralloc(c
, c
->ubo_ranges
,
722 struct vc4_compiler_ubo_range
,
723 c
->ubo_ranges_array_size
);
726 c
->ubo_ranges
[array_id
].dst_offset
= 0;
727 c
->ubo_ranges
[array_id
].src_offset
= start
;
728 c
->ubo_ranges
[array_id
].size
= size
;
729 c
->ubo_ranges
[array_id
].used
= false;
733 ntq_src_is_only_ssa_def_user(nir_src
*src
)
738 if (!list_empty(&src
->ssa
->if_uses
))
741 return (src
->ssa
->uses
.next
== &src
->use_link
&&
742 src
->ssa
->uses
.next
->next
== &src
->ssa
->uses
);
746 * In general, emits a nir_pack_unorm_4x8 as a series of MOVs with the pack
749 * However, as an optimization, it tries to find the instructions generating
750 * the sources to be packed and just emit the pack flag there, if possible.
753 ntq_emit_pack_unorm_4x8(struct vc4_compile
*c
, nir_alu_instr
*instr
)
755 struct qreg result
= qir_get_temp(c
);
756 struct nir_alu_instr
*vec4
= NULL
;
758 /* If packing from a vec4 op (as expected), identify it so that we can
759 * peek back at what generated its sources.
761 if (instr
->src
[0].src
.is_ssa
&&
762 instr
->src
[0].src
.ssa
->parent_instr
->type
== nir_instr_type_alu
&&
763 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
)->op
==
765 vec4
= nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
768 /* If the pack is replicating the same channel 4 times, use the 8888
769 * pack flag. This is common for blending using the alpha
772 if (instr
->src
[0].swizzle
[0] == instr
->src
[0].swizzle
[1] &&
773 instr
->src
[0].swizzle
[0] == instr
->src
[0].swizzle
[2] &&
774 instr
->src
[0].swizzle
[0] == instr
->src
[0].swizzle
[3]) {
775 struct qreg
*dest
= ntq_get_dest(c
, &instr
->dest
.dest
);
776 *dest
= qir_PACK_8888_F(c
,
777 ntq_get_src(c
, instr
->src
[0].src
,
778 instr
->src
[0].swizzle
[0]));
782 for (int i
= 0; i
< 4; i
++) {
783 int swiz
= instr
->src
[0].swizzle
[i
];
786 src
= ntq_get_src(c
, vec4
->src
[swiz
].src
,
787 vec4
->src
[swiz
].swizzle
[0]);
789 src
= ntq_get_src(c
, instr
->src
[0].src
, swiz
);
793 ntq_src_is_only_ssa_def_user(&vec4
->src
[swiz
].src
) &&
794 src
.file
== QFILE_TEMP
&&
795 c
->defs
[src
.index
] &&
796 qir_is_mul(c
->defs
[src
.index
]) &&
797 !c
->defs
[src
.index
]->dst
.pack
) {
798 struct qinst
*rewrite
= c
->defs
[src
.index
];
799 c
->defs
[src
.index
] = NULL
;
800 rewrite
->dst
= result
;
801 rewrite
->dst
.pack
= QPU_PACK_MUL_8A
+ i
;
805 qir_PACK_8_F(c
, result
, src
, i
);
808 struct qreg
*dest
= ntq_get_dest(c
, &instr
->dest
.dest
);
812 /** Handles sign-extended bitfield extracts for 16 bits. */
814 ntq_emit_ibfe(struct vc4_compile
*c
, struct qreg base
, struct qreg offset
,
817 assert(bits
.file
== QFILE_UNIF
&&
818 c
->uniform_contents
[bits
.index
] == QUNIFORM_CONSTANT
&&
819 c
->uniform_data
[bits
.index
] == 16);
821 assert(offset
.file
== QFILE_UNIF
&&
822 c
->uniform_contents
[offset
.index
] == QUNIFORM_CONSTANT
);
823 int offset_bit
= c
->uniform_data
[offset
.index
];
824 assert(offset_bit
% 16 == 0);
826 return qir_UNPACK_16_I(c
, base
, offset_bit
/ 16);
829 /** Handles unsigned bitfield extracts for 8 bits. */
831 ntq_emit_ubfe(struct vc4_compile
*c
, struct qreg base
, struct qreg offset
,
834 assert(bits
.file
== QFILE_UNIF
&&
835 c
->uniform_contents
[bits
.index
] == QUNIFORM_CONSTANT
&&
836 c
->uniform_data
[bits
.index
] == 8);
838 assert(offset
.file
== QFILE_UNIF
&&
839 c
->uniform_contents
[offset
.index
] == QUNIFORM_CONSTANT
);
840 int offset_bit
= c
->uniform_data
[offset
.index
];
841 assert(offset_bit
% 8 == 0);
843 return qir_UNPACK_8_I(c
, base
, offset_bit
/ 8);
847 * If compare_instr is a valid comparison instruction, emits the
848 * compare_instr's comparison and returns the sel_instr's return value based
849 * on the compare_instr's result.
852 ntq_emit_comparison(struct vc4_compile
*c
, struct qreg
*dest
,
853 nir_alu_instr
*compare_instr
,
854 nir_alu_instr
*sel_instr
)
858 switch (compare_instr
->op
) {
884 struct qreg src0
= ntq_get_alu_src(c
, compare_instr
, 0);
885 struct qreg src1
= ntq_get_alu_src(c
, compare_instr
, 1);
887 unsigned unsized_type
=
888 nir_alu_type_get_base_type(nir_op_infos
[compare_instr
->op
].input_types
[0]);
889 if (unsized_type
== nir_type_float
)
890 qir_SF(c
, qir_FSUB(c
, src0
, src1
));
892 qir_SF(c
, qir_SUB(c
, src0
, src1
));
894 switch (sel_instr
->op
) {
899 *dest
= qir_SEL(c
, cond
,
900 qir_uniform_f(c
, 1.0), qir_uniform_f(c
, 0.0));
904 *dest
= qir_SEL(c
, cond
,
905 ntq_get_alu_src(c
, sel_instr
, 1),
906 ntq_get_alu_src(c
, sel_instr
, 2));
910 *dest
= qir_SEL(c
, cond
,
911 qir_uniform_ui(c
, ~0), qir_uniform_ui(c
, 0));
919 * Attempts to fold a comparison generating a boolean result into the
920 * condition code for selecting between two values, instead of comparing the
921 * boolean result against 0 to generate the condition code.
923 static struct qreg
ntq_emit_bcsel(struct vc4_compile
*c
, nir_alu_instr
*instr
,
926 if (!instr
->src
[0].src
.is_ssa
)
928 nir_alu_instr
*compare
=
929 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
934 if (ntq_emit_comparison(c
, &dest
, compare
, instr
))
939 return qir_SEL(c
, QPU_COND_NS
, src
[1], src
[2]);
943 ntq_emit_alu(struct vc4_compile
*c
, nir_alu_instr
*instr
)
945 /* Vectors are special in that they have non-scalarized writemasks,
946 * and just take the first swizzle channel for each argument in order
947 * into each writemask channel.
