2 * Copyright (c) 2012-2019 Etnaviv Project
3 * Copyright (c) 2019 Zodiac Inflight Innovations
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, sub license,
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
13 * next paragraph) shall be included in all copies or substantial portions
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE 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
22 * DEALINGS IN THE SOFTWARE.
25 * Jonathan Marek <jonathan@marek.ca>
26 * Wladimir J. van der Laan <laanwj@gmail.com>
29 #include "etnaviv_compiler.h"
30 #include "etnaviv_asm.h"
31 #include "etnaviv_context.h"
32 #include "etnaviv_debug.h"
33 #include "etnaviv_disasm.h"
34 #include "etnaviv_uniforms.h"
35 #include "etnaviv_util.h"
38 #include "util/u_memory.h"
39 #include "util/register_allocate.h"
40 #include "compiler/nir/nir_builder.h"
41 #include "compiler/nir/nir_worklist.h"
43 #include "tgsi/tgsi_strings.h"
44 #include "util/u_half.h"
48 #define is_fs(c) ((c)->nir->info.stage == MESA_SHADER_FRAGMENT)
49 const struct etna_specs
*specs
;
50 struct etna_shader_variant
*variant
;
52 /* block # to instr index */
56 int inst_ptr
; /* current instruction pointer */
57 struct etna_inst code
[ETNA_MAX_INSTRUCTIONS
* ETNA_INST_SIZE
];
59 /* There was an error during compilation */
63 #define compile_error(ctx, args...) ({ \
69 /* io related lowering
70 * run after lower_int_to_float because it adds i2f/f2i ops
73 etna_lower_io(nir_shader
*shader
, struct etna_shader_variant
*v
)
75 nir_foreach_function(function
, shader
) {
77 nir_builder_init(&b
, function
->impl
);
79 nir_foreach_block(block
, function
->impl
) {
80 nir_foreach_instr_safe(instr
, block
) {
81 if (instr
->type
== nir_instr_type_intrinsic
) {
82 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
84 switch (intr
->intrinsic
) {
85 case nir_intrinsic_load_front_face
: {
86 /* HW front_face is 0.0/1.0, not 0/~0u for bool
87 * lower with a comparison with 0
89 intr
->dest
.ssa
.bit_size
= 32;
91 b
.cursor
= nir_after_instr(instr
);
93 nir_ssa_def
*ssa
= nir_ine(&b
, &intr
->dest
.ssa
, nir_imm_int(&b
, 0));
95 nir_instr_as_alu(ssa
->parent_instr
)->op
= nir_op_ieq
;
97 nir_ssa_def_rewrite_uses_after(&intr
->dest
.ssa
,
101 case nir_intrinsic_store_deref
: {
102 if (shader
->info
.stage
!= MESA_SHADER_FRAGMENT
|| !v
->key
.frag_rb_swap
)
105 nir_deref_instr
*deref
= nir_src_as_deref(intr
->src
[0]);
106 assert(deref
->deref_type
== nir_deref_type_var
);
108 if (deref
->var
->data
.location
!= FRAG_RESULT_COLOR
&&
109 deref
->var
->data
.location
!= FRAG_RESULT_DATA0
)
112 b
.cursor
= nir_before_instr(instr
);
114 nir_ssa_def
*ssa
= nir_mov(&b
, intr
->src
[1].ssa
);
115 nir_alu_instr
*alu
= nir_instr_as_alu(ssa
->parent_instr
);
116 alu
->src
[0].swizzle
[0] = 2;
117 alu
->src
[0].swizzle
[2] = 0;
118 nir_instr_rewrite_src(instr
, &intr
->src
[1], nir_src_for_ssa(ssa
));
120 case nir_intrinsic_load_uniform
: {
121 /* multiply by 16 and convert to int */
122 b
.cursor
= nir_before_instr(instr
);
123 nir_ssa_def
*ssa
= nir_imul(&b
, intr
->src
[0].ssa
, nir_imm_int(&b
, 16));
124 nir_instr_rewrite_src(instr
, &intr
->src
[0], nir_src_for_ssa(ssa
));
131 if (instr
->type
!= nir_instr_type_tex
)
134 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
135 nir_src
*coord
= NULL
;
136 nir_src
*lod_bias
= NULL
;
137 unsigned lod_bias_idx
;
139 assert(tex
