2 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
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:
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
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 * Rob Clark <robclark@freedesktop.org>
27 #include "util/u_atomic.h"
28 #include "util/u_string.h"
29 #include "util/u_memory.h"
30 #include "util/format/u_format.h"
32 #include "drm/freedreno_drmif.h"
34 #include "ir3_shader.h"
35 #include "ir3_compiler.h"
39 ir3_glsl_type_size(const struct glsl_type
*type
, bool bindless
)
41 return glsl_count_attribute_slots(type
, false);
44 /* for vertex shader, the inputs are loaded into registers before the shader
45 * is executed, so max_regs from the shader instructions might not properly
46 * reflect the # of registers actually used, especially in case passthrough
49 * Likewise, for fragment shader, we can have some regs which are passed
50 * input values but never touched by the resulting shader (ie. as result
51 * of dead code elimination or simply because we don't know how to turn
55 fixup_regfootprint(struct ir3_shader_variant
*v
)
59 for (i
= 0; i
< v
->inputs_count
; i
++) {
60 /* skip frag inputs fetch via bary.f since their reg's are
61 * not written by gpu before shader starts (and in fact the
62 * regid's might not even be valid)
64 if (v
->inputs
[i
].bary
)
67 /* ignore high regs that are global to all threads in a warp
68 * (they exist by default) (a5xx+)
70 if (v
->inputs
[i
].regid
>= regid(48,0))
73 if (v
->inputs
[i
].compmask
) {
74 unsigned n
= util_last_bit(v
->inputs
[i
].compmask
) - 1;
75 int32_t regid
= v
->inputs
[i
].regid
+ n
;
76 if (v
->inputs
[i
].half
) {
78 v
->info
.max_half_reg
= MAX2(v
->info
.max_half_reg
, regid
>> 2);
80 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
>> 3);
83 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
>> 2);
88 for (i
= 0; i
< v
->outputs_count
; i
++) {
89 /* for ex, VS shaders with tess don't have normal varying outs: */
90 if (!VALIDREG(v
->outputs
[i
].regid
))
92 int32_t regid
= v
->outputs
[i
].regid
+ 3;
93 if (v
->outputs
[i
].half
) {
95 v
->info
.max_half_reg
= MAX2(v
->info
.max_half_reg
, regid
>> 2);
97 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
>> 3);
100 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
>> 2);
104 for (i
= 0; i
< v
->num_sampler_prefetch
; i
++) {
105 unsigned n
= util_last_bit(v
->sampler_prefetch
[i
].wrmask
) - 1;
106 int32_t regid
= v
->sampler_prefetch
[i
].dst
+ n
;
107 if (v
->sampler_prefetch
[i
].half_precision
) {
108 if (!v
->mergedregs
) {
109 v
->info
.max_half_reg
= MAX2(v
->info
.max_half_reg
, regid
>> 2);
111 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
>> 3);
114 v
->info
.max_reg
= MAX2(v
->info
.max_reg
, regid
>> 2);
119 /* wrapper for ir3_assemble() which does some info fixup based on
120 * shader state. Non-static since used by ir3_cmdline too.
122 void * ir3_shader_assemble(struct ir3_shader_variant
*v
)
124 unsigned gpu_id
= v
->shader
->compiler
->gpu_id
;
127 bin
= ir3_assemble(v
);
132 v
->instrlen
= v
->info
.sizedwords
/ (2 * 16);
134 v
->instrlen
= v
->info
.sizedwords
/ (2 * 4);
137 /* NOTE: if relative addressing is used, we set constlen in
138 * the compiler (to worst-case value) since we don't know in
139 * the assembler what the max addr reg value can be:
141 v
->constlen
= MAX2(v
->constlen
, v
->info
.max_const
+ 1);
143 fixup_regfootprint(v
);
149 assemble_variant(struct ir3_shader_variant
*v
)
151 v
->bin
= ir3_shader_assemble(v
);
153 if (shader_debug_enabled(v
->shader
->type
)) {
154 fprintf(stdout
, "Native code for unnamed %s shader %s:\n",
155 ir3_shader_stage(v
), v
->shader
->nir
->info
.name
);
156 if (v
->shader
->type
== MESA_SHADER_FRAGMENT
)
157 fprintf(stdout
, "SIMD0\n");
158 ir3_shader_disasm(v
, v
->bin
, stdout
);
161 /* no need to keep the ir around beyond this point: */
167 * For creating normal shader variants, 'nonbinning' is NULL. For
168 * creating binning pass shader, it is link to corresponding normal
169 * (non-binning) variant.
