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,
46 const nir_shader_compiler_options
*
47 ir2_get_compiler_options(void)
52 #define OPT(nir, pass, ...) ({ \
53 bool this_progress = false; \
54 NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__); \
57 #define OPT_V(nir, pass, ...) NIR_PASS_V(nir, pass, ##__VA_ARGS__)
60 ir2_optimize_loop(nir_shader
*s
)
66 OPT_V(s
, nir_lower_vars_to_ssa
);
67 progress
|= OPT(s
, nir_opt_copy_prop_vars
);
68 progress
|= OPT(s
, nir_copy_prop
);
69 progress
|= OPT(s
, nir_opt_dce
);
70 progress
|= OPT(s
, nir_opt_cse
);
71 /* progress |= OPT(s, nir_opt_gcm, true); */
72 progress
|= OPT(s
, nir_opt_peephole_select
, UINT_MAX
, true, true);
73 progress
|= OPT(s
, nir_opt_intrinsics
);
74 progress
|= OPT(s
, nir_opt_algebraic
);
75 progress
|= OPT(s
, nir_opt_constant_folding
);
76 progress
|= OPT(s
, nir_opt_dead_cf
);
77 if (OPT(s
, nir_opt_trivial_continues
)) {
79 /* If nir_opt_trivial_continues makes progress, then we need to clean
80 * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
83 OPT(s
, nir_copy_prop
);
86 progress
|= OPT(s
, nir_opt_loop_unroll
, nir_var_all
);
87 progress
|= OPT(s
, nir_opt_if
, false);
88 progress
|= OPT(s
, nir_opt_remove_phis
);
89 progress
|= OPT(s
, nir_opt_undef
);
95 /* trig workarounds is the same as ir3.. but we don't want to include ir3 */
96 bool ir3_nir_apply_trig_workarounds(nir_shader
* shader
);
99 ir2_optimize_nir(nir_shader
*s
, bool lower
)
101 struct nir_lower_tex_options tex_options
= {
106 if (fd_mesa_debug
& FD_DBG_DISASM
) {
107 debug_printf("----------------------\n");
108 nir_print_shader(s
, stdout
);
109 debug_printf("----------------------\n");
112 OPT_V(s
, nir_lower_regs_to_ssa
);
113 OPT_V(s
, nir_lower_vars_to_ssa
);
114 OPT_V(s
, nir_lower_indirect_derefs
, nir_var_shader_in
| nir_var_shader_out
);
117 OPT_V(s
, ir3_nir_apply_trig_workarounds
);
118 OPT_V(s
, nir_lower_tex
, &tex_options
);
121 ir2_optimize_loop(s
);
123 OPT_V(s
, nir_remove_dead_variables
, nir_var_function_temp
);
124 OPT_V(s
, nir_move_load_const
);
126 /* TODO we dont want to get shaders writing to depth for depth textures */
127 if (s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
128 nir_foreach_variable(var
, &s
->outputs
) {
129 if (var
->data
.location
== FRAG_RESULT_DEPTH
)
137 static struct ir2_src
138 load_const(struct ir2_context
*ctx
, float *value_f
, unsigned ncomp
)
140 struct fd2_shader_stateobj
*so
= ctx
->so
;
141 unsigned imm_ncomp
, swiz
, idx
, i
, j
;
142 uint32_t *value
= (uint32_t*) value_f
;
144 /* try to merge with existing immediate (TODO: try with neg) */
145 for (idx
= 0; idx
< so
->num_immediates
; idx
++) {
147 imm_ncomp
= so
->immediates
[idx
].ncomp
;
148 for (i
= 0; i
< ncomp
; i
++) {
149 for (j
= 0; j
< imm_ncomp
; j
++) {
150 if (value
[i
] == so
->immediates
[idx
].val
[j
])
153 if (j
== imm_ncomp
) {
156 so
->immediates
[idx
].val
[imm_ncomp
++] = value
[i
];
158 swiz
|= swiz_set(j
, i
);
160 /* matched all components */
165 /* need to allocate new immediate */
166 if (idx
== so
->num_immediates
) {
169 for (i
= 0; i
< ncomp
; i
++) {
170 for (j
= 0; j
< imm_ncomp
; j
++) {
171 if (value
[i
] == ctx
->so
->immediates
[idx
].val
[j
])
174 if (j
== imm_ncomp
) {
175 so
->immediates
[idx
].val
[imm_ncomp
++] = value
[i
];
177 swiz
|= swiz_set(j
, i
);
179 so
->num_immediates
++;
181 so
->immediates
[idx
].ncomp
= imm_ncomp
;
184 swiz
= swiz_merge(swiz
, IR2_SWIZZLE_XXXX
);
186 return ir2_src(so
->first_immediate
+ idx
, swiz
, IR2_SRC_CONST
);
190 ir2_zero(struct ir2_context
*ctx
)
192 return load_const(ctx
, (float[]) {0.0f
}, 1);
196 update_range(struct ir2_context
*ctx
, struct ir2_reg
*reg
)
198 if (!reg
->initialized
) {
199 reg
->initialized
= true;
200 reg
