2 * Copyright (C) 2018 Jonathan Marek <jonathan@marek.ca>
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 * Jonathan Marek <jonathan@marek.ca>
27 #include "ir2_private.h"
29 #include "freedreno_util.h"
30 #include "fd2_program.h"
32 static const nir_shader_compiler_options options
= {
39 /* .fdot_replicates = true, it is replicated, but it makes things worse */
40 .lower_all_io_to_temps
= true,
41 .vertex_id_zero_based
= true, /* its not implemented anyway */
42 .lower_bitshift
= true,
45 const nir_shader_compiler_options
*
46 ir2_get_compiler_options(void)
51 #define OPT(nir, pass, ...) ({ \
52 bool this_progress = false; \
53 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
56 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
59 ir2_optimize_loop(nir_shader
*s
)
65 OPT_V(s
, nir_lower_vars_to_ssa
);
66 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
67 progress
|= OPT(s
, nir_copy_prop
);
68 progress
|= OPT(s
, nir_opt_dce
);
69 progress
|= OPT(s
, nir_opt_cse
);
70 /* progress |= OPT(s, nir_opt_gcm, true); */
71 progress
|= OPT(s
, nir_opt_peephole_select
, UINT_MAX
, true, true);
72 progress
|= OPT(s
, nir_opt_intrinsics
);
73 progress
|= OPT(s
, nir_opt_algebraic
);
74 progress
|= OPT(s
, nir_opt_constant_folding
);
75 progress
|= OPT(s
, nir_opt_dead_cf
);
76 if (OPT(s
, nir_opt_trivial_continues
)) {
78 /* If nir_opt_trivial_continues makes progress, then we need to clean
79 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
82 OPT(s
, nir_copy_prop
);
85 progress
|= OPT(s
, nir_opt_loop_unroll
, nir_var_all
);
86 progress
|= OPT(s
, nir_opt_if
, false);
87 progress
|= OPT(s
, nir_opt_remove_phis
);
88 progress
|= OPT(s
, nir_opt_undef
);
94 /* trig workarounds is the same as ir3.. but we don't want to include ir3 */
95 bool ir3_nir_apply_trig_workarounds(nir_shader
* shader
);
98 ir2_optimize_nir(nir_shader
*s
, bool lower
)
100 struct nir_lower_tex_options tex_options
= {
105 if (fd_mesa_debug
& FD_DBG_DISASM
) {
106 debug_printf("----------------------\n");
107 nir_print_shader(s
, stdout
);
108 debug_printf("----------------------\n");
111 OPT_V(s
, nir_lower_regs_to_ssa
);
112 OPT_V(s
, nir_lower_vars_to_ssa
);
113 OPT_V(s
, nir_lower_indirect_derefs
, nir_var_shader_in
| nir_var_shader_out
);
116 OPT_V(s
, ir3_nir_apply_trig_workarounds
);
117 OPT_V(s
, nir_lower_tex
, &tex_options
);
120 ir2_optimize_loop(s
);
122 OPT_V(s
, nir_remove_dead_variables
, nir_var_function_temp
);
123 OPT_V(s
, nir_move_load_const
);
125 /* TODO we dont want to get shaders writing to depth for depth textures */
126 if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
127 nir_foreach_variable(var
, &s
->outputs
) {
128 if (var
->data
.location
== FRAG_RESULT_DEPTH
)
136 static struct ir2_src
137 load_const(struct ir2_context
*ctx
, float *value_f
, unsigned ncomp
)
139 struct fd2_shader_stateobj
*so
= ctx
->so
;
140 unsigned imm_ncomp
, swiz
, idx
, i
, j
;
141 uint32_t *value
= (uint32_t*) value_f
;
143 /* try to merge with existing immediate (TODO: try with neg) */
144 for (idx
= 0; idx
< so
->num_immediates
; idx
++) {
146 imm_ncomp
= so
->immediates
[idx
].ncomp
;
147 for (i
= 0; i
< ncomp
; i
++) {
148 for (j
= 0; j
< imm_ncomp
; j
++) {
149 if (value
[i
] == so
->immediates
[idx
].val
[j
])
152 if (j
== imm_ncomp
) {
155 so
->immediates
[idx
].val
[imm_ncomp
++] = value
[i
];
157 swiz
|= swiz_set(j
, i
);
159 /* matched all components */
164 /* need to allocate new immediate */
165 if (idx
== so
->num_immediates
) {
168 for (i
= 0; i
< ncomp
; i
++) {
169 for (j
= 0; j
< imm_ncomp
; j
++) {
170 if (value
[i
] == ctx
->so
->immediates
[idx
].val
[j
])
173 if (j
== imm_ncomp
) {
174 so
->immediates
[idx
].val
[imm_ncomp
++] = value
[i
];
176 swiz
|= swiz_set(j
, i
);
178 so
->num_immediates
++;
180 so
->immediates
[idx
].ncomp
= imm_ncomp
;
183 swiz