949 if (instr
->op
== nir_op_vec2
||
950 instr
->op
== nir_op_vec3
||
951 instr
->op
== nir_op_vec4
) {
953 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
954 srcs
[i
] = ntq_get_src(c
, instr
->src
[i
].src
,
955 instr
->src
[i
].swizzle
[0]);
956 struct qreg
*dest
= ntq_get_dest(c
, &instr
->dest
.dest
);
957 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
962 if (instr
->op
== nir_op_pack_unorm_4x8
) {
963 ntq_emit_pack_unorm_4x8(c
, instr
);
967 if (instr
->op
== nir_op_unpack_unorm_4x8
) {
968 struct qreg src
= ntq_get_src(c
, instr
->src
[0].src
,
969 instr
->src
[0].swizzle
[0]);
970 struct qreg
*dest
= ntq_get_dest(c
, &instr
->dest
.dest
);
971 for (int i
= 0; i
< 4; i
++) {
972 if (instr
->dest
.write_mask
& (1 << i
))
973 dest
[i
] = qir_UNPACK_8_F(c
, src
, i
);
978 /* General case: We can just grab the one used channel per src. */
979 struct qreg src
[nir_op_infos
[instr
->op
].num_inputs
];
980 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
981 src
[i
] = ntq_get_alu_src(c
, instr
, i
);
984 /* Pick the channel to store the output in. */
985 assert(!instr
->dest
.saturate
);
986 struct qreg
*dest
= ntq_get_dest(c
, &instr
->dest
.dest
);
987 assert(util_is_power_of_two(instr
->dest
.write_mask
));
988 dest
+= ffs(instr
->dest
.write_mask
) - 1;
993 *dest
= qir_MOV(c
, src
[0]);
996 *dest
= qir_FMUL(c
, src
[0], src
[1]);
999 *dest
= qir_FADD(c
, src
[0], src
[1]);
1002 *dest
= qir_FSUB(c
, src
[0], src
[1]);
1005 *dest
= qir_FMIN(c
, src
[0], src
[1]);
1008 *dest
= qir_FMAX(c
, src
[0], src
[1]);
1013 *dest
= qir_FTOI(c
, src
[0]);
1017 *dest
= qir_ITOF(c
, src
[0]);
1020 *dest
= qir_AND(c
, src
[0], qir_uniform_f(c
, 1.0));
1023 *dest
= qir_AND(c
, src
[0], qir_uniform_ui(c
, 1));
1028 *dest
= qir_SEL(c
, QPU_COND_ZC
,
1029 qir_uniform_ui(c
, ~0),
1030 qir_uniform_ui(c
, 0));
1034 *dest
= qir_ADD(c
, src
[0], src
[1]);
1037 *dest
= qir_SHR(c
, src
[0], src
[1]);
1040 *dest
= qir_SUB(c
, src
[0], src
[1]);
1043 *dest
= qir_ASR(c
, src
[0], src
[1]);
1046 *dest
= qir_SHL(c
, src
[0], src
[1]);
1049 *dest
= qir_MIN(c
, src
[0], src
[1]);
1052 *dest
= qir_MAX(c
, src
[0], src
[1]);
1055 *dest
= qir_AND(c
, src
[0], src
[1]);
1058 *dest
= qir_OR(c
, src
[0], src
[1]);
1061 *dest
= qir_XOR(c
, src
[0], src
[1]);
1064 *dest
= qir_NOT(c
, src
[0]);
1068 *dest
= ntq_umul(c
, src
[0], src
[1]);
1084 if (!ntq_emit_comparison(c
, dest
, instr
, instr
)) {
1085 fprintf(stderr
, "Bad comparison instruction\n");
1090 *dest
= ntq_emit_bcsel(c
, instr
, src
);
1094 *dest
= qir_SEL(c
, QPU_COND_ZC
, src
[1], src
[2]);
1098 *dest
= ntq_rcp(c
, src
[0]);
1101 *dest
= ntq_rsq(c
, src
[0]);
1104 *dest
= qir_EXP2(c
, src
[0]);
1107 *dest
= qir_LOG2(c
, src
[0]);
1111 *dest
= qir_ITOF(c
, qir_FTOI(c
, src
[0]));
1114 *dest
= ntq_fceil(c
, src
[0]);
1117 *dest
= ntq_ffract(c
, src
[0]);
1120 *dest
= ntq_ffloor(c
, src
[0]);
1124 *dest
= ntq_fsin(c
, src
[0]);
1127 *dest
= ntq_fcos(c
, src
[0]);
1131 *dest
= ntq_fsign(c
, src
[0]);
1135 *dest
= qir_FMAXABS(c
, src
[0], src
[0]);
1138 *dest
= qir_MAX(c
, src
[0],
1139 qir_SUB(c
, qir_uniform_ui(c
, 0), src
[0]));
1142 case nir_op_ibitfield_extract
:
1143 *dest
= ntq_emit_ibfe(c
, src
[0], src
[1], src
[2]);
1146 case nir_op_ubitfield_extract
:
1147 *dest
= ntq_emit_ubfe(c
, src
[0], src
[1], src
[2]);
1150 case nir_op_usadd_4x8
:
1151 *dest
= qir_V8ADDS(c
, src
[0], src
[1]);
1154 case nir_op_ussub_4x8
:
1155 *dest
= qir_V8SUBS(c
, src
[0], src
[1]);
1158 case nir_op_umin_4x8
:
1159 *dest
= qir_V8MIN(c
, src
[0], src
[1]);
1162 case nir_op_umax_4x8
:
1163 *dest
= qir_V8MAX(c
, src
[0], src
[1]);
1166 case nir_op_umul_unorm_4x8
:
1167 *dest
= qir_V8MULD(c
, src
[0], src
[1]);
1171 fprintf(stderr
, "unknown NIR ALU inst: ");
1172 nir_print_instr(&instr
->instr
, stderr
);
1173 fprintf(stderr
, "\n");
1179 emit_frag_end(struct vc4_compile
*c
)
1182 if (c
->output_color_index
!= -1) {
1183 color
= c
->outputs
[c
->output_color_index
];
1185 color
= qir_uniform_ui(c
, 0);
1188 uint32_t discard_cond
= QPU_COND_ALWAYS
;
1189 if (c
->discard
.file
!= QFILE_NULL
) {
1190 qir_SF(c
, c
->discard
);
1191 discard_cond
= QPU_COND_ZS
;
1194 if (c
->fs_key
->stencil_enabled
) {
1195 qir_TLB_STENCIL_SETUP(c
, qir_uniform(c
, QUNIFORM_STENCIL
, 0));
1196 if (c
->fs_key
->stencil_twoside
) {
1197 qir_TLB_STENCIL_SETUP(c
, qir_uniform(c
, QUNIFORM_STENCIL
, 1));
1199 if (c
->fs_key
->stencil_full_writemasks
) {
1200 qir_TLB_STENCIL_SETUP(c
, qir_uniform(c
, QUNIFORM_STENCIL
, 2));
1204 if (c
->output_sample_mask_index
!= -1) {
1205 qir_MS_MASK(c
, c
->outputs
[c
->output_sample_mask_index
]);
1208 if (c
->fs_key
->depth_enabled
) {
1210 if (c
->output_position_index
!= -1) {
1211 z
= qir_FTOI(c
, qir_FMUL(c
, c
->outputs
[c
->output_position_index
+ 2],
1212 qir_uniform_f(c
, 0xffffff)));
1216 struct qinst
*inst
= qir_TLB_Z_WRITE(c
, z
);
1217 inst
->cond
= discard_cond
;
1220 if (!c
->msaa_per_sample_output
) {
1221 struct qinst
*inst
= qir_TLB_COLOR_WRITE(c
, color
);
1222 inst
->cond
= discard_cond
;
1224 for (int i
= 0; i
< VC4_MAX_SAMPLES
; i
++) {
1225 struct qinst