->loop_depth
= ctx
->loop_depth
;
203 if (ctx
->loop_depth
> reg
->loop_depth
) {
204 reg
->block_idx_free
= ctx
->loop_last_block
[reg
->loop_depth
+ 1];
206 reg
->loop_depth
= ctx
->loop_depth
;
207 reg
->block_idx_free
= -1;
210 /* for regs we want to free at the end of the loop in any case
211 * XXX dont do this for ssa
214 reg
->block_idx_free
= ctx
->loop_last_block
[reg
->loop_depth
];
217 static struct ir2_src
218 make_src(struct ir2_context
*ctx
, nir_src src
)
220 struct ir2_src res
= {};
223 nir_const_value
*const_value
= nir_src_as_const_value(src
);
227 float c
[src
.ssa
->num_components
];
228 nir_const_value_to_array(c
, const_value
, src
.ssa
->num_components
, f32
);
229 return load_const(ctx
, c
, src
.ssa
->num_components
);
233 res
.num
= src
.reg
.reg
->index
;
234 res
.type
= IR2_SRC_REG
;
235 reg
= &ctx
->reg
[res
.num
];
237 assert(ctx
->ssa_map
[src
.ssa
->index
] >= 0);
238 res
.num
= ctx
->ssa_map
[src
.ssa
->index
];
239 res
.type
= IR2_SRC_SSA
;
240 reg
= &ctx
->instr
[res
.num
].ssa
;
243 update_range(ctx
, reg
);
248 set_index(struct ir2_context
*ctx
, nir_dest
* dst
,
249 struct ir2_instr
*instr
)
251 struct ir2_reg
*reg
= &instr
->ssa
;
254 ctx
->ssa_map
[dst
->ssa
.index
] = instr
->idx
;
256 assert(instr
->is_ssa
);
257 reg
= &ctx
->reg
[dst
->reg
.reg
->index
];
259 instr
->is_ssa
= false;
262 update_range(ctx
, reg
);
265 static struct ir2_instr
*
266 ir2_instr_create(struct ir2_context
*ctx
, int type
)
268 struct ir2_instr
*instr
;
270 instr
= &ctx
->instr
[ctx
->instr_count
++];
271 instr
->idx
= ctx
->instr_count
- 1;
273 instr
->block_idx
= ctx
->block_idx
;
274 instr
->pred
= ctx
->pred
;
275 instr
->is_ssa
= true;
279 static struct ir2_instr
*
280 instr_create_alu(struct ir2_context
*ctx
, nir_op opcode
, unsigned ncomp
)
282 /* emit_alu will fixup instrs that don't map directly */
283 static const struct ir2_opc
{
284 int8_t scalar
, vector
;
285 } nir_ir2_opc
[nir_num_opcodes
+1] = {
286 [0 ... nir_num_opcodes
- 1] = {-1, -1},
288 [nir_op_mov
] = {MAXs
, MAXv
},
289 [nir_op_fsign
] = {-1, CNDGTEv
},
290 [nir_op_fadd
] = {ADDs
, ADDv
},
291 [nir_op_fsub
] = {ADDs
, ADDv
},
292 [nir_op_fmul
] = {MULs
, MULv
},
293 [nir_op_ffma
] = {-1, MULADDv
},
294 [nir_op_fmax
] = {MAXs
, MAXv
},
295 [nir_op_fmin
] = {MINs
, MINv
},
296 [nir_op_ffloor
] = {FLOORs
, FLOORv
},
297 [nir_op_ffract
] = {FRACs
, FRACv
},
298 [nir_op_ftrunc
] = {TRUNCs
, TRUNCv
},
299 [nir_op_fdot2
] = {-1, DOT2ADDv
},
300 [nir_op_fdot3
] = {-1, DOT3v
},
301 [nir_op_fdot4
] = {-1, DOT4v
},
302 [nir_op_sge
] = {-1, SETGTEv
},
303 [nir_op_slt
] = {-1, SETGTv
},
304 [nir_op_sne
] = {-1, SETNEv
},
305 [nir_op_seq
] = {-1, SETEv
},
306 [nir_op_fcsel
] = {-1, CNDEv
},
307 [nir_op_frsq
] = {RECIPSQ_IEEE
, -1},
308 [nir_op_frcp
] = {RECIP_IEEE
, -1},
309 [nir_op_flog2
] = {LOG_IEEE
, -1},
310 [nir_op_fexp2
] = {EXP_IEEE
, -1},
311 [nir_op_fsqrt
] = {SQRT_IEEE
, -1},
312 [nir_op_fcos
] = {COS
, -1},
313 [nir_op_fsin
] = {SIN
, -1},
314 /* no fsat, fneg, fabs since source mods deal with those */
316 /* so we can use this function with non-nir op */
317 #define ir2_op_cube nir_num_opcodes
318 [ir2_op_cube
] = {-1, CUBEv
},
321 struct ir2_opc op
= nir_ir2_opc
[opcode
];
322 assert(op
.vector
>= 0 || op
.scalar
>= 0);
324 struct ir2_instr
*instr
= ir2_instr_create(ctx
, IR2_ALU
);
325 instr
->alu
.vector_opc
= op
.vector
;
326 instr
->alu
.scalar_opc
= op
.scalar
;
327 instr
->alu
.export
= -1;
328 instr
->alu
.write_mask
= (1 << ncomp
) - 1;
329 instr
->src_count
= opcode
== ir2_op_cube
? 2 :
330 nir_op_infos
[opcode
].num_inputs
;
331 instr
->ssa
.ncomp
= ncomp
;
335 static struct ir2_instr
*
336 instr_create_alu_reg(struct ir2_context
*ctx
, nir_op opcode
,
337 uint8_t write_mask
, struct ir2_instr