->sampler_index
== tex
->texture_index
);
141 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
142 switch (tex
->src
[i
].src_type
) {
143 case nir_tex_src_coord
:
144 coord
= &tex
->src
[i
].src
;
146 case nir_tex_src_bias
:
147 case nir_tex_src_lod
:
149 lod_bias
= &tex
->src
[i
].src
;
158 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) {
159 /* use a dummy load_uniform here to represent texcoord scale */
160 b
.cursor
= nir_before_instr(instr
);
161 nir_intrinsic_instr
*load
=
162 nir_intrinsic_instr_create(b
.shader
, nir_intrinsic_load_uniform
);
163 nir_intrinsic_set_base(load
, ~tex
->sampler_index
);
164 load
->num_components
= 2;
165 load
->src
[0] = nir_src_for_ssa(nir_imm_float(&b
, 0.0f
));
166 nir_ssa_dest_init(&load
->instr
, &load
->dest
, 2, 32, NULL
);
167 nir_intrinsic_set_type(load
, nir_type_float
);
169 nir_builder_instr_insert(&b
, &load
->instr
);
171 nir_ssa_def
*new_coord
= nir_fmul(&b
, coord
->ssa
, &load
->dest
.ssa
);
172 nir_instr_rewrite_src(&tex
->instr
, coord
, nir_src_for_ssa(new_coord
));
175 /* pre HALTI5 needs texture sources in a single source */
177 if (!lod_bias
|| v
->shader
->specs
->halti
>= 5)
180 assert(coord
&& lod_bias
&& tex
->coord_components
< 4);
182 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, nir_op_vec4
);
183 for (unsigned i
= 0; i
< tex
->coord_components
; i
++) {
184 vec
->src
[i
].src
= nir_src_for_ssa(coord
->ssa
);
185 vec
->src
[i
].swizzle
[0] = i
;
187 for (unsigned i
= tex
->coord_components
; i
< 4; i
++)
188 vec
->src
[i
].src
= nir_src_for_ssa(lod_bias
->ssa
);
190 vec
->dest
.write_mask
= 0xf;
191 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, 4, 32, NULL
);
193 nir_tex_instr_remove_src(tex
, lod_bias_idx
);
194 nir_instr_rewrite_src(&tex
->instr
, coord
, nir_src_for_ssa(&vec
->dest
.dest
.ssa
));
195 tex
->coord_components
= 4;
197 nir_instr_insert_before(&tex
->instr
, &vec
->instr
);
204 etna_alu_to_scalar_filter_cb(const nir_instr
*instr
, const void *data
)
206 const struct etna_specs
*specs
= data
;
208 if (instr
->type
!= nir_instr_type_alu
)
211 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
223 /* TODO: can do better than alu_to_scalar for vector compares */
224 case nir_op_b32all_fequal2
:
225 case nir_op_b32all_fequal3
:
226 case nir_op_b32all_fequal4
:
227 case nir_op_b32any_fnequal2
:
228 case nir_op_b32any_fnequal3
:
229 case nir_op_b32any_fnequal4
:
230 case nir_op_b32all_iequal2
:
231 case nir_op_b32all_iequal3
:
232 case nir_op_b32all_iequal4
:
233 case nir_op_b32any_inequal2
:
234 case nir_op_b32any_inequal3
:
235 case nir_op_b32any_inequal4
:
238 if (!specs
->has_halti2_instructions
)
249 etna_lower_alu_impl(nir_function_impl
*impl
, struct etna_compile
*c
)
251 nir_shader
*shader
= impl
->function
->shader
;
254 nir_builder_init(&b
, impl
);
256 /* in a seperate loop so we can apply the multiple-uniform logic to the new fmul */
257 nir_foreach_block(block
, impl
) {
258 nir_foreach_instr_safe(instr
, block
) {
259 if (instr
->type
!= nir_instr_type_alu
)
262 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
263 /* multiply sin/cos src by constant
264 * TODO: do this earlier (but it breaks const_prop opt)
266 if (alu
->op
== nir_op_fsin
|| alu
->op
== nir_op_fcos
) {
267 b
.cursor
= nir_before_instr(instr
);
269 nir_ssa_def
*imm
= c
->specs
->has_new_transcendentals
?