171 static struct ir3_shader_variant
*
172 create_variant(struct ir3_shader
*shader
, const struct ir3_shader_key
*key
,
173 struct ir3_shader_variant
*nonbinning
)
175 struct ir3_shader_variant
*v
= rzalloc_size(shader
, sizeof(*v
));
181 v
->id
= ++shader
->variant_count
;
183 v
->binning_pass
= !!nonbinning
;
184 v
->nonbinning
= nonbinning
;
186 v
->type
= shader
->type
;
188 if (shader
->compiler
->gpu_id
>= 600) {
190 case MESA_SHADER_TESS_CTRL
:
191 case MESA_SHADER_TESS_EVAL
:
192 v
->mergedregs
= false;
194 case MESA_SHADER_VERTEX
:
195 case MESA_SHADER_GEOMETRY
:
196 /* For VS/GS, normally do mergedregs, but if there is tess
197 * we need to not used MERGEDREGS
199 v
->mergedregs
= !key
->tessellation
;
202 v
->mergedregs
= true;
205 v
->mergedregs
= false;
208 ret
= ir3_compile_shader_nir(shader
->compiler
, v
);
210 debug_error("compile failed!");
216 debug_error("assemble failed!");
227 static inline struct ir3_shader_variant
*
228 shader_variant(struct ir3_shader
*shader
, const struct ir3_shader_key
*key
,
231 struct ir3_shader_variant
*v
;
235 for (v
= shader
->variants
; v
; v
= v
->next
)
236 if (ir3_shader_key_equal(key
, &v
->key
))
239 /* compile new variant if it doesn't exist already: */
240 v
= create_variant(shader
, key
, NULL
);
242 v
->next
= shader
->variants
;
243 shader
->variants
= v
;
250 struct ir3_shader_variant
*
251 ir3_shader_get_variant(struct ir3_shader
*shader
, const struct ir3_shader_key
*key
,
252 bool binning_pass
, bool *created
)
254 mtx_lock(&shader
->variants_lock
);
255 struct ir3_shader_variant
*v
=
256 shader_variant(shader
, key
, created
);
258 if (v
&& binning_pass
) {
260 v
->binning
= create_variant(shader
, key
, v
);
263 mtx_unlock(&shader
->variants_lock
);
266 mtx_unlock(&shader
->variants_lock
);
272 ir3_shader_destroy(struct ir3_shader
*shader
)
274 free(shader
->const_state
.immediates
);
275 ralloc_free(shader
->nir
);
276 mtx_destroy(&shader
->variants_lock
);
281 * Creates a bitmask of the used bits of the shader key by this particular
282 * shader. Used by the gallium driver to skip state-dependent recompiles when
286 ir3_setup_used_key(struct ir3_shader
*shader
)
288 nir_shader
*nir
= shader
->nir
;
289 struct shader_info
*info
= &nir
->info
;
290 struct ir3_shader_key
*key
= &shader
->key_mask
;
292 /* This key flag is just used to make for a cheaper ir3_shader_key_equal
293 * check in the common case.