= swiz_merge(swiz
, IR2_SWIZZLE_XXXX
);
185 return ir2_src(so
->first_immediate
+ idx
, swiz
, IR2_SRC_CONST
);
189 ir2_zero(struct ir2_context
*ctx
)
191 return load_const(ctx
, (float[]) {0.0f
}, 1);
195 update_range(struct ir2_context
*ctx
, struct ir2_reg
*reg
)
197 if (!reg
->initialized
) {
198 reg
->initialized
= true;
199 reg
->loop_depth
= ctx
->loop_depth
;
202 if (ctx
->loop_depth
> reg
->loop_depth
) {
203 reg
->block_idx_free
= ctx
->loop_last_block
[reg
->loop_depth
+ 1];
205 reg
->loop_depth
= ctx
->loop_depth
;
206 reg
->block_idx_free
= -1;
209 /* for regs we want to free at the end of the loop in any case
210 * XXX dont do this for ssa
213 reg
->block_idx_free
= ctx
->loop_last_block
[reg
->loop_depth
];
216 static struct ir2_src
217 make_src(struct ir2_context
*ctx
, nir_src src
)
219 struct ir2_src res
= {};
222 nir_const_value
*const_value
= nir_src_as_const_value(src
);
226 float c
[src
.ssa
->num_components
];
227 nir_const_value_to_array(c
, const_value
, src
.ssa
->num_components
, f32
);
228 return load_const(ctx
, c
, src
.ssa
->num_components
);
232 res
.num
= src
.reg
.reg
->index
;
233 res
.type
= IR2_SRC_REG
;
234 reg
= &ctx
->reg
[res
.num
];
236 assert(ctx
->ssa_map
[src
.ssa
->index
] >= 0);
237 res
.num
= ctx
->ssa_map
[src
.ssa
->index
];
238 res
.type
= IR2_SRC_SSA
;
239 reg
= &ctx
->instr
[res
.num
].ssa
;
242 update_range(ctx
, reg
);
247 set_index(struct ir2_context
*ctx
, nir_dest
* dst
,
248 struct ir2_instr
*instr
)
250 struct ir2_reg
*reg
= &instr
->ssa
;
253 ctx
->ssa_map
[dst
->ssa
.index
] = instr
->idx
;
255 assert(instr
->is_ssa
);
256 reg
= &ctx
->reg
[dst
->reg
.reg
->index
];
258 instr
->is_ssa
= false;
261 update_range(ctx
, reg
);
264 static struct ir2_instr
*
265 ir2_instr_create(struct ir2_context
*ctx
, int type
)
267 struct ir2_instr
*instr
;
269 instr
= &ctx
->instr
[ctx
->instr_count
++];
270 instr
->idx
= ctx
->instr_count
- 1;
272 instr
->block_idx
= ctx
->block_idx
;
273 instr
->pred
= ctx
->pred
;
274 instr
->is_ssa
= true;
278 static struct ir2_instr
*
279 instr_create_alu(struct ir2_context
*ctx
, nir_op opcode
, unsigned ncomp
)
281 /* emit_alu will fixup instrs that don't map directly */
282 static const struct ir2_opc
{
283 int8_t scalar
, vector
;
284 } nir_ir2_opc
[nir_num_opcodes
+1] = {
285 [0 ... nir_num_opcodes
- 1] = {-1, -1},
287 [nir_op_mov
] = {MAXs
, MAXv
},
288 [nir_op_fsign
] = {-1, CNDGTEv
},
289 [nir_op_fadd
] = {ADDs
, ADDv
},
290 [nir_op_fsub
] = {ADDs
, ADDv
},
291 [nir_op_fmul
] = {MULs
, MULv
},
292 [nir_op_ffma
] = {-1, MULADDv
},
293 [nir_op_fmax
] = {MAXs
, MAXv
},
294 [nir_op_fmin
] = {MINs
, MINv
},
295 [nir_op_ffloor
] = {FLOORs
, FLOORv
},
296 [nir_op_ffract
] = {FRACs
, FRACv
},
297 [nir_op_ftrunc
] = {TRUNCs
, TRUNCv
},
298 [nir_op_fdot2
] = {-1, DOT2ADDv
},
299 [nir_op_fdot3
] = {-1, DOT3v
},
300 [nir_op_fdot4
] = {-1, DOT4v
},
301 [nir_op_sge
] = {-1, SETGTEv
},
302 [nir_op_slt
] = {-1, SETGTv
},
303 [nir_op_sne
] = {-1, SETNEv
},
304 [nir_op_seq
] = {-1, SETEv
},
305 [nir_op_fcsel
] = {-1, CNDEv
},
306 [nir_op_frsq
] = {RECIPSQ_IEEE
, -1},
307 [nir_op_frcp
] = {RECIP_IEEE
, -1},
308 [nir_op_flog2
] = {LOG_IEEE
, -1},
309 [nir_op_fexp2
] = {EXP_IEEE
, -1},
310 [nir_op_fsqrt
] = {SQRT_IEEE
, -1},
311 [nir_op_fcos
] = {COS
, -1},
312 [nir_op_fsin
] = {SIN
, -1},
313 /* no fsat, fneg, fabs since source mods deal with those */
315 /* so we can use this function with non-nir op */
316 #define ir2_op_cube nir_num_opcodes
317 [ir2_op_cube
] = {-1, CUBEv
},
320 struct ir2_opc op
= nir_ir2_opc
[opcode
];
321 assert(op
.vector
>= 0 || op
.scalar
>= 0);
323 struct ir2_instr
*instr
= ir2_instr_create(ctx
, IR2_ALU
);
324 instr
->alu
.vector_opc
= op
.vector
;
325 instr
->alu
.scalar_opc
= op
.scalar
;
326 instr
->alu
.export
= -1;
327 instr
->alu
.write_mask
= (1 << ncomp
) - 1;
328 instr