*inst
= qir_TLB_COLOR_WRITE_MS(c
, c
->sample_colors
[i
]);
1226 inst
->cond
= discard_cond
;
1232 emit_scaled_viewport_write(struct vc4_compile
*c
, struct qreg rcp_w
)
1234 struct qreg packed
= qir_get_temp(c
);
1236 for (int i
= 0; i
< 2; i
++) {
1238 qir_uniform(c
, QUNIFORM_VIEWPORT_X_SCALE
+ i
, 0);
1240 struct qreg packed_chan
= packed
;
1241 packed_chan
.pack
= QPU_PACK_A_16A
+ i
;
1243 qir_FTOI_dest(c
, packed_chan
,
1246 c
->outputs
[c
->output_position_index
+ i
],
1251 qir_VPM_WRITE(c
, packed
);
1255 emit_zs_write(struct vc4_compile
*c
, struct qreg rcp_w
)
1257 struct qreg zscale
= qir_uniform(c
, QUNIFORM_VIEWPORT_Z_SCALE
, 0);
1258 struct qreg zoffset
= qir_uniform(c
, QUNIFORM_VIEWPORT_Z_OFFSET
, 0);
1260 qir_VPM_WRITE(c
, qir_FADD(c
, qir_FMUL(c
, qir_FMUL(c
,
1261 c
->outputs
[c
->output_position_index
+ 2],
1268 emit_rcp_wc_write(struct vc4_compile
*c
, struct qreg rcp_w
)
1270 qir_VPM_WRITE(c
, rcp_w
);
1274 emit_point_size_write(struct vc4_compile
*c
)
1276 struct qreg point_size
;
1278 if (c
->output_point_size_index
!= -1)
1279 point_size
= c
->outputs
[c
->output_point_size_index
];
1281 point_size
= qir_uniform_f(c
, 1.0);
1283 /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
1286 point_size
= qir_FMAX(c
, point_size
, qir_uniform_f(c
, .125));
1288 qir_VPM_WRITE(c
, point_size
);
1292 * Emits a VPM read of the stub vertex attribute set up by vc4_draw.c.
1294 * The simulator insists that there be at least one vertex attribute, so
1295 * vc4_draw.c will emit one if it wouldn't have otherwise. The simulator also
1296 * insists that all vertex attributes loaded get read by the VS/CS, so we have
1297 * to consume it here.
1300 emit_stub_vpm_read(struct vc4_compile
*c
)
1305 c
->vattr_sizes
[0] = 4;
1306 struct qreg vpm
= { QFILE_VPM
, 0 };
1307 (void)qir_MOV(c
, vpm
);
1312 emit_vert_end(struct vc4_compile
*c
,
1313 struct vc4_varying_slot
*fs_inputs
,
1314 uint32_t num_fs_inputs
)
1316 struct qreg rcp_w
= qir_RCP(c
, c
->outputs
[c
->output_position_index
+ 3]);
1318 emit_stub_vpm_read(c
);
1320 emit_scaled_viewport_write(c
, rcp_w
);
1321 emit_zs_write(c
, rcp_w
);
1322 emit_rcp_wc_write(c
, rcp_w
);
1323 if (c
->vs_key
->per_vertex_point_size
)
1324 emit_point_size_write(c
);
1326 for (int i
= 0; i
< num_fs_inputs
; i
++) {
1327 struct vc4_varying_slot
*input
= &fs_inputs
[i
];
1330 for (j
= 0; j
< c
->num_outputs
; j
++) {
1331 struct vc4_varying_slot
*output
=
1332 &c
->output_slots
[j
];
1334 if (input
->slot
== output
->slot
&&
1335 input
->swizzle
== output
->swizzle
) {
1336 qir_VPM_WRITE(c
, c
->outputs
[j
]);
1340 /* Emit padding if we didn't find a declared VS output for
1343 if (j
== c
->num_outputs
)
1344 qir_VPM_WRITE(c
, qir_uniform_f(c
, 0.0));
1349 emit_coord_end(struct vc4_compile
*c
)
1351 struct qreg rcp_w
= qir_RCP(c
, c
->outputs
[c
->output_position_index
+ 3]);
1353 emit_stub_vpm_read(c
);
1355 for (int i
= 0; i
< 4; i
++)
1356 qir_VPM_WRITE(c
, c
->outputs
[c
->output_position_index
+ i
]);
1358 emit_scaled_viewport_write(c
, rcp_w
);
1359 emit_zs_write(c
, rcp_w
);
1360 emit_rcp_wc_write(c
, rcp_w
);
1361 if (c
->vs_key
->per_vertex_point_size
)
1362 emit_point_size_write(c
);
1366 vc4_optimize_nir(struct nir_shader
*s
)
1373 nir_lower_vars_to_ssa(s
);
1374 nir_lower_alu_to_scalar(s
);
1376 progress
= nir_copy_prop(s
) || progress
;
1377 progress
= nir_opt_dce(s
) || progress
;
1378 progress
= nir_opt_cse(s
) || progress
;
1379 progress
= nir_opt_peephole_select(s
) || progress
;
1380 progress
= nir_opt_algebraic(s
) || progress
;
1381 progress
= nir_opt_constant_folding(s
) || progress
;
1382 progress
= nir_opt_undef(s
) || progress
;
1387 driver_location_compare(const void *in_a
, const void *in_b
)
1389 const nir_variable
*const *a
= in_a
;
1390 const nir_variable
*const *b
= in_b
;
1392 return (*a
)->data
.driver_location
- (*b
)->data
.driver_location
;
1396 ntq_setup_inputs(struct vc4_compile
*c
)
1398 unsigned num_entries
= 0;
1399 nir_foreach_variable(var
, &c
->s
->inputs
)
1402 nir_variable
*vars
[num_entries
];
1405 nir_foreach_variable(var
, &c
->s
->inputs
)
1408 /* Sort the variables so that we emit the input setup in
1409 * driver_location order. This is required for VPM reads, whose data
1410 * is fetched into the VPM in driver_location (TGSI register index)
1413 qsort(&vars
, num_entries
, sizeof(*vars
), driver_location_compare
);
1415 for (unsigned i
= 0; i
< num_entries
; i
++) {
1416 nir_variable
*var
= vars
[i
];
1417 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1418 unsigned loc
= var
->data
.driver_location
;
1420 assert(array_len
== 1);
1422 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1425 if (c
->stage
== QSTAGE_FRAG
) {
1426 if (var
->data
.location
== VARYING_SLOT_POS
) {
1427 emit_fragcoord_input(c
, loc
);
1428 } else if (var
->data
.location
== VARYING_SLOT_FACE
) {
1429 c
->inputs
[loc
* 4 + 0] = qir_FRAG_REV_FLAG(c
);
1430 } else if (var
->data
.location
>= VARYING_SLOT_VAR0
&&
1431 (c
->fs_key
->point_sprite_mask
&
1432 (1 << (var
->data
.location
-
1433 VARYING_SLOT_VAR0
)))) {
1434 c
->inputs
[loc