*share_reg
)
339 struct ir2_instr
*instr
;
342 reg
= share_reg
? share_reg
->reg
: &ctx
->reg
[ctx
->reg_count
++];
343 reg
->ncomp
= MAX2(reg
->ncomp
, util_logbase2(write_mask
) + 1);
345 instr
= instr_create_alu(ctx
, opcode
, util_bitcount(write_mask
));
346 instr
->alu
.write_mask
= write_mask
;
348 instr
->is_ssa
= false;
353 static struct ir2_instr
*
354 instr_create_alu_dest(struct ir2_context
*ctx
, nir_op opcode
, nir_dest
*dst
)
356 struct ir2_instr
*instr
;
357 instr
= instr_create_alu(ctx
, opcode
, nir_dest_num_components(*dst
));
358 set_index(ctx
, dst
, instr
);
362 static struct ir2_instr
*
363 ir2_instr_create_fetch(struct ir2_context
*ctx
, nir_dest
*dst
,
364 instr_fetch_opc_t opc
)
366 struct ir2_instr
*instr
= ir2_instr_create(ctx
, IR2_FETCH
);
367 instr
->fetch
.opc
= opc
;
368 instr
->src_count
= 1;
369 instr
->ssa
.ncomp
= nir_dest_num_components(*dst
);
370 set_index(ctx
, dst
, instr
);
374 static struct ir2_src
375 make_src_noconst(struct ir2_context
*ctx
, nir_src src
)
377 struct ir2_instr
*instr
;
379 if (nir_src_as_const_value(src
)) {
381 instr
= instr_create_alu(ctx
, nir_op_mov
, src
.ssa
->num_components
);
382 instr
->src
[0] = make_src(ctx
, src
);
383 return ir2_src(instr
->idx
, 0, IR2_SRC_SSA
);
386 return make_src(ctx
, src
);
390 emit_alu(struct ir2_context
*ctx
, nir_alu_instr
* alu
)
392 const nir_op_info
*info
= &nir_op_infos
[alu
->op
];
393 nir_dest
*dst
= &alu
->dest
.dest
;
394 struct ir2_instr
*instr
;
398 /* get the number of dst components */
400 ncomp
= dst
->ssa
.num_components
;
403 for (int i
= 0; i
< 4; i
++)
404 ncomp
+= !!(alu
->dest
.write_mask
& 1 << i
);
407 instr
= instr_create_alu(ctx
, alu
->op
, ncomp
);
408 set_index(ctx
, dst
, instr
);
409 instr
->alu
.saturate
= alu
->dest
.saturate
;
410 instr
->alu
.write_mask
= alu
->dest
.write_mask
;
412 for (int i
= 0; i
< info
->num_inputs
; i
++) {
413 nir_alu_src
*src
= &alu
->src
[i
];
415 /* compress swizzle with writemask when applicable */
416 unsigned swiz
= 0, j
= 0;
417 for (int i
= 0; i
< 4; i
++) {
418 if (!(alu
->dest
.write_mask
& 1 << i
) && !info
->output_size
)
420 swiz
|= swiz_set(src
->swizzle
[i
], j
++);
423 instr
->src
[i
] = make_src(ctx
, src
->src
);
424 instr
->src
[i
].swizzle
= swiz_merge(instr
->src
[i
].swizzle
, swiz
);
425 instr
->src
[i
].negate
= src
->negate
;
426 instr
->src
[i
].abs
= src
->abs
;
429 /* workarounds for NIR ops that don't map directly to a2xx ops */
433 instr
->src
[0] = instr
->src
[1];
438 instr
->src
[1] = instr
->src
[2];
442 instr
->src
[1].negate
= !instr
->src
[1].negate
;
445 instr
->src_count
= 3;
446 instr
->src
[2] = ir2_zero(ctx
);
449 /* we need an extra instruction to deal with the zero case */
450 struct ir2_instr
*tmp
;
452 /* tmp = x == 0 ? 0 : 1 */
453 tmp
= instr_create_alu(ctx
, nir_op_fcsel
, ncomp
);
454 tmp
->src
[0] = instr
->src
[0];
455 tmp
->src
[1] = ir2_zero(ctx
);
456 tmp
->src
[2] = load_const(ctx
, (float[]) {1.0f
}, 1);
458 /* result = x >= 0 ? tmp : -tmp */
459 instr
->src
[1] = ir2_src(tmp
->idx
, 0, IR2_SRC_SSA
);
460 instr
->src
[2] = instr
->src
[1];
461 instr
->src
[2].negate
= true;
462 instr
->src_count
= 3;
470 load_input(struct ir2_context
*ctx
, nir_dest
*dst
, unsigned idx
)
472 struct ir2_instr
*instr
;
475 if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
476 instr
= ir2_instr_create_fetch(ctx
, dst
, 0);
477 instr
->src
[0] = ir2_src(0, 0, IR2_SRC_INPUT
);
478 instr
->fetch
.vtx
.const_idx
= 20 + (idx
/ 3);
479 instr
->fetch
.vtx
.const_idx_sel
= idx
% 3;
483 /* get slot from idx */
484 nir_foreach_variable(var
, &ctx
->nir
->inputs
) {
485 if (var
->data
.driver_location
== idx
) {
486 slot
= var
->data
.location
;
493 case VARYING_SLOT_PNTC
:
494 /* need to extract with abs and invert y */