270 nir_imm_float(&b
, 1.0 / M_PI
) :
271 nir_imm_float(&b
, 2.0 / M_PI
);
273 nir_instr_rewrite_src(instr
, &alu
->src
[0].src
,
274 nir_src_for_ssa(nir_fmul(&b
, alu
->src
[0].src
.ssa
, imm
)));
277 /* change transcendental ops to vec2 and insert vec1 mul for the result
278 * TODO: do this earlier (but it breaks with optimizations)
280 if (c
->specs
->has_new_transcendentals
&& (
281 alu
->op
== nir_op_fdiv
|| alu
->op
== nir_op_flog2
||
282 alu
->op
== nir_op_fsin
|| alu
->op
== nir_op_fcos
)) {
283 nir_ssa_def
*ssa
= &alu
->dest
.dest
.ssa
;
285 assert(ssa
->num_components
== 1);
287 nir_alu_instr
*mul
= nir_alu_instr_create(shader
, nir_op_fmul
);
288 mul
->src
[0].src
= mul
->src
[1].src
= nir_src_for_ssa(ssa
);
289 mul
->src
[1].swizzle
[0] = 1;
291 mul
->dest
.write_mask
= 1;
292 nir_ssa_dest_init(&mul
->instr
, &mul
->dest
.dest
, 1, 32, NULL
);
294 ssa
->num_components
= 2;
296 mul
->dest
.saturate
= alu
->dest
.saturate
;
297 alu
->dest
.saturate
= 0;
299 nir_instr_insert_after(instr
, &mul
->instr
);
301 nir_ssa_def_rewrite_uses_after(ssa
, nir_src_for_ssa(&mul
->dest
.dest
.ssa
), &mul
->instr
);
307 static void etna_lower_alu(nir_shader
*shader
, struct etna_compile
*c
)
309 nir_foreach_function(function
, shader
) {
311 etna_lower_alu_impl(function
->impl
, c
);
316 emit_inst(struct etna_compile
*c
, struct etna_inst
*inst
)
318 c
->code
[c
->inst_ptr
++] = *inst
;
321 /* to map nir srcs should to etna_inst srcs */
323 SRC_0_1_2
= (0 << 0) | (1 << 2) | (2 << 4),
324 SRC_0_1_X
= (0 << 0) | (1 << 2) | (3 << 4),
325 SRC_0_X_X
= (0 << 0) | (3 << 2) | (3 << 4),
326 SRC_0_X_1
= (0 << 0) | (3 << 2) | (1 << 4),
327 SRC_0_1_0
= (0 << 0) | (1 << 2) | (0 << 4),
328 SRC_X_X_0
= (3 << 0) | (3 << 2) | (0 << 4),
329 SRC_0_X_0
= (0 << 0) | (3 << 2) | (0 << 4),
332 /* info to translate a nir op to etna_inst */
333 struct etna_op_info
{
334 uint8_t opcode
; /* INST_OPCODE_ */
335 uint8_t src
; /* SRC_ enum */
336 uint8_t cond
; /* INST_CONDITION_ */
337 uint8_t type
; /* INST_TYPE_ */
340 static const struct etna_op_info etna_ops
[] = {
341 [0 ... nir_num_opcodes
- 1] = {0xff},
344 #define OPCT(nir, op, src, cond, type) [nir_op_##nir] = { \
347 INST_CONDITION_##cond, \
350 #define OPC(nir, op, src, cond) OPCT(nir, op, src, cond, F32)
351 #define IOPC(nir, op, src, cond) OPCT(nir, op, src, cond, S32)
352 #define UOPC(nir, op, src, cond) OPCT(nir, op, src, cond, U32)
353 #define OP(nir, op, src) OPC(nir, op, src, TRUE)
354 #define IOP(nir, op, src) IOPC(nir, op, src, TRUE)
355 #define UOP(nir, op, src) UOPC(nir, op, src, TRUE)
356 OP(mov
, MOV
, X_X_0
), OP(fneg
, MOV
, X_X_0
), OP(fabs
, MOV
, X_X_0
), OP(fsat
, MOV
, X_X_0
),
357 OP(fmul
, MUL
, 0_1_X
), OP(fadd
, ADD
, 0_X_1
), OP(ffma
, MAD
, 0_1_2
),
358 OP(fdot2
, DP2
, 0_1_X
), OP(fdot3
, DP3
, 0_1_X
), OP(fdot4
, DP4
, 0_1_X
),
359 OPC(fmin
, SELECT
, 0_1_0
, GT
), OPC(fmax
, SELECT
, 0_1_0
, LT
),
360 OP(ffract
, FRC
, X_X_0
), OP(frcp
, RCP
, X_X_0
), OP(frsq
, RSQ
, X_X_0
),
361 OP(fsqrt
, SQRT
, X_X_0
), OP(fsin
, SIN
, X_X_0
), OP(fcos
, COS
, X_X_0
),
362 OP(fsign
, SIGN
, X_X_0
), OP(ffloor
, FLOOR
, X_X_0
), OP(fceil
, CEIL
, X_X_0
),
363 OP(flog2
, LOG
, X_X_0
), OP(fexp2
, EXP
, X_X_0
),
364 OPC(seq
, SET
, 0_1_X
, EQ
), OPC(sne
, SET
, 0_1_X
, NE
), OPC(sge
, SET
, 0_1_X
, GE
), OPC(slt
, SET
, 0_1_X
, LT
),
365 OPC(fcsel
, SELECT
, 0_1_2
, NZ
),
366 OP(fdiv
, DIV
, 0_1_X
),
367 OP(fddx
, DSX
, 0_X_0
), OP(fddy
, DSY
, 0_X_0
),
370 IOP(i2f32
, I2F
, 0_X_X
),
371 UOP(u2f32
, I2F
