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
];
60 uint64_t consts
[ETNA_MAX_IMM
];
62 /* There was an error during compilation */
66 /* io related lowering
67 * run after lower_int_to_float because it adds i2f/f2i ops
70 etna_lower_io(nir_shader
*shader
, struct etna_shader_variant
*v
)
72 nir_foreach_function(function
, shader
) {
74 nir_builder_init(&b
, function
->impl
);
76 nir_foreach_block(block
, function
->impl
) {
77 nir_foreach_instr_safe(instr
, block
) {
78 if (instr
->type
== nir_instr_type_intrinsic
) {
79 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
81 switch (intr
->intrinsic
) {
82 case nir_intrinsic_load_front_face
: {
83 /* HW front_face is 0.0/1.0, not 0/~0u for bool
84 * lower with a comparison with 0
86 intr
->dest
.ssa
.bit_size
= 32;
88 b
.cursor
= nir_after_instr(instr
);
90 nir_ssa_def
*ssa
= nir_ine(&b
, &intr
->dest
.ssa
, nir_imm_int(&b
, 0));
92 nir_instr_as_alu(ssa
->parent_instr
)->op
= nir_op_ieq
;
94 nir_ssa_def_rewrite_uses_after(&intr
->dest
.ssa
,
98 case nir_intrinsic_store_deref
: {
99 nir_deref_instr
*deref
= nir_src_as_deref(intr
->src
[0]);
100 if (shader
->info
.stage
!= MESA_SHADER_FRAGMENT
|| !v
->key
.frag_rb_swap
)
103 assert(deref
->deref_type
== nir_deref_type_var
);
105 if (deref
->var
->data
.location
!= FRAG_RESULT_COLOR
&&
106 deref
->var
->data
.location
!= FRAG_RESULT_DATA0
)
109 b
.cursor
= nir_before_instr(instr
);
111 nir_ssa_def
*ssa
= nir_mov(&b
, intr
->src
[1].ssa
);
112 nir_alu_instr
*alu
= nir_instr_as_alu(ssa
->parent_instr
);
113 alu
->src
[0].swizzle
[0] = 2;
114 alu
->src
[0].swizzle
[2] = 0;
115 nir_instr_rewrite_src(instr
, &intr
->src
[1], nir_src_for_ssa(ssa
));
117 case nir_intrinsic_load_uniform
: {
118 /* multiply by 16 and convert to int */
119 b
.cursor
= nir_before_instr(instr
);
120 nir_ssa_def
*ssa
= nir_imul(&b
, intr
->src
[0].ssa
, nir_imm_int(&b
, 16));
121 nir_instr_rewrite_src(instr
, &intr
->src
[0], nir_src_for_ssa(ssa
));
123 case nir_intrinsic_load_vertex_id
:
124 case nir_intrinsic_load_instance_id
:
125 /* detect use of vertex_id/instance_id */
126 v
->vs_id_in_reg
= v
->infile
.num_reg
;
133 if (instr
->type
!= nir_instr_type_tex
)
136 nir_tex_instr
*tex
= nir_instr_as_tex(instr
);
137 nir_src
*coord
= NULL
;
138 nir_src
*lod_bias
= NULL
;
139 unsigned lod_bias_idx
;
141 assert(tex
->sampler_index
== tex
->texture_index
);
143 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
144 switch (tex
->src
[i
].src_type
) {
145 case nir_tex_src_coord
:
146 coord
= &tex
->src
[i
].src
;
148 case nir_tex_src_bias
:
149 case nir_tex_src_lod
:
151 lod_bias
= &tex
->src
[i
].src
;
154 case nir_tex_src_comparator
:
162 if (tex
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
) {
163 /* use a dummy load_uniform here to represent texcoord scale */
164 b
.cursor
= nir_before_instr(instr
);
165 nir_intrinsic_instr
*load
=
166 nir_intrinsic_instr_create(b
.shader
, nir_intrinsic_load_uniform
);
167 nir_intrinsic_set_base(load
, ~tex
->sampler_index
);
168 load
->num_components
= 2;
169 load
->src
[0] = nir_src_for_ssa(nir_imm_float(&b
, 0.0f
));
170 nir_ssa_dest_init(&load
->instr
, &load
->dest
, 2, 32, NULL
);
171 nir_intrinsic_set_type(load
, nir_type_float
);
173 nir_builder_instr_insert(&b
, &load
->instr
);
175 nir_ssa_def
