2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
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 #include <sys/types.h>
33 #include "main/mtypes.h"
34 #include "compiler/glsl/glsl_to_nir.h"
35 #include "compiler/nir_types.h"
36 #include "main/imports.h"
37 #include "compiler/nir/nir_builder.h"
38 #include "util/half_float.h"
39 #include "util/u_math.h"
40 #include "util/u_debug.h"
41 #include "util/u_dynarray.h"
42 #include "util/list.h"
43 #include "main/mtypes.h"
46 #include "midgard_nir.h"
47 #include "midgard_compile.h"
48 #include "midgard_ops.h"
52 #include "disassemble.h"
54 static const struct debug_named_value debug_options
[] = {
55 {"msgs", MIDGARD_DBG_MSGS
, "Print debug messages"},
56 {"shaders", MIDGARD_DBG_SHADERS
, "Dump shaders in NIR and MIR"},
57 {"shaderdb", MIDGARD_DBG_SHADERDB
, "Prints shader-db statistics"},
61 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug
, "MIDGARD_MESA_DEBUG", debug_options
, 0)
63 unsigned SHADER_DB_COUNT
= 0;
65 int midgard_debug
= 0;
67 #define DBG(fmt, ...) \
68 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
69 fprintf(stderr, "%s:%d: "fmt, \
70 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
73 midgard_is_branch_unit(unsigned unit
)
75 return (unit
== ALU_ENAB_BRANCH
) || (unit
== ALU_ENAB_BR_COMPACT
);
79 midgard_block_add_successor(midgard_block
*block
, midgard_block
*successor
)
85 for (unsigned i
= 0; i
< block
->nr_successors
; ++i
) {
86 if (block
->successors
[i
] == successor
)
90 block
->successors
[block
->nr_successors
++] = successor
;
91 assert(block
->nr_successors
<= ARRAY_SIZE(block
->successors
));
93 /* Note the predecessor in the other direction */
94 _mesa_set_add(successor
->predecessors
, block
);
97 /* Helpers to generate midgard_instruction's using macro magic, since every
98 * driver seems to do it that way */
100 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
102 #define M_LOAD_STORE(name, store) \
103 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
104 midgard_instruction i = { \
105 .type = TAG_LOAD_STORE_4, \
109 .src = { ~0, ~0, ~0 }, \
112 .op = midgard_op_##name, \
113 .swizzle = SWIZZLE_XYZW, \
119 i.ssa_args.src[0] = ssa; \
121 i.ssa_args.dest = ssa; \
126 #define M_LOAD(name) M_LOAD_STORE(name, false)
127 #define M_STORE(name) M_LOAD_STORE(name, true)
129 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
130 * the corresponding Midgard source */
132 static midgard_vector_alu_src
133 vector_alu_modifiers(nir_alu_src
*src
, bool is_int
, unsigned broadcast_count
,
134 bool half
, bool sext
)
136 if (!src
) return blank_alu_src
;
138 /* Figure out how many components there are so we can adjust the
139 * swizzle. Specifically we want to broadcast the last channel so
140 * things like ball2/3 work
143 if (broadcast_count
) {
144 uint8_t last_component
= src
->swizzle
[broadcast_count
- 1];
146 for (unsigned c
= broadcast_count
; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
147 src
->swizzle
[c
] = last_component
;
151 midgard_vector_alu_src alu_src
= {
155 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
)
159 alu_src
.mod
= midgard_int_normal
;
161 /* Sign/zero-extend if needed */
165 midgard_int_sign_extend
166 : midgard_int_zero_extend
;
169 /* These should have been lowered away */
170 assert(!(src
->abs
|| src
->negate
));
172 alu_src
.mod
= (src
->abs
<< 0) | (src
->negate
<< 1);
178 /* load/store instructions have both 32-bit and 16-bit variants, depending on
179 * whether we are using vectors composed of highp or mediump. At the moment, we
180 * don't support half-floats -- this requires changes in other parts of the
181 * compiler -- therefore the 16-bit versions are commented out. */
183 //M_LOAD(ld_attr_16);
185 //M_LOAD(ld_vary_16);
190 M_LOAD(ld_color_buffer_8
);
191 //M_STORE(st_vary_16);
193 M_LOAD(ld_cubemap_coords
);
194 M_LOAD(ld_compute_id
);
196 static midgard_instruction
197 v_alu_br_compact_cond(midgard_jmp_writeout_op op
, unsigned tag
, signed offset
, unsigned cond
)
199 midgard_branch_cond branch
= {
207 memcpy(&compact
, &branch
, sizeof(branch
));
209 midgard_instruction ins
= {
211 .unit
= ALU_ENAB_BR_COMPACT
,
212 .prepacked_branch
= true,
213 .compact_branch
= true,
214 .br_compact
= compact
,
217 .src
= { ~0, ~0, ~0 },
221 if (op
== midgard_jmp_writeout_op_writeout
)
227 static midgard_instruction
228 v_branch(bool conditional
, bool invert
)
230 midgard_instruction ins
= {
232 .unit
= ALU_ENAB_BRANCH
,
233 .compact_branch
= true,
235 .conditional
= conditional
,
236 .invert_conditional
= invert
240 .src
= { ~0, ~0, ~0 },
247 static midgard_branch_extended
248 midgard_create_branch_extended( midgard_condition cond
,
249 midgard_jmp_writeout_op op
,
251 signed quadword_offset
)
253 /* The condition code is actually a LUT describing a function to
254 * combine multiple condition codes. However, we only support a single
255 * condition code at the moment, so we just duplicate over a bunch of
258 uint16_t duplicated_cond
=
268 midgard_branch_extended branch
= {
270 .dest_tag
= dest_tag
,
271 .offset
= quadword_offset
,
272 .cond
= duplicated_cond
279 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
281 ins
->has_constants
= true;
282 memcpy(&ins
->constants
, constants
, 16);
286 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
288 return glsl_count_attribute_slots(type
, false);
291 /* Lower fdot2 to a vector multiplication followed by channel addition */
293 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
295 if (alu
->op
!= nir_op_fdot2
)
298 b
->cursor
= nir_before_instr(&alu
->instr
);
300 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
301 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
303 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
305 nir_ssa_def
*sum
= nir_fadd(b
,
306 nir_channel(b
, product
, 0),
307 nir_channel(b
, product
, 1));
309 /* Replace the fdot2 with this sum */
310 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
314 midgard_sysval_for_ssbo(nir_intrinsic_instr
*instr
)
316 /* This is way too meta */
317 bool is_store
= instr
->intrinsic
== nir_intrinsic_store_ssbo
;
318 unsigned idx_idx
= is_store
? 1 : 0;
320 nir_src index
= instr
->src
[idx_idx
];
321 assert(nir_src_is_const(index
));
322 uint32_t uindex
= nir_src_as_uint(index
);
324 return PAN_SYSVAL(SSBO
, uindex
);
328 midgard_nir_sysval_for_intrinsic(nir_intrinsic_instr
*instr
)
330 switch (instr
->intrinsic
) {
331 case nir_intrinsic_load_viewport_scale
:
332 return PAN_SYSVAL_VIEWPORT_SCALE
;
333 case nir_intrinsic_load_viewport_offset
:
334 return PAN_SYSVAL_VIEWPORT_OFFSET
;
335 case nir_intrinsic_load_num_work_groups
:
336 return PAN_SYSVAL_NUM_WORK_GROUPS
;
337 case nir_intrinsic_load_ssbo
:
338 case nir_intrinsic_store_ssbo
:
339 return midgard_sysval_for_ssbo(instr
);
345 static int sysval_for_instr(compiler_context
*ctx
, nir_instr
*instr
,
348 nir_intrinsic_instr
*intr
;
349 nir_dest
*dst
= NULL
;
353 bool is_store
= false;
355 switch (instr
->type
) {
356 case nir_instr_type_intrinsic
:
357 intr
= nir_instr_as_intrinsic(instr
);
358 sysval
= midgard_nir_sysval_for_intrinsic(intr
);
360 is_store
|= intr
->intrinsic