->src_count
= opcode
== ir2_op_cube
? 2 :
329 nir_op_infos
[opcode
].num_inputs
;
330 instr
->ssa
.ncomp
= ncomp
;
334 static struct ir2_instr
*
335 instr_create_alu_reg(struct ir2_context
*ctx
, nir_op opcode
,
336 uint8_t write_mask
, struct ir2_instr
*share_reg
)
338 struct ir2_instr
*instr
;
341 reg
= share_reg
? share_reg
->reg
: &ctx
->reg
[ctx
->reg_count
++];
342 reg
->ncomp
= MAX2(reg
->ncomp
, util_logbase2(write_mask
) + 1);
344 instr
= instr_create_alu(ctx
, opcode
, util_bitcount(write_mask
));
345 instr
->alu
.write_mask
= write_mask
;
347 instr
->is_ssa
= false;
352 static struct ir2_instr
*
353 instr_create_alu_dest(struct ir2_context
*ctx
, nir_op opcode
, nir_dest
*dst
)
355 struct ir2_instr
*instr
;
356 instr
= instr_create_alu(ctx
, opcode
, nir_dest_num_components(*dst
));
357 set_index(ctx
, dst
, instr
);
361 static struct ir2_instr
*
362 ir2_instr_create_fetch(struct ir2_context
*ctx
, nir_dest
*dst
,
363 instr_fetch_opc_t opc
)
365 struct ir2_instr
*instr
= ir2_instr_create(ctx
, IR2_FETCH
);
366 instr
->fetch
.opc
= opc
;
367 instr
->src_count
= 1;
368 instr
->ssa
.ncomp
= nir_dest_num_components(*dst
);
369 set_index(ctx
, dst
, instr
);
373 static struct ir2_src
374 make_src_noconst(struct ir2_context
*ctx
, nir_src src
)
376 struct ir2_instr
*instr
;
378 if (nir_src_as_const_value(src
)) {
380 instr
= instr_create_alu(ctx
, nir_op_mov
, src
.ssa
->num_components
);
381 instr
->src
[0] = make_src(ctx
, src
);
382 return ir2_src(instr
->idx
, 0, IR2_SRC_SSA
);
385 return make_src(ctx
, src
);
389 emit_alu(struct ir2_context
*ctx
, nir_alu_instr
* alu
)
391 const nir_op_info
*info
= &nir_op_infos
[alu
->op
];
392 nir_dest
*dst
= &alu
->dest
.dest
;
393 struct ir2_instr
*instr
;
397 /* get the number of dst components */
399 ncomp
= dst
->ssa
.num_components
;
402 for (int i
= 0; i
< 4; i
++)
403 ncomp
+= !!(alu
->dest
.write_mask
& 1 << i
);
406 instr
= instr_create_alu(ctx
, alu
->op
, ncomp
);
407 set_index(ctx
, dst
, instr
);
408 instr
->alu
.saturate
= alu
->dest
.saturate
;
409 instr
->alu
.write_mask
= alu
->dest
.write_mask
;
411 for (int i
= 0; i
< info
->num_inputs
; i
++) {
412 nir_alu_src
*src
= &alu
->src
[i
];
414 /* compress swizzle with writemask when applicable */
415 unsigned swiz
= 0, j
= 0;
416 for (int i
= 0; i
< 4; i
++) {
417 if (!(alu
->dest
.write_mask
& 1 << i
) && !info
->output_size
)
419 swiz
|= swiz_set(src
->swizzle
[i
], j
++);
422 instr
->src
[i
] = make_src(ctx
, src
->src
);
423 instr
->src
[i
].swizzle
= swiz_merge(instr
->src
[i
].swizzle
, swiz
);
424 instr
->src
[i
].negate
= src
->negate
;
425 instr
->src
[i
].abs
= src
->abs
;
428 /* workarounds for NIR ops that don't map directly to a2xx ops */
432 instr
->src
[0] = instr
->src
[1];
437 instr
->src
[1] = instr
->src
[2];
441 instr
->src
[1].negate
= !instr
->src
[1].negate
;
444 instr
->src_count
= 3;
445 instr
->src
[2] = ir2_zero(ctx
);
448 /* we need an extra instruction to deal with the zero case */
449 struct ir2_instr
*tmp
;
451 /* tmp = x == 0 ? 0 : 1 */
452 tmp
= instr_create_alu(ctx
, nir_op_fcsel
, ncomp
);
453 tmp
->src
[0] = instr
->src
[0];
454 tmp
->src
[1] = ir2_zero(ctx
);
455 tmp
->src
[2] = load_const(ctx
, (float[]) {1.0f
}, 1);
457 /* result = x >= 0 ? tmp : -tmp */
458 instr
->src
[1] = ir2_src(tmp
->idx
, 0, IR2_SRC_SSA
);
459 instr
->src
[2] = instr
->src
[1];
460 instr
->src
[2].negate
= true;
461 instr
->src_count
= 3;
469 load_input(struct ir2_context
*ctx
, nir_dest
*dst
, unsigned idx
)
471 struct ir2_instr
*instr
;
474 if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
475 instr
= ir2_instr_create_fetch(ctx
, dst
, 0);
476 instr
->src
[0] = ir2_src(0, 0, IR2_SRC_INPUT
);
477 instr
->fetch
.vtx
.const_idx
= 20 + (idx
/ 3);
478 instr
->fetch
.vtx
.const_idx_sel
= idx
% 3;
482 /* get slot from idx */
483 nir_foreach_variable(var