*new_coord
= nir_fmul(&b
, coord
->ssa
, &load
->dest
.ssa
);
176 nir_instr_rewrite_src(&tex
->instr
, coord
, nir_src_for_ssa(new_coord
));
179 /* pre HALTI5 needs texture sources in a single source */
181 if (!lod_bias
|| v
->shader
->specs
->halti
>= 5)
184 assert(coord
&& lod_bias
&& tex
->coord_components
< 4);
186 nir_alu_instr
*vec
= nir_alu_instr_create(shader
, nir_op_vec4
);
187 for (unsigned i
= 0; i
< tex
->coord_components
; i
++) {
188 vec
->src
[i
].src
= nir_src_for_ssa(coord
->ssa
);
189 vec
->src
[i
].swizzle
[0] = i
;
191 for (unsigned i
= tex
->coord_components
; i
< 4; i
++)
192 vec
->src
[i
].src
= nir_src_for_ssa(lod_bias
->ssa
);
194 vec
->dest
.write_mask
= 0xf;
195 nir_ssa_dest_init(&vec
->instr
, &vec
->dest
.dest
, 4, 32, NULL
);
197 nir_tex_instr_remove_src(tex
, lod_bias_idx
);
198 nir_instr_rewrite_src(&tex
->instr
, coord
, nir_src_for_ssa(&vec
->dest
.dest
.ssa
));
199 tex
->coord_components
= 4;
201 nir_instr_insert_before(&tex
->instr
, &vec
->instr
);
208 etna_alu_to_scalar_filter_cb(const nir_instr
*instr
, const void *data
)
210 const struct etna_specs
*specs
= data
;
212 if (instr
->type
!= nir_instr_type_alu
)
215 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
227 /* TODO: can do better than alu_to_scalar for vector compares */
228 case nir_op_b32all_fequal2
:
229 case nir_op_b32all_fequal3
:
230 case nir_op_b32all_fequal4
:
231 case nir_op_b32any_fnequal2
:
232 case nir_op_b32any_fnequal3
:
233 case nir_op_b32any_fnequal4
:
234 case nir_op_b32all_iequal2
:
235 case nir_op_b32all_iequal3
:
236 case nir_op_b32all_iequal4
:
237 case nir_op_b32any_inequal2
:
238 case nir_op_b32any_inequal3
:
239 case nir_op_b32any_inequal4
:
242 if (!specs
->has_halti2_instructions
)
253 etna_lower_alu_impl(nir_function_impl
*impl
, struct etna_compile
*c
)
255 nir_shader
*shader
= impl
->function
->shader
;
258 nir_builder_init(&b
, impl
);
260 /* in a seperate loop so we can apply the multiple-uniform logic to the new fmul */
261 nir_foreach_block(block
, impl
) {
262 nir_foreach_instr_safe(instr
, block
) {
263 if (instr
->type
!= nir_instr_type_alu
)
266 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
267 /* multiply sin/cos src by constant
268 * TODO: do this earlier (but it breaks const_prop opt)
270 if (alu
->op
== nir_op_fsin
|| alu
->op
== nir_op_fcos
) {
271 b
.cursor
= nir_before_instr(instr
);
273 nir_ssa_def
*imm
= c
->specs
->has_new_transcendentals
?
274 nir_imm_float(&b
, 1.0 / M_PI
) :
275 nir_imm_float(&b
, 2.0 / M_PI
);
277 nir_instr_rewrite_src(instr
, &alu
->src
[0].src
,
278 nir_src_for_ssa(nir_fmul(&b
, alu
->src
[0].src
.ssa
, imm
)));
281 /* change transcendental ops to vec2 and insert vec1 mul for the result
282 * TODO: do this earlier (but it breaks with optimizations)
284 if (c
->specs
->has_new_transcendentals
&& (
285 alu
->op
== nir_op_fdiv
|| alu
->op
== nir_op_flog2
||
286 alu
->op
== nir_op_fsin
|| alu
->op
== nir_op_fcos
)) {
287 nir_ssa_def
*ssa
= &alu
->dest
.dest
.ssa
;
289 assert(ssa
->num_components
== 1);
291 nir_alu_instr
*mul
= nir_alu_instr_create(shader
, nir_op_fmul
);
292 mul
->src
[0].src
= mul
->src
[1].src
= nir_src_for_ssa(ssa
);
293 mul
->src
[1].swizzle
[0] = 1;
295 mul
->dest
.write_mask
= 1;
296 nir_ssa_dest_init(&mul
->instr
, &mul
->dest
.dest