* 4 + 0] = c
->point_x
;
1435 c
->inputs
[loc
* 4 + 1] = c
->point_y
;
1437 emit_fragment_input(c
, loc
, var
->data
.location
);
1440 emit_vertex_input(c
, loc
);
1446 ntq_setup_outputs(struct vc4_compile
*c
)
1448 nir_foreach_variable(var
, &c
->s
->outputs
) {
1449 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1450 unsigned loc
= var
->data
.driver_location
* 4;
1452 assert(array_len
== 1);
1455 for (int i
= 0; i
< 4; i
++)
1456 add_output(c
, loc
+ i
, var
->data
.location
, i
);
1458 if (c
->stage
== QSTAGE_FRAG
) {
1459 switch (var
->data
.location
) {
1460 case FRAG_RESULT_COLOR
:
1461 case FRAG_RESULT_DATA0
:
1462 c
->output_color_index
= loc
;
1464 case FRAG_RESULT_DEPTH
:
1465 c
->output_position_index
= loc
;
1467 case FRAG_RESULT_SAMPLE_MASK
:
1468 c
->output_sample_mask_index
= loc
;
1472 switch (var
->data
.location
) {
1473 case VARYING_SLOT_POS
:
1474 c
->output_position_index
= loc
;
1476 case VARYING_SLOT_PSIZ
:
1477 c
->output_point_size_index
= loc
;
1485 ntq_setup_uniforms(struct vc4_compile
*c
)
1487 nir_foreach_variable(var
, &c
->s
->uniforms
) {
1488 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1489 unsigned array_elem_size
= 4 * sizeof(float);
1491 declare_uniform_range(c
, var
->data
.driver_location
* array_elem_size
,
1492 array_len
* array_elem_size
);
1498 * Sets up the mapping from nir_register to struct qreg *.
1500 * Each nir_register gets a struct qreg per 32-bit component being stored.
1503 ntq_setup_registers(struct vc4_compile
*c
, struct exec_list
*list
)
1505 foreach_list_typed(nir_register
, nir_reg
, node
, list
) {
1506 unsigned array_len
= MAX2(nir_reg
->num_array_elems
, 1);
1507 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
1509 nir_reg
->num_components
);
1511 _mesa_hash_table_insert(c
->def_ht
, nir_reg
, qregs
);
1513 for (int i
= 0; i
< array_len
* nir_reg
->num_components
; i
++)
1514 qregs
[i
] = qir_uniform_ui(c
, 0);
1519 ntq_emit_load_const(struct vc4_compile
*c
, nir_load_const_instr
*instr
)
1521 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1522 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1523 qregs
[i
] = qir_uniform_ui(c
, instr
->value
.u32
[i
]);
1525 _mesa_hash_table_insert(c
->def_ht
, &instr
->def
, qregs
);
1529 ntq_emit_ssa_undef(struct vc4_compile
*c
, nir_ssa_undef_instr
*instr
)
1531 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1533 /* QIR needs there to be *some* value, so pick 0 (same as for
1534 * ntq_setup_registers().
1536 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1537 qregs
[i
] = qir_uniform_ui(c
, 0);
1541 ntq_emit_intrinsic(struct vc4_compile
*c
, nir_intrinsic_instr
*instr
)
1543 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
1544 nir_const_value
*const_offset
;
1546 struct qreg
*dest
= NULL
;
1548 if (info
->has_dest
) {
1549 dest
= ntq_get_dest(c
, &instr
->dest
);
1552 switch (instr
->intrinsic
) {
1553 case nir_intrinsic_load_uniform
:
1554 assert(instr
->num_components
== 1);
1555 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1557 offset
= instr
->const_index
[0] + const_offset
->u32
[0];
1558 assert(offset
% 4 == 0);
1559 /* We need dwords */
1560 offset
= offset
/ 4;
1561 if (offset
< VC4_NIR_STATE_UNIFORM_OFFSET
) {
1562 *dest
= qir_uniform(c
, QUNIFORM_UNIFORM
,
1565 *dest
= qir_uniform(c
, offset
-
1566 VC4_NIR_STATE_UNIFORM_OFFSET
,
1570 *dest
= indirect_uniform_load(c
, instr
);
1574 case nir_intrinsic_load_user_clip_plane
:
1575 *dest
= qir_uniform(c
, QUNIFORM_USER_CLIP_PLANE
,
1576 instr
->const_index
[0]);
1579 case nir_intrinsic_load_sample_mask_in
:
1580 *dest
= qir_uniform(c
, QUNIFORM_SAMPLE_MASK
, 0);
1583 case nir_intrinsic_load_input
:
1584 assert(instr
->num_components
== 1);
1585 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1586 assert(const_offset
&& "vc4 doesn't support indirect inputs");
1587 if (instr
->const_index
[0] >= VC4_NIR_TLB_COLOR_READ_INPUT
) {
1588 assert(const_offset
->u32
[0] == 0);
1589 /* Reads of the per-sample color need to be done in
1592 int sample_index
= (instr
->const_index
[0] -
1593 VC4_NIR_TLB_COLOR_READ_INPUT
);
1594 for (int i
= 0; i
<= sample_index
; i
++) {
1595 if (c
->color_reads
[i
].file
== QFILE_NULL
) {
1597 qir_TLB_COLOR_READ(c
);
1600 *dest
= c
->color_reads
[sample_index
];
1602 offset
= instr
->const_index
[0] + const_offset
->u32
[0];
1603 *dest
= c
->inputs
[offset
];
1607 case nir_intrinsic_store_output
:
1608 const_offset
= nir_src_as_const_value(instr
->src
[1]);
1609 assert(const_offset
&& "vc4 doesn't support indirect outputs");
1610 offset
= instr
->const_index
[0] + const_offset
->u32
[0];
1612 /* MSAA color outputs are the only case where we have an
1613 * output that's not lowered to being a store of a single 32
1616 if (c
->stage
== QSTAGE_FRAG
&& instr
->num_components
== 4) {
1617 assert(offset
== c
->output_color_index
);
1618 for (int i
= 0; i
< 4; i
++) {
1619 c
->sample_colors
[i
] =
1620 qir_MOV(c
, ntq_get_src(c
, instr
->src
[0],
1624 assert(instr
->num_components
== 1);
1625 c
->outputs
[offset
] =
1626 qir_MOV(c
, ntq_get_src(c
, instr
->src
[0], 0));
1627 c