495 instr
= instr_create_alu_dest(ctx
, nir_op_ffma
, dst
);
496 instr
->src
[0] = ir2_src(ctx
->f
->inputs_count
, IR2_SWIZZLE_ZW
, IR2_SRC_INPUT
);
497 instr
->src
[0].abs
= true;
498 instr
->src
[1] = load_const(ctx
, (float[]) {1.0f
, -1.0f
}, 2);
499 instr
->src
[2] = load_const(ctx
, (float[]) {0.0f
, 1.0f
}, 2);
501 case VARYING_SLOT_POS
:
502 /* need to extract xy with abs and add tile offset on a20x
503 * zw from fragcoord input (w inverted in fragment shader)
504 * TODO: only components that are required by fragment shader
506 instr
= instr_create_alu_reg(ctx
,
507 ctx
->so
->is_a20x
? nir_op_fadd
: nir_op_mov
, 3, NULL
);
508 instr
->src
[0] = ir2_src(ctx
->f
->inputs_count
, 0, IR2_SRC_INPUT
);
509 instr
->src
[0].abs
= true;
510 /* on a20x, C64 contains the tile offset */
511 instr
->src
[1] = ir2_src(64, 0, IR2_SRC_CONST
);
513 instr
= instr_create_alu_reg(ctx
, nir_op_mov
, 4, instr
);
514 instr
->src
[0] = ir2_src(ctx
->f
->fragcoord
, 0, IR2_SRC_INPUT
);
516 instr
= instr_create_alu_reg(ctx
, nir_op_frcp
, 8, instr
);
517 instr
->src
[0] = ir2_src(ctx
->f
->fragcoord
, IR2_SWIZZLE_Y
, IR2_SRC_INPUT
);
519 unsigned reg_idx
= instr
->reg
- ctx
->reg
; /* XXX */
520 instr
= instr_create_alu_dest(ctx
, nir_op_mov
, dst
);
521 instr
->src
[0] = ir2_src(reg_idx
, 0, IR2_SRC_REG
);
524 instr
= instr_create_alu_dest(ctx
, nir_op_mov
, dst
);
525 instr
->src
[0] = ir2_src(idx
, 0, IR2_SRC_INPUT
);
531 output_slot(struct ir2_context
*ctx
, nir_intrinsic_instr
*intr
)
534 unsigned idx
= nir_intrinsic_base(intr
);
535 nir_foreach_variable(var
, &ctx
->nir
->outputs
) {
536 if (var
->data
.driver_location
== idx
) {
537 slot
= var
->data
.location
;
546 store_output(struct ir2_context
*ctx
, nir_src src
, unsigned slot
, unsigned ncomp
)
548 struct ir2_instr
*instr
;
551 if (ctx
->so
->type
== MESA_SHADER_VERTEX
) {
553 case VARYING_SLOT_POS
:
554 ctx
->position
= make_src(ctx
, src
);
557 case VARYING_SLOT_PSIZ
:
558 ctx
->so
->writes_psize
= true;
562 /* find matching slot from fragment shader input */
563 for (idx
= 0; idx
< ctx
->f
->inputs_count
; idx
++)
564 if (ctx
->f
->inputs
[idx
].slot
== slot
)
566 if (idx
== ctx
->f
->inputs_count
)
569 } else if (slot
!= FRAG_RESULT_COLOR
&& slot
!= FRAG_RESULT_DATA0
) {
570 /* only color output is implemented */
574 instr
= instr_create_alu(ctx
, nir_op_mov
, ncomp
);
575 instr
->src
[0] = make_src(ctx
, src
);
576 instr
->alu
.export
= idx
;
580 emit_intrinsic(struct ir2_context
*ctx
, nir_intrinsic_instr
*intr
)
582 struct ir2_instr
*instr
;
583 nir_const_value
*const_offset
;
586 switch (intr
->intrinsic
) {
587 case nir_intrinsic_load_input
:
588 load_input(ctx
, &intr
->dest
, nir_intrinsic_base(intr
));
590 case nir_intrinsic_store_output
:
591 store_output(ctx
, intr
->src
[0], output_slot(ctx
, intr
), intr
->num_components
);
593 case nir_intrinsic_load_uniform
:
594 const_offset
= nir_src_as_const_value(intr
->src
[0]);
595 assert(const_offset
); /* TODO can be false in ES2? */
596 idx
= nir_intrinsic_base(intr
);
597 idx
+= (uint32_t) nir_src_as_const_value(intr
->src
[0])[0].f32
;
598 instr
= instr_create_alu_dest(ctx
, nir_op_mov
, &intr
->dest
);
599 instr
->src
[0] = ir2_src(idx
, 0, IR2_SRC_CONST
);
601 case nir_intrinsic_discard
:
602 case nir_intrinsic_discard_if
:
603 instr
= ir2_instr_create(ctx
, IR2_ALU
);
604 instr
->alu
.vector_opc
= VECTOR_NONE
;
605 if (intr
->intrinsic
== nir_intrinsic_discard_if
) {
606 instr
->alu
.scalar_opc
= KILLNEs
;
607 instr
->src
[0] = make_src(ctx
, intr
->src
[0]);
609 instr
->alu
.scalar_opc
= KILLEs
;
610 instr
->src
[0] = ir2_zero(ctx
);
612 instr
->alu
.export
= -1;
613 instr
->src_count
= 1;
614 ctx
->so
->has_kill
= true;
616 case nir_intrinsic_load_front_face