, 1, 32, NULL
);
298 ssa
->num_components
= 2;
300 mul
->dest
.saturate
= alu
->dest
.saturate
;
301 alu
->dest
.saturate
= 0;
303 nir_instr_insert_after(instr
, &mul
->instr
);
305 nir_ssa_def_rewrite_uses_after(ssa
, nir_src_for_ssa(&mul
->dest
.dest
.ssa
), &mul
->instr
);
311 static void etna_lower_alu(nir_shader
*shader
, struct etna_compile
*c
)
313 nir_foreach_function(function
, shader
) {
315 etna_lower_alu_impl(function
->impl
, c
);
320 emit_inst(struct etna_compile
*c
, struct etna_inst
*inst
)
322 c
->code
[c
->inst_ptr
++] = *inst
;
325 /* to map nir srcs should to etna_inst srcs */
327 SRC_0_1_2
= (0 << 0) | (1 << 2) | (2 << 4),
328 SRC_0_1_X
= (0 << 0) | (1 << 2) | (3 << 4),
329 SRC_0_X_X
= (0 << 0) | (3 << 2) | (3 << 4),
330 SRC_0_X_1
= (0 << 0) | (3 << 2) | (1 << 4),
331 SRC_0_1_0
= (0 << 0) | (1 << 2) | (0 << 4),
332 SRC_X_X_0
= (3 << 0) | (3 << 2) | (0 << 4),
333 SRC_0_X_0
= (0 << 0) | (3 << 2) | (0 << 4),
336 /* info to translate a nir op to etna_inst */
337 struct etna_op_info
{
338 uint8_t opcode
; /* INST_OPCODE_ */
339 uint8_t src
; /* SRC_ enum */
340 uint8_t cond
; /* INST_CONDITION_ */
341 uint8_t type
; /* INST_TYPE_ */
344 static const struct etna_op_info etna_ops
[] = {
345 [0 ... nir_num_opcodes
- 1] = {0xff},
348 #define OPCT(nir, op, src, cond, type) [nir_op_##nir] = { \
351 INST_CONDITION_##cond, \
354 #define OPC(nir, op, src, cond) OPCT(nir, op, src, cond, F32)
355 #define IOPC(nir, op, src, cond) OPCT(nir, op, src, cond, S32)
356 #define UOPC(nir, op, src, cond) OPCT(nir, op, src, cond, U32)
357 #define OP(nir, op, src) OPC(nir, op, src, TRUE)
358 #define IOP(nir, op, src) IOPC(nir, op, src, TRUE)
359 #define UOP(nir, op, src) UOPC(nir, op, src, TRUE)
360 OP(mov
, MOV
, X_X_0
), OP(fneg
, MOV
, X_X_0
), OP(fabs
, MOV
, X_X_0
), OP(fsat
, MOV
, X_X_0
),
361 OP(fmul
, MUL
, 0_1_X
), OP(fadd
, ADD
, 0_X_1
), OP(ffma
, MAD
, 0_1_2
),
362 OP(fdot2
, DP2
, 0_1_X
), OP(fdot3
, DP3
, 0_1_X
), OP(fdot4
, DP4
, 0_1_X
),
363 OPC(fmin
, SELECT
, 0_1_0
, GT
), OPC(fmax
, SELECT
, 0_1_0
, LT
),
364 OP(ffract
, FRC
, X_X_0
), OP(frcp
, RCP
, X_X_0
), OP(frsq
, RSQ
, X_X_0
),
365 OP(fsqrt
, SQRT
, X_X_0
), OP(fsin
, SIN
, X_X_0
), OP(fcos
, COS
, X_X_0
),
366 OP(fsign
, SIGN
, X_X_0
), OP(ffloor
, FLOOR
, X_X_0
), OP(fceil
, CEIL
, X_X_0
),
367 OP(flog2
, LOG
, X_X_0
), OP(fexp2
, EXP
, X_X_0
),
368 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
),
369 OPC(fcsel
, SELECT
, 0_1_2
, NZ
),
370 OP(fdiv
, DIV
, 0_1_X
),
371 OP(fddx
, DSX
, 0_X_0
), OP(fddy
, DSY
, 0_X_0
),
374 IOP(i2f32
, I2F
, 0_X_X
),
375 UOP(u2f32
, I2F
, 0_X_X
),
376 IOP(f2i32
, F2I
, 0_X_X
),
377 UOP(f2u32
, F2I
, 0_X_X
),
378 UOP(b2f32
, AND
, 0_X_X
), /* AND with fui(1.0f) */
379 UOP(b2i32
, AND
, 0_X_X
), /* AND with 1 */
380 OPC(f2b32
, CMP
, 0_X_X
, NE
), /* != 0.0 */
381 UOPC(i2b32
, CMP
, 0_X_X
, NE
), /* != 0 */
384 IOP(iadd
, ADD
, 0_X_1
),
385 IOP(imul
, IMULLO0
, 0_1_X
),
386 /* IOP(imad, IMADLO0, 0_1_2), */
387 IOP(ineg
, ADD
, X_X_0
), /* ADD 0, -x */
388 IOP(iabs
, IABS
, X_X_0
),
389 IOP(isign
, SIGN
, X_X_0
),
390 IOPC(imin
, SELECT
, 0_1_0
, GT
),
391 IOPC(imax
, SELECT
, 0_1_0
, LT
),
392 UOPC(umin
, SELECT
, 0_1_0
, GT
),
393 UOPC(umax
, SELECT
, 0_1_0
, LT
),
396 UOPC(b32csel
, SELECT
, 0_1_2
, NZ
),