, &ctx
->nir
->inputs
) {
484 if (var
->data
.driver_location
== idx
) {
485 slot
= var
->data
.location
;
492 case VARYING_SLOT_PNTC
:
493 /* need to extract with abs and invert y */
494 instr
= instr_create_alu_dest(ctx
, nir_op_ffma
, dst
);
495 instr
->src
[0] = ir2_src(ctx
->f
->inputs_count
, IR2_SWIZZLE_ZW
, IR2_SRC_INPUT
);
496 instr
->src
[0].abs
= true;
497 instr
->src
[1] = load_const(ctx
, (float[]) {1.0f
, -1.0f
}, 2);
498 instr
->src
[2] = load_const(ctx
, (float[]) {0.0f
, 1.0f
}, 2);
500 case VARYING_SLOT_POS
:
501 /* need to extract xy with abs and add tile offset on a20x
502 * zw from fragcoord input (w inverted in fragment shader)
503 * TODO: only components that are required by fragment shader
505 instr
= instr_create_alu_reg(ctx
,
506 ctx
->so
->is_a20x
? nir_op_fadd
: nir_op_mov
, 3, NULL
);
507 instr
->src
[0] = ir2_src(ctx
->f
->inputs_count
, 0, IR2_SRC_INPUT
);
508 instr
->src
[0].abs
= true;
509 /* on a20x, C64 contains the tile offset */
510 instr
->src
[1] = ir2_src(64, 0, IR2_SRC_CONST
);
512 instr
= instr_create_alu_reg(ctx
, nir_op_mov
, 4, instr
);
513 instr
->src
[0] = ir2_src(ctx
->f
->fragcoord
, 0, IR2_SRC_INPUT
);
515 instr
= instr_create_alu_reg(ctx
, nir_op_frcp
, 8, instr
);
516 instr
->src
[0] = ir2_src(ctx
->f
->fragcoord
, IR2_SWIZZLE_Y
, IR2_SRC_INPUT
);
518 unsigned reg_idx
= instr
->reg
- ctx
->reg
; /* XXX */
519 instr
= instr_create_alu_dest(ctx
, nir_op_mov
, dst
);
520 instr
->src
[0] = ir2_src(reg_idx
, 0, IR2_SRC_REG
);
523 instr
= instr_create_alu_dest(ctx
, nir_op_mov
, dst
);
524 instr
->src
[0] = ir2_src(idx
, 0, IR2_SRC_INPUT
);
530 output_slot(struct ir2_context
*ctx
, nir_intrinsic_instr
*intr
)
533 unsigned idx
= nir_intrinsic_base(intr
);
534 nir_foreach_variable(var
, &ctx
->nir
->outputs
) {
535 if (var
->data
.driver_location
== idx
) {
536 slot
= var
->data
.location
;
545 store_output(struct ir2_context
*ctx
, nir_src src
, unsigned slot
, unsigned ncomp
)
547 struct ir2_instr
*instr
;
550 if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
552 case VARYING_SLOT_POS
:
553 ctx
->position
= make_src(ctx
, src
);
556 case VARYING_SLOT_PSIZ
:
557 ctx
->so
->writes_psize
= true;
561 /* find matching slot from fragment shader input */
562 for (idx
= 0; idx
< ctx
->f
->inputs_count
; idx
++)
563 if (ctx
->f
->inputs
[idx
].slot
== slot
)
565 if (idx
== ctx
->f
->inputs_count
)
568 } else if (slot
!= FRAG_RESULT_COLOR
&& slot
!= FRAG_RESULT_DATA0
) {
569 /* only color output is implemented */
573 instr
= instr_create_alu(ctx
, nir_op_mov
, ncomp
);
574 instr
->src
[0] = make_src(ctx
, src
);
575 instr
->alu
.export
= idx
;
579 emit_intrinsic(struct ir2_context
*ctx
, nir_intrinsic_instr
*intr
)
581 struct ir2_instr
*instr
;
582 nir_const_value
*const_offset
;
585 switch (intr
->intrinsic
) {
586 case nir_intrinsic_load_input
:
587 load_input(ctx
, &intr
->dest
, nir_intrinsic_base(intr
));
589 case nir_intrinsic_store_output
:
590 store_output(ctx
, intr
->src
[0], output_slot(ctx
, intr
), intr
->num_components
);
592 case nir_intrinsic_load_uniform
:
593 const_offset
= nir_src_as_const_value(intr
->src
[0]);
594 assert(const_offset
); /* TODO can be false in ES2? */
595 idx
= nir_intrinsic_base(intr
);
596 idx
+= (uint32_t) nir_src_as_const_value(intr
->src
[0])[0].f32
;
597 instr
= instr_create_alu_dest(ctx
, nir_op_mov
, &intr
->dest
);
598 instr
->src
[0] = ir2_src(idx
, 0, IR2_SRC_CONST
);
600 case nir_intrinsic_discard
:
601 case nir_intrinsic_discard_if
:
602 instr
= ir2_instr_create(ctx
, IR2_ALU
);
603 instr
->alu
.vector_opc
= VECTOR_NONE
;
604 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
605 instr
->alu
.scalar_opc
= KILLNEs
;
606 instr
->src
[0] = make_src(ctx
, intr
->src
[0]);
608 instr
->alu
.scalar_opc
= KILLEs
;
609 instr
->src
[0] = ir2_zero(ctx
);
611 instr
->alu
.export
= -1;