->num_outputs
= MAX2(c
->num_outputs
, offset
+ 1);
1631 case nir_intrinsic_discard
:
1632 c
->discard
= qir_uniform_ui(c
, ~0);
1635 case nir_intrinsic_discard_if
:
1636 if (c
->discard
.file
== QFILE_NULL
)
1637 c
->discard
= qir_uniform_ui(c
, 0);
1638 c
->discard
= qir_OR(c
, c
->discard
,
1639 ntq_get_src(c
, instr
->src
[0], 0));
1643 fprintf(stderr
, "Unknown intrinsic: ");
1644 nir_print_instr(&instr
->instr
, stderr
);
1645 fprintf(stderr
, "\n");
1651 ntq_emit_if(struct vc4_compile
*c
, nir_if
*if_stmt
)
1653 fprintf(stderr
, "general IF statements not handled.\n");
1657 ntq_emit_instr(struct vc4_compile
*c
, nir_instr
*instr
)
1659 switch (instr
->type
) {
1660 case nir_instr_type_alu
:
1661 ntq_emit_alu(c
, nir_instr_as_alu(instr
));
1664 case nir_instr_type_intrinsic
:
1665 ntq_emit_intrinsic(c
, nir_instr_as_intrinsic(instr
));
1668 case nir_instr_type_load_const
:
1669 ntq_emit_load_const(c
, nir_instr_as_load_const(instr
));
1672 case nir_instr_type_ssa_undef
:
1673 ntq_emit_ssa_undef(c
, nir_instr_as_ssa_undef(instr
));
1676 case nir_instr_type_tex
:
1677 ntq_emit_tex(c
, nir_instr_as_tex(instr
));
1681 fprintf(stderr
, "Unknown NIR instr type: ");
1682 nir_print_instr(instr
, stderr
);
1683 fprintf(stderr
, "\n");
1689 ntq_emit_block(struct vc4_compile
*c
, nir_block
*block
)
1691 nir_foreach_instr(block
, instr
) {
1692 ntq_emit_instr(c
, instr
);
1696 static void ntq_emit_cf_list(struct vc4_compile
*c
, struct exec_list
*list
);
1699 ntq_emit_loop(struct vc4_compile
*c
, nir_loop
*nloop
)
1701 fprintf(stderr
, "LOOPS not fully handled. Rendering errors likely.\n");
1702 ntq_emit_cf_list(c
, &nloop
->body
);
1706 ntq_emit_function(struct vc4_compile
*c
, nir_function_impl
*func
)
1708 fprintf(stderr
, "FUNCTIONS not handled.\n");
1713 ntq_emit_cf_list(struct vc4_compile
*c
, struct exec_list
*list
)
1715 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
1716 switch (node
->type
) {
1717 case nir_cf_node_block
:
1718 ntq_emit_block(c
, nir_cf_node_as_block(node
));
1721 case nir_cf_node_if
:
1722 ntq_emit_if(c
, nir_cf_node_as_if(node
));
1725 case nir_cf_node_loop
:
1726 ntq_emit_loop(c
, nir_cf_node_as_loop(node
));
1729 case nir_cf_node_function
:
1730 ntq_emit_function(c
, nir_cf_node_as_function(node
));
1734 fprintf(stderr
, "Unknown NIR node type\n");
1741 ntq_emit_impl(struct vc4_compile
*c
, nir_function_impl
*impl
)
1743 ntq_setup_registers(c
, &impl
->registers
);
1744 ntq_emit_cf_list(c
, &impl
->body
);
1748 nir_to_qir(struct vc4_compile
*c
)
1750 ntq_setup_inputs(c
);
1751 ntq_setup_outputs(c
);
1752 ntq_setup_uniforms(c
);
1753 ntq_setup_registers(c
, &c
->s
->registers
);
1755 /* Find the main function and emit the body. */
1756 nir_foreach_function(c
->s
, function
) {
1757 assert(strcmp(function
->name
, "main") == 0);
1758 assert(function
->impl
);
1759 ntq_emit_impl(c
, function
->impl
);
1763 static const nir_shader_compiler_options nir_options
= {
1764 .lower_extract_byte
= true,
1765 .lower_extract_word
= true,
1770 .lower_fsqrt
= true,
1771 .lower_negate
= true,
1775 count_nir_instrs_in_block(nir_block
*block
, void *state
)
1777 int *count
= (int *) state
;
1778 nir_foreach_instr(block
, instr
) {
1779 *count
= *count
+ 1;
1785 count_nir_instrs(nir_shader
*nir
)
1788 nir_foreach_function(nir
, function
) {
1789 if (!function
->impl
)
1791 nir_foreach_block(function
->impl
, count_nir_instrs_in_block
, &count
);
1796 static struct vc4_compile
*
1797 vc4_shader_ntq(struct vc4_context
*vc4
, enum qstage stage
,
1798 struct vc4_key
*key
)
1800 struct vc4_compile
*c
= qir_compile_init();
1803 c
->shader_state
= &key
->shader_state
->base
;
1804 c
->program_id
= key
->shader_state
->program_id
;
1805 c
->variant_id
= key
->shader_state
->compiled_variant_count
++;
1810 c
->fs_key
= (struct vc4_fs_key
*)key
;
1811 if (c
->fs_key
->is_points
) {
1812 c
->point_x
= emit_fragment_varying(c
, ~0, 0);
1813 c
->point_y
= emit_fragment_varying(c
, ~0, 0);
1814 } else if (c
->fs_key
->is_lines
) {
1815 c
->line_x
= emit_fragment_varying(c
, ~0, 0);
1819 c
->vs_key
= (struct vc4_vs_key
*)key
;
1822 c
->vs_key
= (struct vc4_vs_key
*)key
;
1826 const struct tgsi_token
*tokens
= key
->shader_state
->base
.tokens
;
1828 if (vc4_debug
& VC4_DEBUG_TGSI
) {
1829 fprintf(stderr
, "%s prog %d/%d TGSI:\n",
1830 qir_get_stage_name(c
->stage
),
1831 c
->program_id
, c
->variant_id
);
1832 tgsi_dump(tokens
, 0);
1835 c
->s
= tgsi_to_nir(tokens
, &nir_options
);
1836 nir_opt_global_to_local(c
->s
);
1837 nir_convert_to_ssa(c
->s
);
1839 if (stage
== QSTAGE_FRAG
)
1840 vc4_nir_lower_blend(c
);
1842 struct nir_lower_tex_options tex_options
= {
1843 /* We would need to implement txs, but we don't want the
1844 * int/float conversions
1846 .lower_rect
= false,
1848 /* We want to use this, but we don't want to newton-raphson
1853 /* Apply swizzles to all samplers. */
1854 .swizzle_result
= ~0,
1857 /* Lower the format swizzle and ARB_texture_swizzle-style swizzle.
1858 * The format swizzling applies before sRGB decode, and
1859 * ARB_texture_swizzle is the last thing before returning the sample.