== nir_intrinsic_store_ssbo
;
362 case nir_instr_type_tex
:
363 tex
= nir_instr_as_tex(instr
);
364 if (tex
->op
!= nir_texop_txs
)
367 sysval
= PAN_SYSVAL(TEXTURE_SIZE
,
368 PAN_TXS_SYSVAL_ID(tex
->texture_index
,
369 nir_tex_instr_dest_size(tex
) -
370 (tex
->is_array
? 1 : 0),
378 if (dest
&& dst
&& !is_store
)
379 *dest
= nir_dest_index(ctx
, dst
);
385 midgard_nir_assign_sysval_body(compiler_context
*ctx
, nir_instr
*instr
)
389 sysval
= sysval_for_instr(ctx
, instr
, NULL
);
393 /* We have a sysval load; check if it's already been assigned */
395 if (_mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
))
398 /* It hasn't -- so assign it now! */
400 unsigned id
= ctx
->sysval_count
++;
401 _mesa_hash_table_u64_insert(ctx
->sysval_to_id
, sysval
, (void *) ((uintptr_t) id
+ 1));
402 ctx
->sysvals
[id
] = sysval
;
406 midgard_nir_assign_sysvals(compiler_context
*ctx
, nir_shader
*shader
)
408 ctx
->sysval_count
= 0;
410 nir_foreach_function(function
, shader
) {
411 if (!function
->impl
) continue;
413 nir_foreach_block(block
, function
->impl
) {
414 nir_foreach_instr_safe(instr
, block
) {
415 midgard_nir_assign_sysval_body(ctx
, instr
);
422 midgard_nir_lower_fdot2(nir_shader
*shader
)
424 bool progress
= false;
426 nir_foreach_function(function
, shader
) {
427 if (!function
->impl
) continue;
430 nir_builder
*b
= &_b
;
431 nir_builder_init(b
, function
->impl
);
433 nir_foreach_block(block
, function
->impl
) {
434 nir_foreach_instr_safe(instr
, block
) {
435 if (instr
->type
!= nir_instr_type_alu
) continue;
437 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
438 midgard_nir_lower_fdot2_body(b
, alu
);
444 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
451 /* Flushes undefined values to zero */
454 optimise_nir(nir_shader
*nir
)
457 unsigned lower_flrp
=
458 (nir
->options
->lower_flrp16
? 16 : 0) |
459 (nir
->options
->lower_flrp32
? 32 : 0) |
460 (nir
->options
->lower_flrp64
? 64 : 0);
462 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
463 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
464 NIR_PASS(progress
, nir
, nir_lower_idiv
);
466 nir_lower_tex_options lower_tex_options
= {
467 .lower_txs_lod
= true,
471 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
476 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
477 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
479 NIR_PASS(progress
, nir
, nir_copy_prop
);
480 NIR_PASS(progress
, nir
, nir_opt_dce
);
481 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
482 NIR_PASS(progress
, nir
, nir_opt_cse
);
483 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
484 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
485 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
487 if (lower_flrp
!= 0) {
488 bool lower_flrp_progress
= false;
489 NIR_PASS(lower_flrp_progress
,
493 false /* always_precise */,
494 nir
->options
->lower_ffma
);
495 if (lower_flrp_progress
) {
496 NIR_PASS(progress
, nir
,
497 nir_opt_constant_folding
);
501 /* Nothing should rematerialize any flrps, so we only
502 * need to do this lowering once.
507 NIR_PASS(progress
, nir
, nir_opt_undef
);
508 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
510 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
513 nir_var_function_temp
);
515 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
518 /* Must be run at the end to prevent creation of fsin/fcos ops */
519 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
524 NIR_PASS(progress
, nir
, nir_opt_dce
);
525 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
526 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
527 NIR_PASS(progress
, nir
, nir_copy_prop
);
530 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
532 /* We implement booleans as 32-bit 0/~0 */
533 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
535 /* Now that booleans are lowered, we can run out late opts */
536 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
538 /* Lower mods for float ops only. Integer ops don't support modifiers
539 * (saturate doesn't make sense on integers, neg/abs require dedicated
542 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
543 NIR_PASS(progress
, nir
, nir_copy_prop
);
544 NIR_PASS(progress
, nir
, nir_opt_dce
);
546 /* Take us out of SSA */
547 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
548 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
550 /* We are a vector architecture; write combine where possible */
551 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
552 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
554 NIR_PASS(progress
, nir
, nir_opt_dce
);
557 /* Do not actually emit a load; instead, cache the constant for inlining */
560 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
562 nir_ssa_def def
= instr
->def
;
564 float *v
= rzalloc_array(NULL
, float, 4);
565 nir_const_value_to_array(v
, instr
->value
, instr
->def
.num_components
, f32
);
567 /* Shifted for SSA, +1 for off-by-one */
568 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, v
);
571 /* Normally constants are embedded implicitly, but for I/O and such we have to
572 * explicitly emit a move with the constant source */
575 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
577 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
579 if (constant_value
) {
580 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, to
);
581 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
582 emit_mir_instruction(ctx
, ins
);
587 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
589 unsigned comp
= src
->swizzle
[0];
591 for (unsigned c
= 1; c
< nr_components
; ++c
) {
592 if (src
->swizzle
[c
] != comp
)
599 /* Midgard puts scalar conditionals in r31.w; move an arbitrary source (the
600 * output of a conditional test) into that register */
603 emit_condition(compiler_context
*ctx
, nir_src
*src
, bool for_branch
, unsigned component
)
605 int condition
= nir_src_index(ctx
, src
);
607 /* Source to swizzle the desired component into w */
609 const midgard_vector_alu_src alu_src
= {
610 .swizzle
= SWIZZLE(component
, component
, component
, component
),
613 /* There is no boolean move instruction. Instead, we simulate a move by
614 * ANDing the condition with itself to get it into r31.w */
616 midgard_instruction ins
= {
619 /* We need to set the conditional as close as possible */
620 .precede_break
= true,
621 .unit
= for_branch
? UNIT_SMUL
: UNIT_SADD
,
622 .mask
= 1 << COMPONENT_W
,
625 .src
= { condition
, condition
, ~0 },
626 .dest
= SSA_FIXED_REGISTER(31),
630 .op
= midgard_alu_op_iand
,
631 .outmod
= midgard_outmod_int_wrap
,
632 .reg_mode
= midgard_reg_mode_32
,
633 .dest_override
= midgard_dest_override_none
,
634 .src1
= vector_alu_srco_unsigned(alu_src
),
635 .src2
= vector_alu_srco_unsigned(alu_src
)
639 emit_mir_instruction(ctx
, ins
);
642 /* Or, for mixed conditions (with csel_v), here's a vector version using all of
646 emit_condition_mixed(compiler_context
*ctx
, nir_alu_src
*src
, unsigned nr_comp
)
648 int condition
= nir_src_index(ctx
, &src
->src
);
650 /* Source to swizzle the desired component into w */
652 const midgard_vector_alu_src alu_src
= {