:
617 /* gl_FrontFacing is in the sign of param.x
618 * rcp required because otherwise we can't differentiate -0.0 and +0.0
620 ctx
->so
->need_param
= true;
622 struct ir2_instr
*tmp
= instr_create_alu(ctx
, nir_op_frcp
, 1);
623 tmp
->src
[0] = ir2_src(ctx
->f
->inputs_count
, 0, IR2_SRC_INPUT
);
625 instr
= instr_create_alu_dest(ctx
, nir_op_sge
, &intr
->dest
);
626 instr
->src
[0] = ir2_src(tmp
->idx
, 0, IR2_SRC_SSA
);
627 instr
->src
[1] = ir2_zero(ctx
);
630 compile_error(ctx
, "unimplemented intr %d\n", intr
->intrinsic
);
636 emit_tex(struct ir2_context
*ctx
, nir_tex_instr
* tex
)
638 bool is_rect
= false, is_cube
= false;
639 struct ir2_instr
*instr
;
640 nir_src
*coord
, *lod_bias
;
642 coord
= lod_bias
= NULL
;
644 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
645 switch (tex
->src
[i
].src_type
) {
646 case nir_tex_src_coord
:
647 coord
= &tex
->src
[i
].src
;
649 case nir_tex_src_bias
:
650 case nir_tex_src_lod
:
652 lod_bias
= &tex
->src
[i
].src
;
655 compile_error(ctx
, "Unhandled NIR tex src type: %d\n",
656 tex
->src
[i
].src_type
);
667 compile_error(ctx
, "unimplemented texop %d\n", tex
->op
);
671 switch (tex
->sampler_dim
) {
672 case GLSL_SAMPLER_DIM_2D
:
674 case GLSL_SAMPLER_DIM_RECT
:
677 case GLSL_SAMPLER_DIM_CUBE
:
681 compile_error(ctx
, "unimplemented sampler %d\n", tex
->sampler_dim
);
685 struct ir2_src src_coord
= make_src_noconst(ctx
, *coord
);
689 * tmp.xy = tmp.xy / |tmp.z| + 1.5
693 struct ir2_instr
*rcp
, *coord_xy
;
696 instr
= instr_create_alu_reg(ctx
, ir2_op_cube
, 15, NULL
);
697 instr
->src
[0] = src_coord
;
698 instr
->src
[0].swizzle
= IR2_SWIZZLE_ZZXY
;
699 instr
->src
[1] = src_coord
;
700 instr
->src
[1].swizzle
= IR2_SWIZZLE_YXZZ
;
702 reg_idx
= instr
->reg
- ctx
->reg
; /* hacky */
704 rcp
= instr_create_alu(ctx
, nir_op_frcp
, 1);
705 rcp
->src
[0] = ir2_src(reg_idx
, IR2_SWIZZLE_Z
, IR2_SRC_REG
);
706 rcp
->src
[0].abs
= true;
708 coord_xy
= instr_create_alu_reg(ctx
, nir_op_ffma
, 3, instr
);
709 coord_xy
->src
[0] = ir2_src(reg_idx
, 0, IR2_SRC_REG
);
710 coord_xy
->src
[1] = ir2_src(rcp
->idx
, IR2_SWIZZLE_XXXX
, IR2_SRC_SSA
);
711 coord_xy
->src
[2] = load_const(ctx
, (float[]) {1.5f
}, 1);
713 src_coord
= ir2_src(reg_idx
, 0, IR2_SRC_REG
);
714 /* TODO: lod/bias transformed by src_coord.z ? */
717 instr
= ir2_instr_create_fetch(ctx
, &tex
->dest
, TEX_FETCH
);
718 instr
->src
[0] = src_coord
;
719 instr
->src
[0].swizzle
= is_cube
? IR2_SWIZZLE_XYW
: 0;
720 instr
->fetch
.tex
.is_cube
= is_cube
;
721 instr
->fetch
.tex
.is_rect
= is_rect
;
722 instr
->fetch
.tex
.samp_id
= tex
->sampler_index
;
724 /* for lod/bias, we insert an extra src for the backend to deal with */
726 instr
->src
[1] = make_src_noconst(ctx
, *lod_bias
);
727 /* backend will use 2-3 components so apply swizzle */
728 swiz_merge_p(&instr
->src
[1].swizzle
, IR2_SWIZZLE_XXXX
);
729 instr
->src_count
= 2;
734 setup_input(struct ir2_context
*ctx
, nir_variable
* in
)
736 struct fd2_shader_stateobj
*so
= ctx
->so
;
737 unsigned array_len
= MAX2(glsl_get_length(in
->type
), 1);
738 unsigned n
= in
->data
.driver_location
;
739 unsigned slot
= in
->data
.location
;
741 assert(array_len
== 1);
744 if (ctx
->so
->type
== MESA_SHADER_VERTEX
)
747 if (ctx
->so
->type
!= MESA_SHADER_FRAGMENT
)
748 compile_error(ctx
, "unknown shader type: %d\n", ctx
->so
->type
);
750 if (slot
== VARYING_SLOT_PNTC
) {
751 so
->need_param
= true;
755 n
= ctx
->f
->inputs_count
++;
757 /* half of fragcoord from param reg, half from a varying */
758 if (slot
== VARYING_SLOT_POS
) {
759 ctx
->f
->fragcoord
= n
;
760 so
->need_param
= true;
763 ctx
->f
->inputs
[n
].slot
= slot
;
764 ctx
->f
->inputs
[n
].ncomp
= glsl_get_components(in
->type
);
766 /* in->data.interpolation?