1861 for (int i
= 0; i
< ARRAY_SIZE(key
->tex
); i
++) {
1862 enum pipe_format format
= c
->key
->tex
[i
].format
;
1867 const uint8_t *format_swizzle
= vc4_get_format_swizzle(format
);
1869 for (int j
= 0; j
< 4; j
++) {
1870 uint8_t arb_swiz
= c
->key
->tex
[i
].swizzle
[j
];
1872 if (arb_swiz
<= 3) {
1873 tex_options
.swizzles
[i
][j
] =
1874 format_swizzle
[arb_swiz
];
1876 tex_options
.swizzles
[i
][j
] = arb_swiz
;
1879 /* If ARB_texture_swizzle is reading from the R, G, or
1880 * B channels of an sRGB texture, then we need to
1881 * apply sRGB decode to this channel at sample time.
1883 if (arb_swiz
< 3 && util_format_is_srgb(format
)) {
1884 c
->tex_srgb_decode
[i
] |= (1 << j
);
1890 nir_lower_tex(c
->s
, &tex_options
);
1892 if (c
->fs_key
&& c
->fs_key
->light_twoside
)
1893 nir_lower_two_sided_color(c
->s
);
1895 if (stage
== QSTAGE_FRAG
)
1896 nir_lower_clip_fs(c
->s
, c
->key
->ucp_enables
);
1898 nir_lower_clip_vs(c
->s
, c
->key
->ucp_enables
);
1900 vc4_nir_lower_io(c
);
1901 vc4_nir_lower_txf_ms(c
);
1902 nir_lower_idiv(c
->s
);
1903 nir_lower_load_const_to_scalar(c
->s
);
1905 vc4_optimize_nir(c
->s
);
1907 nir_remove_dead_variables(c
->s
);
1909 nir_convert_from_ssa(c
->s
, true);
1911 if (vc4_debug
& VC4_DEBUG_SHADERDB
) {
1912 fprintf(stderr
, "SHADER-DB: %s prog %d/%d: %d NIR instructions\n",
1913 qir_get_stage_name(c
->stage
),
1914 c
->program_id
, c
->variant_id
,
1915 count_nir_instrs(c
->s
));
1918 if (vc4_debug
& VC4_DEBUG_NIR
) {
1919 fprintf(stderr
, "%s prog %d/%d NIR:\n",
1920 qir_get_stage_name(c
->stage
),
1921 c
->program_id
, c
->variant_id
);
1922 nir_print_shader(c
->s
, stderr
);
1933 vc4
->prog
.fs
->input_slots
,
1934 vc4
->prog
.fs
->num_inputs
);
1941 if (vc4_debug
& VC4_DEBUG_QIR
) {
1942 fprintf(stderr
, "%s prog %d/%d pre-opt QIR:\n",
1943 qir_get_stage_name(c
->stage
),
1944 c
->program_id
, c
->variant_id
);
1949 qir_lower_uniforms(c
);
1951 qir_schedule_instructions(c
);
1953 if (vc4_debug
& VC4_DEBUG_QIR
) {
1954 fprintf(stderr
, "%s prog %d/%d QIR:\n",
1955 qir_get_stage_name(c
->stage
),
1956 c
->program_id
, c
->variant_id
);
1960 qir_reorder_uniforms(c
);
1961 vc4_generate_code(vc4
, c
);
1963 if (vc4_debug
& VC4_DEBUG_SHADERDB
) {
1964 fprintf(stderr
, "SHADER-DB: %s prog %d/%d: %d instructions\n",
1965 qir_get_stage_name(c
->stage
),
1966 c
->program_id
, c
->variant_id
,
1968 fprintf(stderr
, "SHADER-DB: %s prog %d/%d: %d uniforms\n",
1969 qir_get_stage_name(c
->stage
),
1970 c
->program_id
, c
->variant_id
,
1980 vc4_shader_state_create(struct pipe_context
*pctx
,
1981 const struct pipe_shader_state
*cso
)
1983 struct vc4_context
*vc4
= vc4_context(pctx
);
1984 struct vc4_uncompiled_shader
*so
= CALLOC_STRUCT(vc4_uncompiled_shader
);
1988 so
->base
.tokens
= tgsi_dup_tokens(cso
->tokens
);
1989 so
->program_id
= vc4
->next_uncompiled_program_id
++;
1995 copy_uniform_state_to_shader(struct vc4_compiled_shader
*shader
,
1996 struct vc4_compile
*c
)
1998 int count
= c
->num_uniforms
;
1999 struct vc4_shader_uniform_info
*uinfo
= &shader
->uniforms
;
2001 uinfo
->count
= count
;
2002 uinfo
->data
= ralloc_array(shader
, uint32_t, count
);
2003 memcpy(uinfo
->data
, c
->uniform_data
,
2004 count
* sizeof(*uinfo
->data
));
2005 uinfo
->contents
= ralloc_array(shader
, enum quniform_contents
, count
);
2006 memcpy(uinfo
->contents
, c
->uniform_contents
,
2007 count
* sizeof(*uinfo
->contents
));
2008 uinfo
->num_texture_samples
= c
->num_texture_samples
;
2010 vc4_set_shader_uniform_dirty_flags(shader
);
2013 static struct vc4_compiled_shader
*
2014 vc4_get_compiled_shader(struct vc4_context
*vc4
, enum qstage stage
,
2015 struct vc4_key
*key
)
2017 struct hash_table
*ht
;
2019 if (stage
== QSTAGE_FRAG
) {
2021 key_size
= sizeof(struct vc4_fs_key
);
2024 key_size
= sizeof(struct vc4_vs_key
);
2027 struct vc4_compiled_shader
*shader
;
2028 struct hash_entry
*entry
= _mesa_hash_table_search(ht
, key
);
2032 struct vc4_compile
*c
= vc4_shader_ntq(vc4
, stage
, key
);
2033 shader
= rzalloc(NULL
, struct vc4_compiled_shader
);
2035 shader
->program_id
= vc4
->next_compiled_program_id
++;
2036 if (stage
== QSTAGE_FRAG
) {
2037 bool input_live
[c
->num_input_slots
];
2039 memset(input_live
, 0, sizeof(input_live
));
2040 list_for_each_entry(struct qinst
, inst
, &c
->instructions
, link
) {
2041 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
2042 if (inst
->src
[i
].file
== QFILE_VARY
)
2043 input_live
[inst
->src
[i
].index
] = true;
2047 shader
->input_slots
= ralloc_array(shader
,
2048 struct vc4_varying_slot
,
2049 c
->num_input_slots
);
2051 for (int i
= 0; i
< c
->num_input_slots
; i
++) {
2052 struct vc4_varying_slot
*slot
= &c
->input_slots
[i
];
2057 /* Skip non-VS-output inputs. */
2058 if (slot
->slot
== (uint8_t)~0)
2061 if (slot
->slot
== VARYING_SLOT_COL0
||
2062 slot
->slot
== VARYING_SLOT_COL1
||
2063 slot
->slot
== VARYING_SLOT_BFC0
||
2064 slot
->slot
== VARYING_SLOT_BFC1
) {
2065 shader
->color_inputs
|= (1 << shader
->num_inputs
);
2068 shader
->input_slots
[shader
->num_inputs
] = *slot
;
2069 shader
->num_inputs
++;
2072 shader
->num_inputs
= c
->num_inputs
;
2074 shader
->vattr_offsets
[0] = 0;
2075 for (int i
= 0; i
< 8; i
++) {
2076 shader
->vattr_offsets
[i
+ 1] =
2077 shader
->vattr_offsets
[i
] + c
->vattr_sizes
[i
];
2079 if (c
->vattr_sizes
[i
])
2080 shader
->vattrs_live
|= (1 << i
);
2084 copy_uniform_state_to_shader(shader
, c
);
2085 shader
->bo
= vc4_bo_alloc_shader(vc4
->screen
, c
->qpu_insts
,
2086 c
->qpu_inst_count
* sizeof(uint64_t));
2088 /* Copy the compiler UBO range state to the compiled shader, dropping
2089 * out arrays that were never referenced by an indirect load.