612 instr
->src_count
= 1;
613 ctx
->so
->has_kill
= true;
615 case nir_intrinsic_load_front_face
:
616 /* gl_FrontFacing is in the sign of param.x
617 * rcp required because otherwise we can't differentiate -0.0 and +0.0
619 ctx
->so
->need_param
= true;
621 struct ir2_instr
*tmp
= instr_create_alu(ctx
, nir_op_frcp
, 1);
622 tmp
->src
[0] = ir2_src(ctx
->f
->inputs_count
, 0, IR2_SRC_INPUT
);
624 instr
= instr_create_alu_dest(ctx
, nir_op_sge
, &intr
->dest
);
625 instr
->src
[0] = ir2_src(tmp
->idx
, 0, IR2_SRC_SSA
);
626 instr
->src
[1] = ir2_zero(ctx
);
629 compile_error(ctx
, "unimplemented intr %d\n", intr
->intrinsic
);
635 emit_tex(struct ir2_context
*ctx
, nir_tex_instr
* tex
)
637 bool is_rect
= false, is_cube
= false;
638 struct ir2_instr
*instr
;
639 nir_src
*coord
, *lod_bias
;
641 coord
= lod_bias
= NULL
;
643 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
644 switch (tex
->src
[i
].src_type
) {
645 case nir_tex_src_coord
:
646 coord
= &tex
->src
[i
].src
;
648 case nir_tex_src_bias
:
649 case nir_tex_src_lod
:
651 lod_bias
= &tex
->src
[i
].src
;
654 compile_error(ctx
, "Unhandled NIR tex src type: %d\n",
655 tex
->src
[i
].src_type
);
666 compile_error(ctx
, "unimplemented texop %d\n", tex
->op
);
670 switch (tex
->sampler_dim
) {
671 case GLSL_SAMPLER_DIM_2D
:
673 case GLSL_SAMPLER_DIM_RECT
:
676 case GLSL_SAMPLER_DIM_CUBE
:
680 compile_error(ctx
, "unimplemented sampler %d\n", tex
->sampler_dim
);
684 struct ir2_src src_coord
= make_src_noconst(ctx
, *coord
);
688 * tmp.xy = tmp.xy / |tmp.z| + 1.5
692 struct ir2_instr
*rcp
, *coord_xy
;
695 instr
= instr_create_alu_reg(ctx
, ir2_op_cube
, 15, NULL
);
696 instr
->src
[0] = src_coord
;
697 instr
->src
[0].swizzle
= IR2_SWIZZLE_ZZXY
;
698 instr
->src
[1] = src_coord
;
699 instr
->src
[1].swizzle
= IR2_SWIZZLE_YXZZ
;
701 reg_idx
= instr
->reg
- ctx
->reg
; /* hacky */
703 rcp
= instr_create_alu(ctx
, nir_op_frcp
, 1);
704 rcp
->src
[0] = ir2_src(reg_idx
, IR2_SWIZZLE_Z
, IR2_SRC_REG
);
705 rcp
->src
[0].abs
= true;
707 coord_xy
= instr_create_alu_reg(ctx
, nir_op_ffma
, 3, instr
);
708 coord_xy
->src
[0] = ir2_src(reg_idx
, 0, IR2_SRC_REG
);
709 coord_xy
->src
[1] = ir2_src(rcp
->idx
, IR2_SWIZZLE_XXXX
, IR2_SRC_SSA
);
710 coord_xy
->src
[2] = load_const(ctx
, (float[]) {1.5f
}, 1);
712 src_coord
= ir2_src(reg_idx
, 0, IR2_SRC_REG
);
713 /* TODO: lod/bias transformed by src_coord.z ? */
716 instr
= ir2_instr_create_fetch(ctx
, &tex
->dest
, TEX_FETCH
);
717 instr
->src
[0] = src_coord
;
718 instr
->src
[0].swizzle
= is_cube
? IR2_SWIZZLE_XYW
: 0;
719 instr
->fetch
.tex
.is_cube
= is_cube
;
720 instr
->fetch
.tex
.is_rect
= is_rect
;
721 instr
->fetch
.tex
.samp_id
= tex
->sampler_index
;
723 /* for lod/bias, we insert an extra src for the backend to deal with */
725 instr
->src
[1] = make_src_noconst(ctx
, *lod_bias
);
726 /* backend will use 2-3 components so apply swizzle */
727 swiz_merge_p(&instr
->src
[1].swizzle
, IR2_SWIZZLE_XXXX
);
728 instr
->src_count
= 2;
733 setup_input(struct ir2_context
*ctx
, nir_variable
* in
)
735 struct fd2_shader_stateobj
*so
= ctx
->so
;
736 unsigned array_len
= MAX2(glsl_get_length(in
->type
), 1);
737 unsigned n
= in
->data
.driver_location
;
738 unsigned slot
= in
->data
.location
;
740 assert(array_len
== 1);
743 if (ctx
->so
->type
== MESA_SHADER_VERTEX
)
746 if (ctx
->so
->type
!= MESA_SHADER_FRAGMENT
)
747 compile_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
749 if (slot
== VARYING_SLOT_PNTC
) {
750 so
->need_param
= true;
754 n
= ctx
->f
->inputs_count
++;
756 /* half of fragcoord from param reg, half from a varying */
757 if (slot
== VARYING_SLOT_POS
) {
758 ctx
->f
->fragcoord
= n
;
759 so
->need_param
= true;
762 ctx
->f
->inputs
[n
].slot
= slot
;
763 ctx
->f
->inputs
[n
].ncomp
= glsl_get_components(in
->type
);
765 /* in->data.interpolation?