653 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
),
656 /* There is no boolean move instruction. Instead, we simulate a move by
657 * ANDing the condition with itself to get it into r31.w */
659 midgard_instruction ins
= {
661 .precede_break
= true,
662 .mask
= mask_of(nr_comp
),
664 .src
= { condition
, condition
, ~0 },
665 .dest
= SSA_FIXED_REGISTER(31),
668 .op
= midgard_alu_op_iand
,
669 .outmod
= midgard_outmod_int_wrap
,
670 .reg_mode
= midgard_reg_mode_32
,
671 .dest_override
= midgard_dest_override_none
,
672 .src1
= vector_alu_srco_unsigned(alu_src
),
673 .src2
= vector_alu_srco_unsigned(alu_src
)
677 emit_mir_instruction(ctx
, ins
);
680 #define ALU_CASE(nir, _op) \
682 op = midgard_alu_op_##_op; \
683 assert(src_bitsize == dst_bitsize); \
686 #define ALU_CASE_BCAST(nir, _op, count) \
688 op = midgard_alu_op_##_op; \
689 broadcast_swizzle = count; \
690 assert(src_bitsize == dst_bitsize); \
693 nir_is_fzero_constant(nir_src src
)
695 if (!nir_src_is_const(src
))
698 for (unsigned c
= 0; c
< nir_src_num_components(src
); ++c
) {
699 if (nir_src_comp_as_float(src
, c
) != 0.0)
706 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
707 * special treatment override this anyway. */
709 static midgard_reg_mode
710 reg_mode_for_nir(nir_alu_instr
*instr
)
712 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
714 switch (src_bitsize
) {
716 return midgard_reg_mode_8
;
718 return midgard_reg_mode_16
;
720 return midgard_reg_mode_32
;
722 return midgard_reg_mode_64
;
724 unreachable("Invalid bit size");
729 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
731 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
732 * is handled elsewhere */
734 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
735 midgard_emit_derivatives(ctx
, instr
);
739 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
741 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
.dest
);
742 unsigned nr_components
= nir_dest_num_components(instr
->dest
.dest
);
743 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
745 /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
746 * supported. A few do not and are commented for now. Also, there are a
747 * number of NIR ops which Midgard does not support and need to be
748 * lowered, also TODO. This switch block emits the opcode and calling
749 * convention of the Midgard instruction; actual packing is done in
754 /* Number of components valid to check for the instruction (the rest
755 * will be forced to the last), or 0 to use as-is. Relevant as
756 * ball-type instructions have a channel count in NIR but are all vec4
759 unsigned broadcast_swizzle
= 0;
761 /* What register mode should we operate in? */
762 midgard_reg_mode reg_mode
=
763 reg_mode_for_nir(instr
);
765 /* Do we need a destination override? Used for inline
768 midgard_dest_override dest_override
=
769 midgard_dest_override_none
;
771 /* Should we use a smaller respective source and sign-extend? */
773 bool half_1
= false, sext_1
= false;
774 bool half_2
= false, sext_2
= false;
776 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
777 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
780 ALU_CASE(fadd
, fadd
);
781 ALU_CASE(fmul
, fmul
);
782 ALU_CASE(fmin
, fmin
);
783 ALU_CASE(fmax
, fmax
);
784 ALU_CASE(imin
, imin
);
785 ALU_CASE(imax
, imax
);
786 ALU_CASE(umin
, umin
);
787 ALU_CASE(umax
, umax
);
788 ALU_CASE(ffloor
, ffloor
);
789 ALU_CASE(fround_even
, froundeven
);
790 ALU_CASE(ftrunc
, ftrunc
);
791 ALU_CASE(fceil
, fceil
);
792 ALU_CASE(fdot3
, fdot3
);
793 ALU_CASE(fdot4
, fdot4
);
794 ALU_CASE(iadd
, iadd
);
795 ALU_CASE(isub
, isub
);
796 ALU_CASE(imul
, imul
);
798 /* Zero shoved as second-arg */
799 ALU_CASE(iabs
, iabsdiff
);
803 ALU_CASE(feq32
, feq
);
804 ALU_CASE(fne32
, fne
);
805 ALU_CASE(flt32
, flt
);
806 ALU_CASE(ieq32
, ieq
);
807 ALU_CASE(ine32
, ine
);
808 ALU_CASE(ilt32
, ilt
);
809 ALU_CASE(ult32
, ult
);
811 /* We don't have a native b2f32 instruction. Instead, like many
812 * GPUs, we exploit booleans as 0/~0 for false/true, and
813 * correspondingly AND
814 * by 1.0 to do the type conversion. For the moment, prime us
817 * iand [whatever], #0
819 * At the end of emit_alu (as MIR), we'll fix-up the constant
822 ALU_CASE(b2f32
, iand
);
823 ALU_CASE(b2i32
, iand
);
825 /* Likewise, we don't have a dedicated f2b32 instruction, but
826 * we can do a "not equal to 0.0" test. */
828 ALU_CASE(f2b32
, fne
);
829 ALU_CASE(i2b32
, ine
);
831 ALU_CASE(frcp
, frcp
);
832 ALU_CASE(frsq
, frsqrt
);
833 ALU_CASE(fsqrt
, fsqrt
);
834 ALU_CASE(fexp2
, fexp2
);
835 ALU_CASE(flog2
, flog2
);
837 ALU_CASE(f2i32
, f2i_rtz
);
838 ALU_CASE(f2u32
, f2u_rtz
);
839 ALU_CASE(i2f32
, i2f_rtz
);
840 ALU_CASE(u2f32
, u2f_rtz
);
842 ALU_CASE(f2i16
, f2i_rtz
);
843 ALU_CASE(f2u16
, f2u_rtz
);
844 ALU_CASE(i2f16
, i2f_rtz
);
845 ALU_CASE(u2f16
, u2f_rtz
);
847 ALU_CASE(fsin
, fsin
);
848 ALU_CASE(fcos
, fcos
);
850 /* We'll set invert */
851 ALU_CASE(inot
, imov
);
852 ALU_CASE(iand
, iand
);
854 ALU_CASE(ixor
, ixor
);
855 ALU_CASE(ishl
, ishl
);
856 ALU_CASE(ishr
, iasr
);
857 ALU_CASE(ushr
, ilsr
);
859 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
860 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
861 ALU_CASE(b32all_fequal4
, fball_eq
);
863 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
864 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
865 ALU_CASE(b32any_fnequal4
, fbany_neq
);
867 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
868 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
869 ALU_CASE(b32all_iequal4
, iball_eq
);
871 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
872 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
873 ALU_CASE(b32any_inequal4
, ibany_neq
);
875 /* Source mods will be shoved in later */
876 ALU_CASE(fabs
, fmov
);
877 ALU_CASE(fneg
, fmov
);
878 ALU_CASE(fsat
, fmov
);
880 /* For size conversion, we use a move. Ideally though we would squash
881 * these ops together; maybe that has to happen after in NIR as part of
882 * propagation...? An earlier algebraic pass ensured we step down by
883 * only / exactly one size. If stepping down, we use a dest override to
884 * reduce the size; if stepping up, we use a larger-sized move with a
885 * half source and a sign/zero-extension modifier */
890 /* If we end up upscale, we'll need a sign-extend on the
891 * operand (the second argument) */
898 op
= midgard_alu_op_imov
;
900 if (dst_bitsize
== (src_bitsize
* 2)) {
904 /* Use a greater register mode */
906 } else if (src_bitsize
== (dst_bitsize
* 2)) {
907 /* Converting down */
908 dest_override
= midgard_dest_override_lower
;
915 assert(src_bitsize
== 32);
917 op
= midgard_alu_op_fmov
;
918 dest_override
= midgard_dest_override_lower
;
923 assert(src_bitsize
== 16);
925 op
= midgard_alu_op_fmov
;
932 /* For greater-or-equal, we lower to less-or-equal and flip the
940 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
941 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
942 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
943 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
946 /* Swap via temporary */