398 /* compare with int result */
399 OPC(feq32
, CMP
, 0_1_X
, EQ
),
400 OPC(fne32
, CMP
, 0_1_X
, NE
),
401 OPC(fge32
, CMP
, 0_1_X
, GE
),
402 OPC(flt32
, CMP
, 0_1_X
, LT
),
403 IOPC(ieq32
, CMP
, 0_1_X
, EQ
),
404 IOPC(ine32
, CMP
, 0_1_X
, NE
),
405 IOPC(ige32
, CMP
, 0_1_X
, GE
),
406 IOPC(ilt32
, CMP
, 0_1_X
, LT
),
407 UOPC(uge32
, CMP
, 0_1_X
, GE
),
408 UOPC(ult32
, CMP
, 0_1_X
, LT
),
412 IOP(iand
, AND
, 0_X_1
),
413 IOP(ixor
, XOR
, 0_X_1
),
414 IOP(inot
, NOT
, X_X_0
),
415 IOP(ishl
, LSHIFT
, 0_X_1
),
416 IOP(ishr
, RSHIFT
, 0_X_1
),
417 UOP(ushr
, RSHIFT
, 0_X_1
),
421 etna_emit_block_start(struct etna_compile
*c
, unsigned block
)
423 c
->block_ptr
[block
] = c
->inst_ptr
;
427 etna_emit_alu(struct etna_compile
*c
, nir_op op
, struct etna_inst_dst dst
,
428 struct etna_inst_src src
[3], bool saturate
)
430 struct etna_op_info ei
= etna_ops
[op
];
431 unsigned swiz_scalar
= INST_SWIZ_BROADCAST(ffs(dst
.write_mask
) - 1);
433 assert(ei
.opcode
!= 0xff);
435 struct etna_inst inst
= {
448 if (c
->specs
->has_new_transcendentals
)
456 /* scalar instructions we want src to be in x component */
457 src
[0].swiz
= inst_swiz_compose(src
[0].swiz
, swiz_scalar
);
458 src
[1].swiz
= inst_swiz_compose(src
[1].swiz
, swiz_scalar
);
460 /* deal with instructions which don't have 1:1 mapping */
462 inst
.src
[2] = etna_immediate_float(1.0f
);
465 inst
.src
[2] = etna_immediate_int(1);
468 inst
.src
[1] = etna_immediate_float(0.0f
);
471 inst
.src
[1] = etna_immediate_int(0);
474 inst
.src
[0] = etna_immediate_int(0);
481 /* set the "true" value for CMP instructions */
482 if (inst
.opcode
== INST_OPCODE_CMP
)
483 inst
.src
[2] = etna_immediate_int(-1);
485 for (unsigned j
= 0; j
< 3; j
++) {
486 unsigned i
= ((ei
.src
>> j
*2) & 3);
488 inst
.src
[j
] = src
[i
];
495 etna_emit_tex(struct etna_compile
*c
, nir_texop op
, unsigned texid
, unsigned dst_swiz
,
496 struct etna_inst_dst dst
, struct etna_inst_src coord
,
497 struct etna_inst_src lod_bias
, struct etna_inst_src compare
)
499 struct etna_inst inst
= {
501 .tex
.id
= texid
+ (is_fs(c
) ? 0 : c
->specs
->vertex_sampler_offset
),
502 .tex
.swiz
= dst_swiz
,
507 inst
.src
[1] = lod_bias
;
510 inst
.src
[2] = compare
;
513 case nir_texop_tex
: inst
.opcode
= INST_OPCODE_TEXLD
; break;
514 case nir_texop_txb
: inst
.opcode
= INST_OPCODE_TEXLDB
; break;
515 case nir_texop_txl
: inst
.opcode
= INST_OPCODE_TEXLDL
; break;
524 etna_emit_jump(struct etna_compile
*c
, unsigned block
, struct etna_inst_src condition
)
526 if (!condition
.use
) {
527 emit_inst(c
, &(struct etna_inst
) {.opcode
= INST_OPCODE_BRANCH
, .imm
= block
});
531 struct etna_inst inst
= {
532 .opcode
= INST_OPCODE_BRANCH
,
533 .cond
= INST_CONDITION_NOT
,
534 .type
= INST_TYPE_U32
,
538 inst
.src
[0].swiz
= INST_SWIZ_BROADCAST(inst
.src
[0].swiz
& 3);
543 etna_emit_discard(struct etna_compile
*c
, struct etna_inst_src condition
)
545 if (!condition
.use
) {
546 emit_inst(c
, &(struct etna_inst
) { .opcode
= INST_OPCODE_TEXKILL
});
550 struct etna_inst inst
= {
551 .opcode
= INST_OPCODE_TEXKILL
,
552 .cond
= INST_CONDITION_NZ
,
553 .type
= (c
->specs
->halti
< 2) ? INST_TYPE_F32
: INST_TYPE_U32
,
556 inst
.src
[0].swiz
= INST_SWIZ_BROADCAST(inst
.src
[0].swiz
& 3);
561 etna_emit_output(struct etna_compile
*c
, nir_variable