766 * opengl ES 2.0 can't do flat mode, but we still get it from GALLIUM_HUD
771 emit_undef(struct ir2_context
*ctx
, nir_ssa_undef_instr
* undef
)
773 /* TODO we don't want to emit anything for undefs */
775 struct ir2_instr
*instr
;
777 instr
= instr_create_alu_dest(ctx
, nir_op_mov
,
778 &(nir_dest
) {.ssa
= undef
->def
,.is_ssa
= true});
779 instr
->src
[0] = ir2_src(0, 0, IR2_SRC_CONST
);
783 emit_instr(struct ir2_context
*ctx
, nir_instr
* instr
)
785 switch (instr
->type
) {
786 case nir_instr_type_alu
:
787 emit_alu(ctx
, nir_instr_as_alu(instr
));
789 case nir_instr_type_deref
:
790 /* ignored, handled as part of the intrinsic they are src to */
792 case nir_instr_type_intrinsic
:
793 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
795 case nir_instr_type_load_const
:
796 /* dealt with when using nir_src */
798 case nir_instr_type_tex
:
799 emit_tex(ctx
, nir_instr_as_tex(instr
));
801 case nir_instr_type_jump
:
802 ctx
->block_has_jump
[ctx
->block_idx
] = true;
804 case nir_instr_type_ssa_undef
:
805 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
812 /* fragcoord.zw and a20x hw binning outputs */
814 extra_position_exports(struct ir2_context
*ctx
, bool binning
)
816 struct ir2_instr
*instr
, *rcp
, *sc
, *wincoord
, *off
;
818 if (ctx
->f
->fragcoord
< 0 && !binning
)
821 instr
= instr_create_alu(ctx
, nir_op_fmax
, 1);
822 instr
->src
[0] = ctx
->position
;
823 instr
->src
[0].swizzle
= IR2_SWIZZLE_W
;
824 instr
->src
[1] = ir2_zero(ctx
);
826 rcp
= instr_create_alu(ctx
, nir_op_frcp
, 1);
827 rcp
->src
[0] = ir2_src(instr
->idx
, 0, IR2_SRC_SSA
);
829 sc
= instr_create_alu(ctx
, nir_op_fmul
, 4);
830 sc
->src
[0] = ctx
->position
;
831 sc
->src
[1] = ir2_src(rcp
->idx
, IR2_SWIZZLE_XXXX
, IR2_SRC_SSA
);
833 wincoord
= instr_create_alu(ctx
, nir_op_ffma
, 4);
834 wincoord
->src
[0] = ir2_src(66, 0, IR2_SRC_CONST
);
835 wincoord
->src
[1] = ir2_src(sc
->idx
, 0, IR2_SRC_SSA
);
836 wincoord
->src
[2] = ir2_src(65, 0, IR2_SRC_CONST
);
839 if (ctx
->f
->fragcoord
>= 0 && !binning
) {
840 instr
= instr_create_alu(ctx
, nir_op_mov
, 1);
841 instr
->src
[0] = ir2_src(wincoord
->idx
, IR2_SWIZZLE_Z
, IR2_SRC_SSA
);
842 instr
->alu
.export
= ctx
->f
->fragcoord
;
844 instr
= instr_create_alu(ctx
, nir_op_mov
, 1);
845 instr
->src
[0] = ctx
->position
;
846 instr
->src
[0].swizzle
= IR2_SWIZZLE_W
;
847 instr
->alu
.export
= ctx
->f
->fragcoord
;
848 instr
->alu
.write_mask
= 2;
854 off
= instr_create_alu(ctx
, nir_op_fadd
, 1);
855 off
->src
[0] = ir2_src(64, 0, IR2_SRC_CONST
);
856 off
->src
[1] = ir2_src(2, 0, IR2_SRC_INPUT
);
858 /* 8 max set in freedreno_screen.. unneeded instrs patched out */
859 for (int i
= 0; i
< 8; i
++) {
860 instr
= instr_create_alu(ctx
, nir_op_ffma
, 4);
861 instr
->src
[0] = ir2_src(1, IR2_SWIZZLE_WYWW
, IR2_SRC_CONST
);
862 instr
->src
[1] = ir2_src(off
->idx
, IR2_SWIZZLE_XXXX
, IR2_SRC_SSA
);
863 instr
->src
[2] = ir2_src(3 + i
, 0, IR2_SRC_CONST
);
864 instr
->alu
.export
= 32;
866 instr
= instr_create_alu(ctx
, nir_op_ffma
, 4);
867 instr
->src
[0] = ir2_src(68 + i
* 2, 0, IR2_SRC_CONST
);
868 instr
->src
[1] = ir2_src(wincoord
->idx
, 0, IR2_SRC_SSA
);
869 instr
->src
[2] = ir2_src(67 + i
* 2, 0, IR2_SRC_CONST
);
870 instr
->alu
.export
= 33;
874 static bool emit_cf_list(struct ir2_context
*ctx
, struct exec_list
*list
);
877 emit_block(struct ir2_context
*ctx
, nir_block
* block
)
879 struct ir2_instr
*instr
;
880 nir_block
*succs
= block
->successors
[0];
882 ctx
->block_idx