947 nir_alu_src temp
= instr
->src
[1];
948 instr
->src
[1] = instr
->src
[0];
949 instr
->src
[0] = temp
;
954 case nir_op_b32csel
: {
955 /* Midgard features both fcsel and icsel, depending on
956 * the type of the arguments/output. However, as long
957 * as we're careful we can _always_ use icsel and
958 * _never_ need fcsel, since the latter does additional
959 * floating-point-specific processing whereas the
960 * former just moves bits on the wire. It's not obvious
961 * why these are separate opcodes, save for the ability
962 * to do things like sat/pos/abs/neg for free */
964 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
965 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
967 /* csel works as a two-arg in Midgard, since the condition is hardcoded in r31.w */
970 /* Emit the condition into r31 */
973 emit_condition_mixed(ctx
, &instr
->src
[0], nr_components
);
975 emit_condition(ctx
, &instr
->src
[0].src
, false, instr
->src
[0].swizzle
[0]);
977 /* The condition is the first argument; move the other
978 * arguments up one to be a binary instruction for
981 memmove(instr
->src
, instr
->src
+ 1, 2 * sizeof(nir_alu_src
));
986 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
991 /* Midgard can perform certain modifiers on output of an ALU op */
994 if (midgard_is_integer_out_op(op
)) {
995 outmod
= midgard_outmod_int_wrap
;
997 bool sat
= instr
->dest
.saturate
|| instr
->op
== nir_op_fsat
;
998 outmod
= sat
? midgard_outmod_sat
: midgard_outmod_none
;
1001 /* fmax(a, 0.0) can turn into a .pos modifier as an optimization */
1003 if (instr
->op
== nir_op_fmax
) {
1004 if (nir_is_fzero_constant(instr
->src
[0].src
)) {
1005 op
= midgard_alu_op_fmov
;
1007 outmod
= midgard_outmod_pos
;
1008 instr
->src
[0] = instr
->src
[1];
1009 } else if (nir_is_fzero_constant(instr
->src
[1].src
)) {
1010 op
= midgard_alu_op_fmov
;
1012 outmod
= midgard_outmod_pos
;
1016 /* Fetch unit, quirks, etc information */
1017 unsigned opcode_props
= alu_opcode_props
[op
].props
;
1018 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
1020 /* src0 will always exist afaik, but src1 will not for 1-argument
1021 * instructions. The latter can only be fetched if the instruction
1022 * needs it, or else we may segfault. */
1024 unsigned src0
= nir_alu_src_index(ctx
, &instr
->src
[0]);
1025 unsigned src1
= nr_inputs
== 2 ? nir_alu_src_index(ctx
, &instr
->src
[1]) : ~0;
1027 /* Rather than use the instruction generation helpers, we do it
1028 * ourselves here to avoid the mess */
1030 midgard_instruction ins
= {
1034 quirk_flipped_r24
? ~0 : src0
,
1035 quirk_flipped_r24
? src0
: src1
,
1042 nir_alu_src
*nirmods
[2] = { NULL
};
1044 if (nr_inputs
== 2) {
1045 nirmods
[0] = &instr
->src
[0];
1046 nirmods
[1] = &instr
->src
[1];
1047 } else if (nr_inputs
== 1) {
1048 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
1053 /* These were lowered to a move, so apply the corresponding mod */
1055 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
1056 nir_alu_src
*s
= nirmods
[quirk_flipped_r24
];
1058 if (instr
->op
== nir_op_fneg
)
1059 s
->negate
= !s
->negate
;
1061 if (instr
->op
== nir_op_fabs
)
1065 bool is_int
= midgard_is_integer_op(op
);
1067 ins
.mask
= mask_of(nr_components
);
1069 midgard_vector_alu alu
= {
1071 .reg_mode
= reg_mode
,
1072 .dest_override
= dest_override
,
1075 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, sext_1
)),
1076 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
, broadcast_swizzle
, half_2
, sext_2
)),
1079 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
1082 ins
.mask
&= instr
->dest
.write_mask
;
1086 /* Late fixup for emulated instructions */
1088 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1089 /* Presently, our second argument is an inline #0 constant.
1090 * Switch over to an embedded 1.0 constant (that can't fit
1091 * inline, since we're 32-bit, not 16-bit like the inline
1094 ins
.ssa_args
.inline_constant
= false;
1095 ins
.ssa_args
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1096 ins
.has_constants
= true;
1098 if (instr
->op
== nir_op_b2f32
) {
1099 ins
.constants
[0] = 1.0f
;
1101 /* Type pun it into place */
1103 memcpy(&ins
.constants
[0], &one
, sizeof(uint32_t));
1106 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1107 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1108 /* Lots of instructions need a 0 plonked in */
1109 ins
.ssa_args
.inline_constant
= false;
1110 ins
.ssa_args
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1111 ins
.has_constants
= true;
1112 ins
.constants
[0] = 0.0f
;
1113 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1114 } else if (instr
->op
== nir_op_inot
) {
1118 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1119 /* To avoid duplicating the lookup tables (probably), true LUT
1120 * instructions can only operate as if they were scalars. Lower
1121 * them here by changing the component. */
1123 uint8_t original_swizzle
[4];
1124 memcpy(original_swizzle
, nirmods
[0]->swizzle
, sizeof(nirmods
[0]->swizzle
));
1125 unsigned orig_mask
= ins
.mask
;
1127 for (int i
= 0; i
< nr_components
; ++i
) {
1128 /* Mask the associated component, dropping the
1129 * instruction if needed */
1132 ins
.mask
&= orig_mask
;
1137 for (int j
= 0; j
< 4; ++j
)
1138 nirmods
[0]->swizzle
[j
] = original_swizzle
[i
]; /* Pull from the correct component */
1140 ins
.alu
.src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, false));
1141 emit_mir_instruction(ctx
, ins
);
1144 emit_mir_instruction(ctx
, ins
);
1151 mir_mask_for_intr(nir_instr
*instr
, bool is_read
)
1153 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1156 return mask_of(nir_intrinsic_dest_components(intr
));
1158 return nir_intrinsic_write_mask(intr
);
1161 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1162 * optimized) versions of UBO #0 */
1164 midgard_instruction
*
1166 compiler_context
*ctx
,
1170 nir_src
*indirect_offset
,
1173 /* TODO: half-floats */
1175 midgard_instruction ins
= m_ld_ubo_int4(dest
, offset
);
1177 assert((offset
& 0xF) == 0);
1180 /* TODO: Don't split */
1181 ins
.load_store
.varying_parameters
= (offset
& 7) << 7;
1182 ins
.load_store
.address
= offset
>> 3;
1183 ins
.mask
= mir_mask_for_intr(instr
, true);
1185 if (indirect_offset
) {
1186 ins
.ssa_args
.src
[1] = nir_src_index(ctx
, indirect_offset
);
1187 ins
.load_store
.arg_2
= 0x80;
1189 ins
.load_store
.arg_2
= 0x1E;
1192 ins
.load_store
.arg_1
= index
;
1194 return emit_mir_instruction(ctx
, ins
);
1197 /* SSBO reads are like UBO reads if you squint */
1201 compiler_context
*ctx
,
1206 nir_src
*indirect_offset
,
1211 midgard_instruction ins
;
1214 ins
= m_ld_int4(srcdest
, offset
);
1216 ins
= m_st_int4(srcdest
, offset
);
1218 /* SSBO reads use a generic memory read interface, so we need the
1219 * address of the SSBO as the first argument. This is a sysval. */
1221 unsigned addr
= make_compiler_temp(ctx
);
1222 emit_sysval_read(ctx
, instr
, addr
, 2);
1224 /* The source array is a bit of a leaky abstraction for SSBOs.
1225 * Nevertheless, for loads:
1231 * Whereas for stores:
1237 * We would like arg_1 = the address and
1238 * arg_2 = the offset.