767 * opengl ES 2.0 can't do flat mode, but we still get it from GALLIUM_HUD
772 emit_undef(struct ir2_context
*ctx
, nir_ssa_undef_instr
* undef
)
774 /* TODO we don't want to emit anything for undefs */
776 struct ir2_instr
*instr
;
778 instr
= instr_create_alu_dest(ctx
, nir_op_mov
,
779 &(nir_dest
) {.ssa
= undef
->def
,.is_ssa
= true});
780 instr
->src
[0] = ir2_src(0, 0, IR2_SRC_CONST
);
784 emit_instr(struct ir2_context
*ctx
, nir_instr
* instr
)
786 switch (instr
->type
) {
787 case nir_instr_type_alu
:
788 emit_alu(ctx
, nir_instr_as_alu(instr
));
790 case nir_instr_type_deref
:
791 /* ignored, handled as part of the intrinsic they are src to */
793 case nir_instr_type_intrinsic
:
794 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
796 case nir_instr_type_load_const
:
797 /* dealt with when using nir_src */
799 case nir_instr_type_tex
:
800 emit_tex(ctx
, nir_instr_as_tex(instr
));
802 case nir_instr_type_jump
:
803 ctx
->block_has_jump
[ctx
->block_idx
] = true;
805 case nir_instr_type_ssa_undef
:
806 emit_undef(ctx
, nir_instr_as_ssa_undef(instr
));
813 /* fragcoord.zw and a20x hw binning outputs */
815 extra_position_exports(struct ir2_context
*ctx
, bool binning
)
817 struct ir2_instr
*instr
, *rcp
, *sc
, *wincoord
, *off
;
819 if (ctx
->f
->fragcoord
< 0 && !binning
)
822 instr
= instr_create_alu(ctx
, nir_op_fmax
, 1);
823 instr
->src
[0] = ctx
->position
;
824 instr
->src
[0].swizzle
= IR2_SWIZZLE_W
;
825 instr
->src
[1] = ir2_zero(ctx
);
827 rcp
= instr_create_alu(ctx
, nir_op_frcp
, 1);
828 rcp
->src
[0] = ir2_src(instr
->idx
, 0, IR2_SRC_SSA
);
830 sc
= instr_create_alu(ctx
, nir_op_fmul
, 4);
831 sc
->src
[0] = ctx
->position
;
832 sc
->src
[1] = ir2_src(rcp
->idx
, IR2_SWIZZLE_XXXX
, IR2_SRC_SSA
);
834 wincoord
= instr_create_alu(ctx
, nir_op_ffma
, 4);
835 wincoord
->src
[0] = ir2_src(66, 0, IR2_SRC_CONST
);
836 wincoord
->src
[1] = ir2_src(sc
->idx
, 0, IR2_SRC_SSA
);
837 wincoord
->src
[2] = ir2_src(65, 0, IR2_SRC_CONST
);
840 if (ctx
->f
->fragcoord
>= 0 && !binning
) {
841 instr
= instr_create_alu(ctx
, nir_op_mov
, 1);
842 instr
->src
[0] = ir2_src(wincoord
->idx
, IR2_SWIZZLE_Z
, IR2_SRC_SSA
);
843 instr
->alu
.export
= ctx
->f
->fragcoord
;
845 instr
= instr_create_alu(ctx
, nir_op_mov
, 1);
846 instr
->src
[0] = ctx
->position
;
847 instr
->src
[0].swizzle
= IR2_SWIZZLE_W
;
848 instr
->alu
.export
= ctx
->f
->fragcoord
;
849 instr
->alu
.write_mask
= 2;
855 off
= instr_create_alu(ctx
, nir_op_fadd
, 1);
856 off
->src
[0] = ir2_src(64, 0, IR2_SRC_CONST
);
857 off
->src
[1] = ir2_src(2, 0, IR2_SRC_INPUT
);
859 /* 8 max set in freedreno_screen.. unneeded instrs patched out */
860 for (int i
= 0; i
< 8; i
++) {
861 instr
= instr_create_alu(ctx
, nir_op_ffma
, 4);
862 instr
->src
[0] = ir2_src(1, IR2_SWIZZLE_WYWW
, IR2_SRC_CONST
);
863 instr
->src
[1] = ir2_src(off
->idx
, IR2_SWIZZLE_XXXX
, IR2_SRC_SSA
);
864 instr
->src
[2] = ir2_src(3 + i
, 0, IR2_SRC_CONST
);
865 instr
->alu
.export
= 32;
867 instr
= instr_create_alu(ctx
, nir_op_ffma
, 4);
868 instr
->src
[0] = ir2_src(68 + i
* 2, 0, IR2_SRC_CONST
);
869 instr
->src
[1] = ir2_src(wincoord
->idx
, 0, IR2_SRC_SSA
);
870 instr
->src
[2] = ir2_src(67 + i
* 2, 0, IR2_SRC_CONST
);
871 instr
->alu
.export
= 33;
875 static bool emit_cf_list(struct ir2_context
*ctx
, struct exec_list
*list
);
878 emit_block(struct ir2_context
*ctx
, nir_block
* block
)
880 struct ir2_instr
*instr
;
881 nir_block
*succs
= block
->successors
[0];
883 ctx
->block_idx