295 key
->has_per_samp
= true;
297 if (info
->stage
== MESA_SHADER_FRAGMENT
) {
298 key
->fsaturate_s
= ~0;
299 key
->fsaturate_t
= ~0;
300 key
->fsaturate_r
= ~0;
301 key
->fastc_srgb
= ~0;
304 if (info
->inputs_read
& VARYING_BITS_COLOR
) {
305 key
->rasterflat
= true;
306 key
->color_two_side
= true;
309 if ((info
->outputs_written
& ~(FRAG_RESULT_DEPTH
|
310 FRAG_RESULT_STENCIL
|
311 FRAG_RESULT_SAMPLE_MASK
)) != 0) {
312 key
->fclamp_color
= true;
315 /* Only used for deciding on behavior of
316 * nir_intrinsic_load_barycentric_sample
318 key
->msaa
= info
->fs
.uses_sample_qualifier
;
320 key
->tessellation
= ~0;
323 if (info
->outputs_written
& VARYING_BITS_COLOR
)
324 key
->vclamp_color
= true;
326 if (info
->stage
== MESA_SHADER_VERTEX
) {
327 key
->vsaturate_s
= ~0;
328 key
->vsaturate_t
= ~0;
329 key
->vsaturate_r
= ~0;
330 key
->vastc_srgb
= ~0;
337 ir3_shader_from_nir(struct ir3_compiler
*compiler
, nir_shader
*nir
,
338 unsigned reserved_user_consts
, struct ir3_stream_output_info
*stream_output
)
340 struct ir3_shader
*shader
= rzalloc_size(NULL
, sizeof(*shader
));
342 mtx_init(&shader
->variants_lock
, mtx_plain
);
343 shader
->compiler
= compiler
;
344 shader
->id
= p_atomic_inc_return(&shader
->compiler
->shader_count
);
345 shader
->type
= nir
->info
.stage
;
347 memcpy(&shader
->stream_output
, stream_output
, sizeof(shader
->stream_output
));
348 shader
->const_state
.num_reserved_user_consts
= reserved_user_consts
;
350 if (nir
->info
.stage
== MESA_SHADER_GEOMETRY
)
351 NIR_PASS_V(nir
, ir3_nir_lower_gs
);
353 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, ir3_glsl_type_size
,
354 (nir_lower_io_options
)0);
356 if (compiler
->gpu_id
>= 600 &&
357 nir
->info
.stage
== MESA_SHADER_FRAGMENT
&&
358 !(ir3_shader_debug
& IR3_DBG_NOFP16
))
359 NIR_PASS_V(nir
, nir_lower_mediump_outputs
);
361 if (nir
->info
.stage
== MESA_SHADER_FRAGMENT
) {
362 /* NOTE: lower load_barycentric_at_sample first, since it
363 * produces load_barycentric_at_offset:
365 NIR_PASS_V(nir
, ir3_nir_lower_load_barycentric_at_sample
);
366 NIR_PASS_V(nir
, ir3_nir_lower_load_barycentric_at_offset
);
368 NIR_PASS_V(nir
, ir3_nir_move_varying_inputs
);
371 NIR_PASS_V(nir
, nir_lower_io_arrays_to_elements_no_indirects
, false);
373 NIR_PASS_V(nir
, nir_lower_amul
, ir3_glsl_type_size
);
375 /* do first pass optimization, ignoring the key: */
376 ir3_optimize_nir(shader
, nir
);
379 if (ir3_shader_debug
& IR3_DBG_DISASM
) {
380 printf("dump nir%d: type=%d", shader
->id
, shader
->type
);
381 nir_print_shader(shader
->nir
, stdout
);
384 ir3_setup_used_key(shader
);
389 static void dump_reg(FILE *out
, const char *name
, uint32_t r
)
391 if (r
!= regid(63,0)) {
392 const char *reg_type