2091 * (Note that QIR dead code elimination of an array access still
2092 * leaves that array alive, though)
2094 if (c
->num_ubo_ranges
) {
2095 shader
->num_ubo_ranges
= c
->num_ubo_ranges
;
2096 shader
->ubo_ranges
= ralloc_array(shader
, struct vc4_ubo_range
,
2099 for (int i
= 0; i
< c
->num_uniform_ranges
; i
++) {
2100 struct vc4_compiler_ubo_range
*range
=
2105 shader
->ubo_ranges
[j
].dst_offset
= range
->dst_offset
;
2106 shader
->ubo_ranges
[j
].src_offset
= range
->src_offset
;
2107 shader
->ubo_ranges
[j
].size
= range
->size
;
2108 shader
->ubo_size
+= c
->ubo_ranges
[i
].size
;
2112 if (shader
->ubo_size
) {
2113 if (vc4_debug
& VC4_DEBUG_SHADERDB
) {
2114 fprintf(stderr
, "SHADER-DB: %s prog %d/%d: %d UBO uniforms\n",
2115 qir_get_stage_name(c
->stage
),
2116 c
->program_id
, c
->variant_id
,
2117 shader
->ubo_size
/ 4);
2121 qir_compile_destroy(c
);
2123 struct vc4_key
*dup_key
;
2124 dup_key
= ralloc_size(shader
, key_size
);
2125 memcpy(dup_key
, key
, key_size
);
2126 _mesa_hash_table_insert(ht
, dup_key
, shader
);
2132 vc4_setup_shared_key(struct vc4_context
*vc4
, struct vc4_key
*key
,
2133 struct vc4_texture_stateobj
*texstate
)
2135 for (int i
= 0; i
< texstate
->num_textures
; i
++) {
2136 struct pipe_sampler_view
*sampler
= texstate
->textures
[i
];
2137 struct pipe_sampler_state
*sampler_state
=
2138 texstate
->samplers
[i
];
2143 key
->tex
[i
].format
= sampler
->format
;
2144 key
->tex
[i
].swizzle
[0] = sampler
->swizzle_r
;
2145 key
->tex
[i
].swizzle
[1] = sampler
->swizzle_g
;
2146 key
->tex
[i
].swizzle
[2] = sampler
->swizzle_b
;
2147 key
->tex
[i
].swizzle
[3] = sampler
->swizzle_a
;
2149 if (sampler
->texture
->nr_samples
> 1) {
2150 key
->tex
[i
].msaa_width
= sampler
->texture
->width0
;
2151 key
->tex
[i
].msaa_height
= sampler
->texture
->height0
;
2152 } else if (sampler
){
2153 key
->tex
[i
].compare_mode
= sampler_state
->compare_mode
;
2154 key
->tex
[i
].compare_func
= sampler_state
->compare_func
;
2155 key
->tex
[i
].wrap_s
= sampler_state
->wrap_s
;
2156 key
->tex
[i
].wrap_t
= sampler_state
->wrap_t
;
2160 key
->ucp_enables
= vc4
->rasterizer
->base
.clip_plane_enable
;
2164 vc4_update_compiled_fs(struct vc4_context
*vc4
, uint8_t prim_mode
)
2166 struct vc4_fs_key local_key
;
2167 struct vc4_fs_key
*key
= &local_key
;
2169 if (!(vc4
->dirty
& (VC4_DIRTY_PRIM_MODE
|
2171 VC4_DIRTY_FRAMEBUFFER
|
2173 VC4_DIRTY_RASTERIZER
|
2175 VC4_DIRTY_TEXSTATE
|
2176 VC4_DIRTY_UNCOMPILED_FS
))) {
2180 memset(key
, 0, sizeof(*key
));
2181 vc4_setup_shared_key(vc4
, &key
->base
, &vc4
->fragtex
);
2182 key
->base
.shader_state
= vc4
->prog
.bind_fs
;
2183 key
->is_points
= (prim_mode
== PIPE_PRIM_POINTS
);
2184 key
->is_lines
= (prim_mode
>= PIPE_PRIM_LINES
&&
2185 prim_mode
<= PIPE_PRIM_LINE_STRIP
);
2186 key
->blend
= vc4
->blend
->rt
[0];
2187 if (vc4
->blend
->logicop_enable
) {
2188 key
->logicop_func
= vc4
->blend
->logicop_func
;
2190 key
->logicop_func
= PIPE_LOGICOP_COPY
;
2192 key
->msaa
= vc4
->rasterizer
->base
.multisample
;
2193 key
->sample_coverage
= (vc4
->rasterizer
->base
.multisample
&&
2194 vc4
->sample_mask
!= (1 << VC4_MAX_SAMPLES
) - 1);
2195 key
->sample_alpha_to_coverage
= vc4
->blend
->alpha_to_coverage
;
2196 key
->sample_alpha_to_one
= vc4
->blend
->alpha_to_one
;
2197 if (vc4
->framebuffer
.cbufs
[0])
2198 key
->color_format
= vc4
->framebuffer
.cbufs
[0]->format
;
2200 key
->stencil_enabled
= vc4
->zsa
->stencil_uniforms
[0] != 0;
2201 key
->stencil_twoside
= vc4
->zsa
->stencil_uniforms
[1] != 0;
2202 key
->stencil_full_writemasks
= vc4
->zsa
->stencil_uniforms
[2] != 0;
2203 key
->depth_enabled
= (vc4
->zsa
->base
.depth
.enabled
||
2204 key
->stencil_enabled
);
2205 if (vc4
->zsa
->base
.alpha
.enabled
) {
2206 key
->alpha_test
= true;
2207 key
->alpha_test_func
= vc4
->zsa
->base
.alpha
.func
;
2210 if (key
->is_points
) {
2211 key
->point_sprite_mask
=
2212 vc4
->rasterizer
->base
.sprite_coord_enable
;
2213 key
->point_coord_upper_left
=
2214 (vc4
->rasterizer
->base
.sprite_coord_mode
==
2215 PIPE_SPRITE_COORD_UPPER_LEFT
);
2218 key
->light_twoside
= vc4
->rasterizer
->base
.light_twoside
;
2220 struct vc4_compiled_shader
*old_fs
= vc4
->prog
.fs
;
2221 vc4
->prog
.fs
= vc4_get_compiled_shader(vc4
, QSTAGE_FRAG
, &key
->base
);
2222 if (vc4
->prog
.fs
== old_fs
)
2225 vc4
->dirty
|= VC4_DIRTY_COMPILED_FS
;
2226 if (vc4
->rasterizer