1241 ins
.ssa_args
.src
[is_read
? 0 : 1] = addr
;
1243 /* TODO: What is this? It looks superficially like a shift << 5, but
1244 * arg_1 doesn't take a shift Should it be E0 or A0? */
1245 if (indirect_offset
)
1246 ins
.load_store
.arg_1
|= 0xE0;
1248 /* We also need to emit the indirect offset */
1250 if (indirect_offset
)
1251 ins
.ssa_args
.src
[is_read
? 1 : 2] = nir_src_index(ctx
, indirect_offset
);
1253 ins
.load_store
.arg_2
= 0x7E;
1255 /* TODO: Bounds check */
1257 /* Finally, we emit the direct offset */
1259 ins
.load_store
.varying_parameters
= (offset
& 0x1FF) << 1;
1260 ins
.load_store
.address
= (offset
>> 9);
1261 ins
.mask
= mir_mask_for_intr(instr
, is_read
);
1263 emit_mir_instruction(ctx
, ins
);
1268 compiler_context
*ctx
,
1269 unsigned dest
, unsigned offset
,
1270 unsigned nr_comp
, unsigned component
,
1271 nir_src
*indirect_offset
, nir_alu_type type
)
1273 /* XXX: Half-floats? */
1274 /* TODO: swizzle, mask */
1276 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1277 ins
.mask
= mask_of(nr_comp
);
1278 ins
.load_store
.swizzle
= SWIZZLE_XYZW
>> (2 * component
);
1280 midgard_varying_parameter p
= {
1282 .interpolation
= midgard_interp_default
,
1283 .flat
= /*var->data.interpolation == INTERP_MODE_FLAT*/ 0
1287 memcpy(&u
, &p
, sizeof(p
));
1288 ins
.load_store
.varying_parameters
= u
;
1290 if (indirect_offset
)
1291 ins
.ssa_args
.src
[1] = nir_src_index(ctx
, indirect_offset
);
1293 ins
.load_store
.arg_2
= 0x1E;
1295 ins
.load_store
.arg_1
= 0x9E;
1297 /* Use the type appropriate load */
1301 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1304 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1306 case nir_type_float
:
1307 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1310 unreachable("Attempted to load unknown type");
1314 emit_mir_instruction(ctx
, ins
);
1318 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
, signed dest_override
,
1319 unsigned nr_components
)
1323 /* Figure out which uniform this is */
1324 int sysval
= sysval_for_instr(ctx
, instr
, &dest
);
1325 void *val
= _mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
);
1327 if (dest_override
>= 0)
1328 dest
= dest_override
;
1330 /* Sysvals are prefix uniforms */
1331 unsigned uniform
= ((uintptr_t) val
) - 1;
1333 /* Emit the read itself -- this is never indirect */
1334 midgard_instruction
*ins
=
1335 emit_ubo_read(ctx
, instr
, dest
, uniform
* 16, NULL
, 0);
1337 ins
->mask
= mask_of(nr_components
);
1341 compute_builtin_arg(nir_op op
)
1344 case nir_intrinsic_load_work_group_id
:
1346 case nir_intrinsic_load_local_invocation_id
:
1349 unreachable("Invalid compute paramater loaded");
1353 /* Emit store for a fragment shader, which is encoded via a fancy branch. TODO:
1354 * Handle MRT here */
1357 emit_fragment_store(compiler_context
*ctx
, unsigned src
, unsigned rt
)
1359 /* First, move in whatever we're outputting */
1360 midgard_instruction move
= v_mov(src
, blank_alu_src
, SSA_FIXED_REGISTER(0));
1362 /* Force a tight schedule. TODO: Make the scheduler MRT aware */
1363 move
.unit
= UNIT_VMUL
;
1364 move
.precede_break
= true;
1365 move
.dont_eliminate
= true;
1368 emit_mir_instruction(ctx
, move
);
1370 /* If we're doing MRT, we need to specify the render target */
1372 midgard_instruction rt_move
= {
1379 /* We'll write to r1.z */
1380 rt_move
= v_mov(~0, blank_alu_src
, SSA_FIXED_REGISTER(1));
1381 rt_move
.mask
= 1 << COMPONENT_Z
;
1382 rt_move
.unit
= UNIT_SADD
;
1384 /* r1.z = (rt * 0x100) */
1385 rt_move
.ssa_args
.inline_constant
= true;
1386 rt_move
.inline_constant
= (rt
* 0x100);
1389 ctx
->work_registers
= MAX2(ctx
->work_registers
, 1);
1392 emit_mir_instruction(ctx
, rt_move
);
1395 /* Next, generate the branch. For R render targets in the writeout, the
1396 * i'th render target jumps to pseudo-offset [2(R-1) + i] */
1398 unsigned offset
= (2 * (ctx
->nir
->num_outputs
- 1)) + rt
;
1400 struct midgard_instruction ins
=
1401 v_alu_br_compact_cond(midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, offset
, midgard_condition_always
);
1403 /* Add dependencies */
1404 ins
.ssa_args
.src
[0] = move
.ssa_args
.dest
;
1405 ins
.ssa_args
.src
[1] = rt_move
.ssa_args
.dest
;
1407 /* Emit the branch */
1408 emit_mir_instruction(ctx
, ins
);
1412 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1414 unsigned reg
= nir_dest_index(ctx
, &instr
->dest
);
1415 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1416 ins
.mask
= mask_of(3);
1417 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1418 emit_mir_instruction(ctx
, ins
);
1421 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1423 unsigned offset
= 0, reg
;
1425 switch (instr
->intrinsic
) {
1426 case nir_intrinsic_discard_if
:
1427 emit_condition(ctx
, &instr
->src
[0], true, COMPONENT_X
);
1431 case nir_intrinsic_discard
: {
1432 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1433 struct midgard_instruction discard
= v_branch(conditional
, false);
1434 discard
.branch
.target_type
= TARGET_DISCARD
;
1435 emit_mir_instruction(ctx
, discard
);
1439 case nir_intrinsic_load_uniform
:
1440 case nir_intrinsic_load_ubo
:
1441 case nir_intrinsic_load_ssbo
:
1442 case nir_intrinsic_load_input
: {
1443 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1444 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1445 bool is_ssbo
= instr
->intrinsic
== nir_intrinsic_load_ssbo
;
1447 /* Get the base type of the intrinsic */
1448 /* TODO: Infer type? Does it matter? */
1450 (is_ubo
|| is_ssbo
) ? nir_type_uint
: nir_intrinsic_type(instr
);
1451 t
= nir_alu_type_get_base_type(t
);
1453 if (!(is_ubo
|| is_ssbo
)) {
1454 offset
= nir_intrinsic_base(instr
);
1457 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1459 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1461 bool direct
= nir_src_is_const(*src_offset
);
1462 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1465 offset
+= nir_src_as_uint(*src_offset
);
1467 /* We may need to apply a fractional offset */
1468 int component
= instr
->intrinsic
== nir_intrinsic_load_input
?