*var
, struct etna_inst_src src
)
563 struct etna_shader_io_file
*sf
= &c
->variant
->outfile
;
566 switch (var
->data
.location
) {
567 case FRAG_RESULT_COLOR
:
568 case FRAG_RESULT_DATA0
: /* DATA0 is used by gallium shaders for color */
569 c
->variant
->ps_color_out_reg
= src
.reg
;
571 case FRAG_RESULT_DEPTH
:
572 c
->variant
->ps_depth_out_reg
= src
.reg
;
575 unreachable("Unsupported fs output");
580 switch (var
->data
.location
) {
581 case VARYING_SLOT_POS
:
582 c
->variant
->vs_pos_out_reg
= src
.reg
;
584 case VARYING_SLOT_PSIZ
:
585 c
->variant
->vs_pointsize_out_reg
= src
.reg
;
588 sf
->reg
[sf
->num_reg
].reg
= src
.reg
;
589 sf
->reg
[sf
->num_reg
].slot
= var
->data
.location
;
590 sf
->reg
[sf
->num_reg
].num_components
= glsl_get_components(var
->type
);
597 etna_emit_load_ubo(struct etna_compile
*c
, struct etna_inst_dst dst
,
598 struct etna_inst_src src
, struct etna_inst_src base
)
600 /* convert float offset back to integer */
601 if (c
->specs
->halti
< 2) {
602 emit_inst(c
, &(struct etna_inst
) {
603 .opcode
= INST_OPCODE_F2I
,
604 .type
= INST_TYPE_U32
,
609 emit_inst(c
, &(struct etna_inst
) {
610 .opcode
= INST_OPCODE_LOAD
,
611 .type
= INST_TYPE_U32
,
615 .rgroup
= INST_RGROUP_TEMP
,
617 .swiz
= INST_SWIZ_BROADCAST(ffs(dst
.write_mask
) - 1)
625 emit_inst(c
, &(struct etna_inst
) {
626 .opcode
= INST_OPCODE_LOAD
,
627 .type
= INST_TYPE_U32
,
634 #define OPT(nir, pass, ...) ({ \
635 bool this_progress = false; \
636 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
639 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
642 etna_optimize_loop(nir_shader
*s
)
648 OPT_V(s
, nir_lower_vars_to_ssa
);
649 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
650 progress
|= OPT(s
, nir_copy_prop
);
651 progress
|= OPT(s
, nir_opt_dce
);
652 progress
|= OPT(s
, nir_opt_cse
);
653 progress
|= OPT(s
, nir_opt_peephole_select
, 16, true, true);
654 progress
|= OPT(s
, nir_opt_intrinsics
);
655 progress
|= OPT(s
, nir_opt_algebraic
);
656 progress
|= OPT(s
, nir_opt_constant_folding
);
657 progress
|= OPT(s
, nir_opt_dead_cf
);
658 if (OPT(s
, nir_opt_trivial_continues
)) {
660 /* If nir_opt_trivial_continues makes progress, then we need to clean
661 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
664 OPT(s
, nir_copy_prop
);
667 progress
|= OPT(s
, nir_opt_loop_unroll
, nir_var_all
);
668 progress
|= OPT(s
, nir_opt_if
, false);
669 progress
|= OPT(s
, nir_opt_remove_phis
);
670 progress
|= OPT(s
, nir_opt_undef
);
676 etna_glsl_type_size(const struct glsl_type
*type
, bool bindless
)
678 return glsl_count_attribute_slots(type
, false);
682 copy_uniform_state_to_shader(struct etna_shader_variant
*sobj
, uint64_t *consts
, unsigned count
)
684 struct etna_shader_uniform_info
*uinfo
= &sobj
->uniforms
;
686 uinfo
->imm_count
= count
* 4;
687 uinfo
->imm_data
= MALLOC(uinfo
->imm_count
* sizeof(*uinfo
->imm_data
));
688 uinfo
->imm_contents
= MALLOC(uinfo
->imm_count
* sizeof(*uinfo
->imm_contents
));
690 for (unsigned i
= 0; i
< uinfo
->imm_count
; i
++) {
691 uinfo
->imm_data
[i
] = consts
[i
];
692 uinfo
->imm_contents
[i
] = consts
[i
] >> 32;
695 etna_set_shader_uniforms_dirty_flags(sobj
);
698 #include "etnaviv_compiler_nir_emit.h"
701 etna_compile_shader_nir(struct etna_shader_variant
*v