= block
->index
;
884 nir_foreach_instr(instr
, block
)
885 emit_instr(ctx
, instr
);
887 if (!succs
|| !succs
->index
)
890 /* we want to be smart and always jump and have the backend cleanup
891 * but we are not, so there are two cases where jump is needed:
892 * loops (succs index lower)
893 * jumps (jump instruction seen in block)
895 if (succs
->index
> block
->index
&& !ctx
->block_has_jump
[block
->index
])
898 assert(block
->successors
[1] == NULL
);
900 instr
= ir2_instr_create(ctx
, IR2_CF
);
901 instr
->cf
.block_idx
= succs
->index
;
902 /* XXX can't jump to a block with different predicate */
907 emit_if(struct ir2_context
*ctx
, nir_if
* nif
)
909 unsigned pred
= ctx
->pred
, pred_idx
= ctx
->pred_idx
;
910 struct ir2_instr
*instr
;
912 /* XXX: blob seems to always use same register for condition */
914 instr
= ir2_instr_create(ctx
, IR2_ALU
);
915 instr
->src
[0] = make_src(ctx
, nif
->condition
);
916 instr
->src_count
= 1;
917 instr
->ssa
.ncomp
= 1;
918 instr
->alu
.vector_opc
= VECTOR_NONE
;
919 instr
->alu
.scalar_opc
= SCALAR_NONE
;
920 instr
->alu
.export
= -1;
921 instr
->alu
.write_mask
= 1;
924 /* if nested, use PRED_SETNE_PUSHv */
926 instr
->alu
.vector_opc
= PRED_SETNE_PUSHv
;
927 instr
->src
[1] = instr
->src
[0];
928 instr
->src
[0] = ir2_src(pred_idx
, 0, IR2_SRC_SSA
);
929 instr
->src
[0].swizzle
= IR2_SWIZZLE_XXXX
;
930 instr
->src
[1].swizzle
= IR2_SWIZZLE_XXXX
;
931 instr
->src_count
= 2;
933 instr
->alu
.scalar_opc
= PRED_SETNEs
;
936 ctx
->pred_idx
= instr
->idx
;
939 emit_cf_list(ctx
, &nif
->then_list
);
941 /* TODO: if these is no else branch we don't need this
942 * and if the else branch is simple, can just flip ctx->pred instead
944 instr
= ir2_instr_create(ctx
, IR2_ALU
);
945 instr
->src
[0] = ir2_src(ctx
->pred_idx
, 0, IR2_SRC_SSA
);
946 instr
->src_count
= 1;
947 instr
->ssa
.ncomp
= 1;
948 instr
->alu
.vector_opc
= VECTOR_NONE
;
949 instr
->alu
.scalar_opc
= PRED_SET_INVs
;
950 instr
->alu
.export
= -1;
951 instr
->alu
.write_mask
= 1;
953 ctx
->pred_idx
= instr
->idx
;
955 emit_cf_list(ctx
, &nif
->else_list
);
957 /* restore predicate for nested predicates */
959 instr
= ir2_instr_create(ctx
, IR2_ALU
);
960 instr
->src
[0] = ir2_src(ctx
->pred_idx
, 0, IR2_SRC_SSA
);
961 instr
->src_count
= 1;
962 instr
->ssa
.ncomp
= 1;
963 instr
->alu
.vector_opc
= VECTOR_NONE
;
964 instr
->alu
.scalar_opc
= PRED_SET_POPs
;
965 instr
->alu
.export
= -1;
966 instr
->alu
.write_mask
= 1;
968 ctx
->pred_idx
= instr
->idx
;
971 /* restore ctx->pred */
975 /* get the highest block idx in the loop, so we know when
976 * we can free registers that are allocated outside the loop
979 loop_last_block(struct exec_list
*list
)
982 exec_node_data(nir_cf_node
, exec_list_get_tail(list
), node
);
983 switch (node
->type
) {
984 case nir_cf_node_block
:
985 return nir_cf_node_as_block(node
)->index
;
987 assert(0); /* XXX could this ever happen? */
989 case nir_cf_node_loop
:
990 return loop_last_block(&nir_cf_node_as_loop(node
)->body
);
992 compile_error(ctx
, "Not supported\n");
998 emit_loop(struct ir2_context
*ctx
, nir_loop
*nloop
)
1000 ctx
->loop_last_block
[++ctx
->loop_depth
] = loop_last_block(&nloop
->body
);
1001 emit_cf_list(ctx
, &nloop
->body
);
1006 emit_cf_list(struct ir2_context
*ctx
, struct exec_list
*list
)
1009 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
1011 switch (node
->type
) {
1012 case nir_cf_node_block
:
1013 ret
= emit_block(ctx
, nir_cf_node_as_block(node
));
1015 case nir_cf_node_if