, 0_X_X
),
372 IOP(f2i32
, F2I
, 0_X_X
),
373 UOP(f2u32
, F2I
, 0_X_X
),
374 UOP(b2f32
, AND
, 0_X_X
), /* AND with fui(1.0f) */
375 UOP(b2i32
, AND
, 0_X_X
), /* AND with 1 */
376 OPC(f2b32
, CMP
, 0_X_X
, NE
), /* != 0.0 */
377 UOPC(i2b32
, CMP
, 0_X_X
, NE
), /* != 0 */
380 IOP(iadd
, ADD
, 0_X_1
),
381 IOP(imul
, IMULLO0
, 0_1_X
),
382 /* IOP(imad, IMADLO0, 0_1_2), */
383 IOP(ineg
, ADD
, X_X_0
), /* ADD 0, -x */
384 IOP(iabs
, IABS
, X_X_0
),
385 IOP(isign
, SIGN
, X_X_0
),
386 IOPC(imin
, SELECT
, 0_1_0
, GT
),
387 IOPC(imax
, SELECT
, 0_1_0
, LT
),
388 UOPC(umin
, SELECT
, 0_1_0
, GT
),
389 UOPC(umax
, SELECT
, 0_1_0
, LT
),
392 UOPC(b32csel
, SELECT
, 0_1_2
, NZ
),
394 /* compare with int result */
395 OPC(feq32
, CMP
, 0_1_X
, EQ
),
396 OPC(fne32
, CMP
, 0_1_X
, NE
),
397 OPC(fge32
, CMP
, 0_1_X
, GE
),
398 OPC(flt32
, CMP
, 0_1_X
, LT
),
399 IOPC(ieq32
, CMP
, 0_1_X
, EQ
),
400 IOPC(ine32
, CMP
, 0_1_X
, NE
),
401 IOPC(ige32
, CMP
, 0_1_X
, GE
),
402 IOPC(ilt32
, CMP
, 0_1_X
, LT
),
403 UOPC(uge32
, CMP
, 0_1_X
, GE
),
404 UOPC(ult32
, CMP
, 0_1_X
, LT
),
408 IOP(iand
, AND
, 0_X_1
),
409 IOP(ixor
, XOR
, 0_X_1
),
410 IOP(inot
, NOT
, X_X_0
),
411 IOP(ishl
, LSHIFT
, 0_X_1
),
412 IOP(ishr
, RSHIFT
, 0_X_1
),
413 UOP(ushr
, RSHIFT
, 0_X_1
),
417 etna_emit_block_start(struct etna_compile
*c
, unsigned block
)
419 c
->block_ptr
[block
] = c
->inst_ptr
;
423 etna_emit_alu(struct etna_compile
*c
, nir_op op
, struct etna_inst_dst dst
,
424 struct etna_inst_src src
[3], bool saturate
)
426 struct etna_op_info ei
= etna_ops
[op
];
427 unsigned swiz_scalar
= INST_SWIZ_BROADCAST(ffs(dst
.write_mask
) - 1);
429 assert(ei
.opcode
!= 0xff);
431 struct etna_inst inst
= {
444 if (c
->specs
->has_new_transcendentals
)
452 /* scalar instructions we want src to be in x component */
453 src
[0].swiz
= inst_swiz_compose(src
[0].swiz
, swiz_scalar
);
454 src
[1].swiz
= inst_swiz_compose(src
[1].swiz
, swiz_scalar
);
456 /* deal with instructions which don't have 1:1 mapping */
458 inst
.src
[2] = etna_immediate_float(1.0f
);
461 inst
.src
[2] = etna_immediate_int(1);
464 inst
.src
[1] = etna_immediate_float(0.0f
);
467 inst
.src
[1] = etna_immediate_int(0);
470 inst
.src
[0] = etna_immediate_int(0);
477 /* set the "true" value for CMP instructions */
478 if (inst
.opcode
== INST_OPCODE_CMP
)
479 inst
.src
[2] = etna_immediate_int(-1);
481 for (unsigned j
= 0; j
< 3; j
++) {
482 unsigned i
= ((ei
.src
>> j
*2) & 3);
484 inst
.src
[j
] = src
[i
];
491 etna_emit_tex(struct etna_compile
*c
, nir_texop op
, unsigned texid
, unsigned dst_swiz
,
492 struct etna_inst_dst dst
, struct etna_inst_src coord
,
493 struct etna_inst_src lod_bias
)
495 struct etna_inst inst
= {
497 .tex
.id
= texid
+ (is_fs(c
) ? 0 : c
->specs
->vertex_sampler_offset
),
498 .tex
.swiz
= dst_swiz
,
503 inst
.src
[1] = lod_bias
;
506 case nir_texop_tex
: inst
.opcode
= INST_OPCODE_TEXLD
; break;
507 case nir_texop_txb
: inst
.opcode
= INST_OPCODE_TEXLDB
; break;
508 case nir_texop_txl
: inst
.opcode
= INST_OPCODE_TEXLDL
; break;
517 etna_emit_jump(struct etna_compile
*c
, unsigned block
, struct etna_inst_src condition
)
519 if (!condition
.use
) {
520 emit_inst(c
, &(struct etna_inst
) {.opcode
= INST_OPCODE_BRANCH
, .imm
= block
});
524 struct etna_inst inst
= {
525 .opcode
= INST_OPCODE_BRANCH
,
526 .cond
= INST_CONDITION_NOT
,
527 .type
= INST_TYPE_U32
,
531 inst
.src
[0].swiz
= INST_SWIZ_BROADCAST(inst
.src
[0].swiz
& 3);
536 etna_emit_discard(struct etna_compile
*c
, struct etna_inst_src condition