)
706 struct etna_compile
*c
= CALLOC_STRUCT(etna_compile
);
711 c
->specs
= v
->shader
->specs
;
712 c
->nir
= nir_shader_clone(NULL
, v
->shader
->nir
);
714 nir_shader
*s
= c
->nir
;
715 const struct etna_specs
*specs
= c
->specs
;
717 v
->stage
= s
->info
.stage
;
718 v
->num_loops
= 0; /* TODO */
719 v
->vs_id_in_reg
= -1;
720 v
->vs_pos_out_reg
= -1;
721 v
->vs_pointsize_out_reg
= -1;
722 v
->ps_color_out_reg
= 0; /* 0 for shader that doesn't write fragcolor.. */
723 v
->ps_depth_out_reg
= -1;
725 /* setup input linking */
726 struct etna_shader_io_file
*sf
= &v
->infile
;
727 if (s
->info
.stage
== MESA_SHADER_VERTEX
) {
728 nir_foreach_variable(var
, &s
->inputs
) {
729 unsigned idx
= var
->data
.driver_location
;
730 sf
->reg
[idx
].reg
= idx
;
731 sf
->reg
[idx
].slot
= var
->data
.location
;
732 sf
->reg
[idx
].num_components
= glsl_get_components(var
->type
);
733 sf
->num_reg
= MAX2(sf
->num_reg
, idx
+1);
737 nir_foreach_variable(var
, &s
->inputs
) {
738 unsigned idx
= var
->data
.driver_location
;
739 sf
->reg
[idx
].reg
= idx
+ 1;
740 sf
->reg
[idx
].slot
= var
->data
.location
;
741 sf
->reg
[idx
].num_components
= glsl_get_components(var
->type
);
742 sf
->num_reg
= MAX2(sf
->num_reg
, idx
+1);
745 assert(sf
->num_reg
== count
);
748 NIR_PASS_V(s
, nir_lower_io
, ~nir_var_shader_out
, etna_glsl_type_size
,
749 (nir_lower_io_options
)0);
751 OPT_V(s
, nir_lower_regs_to_ssa
);
752 OPT_V(s
, nir_lower_vars_to_ssa
);
753 OPT_V(s
, nir_lower_indirect_derefs
, nir_var_all
);
754 OPT_V(s
, nir_lower_tex
, &(struct nir_lower_tex_options
) { .lower_txp
= ~0u });
755 OPT_V(s
, nir_lower_alu_to_scalar
, etna_alu_to_scalar_filter_cb
, specs
);
757 etna_optimize_loop(s
);
759 OPT_V(s
, etna_lower_io
, v
);
761 if (v
->shader
->specs
->vs_need_z_div
)
762 NIR_PASS_V(s
, nir_lower_clip_halfz
);
764 /* lower pre-halti2 to float (halti0 has integers, but only scalar..) */
765 if (c
->specs
->halti
< 2) {
766 /* use opt_algebraic between int_to_float and boot_to_float because
767 * int_to_float emits ftrunc, and ftrunc lowering generates bool ops
769 OPT_V(s
, nir_lower_int_to_float
);
770 OPT_V(s
, nir_opt_algebraic
);
771 OPT_V(s
, nir_lower_bool_to_float
);
773 OPT_V(s
, nir_lower_idiv
, nir_lower_idiv_fast
);
774 OPT_V(s
, nir_lower_bool_to_int32
);
777 etna_optimize_loop(s
);
779 if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS
))
780 nir_print_shader(s
, stdout
);
782 while( OPT(s
, nir_opt_vectorize
) );
783 OPT_V(s
, nir_lower_alu_to_scalar
, etna_alu_to_scalar_filter_cb
, specs
);
785 NIR_PASS_V(s
, nir_remove_dead_variables
, nir_var_function_temp
);
786 NIR_PASS_V(s
, nir_opt_algebraic_late
);
788 NIR_PASS_V(s
, nir_move_vec_src_uses_to_dest
);
789 NIR_PASS_V(s
, nir_copy_prop
);
790 /* only HW supported integer source mod is ineg for iadd instruction (?) */
791 NIR_PASS_V(s
, nir_lower_to_source_mods
, ~nir_lower_int_source_mods
);
792 /* need copy prop after uses_to_dest, and before src mods: see
793 * dEQP-GLES2.functional.shaders.random.all_features.fragment.95
796 NIR_PASS_V(s
, nir_opt_dce
);
798 NIR_PASS_V(s
, etna_lower_alu
, c
);
800 if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS
))
801 nir_print_shader(s
, stdout
);
803 unsigned block_ptr
[nir_shader_get_entrypoint(s
)->num_blocks
];
804 c