= (r
& HALF_REG_ID
) ? "hr" : "r";
393 fprintf(out
, "; %s: %s%d.%c\n", name
, reg_type
,
394 (r
& ~HALF_REG_ID
) >> 2, "xyzw"[r
& 0x3]);
398 static void dump_output(FILE *out
, struct ir3_shader_variant
*so
,
399 unsigned slot
, const char *name
)
402 regid
= ir3_find_output_regid(so
, slot
);
403 dump_reg(out
, name
, regid
);
407 input_name(struct ir3_shader_variant
*so
, int i
)
409 if (so
->inputs
[i
].sysval
) {
410 return gl_system_value_name(so
->inputs
[i
].slot
);
411 } else if (so
->type
== MESA_SHADER_VERTEX
) {
412 return gl_vert_attrib_name(so
->inputs
[i
].slot
);
414 return gl_varying_slot_name(so
->inputs
[i
].slot
);
419 output_name(struct ir3_shader_variant
*so
, int i
)
421 if (so
->type
== MESA_SHADER_FRAGMENT
) {
422 return gl_frag_result_name(so
->outputs
[i
].slot
);
424 switch (so
->outputs
[i
].slot
) {
425 case VARYING_SLOT_GS_HEADER_IR3
:
427 case VARYING_SLOT_GS_VERTEX_FLAGS_IR3
:
428 return "GS_VERTEX_FLAGS";
429 case VARYING_SLOT_TCS_HEADER_IR3
:
432 return gl_varying_slot_name(so
->outputs
[i
].slot
);
438 ir3_shader_disasm(struct ir3_shader_variant
*so
, uint32_t *bin
, FILE *out
)
440 struct ir3
*ir
= so
->ir
;
441 struct ir3_register
*reg
;
442 const char *type
= ir3_shader_stage(so
);
446 foreach_input_n (instr
, i
, ir
) {
447 reg
= instr
->regs
[0];
449 fprintf(out
, "@in(%sr%d.%c)\tin%d",
450 (reg
->flags
& IR3_REG_HALF
) ? "h" : "",
451 (regid
>> 2), "xyzw"[regid
& 0x3], i
);
453 if (reg
->wrmask
> 0x1)
454 fprintf(out
, " (wrmask=0x%x)", reg
->wrmask
);
458 /* print pre-dispatch texture fetches: */
459 for (i
= 0; i
< so
->num_sampler_prefetch
; i
++) {
460 const struct ir3_sampler_prefetch
*fetch
= &so
->sampler_prefetch
[i
];
461 fprintf(out
, "@tex(%sr%d.%c)\tsrc=%u, samp=%u, tex=%u, wrmask=0x%x, cmd=%u\n",
462 fetch
->half_precision
? "h" : "",
463 fetch
->dst
>> 2, "xyzw"[fetch
->dst
& 0x3],
464 fetch
->src
, fetch
->samp_id
, fetch
->tex_id
,
465 fetch
->wrmask
, fetch
->cmd
);
468 foreach_output_n (instr
, i
, ir
) {
469 reg
= instr
->regs
[0];
471 fprintf(out
, "@out(%sr%d.%c)\tout%d",
472 (reg
->flags
& IR3_REG_HALF
) ? "h" : "",
473 (regid
>> 2), "xyzw"[regid
& 0x3], i
);
474 if (reg
->wrmask
> 0x1)
475 fprintf(out
, " (wrmask=0x%x)", reg
->wrmask
);
479 const struct ir3_const_state
*const_state
= ir3_const_state(so
);
480 for (i
= 0; i
< const_state
->immediates_count
; i
++) {
481 fprintf(out
, "@const(c%d.x)\t", const_state
->offsets
.immediate
+ i
);
482 fprintf(out
, "0x%08x, 0x%08x, 0x%08x, 0x%08x\n",
483 const_state
->immediates
[i
].val
[0],
484 const_state
->immediates
[i
].val
[1],
485 const_state
->immediates
[i
].val
[2],
486 const_state
->immediates
[i
].val
[3]);
489 disasm_a3xx(bin
, so
->info
.sizedwords