)
538 if (!condition
.use
) {
539 emit_inst(c
, &(struct etna_inst
) { .opcode
= INST_OPCODE_TEXKILL
});
543 struct etna_inst inst
= {
544 .opcode
= INST_OPCODE_TEXKILL
,
545 .cond
= INST_CONDITION_NZ
,
546 .type
= (c
->specs
->halti
< 2) ? INST_TYPE_F32
: INST_TYPE_U32
,
549 inst
.src
[0].swiz
= INST_SWIZ_BROADCAST(inst
.src
[0].swiz
& 3);
554 etna_emit_output(struct etna_compile
*c
, nir_variable
*var
, struct etna_inst_src src
)
556 struct etna_shader_io_file
*sf
= &c
->variant
->outfile
;
559 switch (var
->data
.location
) {
560 case FRAG_RESULT_COLOR
:
561 case FRAG_RESULT_DATA0
: /* DATA0 is used by gallium shaders for color */
562 c
->variant
->ps_color_out_reg
= src
.reg
;
564 case FRAG_RESULT_DEPTH
:
565 c
->variant
->ps_depth_out_reg
= src
.reg
;
568 unreachable("Unsupported fs output");
573 switch (var
->data
.location
) {
574 case VARYING_SLOT_POS
:
575 c
->variant
->vs_pos_out_reg
= src
.reg
;
577 case VARYING_SLOT_PSIZ
:
578 c
->variant
->vs_pointsize_out_reg
= src
.reg
;
581 sf
->reg
[sf
->num_reg
].reg
= src
.reg
;
582 sf
->reg
[sf
->num_reg
].slot
= var
->data
.location
;
583 sf
->reg
[sf
->num_reg
].num_components
= glsl_get_components(var
->type
);
590 etna_emit_load_ubo(struct etna_compile
*c
, struct etna_inst_dst dst
,
591 struct etna_inst_src src
, struct etna_inst_src base
)
593 /* convert float offset back to integer */
594 if (c
->specs
->halti
< 2) {
595 emit_inst(c
, &(struct etna_inst
) {
596 .opcode
= INST_OPCODE_F2I
,
597 .type
= INST_TYPE_U32
,
602 emit_inst(c
, &(struct etna_inst
) {
603 .opcode
= INST_OPCODE_LOAD
,
604 .type
= INST_TYPE_U32
,
608 .rgroup
= INST_RGROUP_TEMP
,
610 .swiz
= INST_SWIZ_BROADCAST(ffs(dst
.write_mask
) - 1)
618 emit_inst(c
, &(struct etna_inst
) {
619 .opcode
= INST_OPCODE_LOAD
,
620 .type
= INST_TYPE_U32
,
627 #define OPT(nir, pass, ...) ({ \
628 bool this_progress = false; \
629 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
632 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
635 etna_optimize_loop(nir_shader
*s
)
641 OPT_V(s
, nir_lower_vars_to_ssa
);
642 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
643 progress
|= OPT(s
, nir_copy_prop
);
644 progress
|= OPT(s
, nir_opt_dce
);
645 progress
|= OPT(s
, nir_opt_cse
);
646 progress
|= OPT(s
, nir_opt_peephole_select
, 16, true, true);
647 progress
|= OPT(s
, nir_opt_intrinsics
);
648 progress
|= OPT(s
, nir_opt_algebraic
);
649 progress
|= OPT(s
, nir_opt_constant_folding
);
650 progress
|= OPT(s
, nir_opt_dead_cf
);
651 if (OPT(s
, nir_opt_trivial_continues
)) {
653 /* If nir_opt_trivial_continues makes progress, then we need to clean
654 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
657 OPT(s
, nir_copy_prop
);
660 progress
|= OPT(s
, nir_opt_loop_unroll
, nir_var_all
);
661 progress
|= OPT(s
, nir_opt_if
, false);
662 progress
|= OPT(s
, nir_opt_remove_phis
);
663 progress
|= OPT(s
, nir_opt_undef
);
669 etna_glsl_type_size(const struct glsl_type
*type
, bool bindless
)
671 return glsl_count_attribute_slots(type
, false);
675 copy_uniform_state_to_shader(struct etna_shader_variant
*sobj
, uint64_t *consts
, unsigned count
)
677 struct etna_shader_uniform_info
*uinfo
= &sobj
->uniforms
;
679 uinfo
->imm_count
= count
* 4;
680 uinfo
->imm_data
= MALLOC(uinfo
->imm_count
* sizeof(*uinfo
->imm_data
));
681 uinfo
->imm_contents
= MALLOC(uinfo
->imm_count
* sizeof(*uinfo
->imm_contents
));
683 for (unsigned i
= 0; i
< uinfo
->imm_count
; i
++) {
684 uinfo
->imm_data
[i
] = consts
[i
];
685 uinfo