->block_ptr
= block_ptr
;
807 ASSERTED
bool ok
= emit_shader(c
, &v
->num_temps
, &num_consts
);
810 /* empty shader, emit NOP */
812 emit_inst(c
, &(struct etna_inst
) { .opcode
= INST_OPCODE_NOP
});
814 /* assemble instructions, fixing up labels */
815 uint32_t *code
= MALLOC(c
->inst_ptr
* 16);
816 for (unsigned i
= 0; i
< c
->inst_ptr
; i
++) {
817 struct etna_inst
*inst
= &c
->code
[i
];
818 if (inst
->opcode
== INST_OPCODE_BRANCH
)
819 inst
->imm
= block_ptr
[inst
->imm
];
821 inst
->halti5
= specs
->halti
>= 5;
822 etna_assemble(&code
[i
* 4], inst
);
825 v
->code_size
= c
->inst_ptr
* 4;
827 v
->needs_icache
= c
->inst_ptr
> specs
->max_instructions
;
829 copy_uniform_state_to_shader(v
, c
->consts
, num_consts
);
831 if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
832 v
->input_count_unk8
= 31; /* XXX what is this */
833 assert(v
->ps_depth_out_reg
<= 0);
839 v
->input_count_unk8
= DIV_ROUND_UP(v
->infile
.num_reg
+ 4, 16); /* XXX what is this */
841 /* fill in "mystery meat" load balancing value. This value determines how
842 * work is scheduled between VS and PS
843 * in the unified shader architecture. More precisely, it is determined from
844 * the number of VS outputs, as well as chip-specific
845 * vertex output buffer size, vertex cache size, and the number of shader
848 * XXX this is a conservative estimate, the "optimal" value is only known for
849 * sure at link time because some
850 * outputs may be unused and thus unmapped. Then again, in the general use
851 * case with GLSL the vertex and fragment
852 * shaders are linked already before submitting to Gallium, thus all outputs
855 * note: TGSI compiler counts all outputs (including position and pointsize), here
856 * v->outfile.num_reg only counts varyings, +1 to compensate for the position output
857 * TODO: might have a problem that we don't count pointsize when it is used
860 int half_out
= v
->outfile
.num_reg
/ 2 + 1;
863 uint32_t b
= ((20480 / (specs
->vertex_output_buffer_size
-
864 2 * half_out
* specs
->vertex_cache_size
)) +
867 uint32_t a
= (b
+ 256 / (specs
->shader_core_count
* half_out
)) / 2;
868 v
->vs_load_balancing
= VIVS_VS_LOAD_BALANCING_A(MIN2(a
, 255)) |
869 VIVS_VS_LOAD_BALANCING_B(MIN2(b
, 255)) |
870 VIVS_VS_LOAD_BALANCING_C(0x3f) |
871 VIVS_VS_LOAD_BALANCING_D(0x0f);
879 etna_destroy_shader_nir(struct etna_shader_variant
*shader
)
884 FREE(shader
->uniforms
.imm_data
);
885 FREE(shader
->uniforms
.imm_contents
);
889 extern const char *tgsi_swizzle_names
[];
891 etna_dump_shader_nir(const struct etna_shader_variant
*shader
)
893 if (shader
->stage
== MESA_SHADER_VERTEX
)
898 etna_disasm(shader
->code
, shader
->code_size
, PRINT_RAW
);
900 printf("num loops: %i\n", shader
->num_loops
);
901 printf("num temps: %i\n", shader
->num_temps
);
902 printf("immediates:\n");
903 for (int idx
= 0; idx
< shader
->uniforms
.imm_count
; ++idx
) {
904 printf(" [%i].%s = %f (0x%08x) (%d)\n",
906 tgsi_swizzle_names
[idx
% 4],
907 *((float *)&shader
->uniforms
.imm_data
[idx
]),
908 shader
->uniforms
.imm_data
[idx
],
909 shader
->uniforms
.imm_contents
[idx
]);
912 for (int idx
= 0; idx
< shader
->infile
.num_reg
; ++idx
) {
913 printf(" [%i] name=%s comps=%i\n", shader
->infile
.reg
[idx
].reg
,
914 (shader
->stage
== MESA_SHADER_VERTEX
) ?