= block
->index
;
885 nir_foreach_instr(instr
, block
)
886 emit_instr(ctx
, instr
);
888 if (!succs
|| !succs
->index
)
891 /* we want to be smart and always jump and have the backend cleanup
892 * but we are not, so there are two cases where jump is needed:
893 * loops (succs index lower)
894 * jumps (jump instruction seen in block)
896 if (succs
->index
> block
->index
&& !ctx
->block_has_jump
[block
->index
])
899 assert(block
->successors
[1] == NULL
);
901 instr
= ir2_instr_create(ctx
, IR2_CF
);
902 instr
->cf
.block_idx
= succs
->index
;
903 /* XXX can't jump to a block with different predicate */
908 emit_if(struct ir2_context
*ctx
, nir_if
* nif
)
910 unsigned pred
= ctx
->pred
, pred_idx
= ctx
->pred_idx
;
911 struct ir2_instr
*instr
;
913 /* XXX: blob seems to always use same register for condition */
915 instr
= ir2_instr_create(ctx
, IR2_ALU
);
916 instr
->src
[0] = make_src(ctx
, nif
->condition
);
917 instr
->src_count
= 1;
918 instr
->ssa
.ncomp
= 1;
919 instr
->alu
.vector_opc
= VECTOR_NONE
;
920 instr
->alu
.scalar_opc
= SCALAR_NONE
;
921 instr
->alu
.export
= -1;
922 instr
->alu
.write_mask
= 1;
925 /* if nested, use PRED_SETNE_PUSHv */
927 instr
->alu
.vector_opc
= PRED_SETNE_PUSHv
;
928 instr
->src
[1] = instr
->src
[0];
929 instr
->src
[0] = ir2_src(pred_idx
, 0, IR2_SRC_SSA
);
930 instr
->src
[0].swizzle
= IR2_SWIZZLE_XXXX
;
931 instr
->src
[1].swizzle
= IR2_SWIZZLE_XXXX
;
932 instr
->src_count
= 2;
934 instr
->alu
.scalar_opc
= PRED_SETNEs
;
937 ctx
->pred_idx
= instr
->idx
;
940 emit_cf_list(ctx
, &nif
->then_list
);
942 /* TODO: if these is no else branch we don't need this
943 * and if the else branch is simple, can just flip ctx->pred instead
945 instr
= ir2_instr_create(ctx
, IR2_ALU
);
946 instr
->src
[0] = ir2_src(ctx
->pred_idx
, 0, IR2_SRC_SSA
);
947 instr
->src_count
= 1;
948 instr
->ssa
.ncomp
= 1;
949 instr
->alu
.vector_opc
= VECTOR_NONE
;
950 instr
->alu
.scalar_opc
= PRED_SET_INVs
;
951 instr
->alu
.export
= -1;
952 instr
->alu
.write_mask
= 1;
954 ctx
->pred_idx
= instr
->idx
;
956 emit_cf_list(ctx
, &nif
->else_list
);
958 /* restore predicate for nested predicates */
960 instr
= ir2_instr_create(ctx
, IR2_ALU
);
961 instr
->src
[0] = ir2_src(ctx
->pred_idx
, 0, IR2_SRC_SSA
);
962 instr
->src_count
= 1;
963 instr
->ssa
.ncomp
= 1;
964 instr
->alu
.vector_opc
= VECTOR_NONE
;
965 instr
->alu
.scalar_opc
= PRED_SET_POPs
;
966 instr
->alu
.export
= -1;
967 instr
->alu
.write_mask
= 1;
969 ctx
->pred_idx
= instr
->idx
;
972 /* restore ctx->pred */
976 /* get the highest block idx in the loop, so we know when
977 * we can free registers that are allocated outside the loop
980 loop_last_block(struct exec_list
*list
)
983 exec_node_data(nir_cf_node
, exec_list_get_tail(list
), node
);
984 switch (node
->type
) {
985 case nir_cf_node_block
:
986 return nir_cf_node_as_block(node
)->index
;
988 assert(0); /* XXX could this ever happen? */
990 case nir_cf_node_loop
:
991 return loop_last_block(&nir_cf_node_as_loop(node
)->body
);
993 compile_error(ctx
, "Not supported\n");
999 emit_loop(struct ir2_context
*ctx
, nir_loop
*nloop
)
1001 ctx
->loop_last_block
[++ctx
->loop_depth
] = loop_last_block(&nloop
->body
);
1002 emit_cf_list(ctx
, &nloop
->body
);
1007 emit_cf_list(struct ir2_context
*ctx
, struct exec_list
*list
)
1010 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
1012 switch (node
->type
) {
1013 case nir_cf_node_block
:
1014 ret
= emit_block(ctx
, nir_cf_node_as_block(node
));