->base
.flatshade
&&
2227 old_fs
&& vc4
->prog
.fs
->color_inputs
!= old_fs
->color_inputs
) {
2228 vc4
->dirty
|= VC4_DIRTY_FLAT_SHADE_FLAGS
;
2233 vc4_update_compiled_vs(struct vc4_context
*vc4
, uint8_t prim_mode
)
2235 struct vc4_vs_key local_key
;
2236 struct vc4_vs_key
*key
= &local_key
;
2238 if (!(vc4
->dirty
& (VC4_DIRTY_PRIM_MODE
|
2239 VC4_DIRTY_RASTERIZER
|
2241 VC4_DIRTY_TEXSTATE
|
2242 VC4_DIRTY_VTXSTATE
|
2243 VC4_DIRTY_UNCOMPILED_VS
|
2244 VC4_DIRTY_COMPILED_FS
))) {
2248 memset(key
, 0, sizeof(*key
));
2249 vc4_setup_shared_key(vc4
, &key
->base
, &vc4
->verttex
);
2250 key
->base
.shader_state
= vc4
->prog
.bind_vs
;
2251 key
->compiled_fs_id
= vc4
->prog
.fs
->program_id
;
2253 for (int i
= 0; i
< ARRAY_SIZE(key
->attr_formats
); i
++)
2254 key
->attr_formats
[i
] = vc4
->vtx
->pipe
[i
].src_format
;
2256 key
->per_vertex_point_size
=
2257 (prim_mode
== PIPE_PRIM_POINTS
&&
2258 vc4
->rasterizer
->base
.point_size_per_vertex
);
2260 struct vc4_compiled_shader
*vs
=
2261 vc4_get_compiled_shader(vc4
, QSTAGE_VERT
, &key
->base
);
2262 if (vs
!= vc4
->prog
.vs
) {
2264 vc4
->dirty
|= VC4_DIRTY_COMPILED_VS
;
2267 key
->is_coord
= true;
2268 struct vc4_compiled_shader
*cs
=
2269 vc4_get_compiled_shader(vc4
, QSTAGE_COORD
, &key
->base
);
2270 if (cs
!= vc4
->prog
.cs
) {
2272 vc4
->dirty
|= VC4_DIRTY_COMPILED_CS
;
2277 vc4_update_compiled_shaders(struct vc4_context
*vc4
, uint8_t prim_mode
)
2279 vc4_update_compiled_fs(vc4
, prim_mode
);
2280 vc4_update_compiled_vs(vc4
, prim_mode
);
2284 fs_cache_hash(const void *key
)
2286 return _mesa_hash_data(key
, sizeof(struct vc4_fs_key
));
2290 vs_cache_hash(const void *key
)
2292 return _mesa_hash_data(key
, sizeof(struct vc4_vs_key
));
2296 fs_cache_compare(const void *key1
, const void *key2
)
2298 return memcmp(key1
, key2
, sizeof(struct vc4_fs_key
)) == 0;
2302 vs_cache_compare(const void *key1
, const void *key2
)
2304 return memcmp(key1
, key2
, sizeof(struct vc4_vs_key
)) == 0;
2308 delete_from_cache_if_matches(struct hash_table
*ht
,
2309 struct hash_entry
*entry
,
2310 struct vc4_uncompiled_shader
*so
)
2312 const struct vc4_key
*key
= entry
->key
;
2314 if (key
->shader_state
== so
) {
2315 struct vc4_compiled_shader
*shader
= entry
->data
;
2316 _mesa_hash_table_remove(ht
, entry
);
2317 vc4_bo_unreference(&shader
->bo
);
2318 ralloc_free(shader
);
2323 vc4_shader_state_delete(struct pipe_context
*pctx
, void *hwcso
)
2325 struct vc4_context
*vc4
= vc4_context(pctx
);
2326 struct vc4_uncompiled_shader
*so
= hwcso
;
2328 struct hash_entry
*entry
;
2329 hash_table_foreach(vc4
->fs_cache
, entry
)
2330 delete_from_cache_if_matches(vc4
->fs_cache
, entry
, so
);
2331 hash_table_foreach(vc4
->vs_cache
, entry
)
2332 delete_from_cache_if_matches(vc4
->vs_cache
, entry
, so
);
2334 free((void *)so
->base
.tokens
);
2339 vc4_fp_state_bind(struct pipe_context
*pctx
, void *hwcso
)
2341 struct vc4_context
*vc4
= vc4_context(pctx
);
2342 vc4
->prog
.bind_fs
= hwcso
;
2343 vc4
->dirty
|= VC4_DIRTY_UNCOMPILED_FS
;
2347 vc4_vp_state_bind(struct pipe_context
*pctx
, void *hwcso
)
2349 struct vc4_context
*vc4
= vc4_context(pctx
);
2350 vc4
->prog
.bind_vs
= hwcso
;
2351 vc4
->dirty
|= VC4_DIRTY_UNCOMPILED_VS
;
2355 vc4_program_init(struct pipe_context
*pctx
)
2357 struct vc4_context
*vc4
= vc4_context(pctx
);
2359 pctx
->create_vs_state
= vc4_shader_state_create
;
2360 pctx
->delete_vs_state
= vc4_shader_state_delete
;
2362 pctx
->create_fs_state
= vc4_shader_state_create
;
2363 pctx
->delete_fs_state
= vc4_shader_state_delete
;
2365 pctx
->bind_fs_state
= vc4_fp_state_bind
;
2366 pctx
->bind_vs_state
= vc4_vp_state_bind
;
2368 vc4
->fs_cache
= _mesa_hash_table_create(pctx
, fs_cache_hash
,
2370 vc4
->vs_cache
= _mesa_hash_table_create(pctx
, vs_cache_hash
,
2375 vc4_program_fini(struct pipe_context
*pctx
)
2377 struct vc4_context
*vc4
= vc4_context(pctx
);
2379 struct hash_entry
*entry
;
2380 hash_table_foreach(vc4
->fs_cache
, entry
) {
2381 struct vc4_compiled_shader
*shader
= entry
->data
;
2382 vc4_bo_unreference(&shader
->bo
);
2383 ralloc_free(shader
);
2384 _mesa_hash_table_remove(vc4
->fs_cache
, entry
);
2387 hash_table_foreach(vc4
->vs_cache
, entry
) {
2388 struct vc4_compiled_shader
*shader
= entry
->data
;
2389 vc4_bo_unreference(&shader
->bo
);
2390 ralloc_free(shader
);
2391 _mesa_hash_table_remove(vc4
->vs_cache
, entry
);