->imm_contents
[i
] = consts
[i
] >> 32;
688 etna_set_shader_uniforms_dirty_flags(sobj
);
691 #include "etnaviv_compiler_nir_emit.h"
694 etna_compile_shader_nir(struct etna_shader_variant
*v
)
699 struct etna_compile
*c
= CALLOC_STRUCT(etna_compile
);
704 c
->specs
= v
->shader
->specs
;
705 c
->nir
= nir_shader_clone(NULL
, v
->shader
->nir
);
707 nir_shader
*s
= c
->nir
;
708 const struct etna_specs
*specs
= c
->specs
;
710 v
->stage
= s
->info
.stage
;
711 v
->num_loops
= 0; /* TODO */
712 v
->vs_id_in_reg
= -1;
713 v
->vs_pos_out_reg
= -1;
714 v
->vs_pointsize_out_reg
= -1;
715 v
->ps_color_out_reg
= 0; /* 0 for shader that doesn't write fragcolor.. */
716 v
->ps_depth_out_reg
= -1;
718 /* setup input linking */
719 struct etna_shader_io_file
*sf
= &v
->infile
;
720 if (s
->info
.stage
== MESA_SHADER_VERTEX
) {
721 nir_foreach_variable(var
, &s
->inputs
) {
722 unsigned idx
= var
->data
.driver_location
;
723 sf
->reg
[idx
].reg
= idx
;
724 sf
->reg
[idx
].slot
= var
->data
.location
;
725 sf
->reg
[idx
].num_components
= glsl_get_components(var
->type
);
726 sf
->num_reg
= MAX2(sf
->num_reg
, idx
+1);
730 nir_foreach_variable(var
, &s
->inputs
) {
731 unsigned idx
= var
->data
.driver_location
;
732 sf
->reg
[idx
].reg
= idx
+ 1;
733 sf
->reg
[idx
].slot
= var
->data
.location
;
734 sf
->reg
[idx
].num_components
= glsl_get_components(var
->type
);
735 sf
->num_reg
= MAX2(sf
->num_reg
, idx
+1);
738 assert(sf
->num_reg
== count
);
741 NIR_PASS_V(s
, nir_lower_io
, ~nir_var_shader_out
, etna_glsl_type_size
,
742 (nir_lower_io_options
)0);
744 OPT_V(s
, nir_lower_regs_to_ssa
);
745 OPT_V(s
, nir_lower_vars_to_ssa
);
746 OPT_V(s
, nir_lower_indirect_derefs
, nir_var_all
);
747 OPT_V(s
, nir_lower_tex
, &(struct nir_lower_tex_options
) { .lower_txp
= ~0u });
748 OPT_V(s
, nir_lower_alu_to_scalar
, etna_alu_to_scalar_filter_cb
, specs
);
750 etna_optimize_loop(s
);
752 OPT_V(s
, etna_lower_io
, v
);
754 /* lower pre-halti2 to float (halti0 has integers, but only scalar..) */
755 if (c
->specs
->halti
< 2) {
756 /* use opt_algebraic between int_to_float and boot_to_float because
757 * int_to_float emits ftrunc, and ftrunc lowering generates bool ops
759 OPT_V(s
, nir_lower_int_to_float
);
760 OPT_V(s
, nir_opt_algebraic
);
761 OPT_V(s
, nir_lower_bool_to_float
);
763 OPT_V(s
, nir_lower_idiv
);
764 OPT_V(s
, nir_lower_bool_to_int32
);
767 etna_optimize_loop(s
);
769 if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS
))
770 nir_print_shader(s
, stdout
);
772 while( OPT(s
, nir_opt_vectorize
) );
773 OPT_V(s
, nir_lower_alu_to_scalar
, etna_alu_to_scalar_filter_cb
, specs
);
775 NIR_PASS_V(s
, nir_remove_dead_variables
, nir_var_function_temp
);
776 NIR_PASS_V(s
, nir_opt_algebraic_late
);
778 NIR_PASS_V(s
, nir_move_vec_src_uses_to_dest
);
779 NIR_PASS_V(s
, nir_copy_prop
);
780 /* only HW supported integer source mod is ineg for iadd instruction (?) */
781 NIR_PASS_V(s
, nir_lower_to_source_mods
, ~nir_lower_int_source_mods
);
782 /* need copy prop after uses_to_dest, and before src mods: see
783 * dEQP-GLES2.functional.shaders.random.all_features.fragment.95
786 NIR_PASS_V(s
, nir_opt_dce
);
788 NIR_PASS_V(s
, etna_lower_alu
, c
);
790 if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS
))
791 nir_print_shader(s
, stdout
);
793 uint64_t consts
[ETNA_MAX_IMM
] = {};
795 unsigned block_ptr
[nir_shader_get_entrypoint(s
)->num_blocks
];
796 c
->block_ptr
= block_ptr
;
797 struct emit_options options
= {
798 .max_temps
= ETNA_MAX_TEMPS
,