1016 case nir_cf_node_if
:
1017 emit_if(ctx
, nir_cf_node_as_if(node
));
1019 case nir_cf_node_loop
:
1020 emit_loop(ctx
, nir_cf_node_as_loop(node
));
1022 case nir_cf_node_function
:
1023 compile_error(ctx
, "Not supported\n");
1030 static void cleanup_binning(struct ir2_context
*ctx
)
1032 assert(ctx
->so
->type
== MESA_SHADER_VERTEX
);
1034 /* kill non-position outputs for binning variant */
1035 nir_foreach_block(block
, nir_shader_get_entrypoint(ctx
->nir
)) {
1036 nir_foreach_instr_safe(instr
, block
) {
1037 if (instr
->type
!= nir_instr_type_intrinsic
)
1040 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1041 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
1044 if (output_slot(ctx
, intr
) != VARYING_SLOT_POS
)
1045 nir_instr_remove(instr
);
1049 ir2_optimize_nir(ctx
->nir
, false);
1053 ir2_nir_compile(struct ir2_context
*ctx
, bool binning
)
1055 struct fd2_shader_stateobj
*so
= ctx
->so
;
1057 memset(ctx
->ssa_map
, 0xff, sizeof(ctx
->ssa_map
));
1059 ctx
->nir
= nir_shader_clone(NULL
, so
->nir
);
1062 cleanup_binning(ctx
);
1065 OPT_V(ctx
->nir
, nir_opt_algebraic_late
);
1067 OPT_V(ctx
->nir
, nir_lower_to_source_mods
, nir_lower_all_source_mods
);
1068 OPT_V(ctx
->nir
, nir_copy_prop
);
1069 OPT_V(ctx
->nir
, nir_opt_dce
);
1070 OPT_V(ctx
->nir
, nir_opt_move_comparisons
);
1072 OPT_V(ctx
->nir
, nir_lower_bool_to_float
);
1073 OPT_V(ctx
->nir
, nir_lower_int_to_float
);
1075 /* lower to scalar instructions that can only be scalar on a2xx */
1076 OPT_V(ctx
->nir
, ir2_nir_lower_scalar
);
1078 OPT_V(ctx
->nir
, nir_lower_locals_to_regs
);
1080 OPT_V(ctx
->nir
, nir_convert_from_ssa
, true);
1082 OPT_V(ctx
->nir
, nir_move_vec_src_uses_to_dest
);
1083 OPT_V(ctx
->nir
, nir_lower_vec_to_movs
);
1085 OPT_V(ctx
->nir
, nir_opt_dce
);
1087 nir_sweep(ctx
->nir
);
1089 if (fd_mesa_debug
& FD_DBG_DISASM
) {
1090 debug_printf("----------------------\n");
1091 nir_print_shader(ctx
->nir
, stdout
);
1092 debug_printf("----------------------\n");
1095 /* fd2_shader_stateobj init */
1096 if (so
->type
== MESA_SHADER_FRAGMENT
) {
1097 ctx
->f
->fragcoord
= -1;
1098 ctx
->f
->inputs_count
= 0;
1099 memset(ctx
->f
->inputs
, 0, sizeof(ctx
->f
->inputs
));
1103 nir_foreach_variable(in
, &ctx
->nir
->inputs
)
1104 setup_input(ctx
, in
);
1106 if (so
->type
== MESA_SHADER_FRAGMENT
) {
1108 for (idx
= 0; idx
< ctx
->f
->inputs_count
; idx
++) {
1109 ctx
->input
[idx
].ncomp
= ctx
->f
->inputs
[idx
].ncomp
;
1110 update_range(ctx
, &ctx
->input
[idx
]);
1112 /* assume we have param input and kill it later if not */
1113 ctx
->input
[idx
].ncomp
= 4;
1114 update_range(ctx
, &ctx
->input
[idx
]);
1116 ctx
->input
[0].ncomp
= 1;
1117 ctx
->input
[2].ncomp
= 1;
1118 update_range(ctx
, &ctx
->input
[0]);
1119 update_range(ctx
, &ctx
->input
[2]);
1122 /* And emit the body: */
1123 nir_function_impl
*fxn
= nir_shader_get_entrypoint(ctx
->nir
);
1125 nir_foreach_register(reg
, &fxn
->registers
) {
1126 ctx
->reg
[reg
->index
].ncomp
= reg
->num_components
;
1127 ctx
->reg_count
= MAX2(ctx
->reg_count
, reg
->index
+ 1);
1130 nir_metadata_require(fxn
, nir_metadata_block_index
);
1131 emit_cf_list(ctx
, &fxn
->body
);
1132 /* TODO emit_block(ctx, fxn->end_block); */
1134 if (so
->type
== MESA_SHADER_VERTEX
)
1135 extra_position_exports(ctx
, binning
);
1137 ralloc_free(ctx
->nir
);
1139 /* kill unused param input */
1140 if (so
->type
== MESA_SHADER_FRAGMENT
&& !so
->need_param
)
1141 ctx
->input
[ctx
->f
->inputs_count
].initialized
= false;