:
1016 emit_if(ctx
, nir_cf_node_as_if(node
));
1018 case nir_cf_node_loop
:
1019 emit_loop(ctx
, nir_cf_node_as_loop(node
));
1021 case nir_cf_node_function
:
1022 compile_error(ctx
, "Not supported\n");
1029 static void cleanup_binning(struct ir2_context
*ctx
)
1031 assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
1033 /* kill non-position outputs for binning variant */
1034 nir_foreach_block(block
, nir_shader_get_entrypoint(ctx
->nir
)) {
1035 nir_foreach_instr_safe(instr
, block
) {
1036 if (instr
->type
!= nir_instr_type_intrinsic
)
1039 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1040 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
1043 if (output_slot(ctx
, intr
) != VARYING_SLOT_POS
)
1044 nir_instr_remove(instr
);
1048 ir2_optimize_nir(ctx
->nir
, false);
1052 ir2_nir_compile(struct ir2_context
*ctx
, bool binning
)
1054 struct fd2_shader_stateobj
*so
= ctx
->so
;
1056 memset(ctx
->ssa_map
, 0xff, sizeof(ctx
->ssa_map
));
1058 ctx
->nir
= nir_shader_clone(NULL
, so
->nir
);
1061 cleanup_binning(ctx
);
1064 OPT_V(ctx
->nir
, nir_opt_algebraic_late
);
1066 OPT_V(ctx
->nir
, nir_lower_to_source_mods
, nir_lower_all_source_mods
);
1067 OPT_V(ctx
->nir
, nir_copy_prop
);
1068 OPT_V(ctx
->nir
, nir_opt_dce
);
1069 OPT_V(ctx
->nir
, nir_opt_move_comparisons
);
1071 OPT_V(ctx
->nir
, nir_lower_bool_to_float
);
1072 OPT_V(ctx
->nir
, nir_lower_int_to_float
);
1074 /* lower to scalar instructions that can only be scalar on a2xx */
1075 OPT_V(ctx
->nir
, ir2_nir_lower_scalar
);
1077 OPT_V(ctx
->nir
, nir_lower_locals_to_regs
);
1079 OPT_V(ctx
->nir
, nir_convert_from_ssa
, true);
1081 OPT_V(ctx
->nir
, nir_move_vec_src_uses_to_dest
);
1082 OPT_V(ctx
->nir
, nir_lower_vec_to_movs
);
1084 OPT_V(ctx
->nir
, nir_opt_dce
);
1086 nir_sweep(ctx
->nir
);
1088 if (fd_mesa_debug
& FD_DBG_DISASM
) {
1089 debug_printf("----------------------\n");
1090 nir_print_shader(ctx
->nir
, stdout
);
1091 debug_printf("----------------------\n");
1094 /* fd2_shader_stateobj init */
1095 if (so
->type
== MESA_SHADER_FRAGMENT
) {
1096 ctx
->f
->fragcoord
= -1;
1097 ctx
->f
->inputs_count
= 0;
1098 memset(ctx
->f
->inputs
, 0, sizeof(ctx
->f
->inputs
));
1102 nir_foreach_variable(in
, &ctx
->nir
->inputs
)
1103 setup_input(ctx
, in
);
1105 if (so
->type
== MESA_SHADER_FRAGMENT
) {
1107 for (idx
= 0; idx
< ctx
->f
->inputs_count
; idx
++) {
1108 ctx
->input
[idx
].ncomp
= ctx
->f
->inputs
[idx
].ncomp
;
1109 update_range(ctx
, &ctx
->input
[idx
]);
1111 /* assume we have param input and kill it later if not */
1112 ctx
->input
[idx
].ncomp
= 4;
1113 update_range(ctx
, &ctx
->input
[idx
]);
1115 ctx
->input
[0].ncomp
= 1;
1116 ctx
->input
[2].ncomp
= 1;
1117 update_range(ctx
, &ctx
->input
[0]);
1118 update_range(ctx
, &ctx
->input
[2]);
1121 /* And emit the body: */
1122 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->nir
);
1124 nir_foreach_register(reg
, &fxn
->registers
) {
1125 ctx
->reg
[reg
->index
].ncomp
= reg
->num_components
;
1126 ctx
->reg_count
= MAX2(ctx
->reg_count
, reg
->index
+ 1);
1129 nir_metadata_require(fxn
, nir_metadata_block_index
);
1130 emit_cf_list(ctx
, &fxn
->body
);
1131 /* TODO emit_block(ctx, fxn->end_block); */
1133 if (so
->type
== MESA_SHADER_VERTEX
)
1134 extra_position_exports(ctx
, binning
);
1136 ralloc_free(ctx
->nir
);
1138 /* kill unused param input */
1139 if (so
->type
== MESA_SHADER_FRAGMENT
&& !so
->need_param
)
1140 ctx
->input
[ctx
->f
->inputs_count
].initialized
= false;