915 gl_vert_attrib_name(shader
->infile
.reg
[idx
].slot
) :
916 gl_varying_slot_name(shader
->infile
.reg
[idx
].slot
),
917 shader
->infile
.reg
[idx
].num_components
);
919 printf("outputs:\n");
920 for (int idx
= 0; idx
< shader
->outfile
.num_reg
; ++idx
) {
921 printf(" [%i] name=%s comps=%i\n", shader
->outfile
.reg
[idx
].reg
,
922 (shader
->stage
== MESA_SHADER_VERTEX
) ?
923 gl_varying_slot_name(shader
->outfile
.reg
[idx
].slot
) :
924 gl_frag_result_name(shader
->outfile
.reg
[idx
].slot
),
925 shader
->outfile
.reg
[idx
].num_components
);
927 printf("special:\n");
928 if (shader
->stage
== MESA_SHADER_VERTEX
) {
929 printf(" vs_pos_out_reg=%i\n", shader
->vs_pos_out_reg
);
930 printf(" vs_pointsize_out_reg=%i\n", shader
->vs_pointsize_out_reg
);
931 printf(" vs_load_balancing=0x%08x\n", shader
->vs_load_balancing
);
933 printf(" ps_color_out_reg=%i\n", shader
->ps_color_out_reg
);
934 printf(" ps_depth_out_reg=%i\n", shader
->ps_depth_out_reg
);
936 printf(" input_count_unk8=0x%08x\n", shader
->input_count_unk8
);
939 static const struct etna_shader_inout
*
940 etna_shader_vs_lookup(const struct etna_shader_variant
*sobj
,
941 const struct etna_shader_inout
*in
)
943 for (int i
= 0; i
< sobj
->outfile
.num_reg
; i
++)
944 if (sobj
->outfile
.reg
[i
].slot
== in
->slot
)
945 return &sobj
->outfile
.reg
[i
];
951 etna_link_shader_nir(struct etna_shader_link_info
*info
,
952 const struct etna_shader_variant
*vs
,
953 const struct etna_shader_variant
*fs
)
956 /* For each fragment input we need to find the associated vertex shader
957 * output, which can be found by matching on semantic name and index. A
958 * binary search could be used because the vs outputs are sorted by their
959 * semantic index and grouped by semantic type by fill_in_vs_outputs.
961 assert(fs
->infile
.num_reg
< ETNA_NUM_INPUTS
);
962 info
->pcoord_varying_comp_ofs
= -1;
964 for (int idx
= 0; idx
< fs
->infile
.num_reg
; ++idx
) {
965 const struct etna_shader_inout
*fsio
= &fs
->infile
.reg
[idx
];
966 const struct etna_shader_inout
*vsio
= etna_shader_vs_lookup(vs
, fsio
);
967 struct etna_varying
*varying
;
968 bool interpolate_always
= true;
970 assert(fsio
->reg
> 0 && fsio
->reg
<= ARRAY_SIZE(info
->varyings
));
972 if (fsio
->reg
> info
->num_varyings
)
973 info
->num_varyings
= fsio
->reg
;
975 varying
= &info
->varyings
[fsio
->reg
- 1];
976 varying
->num_components
= fsio
->num_components
;
978 if (!interpolate_always
) /* colors affected by flat shading */
979 varying
->pa_attributes
= 0x200;
980 else /* texture coord or other bypasses flat shading */
981 varying
->pa_attributes
= 0x2f1;
983 varying
->use
[0] = VARYING_COMPONENT_USE_UNUSED
;
984 varying
->use
[1] = VARYING_COMPONENT_USE_UNUSED
;
985 varying
->use
[2] = VARYING_COMPONENT_USE_UNUSED
;
986 varying
->use
[3] = VARYING_COMPONENT_USE_UNUSED
;
988 /* point coord is an input to the PS without matching VS output,
989 * so it gets a varying slot without being assigned a VS register.
991 if (fsio
->slot
== VARYING_SLOT_PNTC
) {
992 varying
->use
[0] = VARYING_COMPONENT_USE_POINTCOORD_X
;
993 varying
->use
[1] = VARYING_COMPONENT_USE_POINTCOORD_Y
;
995 info
->pcoord_varying_comp_ofs
= comp_ofs
;
997 if (vsio
== NULL
) { /* not found -- link error */
998 BUG("Semantic value not found in vertex shader outputs\n");
1001 varying
->reg
= vsio
->reg
;
1004 comp_ofs
+= varying
->num_components
;
1007 assert(info
->num_varyings
== fs
->infile
.num_reg
);