1469 nir_intrinsic_component(instr
) : 0;
1470 reg
= nir_dest_index(ctx
, &instr
->dest
);
1472 if (is_uniform
&& !ctx
->is_blend
) {
1473 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysval_count
+ offset
) * 16, indirect_offset
, 0);
1474 } else if (is_ubo
) {
1475 nir_src index
= instr
->src
[0];
1477 /* We don't yet support indirect UBOs. For indirect
1478 * block numbers (if that's possible), we don't know
1479 * enough about the hardware yet. For indirect sources,
1480 * we know what we need but we need to add some NIR
1481 * support for lowering correctly with respect to
1484 assert(nir_src_is_const(index
));
1485 assert(nir_src_is_const(*src_offset
));
1487 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1488 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, NULL
, uindex
);
1489 } else if (is_ssbo
) {
1490 nir_src index
= instr
->src
[0];
1491 assert(nir_src_is_const(index
));
1492 uint32_t uindex
= nir_src_as_uint(index
);
1494 emit_ssbo_access(ctx
, &instr
->instr
, true, reg
, offset
, indirect_offset
, uindex
);
1495 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1496 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, !direct
? &instr
->src
[0] : NULL
, t
);
1497 } else if (ctx
->is_blend
) {
1498 /* For blend shaders, load the input color, which is
1499 * preloaded to r0 */
1501 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), blank_alu_src
, reg
);
1502 emit_mir_instruction(ctx
, move
);
1503 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1504 midgard_instruction ins
= m_ld_attr_32(reg
, offset
);
1505 ins
.load_store
.arg_1
= 0x1E;
1506 ins
.load_store
.arg_2
= 0x1E;
1507 ins
.mask
= mask_of(nr_comp
);
1509 /* Use the type appropriate load */
1513 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1516 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1518 case nir_type_float
:
1519 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1522 unreachable("Attempted to load unknown type");
1526 emit_mir_instruction(ctx
, ins
);
1528 DBG("Unknown load\n");
1535 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1537 case nir_intrinsic_load_raw_output_pan
:
1538 reg
= nir_dest_index(ctx
, &instr
->dest
);
1539 assert(ctx
->is_blend
);
1541 midgard_instruction ins
= m_ld_color_buffer_8(reg
, 0);
1542 emit_mir_instruction(ctx
, ins
);
1545 case nir_intrinsic_load_blend_const_color_rgba
: {
1546 assert(ctx
->is_blend
);
1547 reg
= nir_dest_index(ctx
, &instr
->dest
);
1549 /* Blend constants are embedded directly in the shader and
1550 * patched in, so we use some magic routing */
1552 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, reg
);
1553 ins
.has_constants
= true;
1554 ins
.has_blend_constant
= true;
1555 emit_mir_instruction(ctx
, ins
);
1559 case nir_intrinsic_store_output
:
1560 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1562 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1564 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1566 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1567 /* Determine number of render targets */
1568 emit_fragment_store(ctx
, reg
, offset
);
1569 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1570 /* We should have been vectorized, though we don't
1571 * currently check that st_vary is emitted only once
1572 * per slot (this is relevant, since there's not a mask
1573 * parameter available on the store [set to 0 by the
1574 * blob]). We do respect the component by adjusting the
1575 * swizzle. If this is a constant source, we'll need to
1576 * emit that explicitly. */
1578 emit_explicit_constant(ctx
, reg
, reg
);
1580 unsigned component
= nir_intrinsic_component(instr
);
1581 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1583 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1584 st
.load_store
.arg_1
= 0x9E;
1585 st
.load_store
.arg_2
= 0x1E;
1586 st
.load_store
.swizzle
= swizzle_of(nr_comp
) << (2*component
);
1587 emit_mir_instruction(ctx
, st
);
1589 DBG("Unknown store\n");
1595 /* Special case of store_output for lowered blend shaders */
1596 case nir_intrinsic_store_raw_output_pan
:
1597 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1598 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1599 emit_fragment_store(ctx
, reg
, 0);
1603 case nir_intrinsic_store_ssbo
:
1604 assert(nir_src_is_const(instr
->src
[1]));
1606 bool direct_offset
= nir_src_is_const(instr
->src
[2]);
1607 offset
= direct_offset
? nir_src_as_uint(instr
->src
[2]) : 0;
1608 nir_src
*indirect_offset
= direct_offset
? NULL
: &instr
->src
[2];
1609 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1611 uint32_t uindex
= nir_src_as_uint(instr
->src
[1]);
1613 emit_explicit_constant(ctx
, reg
, reg
);
1614 emit_ssbo_access(ctx
, &instr
->instr
, false, reg
, offset
, indirect_offset
, uindex
);
1617 case nir_intrinsic_load_alpha_ref_float
:
1618 assert(instr
->dest
.is_ssa
);
1620 float ref_value
= ctx
->alpha_ref
;
1622 /* See emit_load_const */
1623 float *v
= ralloc_array(NULL
, float, 4);
1624 memcpy(v
, &ref_value
, sizeof(float));
1625 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (instr
->dest
.ssa
.index
<< 1) + 1, v
);
1628 case nir_intrinsic_load_viewport_scale
:
1629 case nir_intrinsic_load_viewport_offset
:
1630 case nir_intrinsic_load_num_work_groups
:
1631 emit_sysval_read(ctx
, &instr
->instr
, ~0, 3);
1634 case nir_intrinsic_load_work_group_id
:
1635 case nir_intrinsic_load_local_invocation_id
:
1636 emit_compute_builtin(ctx
, instr
);
1640 printf ("Unhandled intrinsic\n");
1647 midgard_tex_format(enum glsl_sampler_dim dim
)
1650 case GLSL_SAMPLER_DIM_1D
:
1651 case GLSL_SAMPLER_DIM_BUF
:
1654 case GLSL_SAMPLER_DIM_2D
:
1655 case GLSL_SAMPLER_DIM_EXTERNAL
:
1656 case GLSL_SAMPLER_DIM_RECT
:
1659 case GLSL_SAMPLER_DIM_3D
:
1662 case GLSL_SAMPLER_DIM_CUBE
:
1663 return MALI_TEX_CUBE
;
1666 DBG("Unknown sampler dim type\n");
1672 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1676 pan_attach_constant_bias(
1677 compiler_context
*ctx
,
1679 midgard_texture_word
*word
)
1681 /* To attach as constant, it has to *be* constant */
1683 if (!nir_src_is_const(lod
))
1686 float f
= nir_src_as_float(lod
);
1688 /* Break into fixed-point */
1690 float lod_frac
= f
- lod_int
;
1692 /* Carry over negative fractions */
1693 if (lod_frac
< 0.0) {
1699 word
->bias
= float_to_ubyte(lod_frac
);
1700 word
->bias_int
= lod_int
;
1705 static enum mali_sampler_type
1706 midgard_sampler_type(nir_alu_type t
) {
1707 switch (nir_alu_type_get_base_type(t
))
1709 case nir_type_float
:
1710 return MALI_SAMPLER_FLOAT
;
1712 return MALI_SAMPLER_SIGNED
;
1714 return MALI_SAMPLER_UNSIGNED
;
1716 unreachable("Unknown sampler type");
1721 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1722 unsigned midgard_texop
)
1725 //assert (!instr->sampler);
1726 //assert (!instr->texture_array_size);
1728 int texture_index
= instr
->texture_index
;
1729 int sampler_index
= texture_index
;
1731 /* No helper to build texture words -- we do it all here */
1732 midgard_instruction ins
= {
1733 .type
= TAG_TEXTURE_4
,
1736 .dest
= nir_dest_index(ctx
, &instr
->dest
),
1737 .src
= { ~0, ~0, ~0 },
1740 .op
= midgard_texop
,
1741 .format
= midgard_tex_format(instr
->sampler_dim
),
1742 .texture_handle
= texture_index
,
1743 .sampler_handle
= sampler_index
,
1744 .swizzle
= SWIZZLE_XYZW
,
1745 .in_reg_swizzle
= SWIZZLE_XYZW
,
1751 .sampler_type
= midgard_sampler_type(instr
->dest_type
),
1755 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1756 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1757 midgard_vector_alu_src alu_src
= blank_alu_src
;
1758 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1760 switch (instr
->src
[i
].src_type
) {
1761 case nir_tex_src_coord
: {
1762 emit_explicit_constant(ctx
, index
, index
);
1764 /* Texelfetch coordinates uses all four elements
1765 * (xyz/index) regardless of texture dimensionality,
1766 * which means it's necessary to zero the unused
1767 * components to keep everything happy */
1769 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1770 unsigned old_index
= index
;
1772 index
= make_compiler_temp(ctx
);
1774 /* mov index, old_index */
1775 midgard_instruction mov
= v_mov(old_index
, blank_alu_src
, index
);
1777 emit_mir_instruction(ctx
, mov
);
1779 /* mov index.zw, #0 */
1780 mov
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
),
1781 blank_alu_src
, index
);
1782 mov
.has_constants
= true;
1783 mov
.mask
= (1 << COMPONENT_Z
) | (1 << COMPONENT_W
);
1784 emit_mir_instruction(ctx
, mov
);