799 .max_consts
= ETNA_MAX_IMM
/ 4,
800 .id_reg
= sf
->num_reg
,
801 .single_const_src
= c
->specs
->halti
< 5,
802 .etna_new_transcendentals
= c
->specs
->has_new_transcendentals
,
803 .no_integers
= c
->specs
->halti
< 2,
809 ASSERTED
bool ok
= emit_shader(c
->nir
, &options
, &v
->num_temps
, &num_consts
);
812 /* empty shader, emit NOP */
814 emit_inst(c
, &(struct etna_inst
) { .opcode
= INST_OPCODE_NOP
});
816 /* assemble instructions, fixing up labels */
817 uint32_t *code
= MALLOC(c
->inst_ptr
* 16 + 1024);
818 for (unsigned i
= 0; i
< c
->inst_ptr
; i
++) {
819 struct etna_inst
*inst
= &c
->code
[i
];
820 if (inst
->opcode
== INST_OPCODE_BRANCH
)
821 inst
->imm
= block_ptr
[inst
->imm
];
823 inst
->halti5
= specs
->halti
>= 5;
824 etna_assemble(&code
[i
* 4], inst
);
827 v
->code_size
= c
->inst_ptr
* 4;
829 v
->needs_icache
= c
->inst_ptr
> specs
->max_instructions
;
831 copy_uniform_state_to_shader(v
, consts
, num_consts
);
833 if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
834 v
->input_count_unk8
= 31; /* XXX what is this */
835 assert(v
->ps_depth_out_reg
<= 0);
841 v
->input_count_unk8
= DIV_ROUND_UP(v
->infile
.num_reg
+ 4, 16); /* XXX what is this */
843 /* fill in "mystery meat" load balancing value. This value determines how
844 * work is scheduled between VS and PS
845 * in the unified shader architecture. More precisely, it is determined from
846 * the number of VS outputs, as well as chip-specific
847 * vertex output buffer size, vertex cache size, and the number of shader
850 * XXX this is a conservative estimate, the "optimal" value is only known for
851 * sure at link time because some
852 * outputs may be unused and thus unmapped. Then again, in the general use
853 * case with GLSL the vertex and fragment
854 * shaders are linked already before submitting to Gallium, thus all outputs
857 * note: TGSI compiler counts all outputs (including position and pointsize), here
858 * v->outfile.num_reg only counts varyings, +1 to compensate for the position output
859 * TODO: might have a problem that we don't count pointsize when it is used
862 int half_out
= v
->outfile
.num_reg
/ 2 + 1;
865 uint32_t b
= ((20480 / (specs
->vertex_output_buffer_size
-
866 2 * half_out
* specs
->vertex_cache_size
)) +
869 uint32_t a
= (b
+ 256 / (specs
->shader_core_count
* half_out
)) / 2;
870 v
->vs_load_balancing
= VIVS_VS_LOAD_BALANCING_A(MIN2(a
, 255)) |
871 VIVS_VS_LOAD_BALANCING_B(MIN2(b
, 255)) |
872 VIVS_VS_LOAD_BALANCING_C(0x3f) |
873 VIVS_VS_LOAD_BALANCING_D(0x0f);
881 etna_destroy_shader_nir(struct etna_shader_variant
*shader
)
886 FREE(shader
->uniforms
.imm_data
);
887 FREE(shader
->uniforms
.imm_contents
);
891 extern const char *tgsi_swizzle_names
[];
893 etna_dump_shader_nir(const struct etna_shader_variant
*shader
)
895 if (shader
->stage
== MESA_SHADER_VERTEX
)
900 etna_disasm(shader
->code
, shader
->code_size
, PRINT_RAW
);
902 printf("num loops: %i\n", shader
->num_loops
);
903 printf("num temps: %i\n", shader
->num_temps
);
904 printf("immediates:\n");
905 for (int idx
= 0; idx
< shader
->uniforms
.imm_count
; ++idx
) {
906 printf(" [%i].%s = %f (0x%08x) (%d)\n",
908 tgsi_swizzle_names
[idx
% 4],
909 *((float *)&shader
->uniforms
.imm_data
[idx
]),
910 shader
->uniforms
.imm_data
[idx
],
911 shader
->uniforms
.imm_contents
[idx
]);
914 for (int idx
= 0; idx
< shader
->infile
.num_reg
; ++idx
) {
915 printf(" [%i] name=%s comps=%i\n", shader
->infile
.reg
[idx
].reg
,
916 (shader
->stage
== MESA_SHADER_VERTEX
) ?