, 0, out
, ir
->compiler
->gpu_id
);
491 fprintf(out
, "; %s: outputs:", type
);
492 for (i
= 0; i
< so
->outputs_count
; i
++) {
493 uint8_t regid
= so
->outputs
[i
].regid
;
494 const char *reg_type
= so
->outputs
[i
].half
? "hr" : "r";
495 fprintf(out
, " %s%d.%c (%s)",
496 reg_type
, (regid
>> 2), "xyzw"[regid
& 0x3],
501 fprintf(out
, "; %s: inputs:", type
);
502 for (i
= 0; i
< so
->inputs_count
; i
++) {
503 uint8_t regid
= so
->inputs
[i
].regid
;
504 fprintf(out
, " r%d.%c (%s slot=%d cm=%x,il=%u,b=%u)",
505 (regid
>> 2), "xyzw"[regid
& 0x3],
508 so
->inputs
[i
].compmask
,
514 /* print generic shader info: */
515 fprintf(out
, "; %s prog %d/%d: %u instr, %u nops, %u non-nops, %u mov, %u cov, %u dwords\n",
516 type
, so
->shader
->id
, so
->id
,
517 so
->info
.instrs_count
,
519 so
->info
.instrs_count
- so
->info
.nops_count
,
520 so
->info
.mov_count
, so
->info
.cov_count
,
521 so
->info
.sizedwords
);
523 fprintf(out
, "; %s prog %d/%d: %u last-baryf, %d half, %d full, %u constlen\n",
524 type
, so
->shader
->id
, so
->id
,
526 so
->info
.max_half_reg
+ 1,
527 so
->info
.max_reg
+ 1,
530 fprintf(out
, "; %s prog %d/%d: %u sstall, %u (ss), %u (sy), %d max_sun, %d loops\n",
531 type
, so
->shader
->id
, so
->id
,
538 /* print shader type specific info: */
540 case MESA_SHADER_VERTEX
:
541 dump_output(out
, so
, VARYING_SLOT_POS
, "pos");
542 dump_output(out
, so
, VARYING_SLOT_PSIZ
, "psize");
544 case MESA_SHADER_FRAGMENT
:
545 dump_reg(out
, "pos (ij_pixel)",
546 ir3_find_sysval_regid(so
, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL
));
547 dump_reg(out
, "pos (ij_centroid)",
548 ir3_find_sysval_regid(so
, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID
));
549 dump_reg(out
, "pos (ij_size)",
550 ir3_find_sysval_regid(so
, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE
));
551 dump_output(out
, so
, FRAG_RESULT_DEPTH
, "posz");
552 if (so
->color0_mrt
) {
553 dump_output(out
, so
, FRAG_RESULT_COLOR
, "color");
555 dump_output(out
, so
, FRAG_RESULT_DATA0
, "data0");
556 dump_output(out
, so
, FRAG_RESULT_DATA1
, "data1");
557 dump_output(out
, so
, FRAG_RESULT_DATA2
, "data2");
558 dump_output(out
, so
, FRAG_RESULT_DATA3
, "data3");
559 dump_output(out
, so
, FRAG_RESULT_DATA4
, "data4");
560 dump_output(out
, so
, FRAG_RESULT_DATA5
, "data5");
561 dump_output(out
, so
, FRAG_RESULT_DATA6
, "data6");
562 dump_output(out
, so
, FRAG_RESULT_DATA7
, "data7");
564 dump_reg(out
, "fragcoord",
565 ir3_find_sysval_regid(so
, SYSTEM_VALUE_FRAG_COORD
));
566 dump_reg(out
, "fragface",
567 ir3_find_sysval_regid(so
, SYSTEM_VALUE_FRONT_FACE
));
578 ir3_shader_outputs(const struct ir3_shader
*so
)
580 return so
->nir
->info
.outputs_written
;