1787 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1788 /* texelFetch is undefined on samplerCube */
1789 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1791 /* For cubemaps, we use a special ld/st op to
1792 * select the face and copy the xy into the
1793 * texture register */
1795 unsigned temp
= make_compiler_temp(ctx
);
1796 midgard_instruction ld
= m_ld_cubemap_coords(temp
, 0);
1797 ld
.ssa_args
.src
[0] = index
;
1798 ld
.mask
= 0x3; /* xy */
1799 ld
.load_store
.arg_1
= 0x20;
1800 ld
.load_store
.swizzle
= alu_src
.swizzle
;
1801 emit_mir_instruction(ctx
, ld
);
1803 ins
.ssa_args
.src
[0] = temp
;
1804 ins
.texture
.in_reg_swizzle
= SWIZZLE_XYXX
;
1806 ins
.ssa_args
.src
[0] = index
;
1809 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1810 /* Array component in w but NIR wants it in z */
1811 if (nr_components
== 3)
1812 ins
.texture
.in_reg_swizzle
= SWIZZLE_XYZZ
;
1813 else if (nr_components
== 2)
1814 ins
.texture
.in_reg_swizzle
= SWIZZLE_XYXX
;
1816 unreachable("Invalid texture 2D components");
1822 case nir_tex_src_bias
:
1823 case nir_tex_src_lod
: {
1824 /* Try as a constant if we can */
1826 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1827 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1830 ins
.texture
.lod_register
= true;
1831 ins
.ssa_args
.src
[1] = index
;
1832 emit_explicit_constant(ctx
, index
, index
);
1838 unreachable("Unknown texture source type\n");
1842 emit_mir_instruction(ctx
, ins
);
1844 /* Used for .cont and .last hinting */
1845 ctx
->texture_op_count
++;
1849 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1851 /* Fixup op, since only textureLod is permitted in VS but NIR can give
1852 * generic tex in some cases (which confuses the hardware) */
1854 bool is_vertex
= ctx
->stage
== MESA_SHADER_VERTEX
;
1856 if (is_vertex
&& instr
->op
== nir_texop_tex
)
1857 instr
->op
= nir_texop_txl
;
1859 switch (instr
->op
) {
1862 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1865 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1868 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1871 emit_sysval_read(ctx
, &instr
->instr
, ~0, 4);
1874 unreachable("Unhanlded texture op");
1879 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1881 switch (instr
->type
) {
1882 case nir_jump_break
: {
1883 /* Emit a branch out of the loop */
1884 struct midgard_instruction br
= v_branch(false, false);
1885 br
.branch
.target_type
= TARGET_BREAK
;
1886 br
.branch
.target_break
= ctx
->current_loop_depth
;
1887 emit_mir_instruction(ctx
, br
);
1892 DBG("Unknown jump type %d\n", instr
->type
);
1898 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1900 switch (instr
->type
) {
1901 case nir_instr_type_load_const
:
1902 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1905 case nir_instr_type_intrinsic
:
1906 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1909 case nir_instr_type_alu
:
1910 emit_alu(ctx
, nir_instr_as_alu(instr
));
1913 case nir_instr_type_tex
:
1914 emit_tex(ctx
, nir_instr_as_tex(instr
));
1917 case nir_instr_type_jump
:
1918 emit_jump(ctx
, nir_instr_as_jump(instr
));
1921 case nir_instr_type_ssa_undef
:
1926 DBG("Unhandled instruction type\n");
1932 /* ALU instructions can inline or embed constants, which decreases register
1933 * pressure and saves space. */
1935 #define CONDITIONAL_ATTACH(src) { \
1936 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->ssa_args.src + 1); \
1939 attach_constants(ctx, alu, entry, alu->ssa_args.src + 1); \
1940 alu->ssa_args.src = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
1945 inline_alu_constants(compiler_context
*ctx
)
1947 mir_foreach_instr(ctx
, alu
) {
1948 /* Other instructions cannot inline constants */
1949 if (alu
->type
!= TAG_ALU_4
) continue;
1951 /* If there is already a constant here, we can do nothing */
1952 if (alu
->has_constants
) continue;
1954 CONDITIONAL_ATTACH(src
[0]);
1956 if (!alu
->has_constants
) {
1957 CONDITIONAL_ATTACH(src
[1])
1958 } else if (!alu
->inline_constant
) {
1959 /* Corner case: _two_ vec4 constants, for instance with a
1960 * csel. For this case, we can only use a constant
1961 * register for one, we'll have to emit a move for the
1962 * other. Note, if both arguments are constants, then
1963 * necessarily neither argument depends on the value of
1964 * any particular register. As the destination register
1965 * will be wiped, that means we can spill the constant
1966 * to the destination register.
1969 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->ssa_args
.src
[1] + 1);
1970 unsigned scratch
= alu
->ssa_args
.dest
;
1973 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, scratch
);
1974 attach_constants(ctx
, &ins
, entry
, alu
->ssa_args
.src
[1] + 1);
1976 /* Force a break XXX Defer r31 writes */
1977 ins
.unit
= UNIT_VLUT
;
1979 /* Set the source */
1980 alu
->ssa_args
.src
[1] = scratch
;
1982 /* Inject us -before- the last instruction which set r31 */
1983 mir_insert_instruction_before(mir_prev_op(alu
), ins
);
1989 /* Being a little silly with the names, but returns the op that is the bitwise
1990 * inverse of the op with the argument switched. I.e. (f and g are
1993 * f(a, b) = ~g(b, a)
1995 * Corollary: if g is the contrapositve of f, f is the contrapositive of g:
1997 * f(a, b) = ~g(b, a)
1998 * ~f(a, b) = g(b, a)
1999 * ~f(a, b) = ~h(a, b) where h is the contrapositive of g
2002 * Thus we define this function in pairs.
2005 static inline midgard_alu_op
2006 mir_contrapositive(midgard_alu_op op
)
2009 case midgard_alu_op_flt
:
2010 return midgard_alu_op_fle
;
2011 case midgard_alu_op_fle
:
2012 return midgard_alu_op_flt
;
2014 case midgard_alu_op_ilt
:
2015 return midgard_alu_op_ile
;
2016 case midgard_alu_op_ile
:
2017 return midgard_alu_op_ilt
;
2020 unreachable("No known contrapositive");
2024 /* Midgard supports two types of constants, embedded constants (128-bit) and
2025 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2026 * constants can be demoted to inline constants, for space savings and
2027 * sometimes a performance boost */
2030 embedded_to_inline_constant(compiler_context
*ctx
)
2032 mir_foreach_instr(ctx
, ins
) {
2033 if (!ins
->has_constants
) continue;
2035 if (ins
->ssa_args
.inline_constant
) continue;
2037 /* Blend constants must not be inlined by definition */
2038 if (ins
->has_blend_constant
) continue;
2040 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2041 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2042 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2044 if (!(is_16
|| is_32
))
2047 /* src1 cannot be an inline constant due to encoding
2048 * restrictions. So, if possible we try to flip the arguments
2051 int op
= ins
->alu
.op
;
2053 if (ins
->ssa_args
.src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2054 bool flip
= alu_opcode_props
[op
].props
& OP_COMMUTES
;
2057 /* Conditionals can be inverted */
2058 case midgard_alu_op_flt
:
2059 case midgard_alu_op_ilt
:
2060 case midgard_alu_op_fle
:
2061 case midgard_alu_op_ile
:
2062 ins
->alu
.op
= mir_contrapositive(ins
->alu
.op
);
2067 case midgard_alu_op_fcsel
:
2068 case midgard_alu_op_icsel
:
2069 DBG("Missed non-commutative flip (%s)\n", alu_opcode_props
[op
].name
);
2075 /* Flip the SSA numbers */
2076 ins
->ssa_args
.src
[0] = ins
->ssa_args
.src
[1];
2077 ins
->ssa_args
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
2079 /* And flip the modifiers */
2083 src_temp
= ins
->alu
.src2
;
2084 ins
->alu
.src2
= ins
->alu
.src1
;
2085 ins
->alu
.src1
= src_temp
;
2089 if (ins
->ssa_args
.src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2090 /* Extract the source information */
2092 midgard_vector_alu_src
*src
;
2093 int q
= ins
->alu
.src2
;
2094 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2097 /* Component is from the swizzle, e.g. r26.w -> w component. TODO: What if x is masked out? */
2098 int component
= src
->swizzle
& 3;
2100 /* Scale constant appropriately, if we can legally */
2101 uint16_t scaled_constant
= 0;
2103 if (midgard_is_integer_op(op
) || is_16
) {
2104 unsigned int *iconstants
= (unsigned int *) ins
->constants
;
2105 scaled_constant
= (uint16_t) iconstants
[component
];
2107 /* Constant overflow after resize */
2108 if (scaled_constant
!= iconstants
[component
])
2111 float original
= (float) ins
->constants
[component
];
2112 scaled_constant
= _mesa_float_to_half(original
);
2114 /* Check for loss of precision. If this is
2115 * mediump, we don't care, but for a highp
2116 * shader, we need to pay attention. NIR
2117 * doesn't yet tell us which mode we're in!