917 gl_vert_attrib_name(shader
->infile
.reg
[idx
].slot
) :
918 gl_varying_slot_name(shader
->infile
.reg
[idx
].slot
),
919 shader
->infile
.reg
[idx
].num_components
);
921 printf("outputs:\n");
922 for (int idx
= 0; idx
< shader
->outfile
.num_reg
; ++idx
) {
923 printf(" [%i] name=%s comps=%i\n", shader
->outfile
.reg
[idx
].reg
,
924 (shader
->stage
== MESA_SHADER_VERTEX
) ?
925 gl_varying_slot_name(shader
->outfile
.reg
[idx
].slot
) :
926 gl_frag_result_name(shader
->outfile
.reg
[idx
].slot
),
927 shader
->outfile
.reg
[idx
].num_components
);
929 printf("special:\n");
930 if (shader
->stage
== MESA_SHADER_VERTEX
) {
931 printf(" vs_pos_out_reg=%i\n", shader
->vs_pos_out_reg
);
932 printf(" vs_pointsize_out_reg=%i\n", shader
->vs_pointsize_out_reg
);
933 printf(" vs_load_balancing=0x%08x\n", shader
->vs_load_balancing
);
935 printf(" ps_color_out_reg=%i\n", shader
->ps_color_out_reg
);
936 printf(" ps_depth_out_reg=%i\n", shader
->ps_depth_out_reg
);
938 printf(" input_count_unk8=0x%08x\n", shader
->input_count_unk8
);
941 static const struct etna_shader_inout
*
942 etna_shader_vs_lookup(const struct etna_shader_variant
*sobj
,
943 const struct etna_shader_inout
*in
)
945 for (int i
= 0; i
< sobj
->outfile
.num_reg
; i
++)
946 if (sobj
->outfile
.reg
[i
].slot
== in
->slot
)
947 return &sobj
->outfile
.reg
[i
];
953 etna_link_shader_nir(struct etna_shader_link_info
*info
,
954 const struct etna_shader_variant
*vs
,
955 const struct etna_shader_variant
*fs
)
958 /* For each fragment input we need to find the associated vertex shader
959 * output, which can be found by matching on semantic name and index. A
960 * binary search could be used because the vs outputs are sorted by their
961 * semantic index and grouped by semantic type by fill_in_vs_outputs.
963 assert(fs
->infile
.num_reg
< ETNA_NUM_INPUTS
);
964 info
->pcoord_varying_comp_ofs
= -1;
966 for (int idx
= 0; idx
< fs
->infile
.num_reg
; ++idx
) {
967 const struct etna_shader_inout
*fsio
= &fs
->infile
.reg
[idx
];
968 const struct etna_shader_inout
*vsio
= etna_shader_vs_lookup(vs
, fsio
);
969 struct etna_varying
*varying
;
970 bool interpolate_always
= true;
972 assert(fsio
->reg
> 0 && fsio
->reg
<= ARRAY_SIZE(info
->varyings
));
974 if (fsio
->reg
> info
->num_varyings
)
975 info
->num_varyings
= fsio
->reg
;
977 varying
= &info
->varyings
[fsio
->reg
- 1];
978 varying
->num_components
= fsio
->num_components
;
980 if (!interpolate_always
) /* colors affected by flat shading */
981 varying
->pa_attributes
= 0x200;
982 else /* texture coord or other bypasses flat shading */
983 varying
->pa_attributes
= 0x2f1;
985 varying
->use
[0] = VARYING_COMPONENT_USE_UNUSED
;
986 varying
->use
[1] = VARYING_COMPONENT_USE_UNUSED
;
987 varying
->use
[2] = VARYING_COMPONENT_USE_UNUSED
;
988 varying
->use
[3] = VARYING_COMPONENT_USE_UNUSED
;
990 /* point coord is an input to the PS without matching VS output,
991 * so it gets a varying slot without being assigned a VS register.
993 if (fsio
->slot
== VARYING_SLOT_PNTC
) {
994 varying
->use
[0] = VARYING_COMPONENT_USE_POINTCOORD_X
;
995 varying
->use
[1] = VARYING_COMPONENT_USE_POINTCOORD_Y
;
997 info
->pcoord_varying_comp_ofs
= comp_ofs
;
999 if (vsio
== NULL
) { /* not found -- link error */
1000 BUG("Semantic value not found in vertex shader outputs\n");
1003 varying
->reg
= vsio
->reg
;
1006 comp_ofs
+= varying
->num_components
;
1009 assert(info
->num_varyings
== fs
->infile
.num_reg
);