2118 * Practically this prevents most constants
2119 * from being inlined, sadly. */
2121 float fp32
= _mesa_half_to_float(scaled_constant
);
2123 if (fp32
!= original
)
2127 /* We don't know how to handle these with a constant */
2129 if (mir_nontrivial_source2_mod_simple(ins
) || src
->rep_low
|| src
->rep_high
) {
2130 DBG("Bailing inline constant...\n");
2134 /* Make sure that the constant is not itself a
2135 * vector by checking if all accessed values
2136 * (by the swizzle) are the same. */
2138 uint32_t *cons
= (uint32_t *) ins
->constants
;
2139 uint32_t value
= cons
[component
];
2141 bool is_vector
= false;
2142 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2144 for (int c
= 1; c
< 4; ++c
) {
2145 /* We only care if this component is actually used */
2146 if (!(mask
& (1 << c
)))
2149 uint32_t test
= cons
[(src
->swizzle
>> (2 * c
)) & 3];
2151 if (test
!= value
) {
2160 /* Get rid of the embedded constant */
2161 ins
->has_constants
= false;
2162 ins
->ssa_args
.src
[1] = ~0;
2163 ins
->ssa_args
.inline_constant
= true;
2164 ins
->inline_constant
= scaled_constant
;
2169 /* Dead code elimination for branches at the end of a block - only one branch
2170 * per block is legal semantically */
2173 midgard_opt_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2175 bool branched
= false;
2177 mir_foreach_instr_in_block_safe(block
, ins
) {
2178 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2180 /* We ignore prepacked branches since the fragment epilogue is
2181 * just generally special */
2182 if (ins
->prepacked_branch
) continue;
2184 /* Discards are similarly special and may not correspond to the
2187 if (ins
->branch
.target_type
== TARGET_DISCARD
) continue;
2190 /* We already branched, so this is dead */
2191 mir_remove_instruction(ins
);
2198 /* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
2199 * the move can be propagated away entirely */
2202 mir_compose_float_outmod(midgard_outmod_float
*outmod
, midgard_outmod_float comp
)
2205 if (comp
== midgard_outmod_none
)
2208 if (*outmod
== midgard_outmod_none
) {
2213 /* TODO: Compose rules */
2218 midgard_opt_pos_propagate(compiler_context
*ctx
, midgard_block
*block
)
2220 bool progress
= false;
2222 mir_foreach_instr_in_block_safe(block
, ins
) {
2223 if (ins
->type
!= TAG_ALU_4
) continue;
2224 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2225 if (ins
->alu
.outmod
!= midgard_outmod_pos
) continue;
2227 /* TODO: Registers? */
2228 unsigned src
= ins
->ssa_args
.src
[1];
2229 if (src
& IS_REG
) continue;
2230 assert(!mir_has_multiple_writes(ctx
, src
));
2232 /* There might be a source modifier, too */
2233 if (mir_nontrivial_source2_mod(ins
)) continue;
2235 /* Backpropagate the modifier */
2236 mir_foreach_instr_in_block_from_rev(block
, v
, mir_prev_op(ins
)) {
2237 if (v
->type
!= TAG_ALU_4
) continue;
2238 if (v
->ssa_args
.dest
!= src
) continue;
2240 /* Can we even take a float outmod? */
2241 if (midgard_is_integer_out_op(v
->alu
.op
)) continue;
2243 midgard_outmod_float temp
= v
->alu
.outmod
;
2244 progress
|= mir_compose_float_outmod(&temp
, ins
->alu
.outmod
);
2246 /* Throw in the towel.. */
2247 if (!progress
) break;
2249 /* Otherwise, transfer the modifier */
2250 v
->alu
.outmod
= temp
;
2251 ins
->alu
.outmod
= midgard_outmod_none
;
2261 emit_fragment_epilogue(compiler_context
*ctx
)
2263 /* Just emit the last chunk with the branch */
2264 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, ~0, midgard_condition_always
);
2267 static midgard_block
*
2268 create_empty_block(compiler_context
*ctx
)
2270 midgard_block
*blk
= rzalloc(ctx
, midgard_block
);
2272 blk
->predecessors
= _mesa_set_create(blk
,
2274 _mesa_key_pointer_equal
);
2276 blk
->source_id
= ctx
->block_source_count
++;
2281 static midgard_block
*
2282 emit_block(compiler_context
*ctx
, nir_block
*block
)
2284 midgard_block
*this_block
= ctx
->after_block
;
2285 ctx
->after_block
= NULL
;
2288 this_block
= create_empty_block(ctx
);
2290 list_addtail(&this_block
->link
, &ctx
->blocks
);
2292 this_block
->is_scheduled
= false;
2295 ctx
->texture_index
[0] = ~0;
2296 ctx
->texture_index
[1] = ~0;
2298 /* Set up current block */
2299 list_inithead(&this_block
->instructions
);
2300 ctx
->current_block
= this_block
;
2302 nir_foreach_instr(instr
, block
) {
2303 emit_instr(ctx
, instr
);
2304 ++ctx
->instruction_count
;
2307 inline_alu_constants(ctx
);
2308 embedded_to_inline_constant(ctx
);
2310 /* Append fragment shader epilogue (value writeout) */
2311 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
2312 if (block
== nir_impl_last_block(ctx
->func
->impl
)) {
2313 emit_fragment_epilogue(ctx
);
2317 /* Allow the next control flow to access us retroactively, for
2319 ctx
->current_block
= this_block
;
2324 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2327 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2329 midgard_block
*before_block
= ctx
->current_block
;
2331 /* Conditional branches expect the condition in r31.w; emit a move for
2332 * that in the _previous_ block (which is the current block). */
2333 emit_condition(ctx
, &nif
->condition
, true, COMPONENT_X
);
2335 /* Speculatively emit the branch, but we can't fill it in until later */
2336 EMIT(branch
, true, true);
2337 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2339 /* Emit the two subblocks. */
2340 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2341 midgard_block
*end_then_block
= ctx
->current_block
;
2343 /* Emit a jump from the end of the then block to the end of the else */
2344 EMIT(branch
, false, false);
2345 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2347 /* Emit second block, and check if it's empty */
2349 int else_idx
= ctx
->block_count
;
2350 int count_in
= ctx
->instruction_count
;
2351 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2352 midgard_block
*end_else_block
= ctx
->current_block
;
2353 int after_else_idx
= ctx
->block_count
;
2355 /* Now that we have the subblocks emitted, fix up the branches */
2360 if (ctx
->instruction_count
== count_in
) {
2361 /* The else block is empty, so don't emit an exit jump */
2362 mir_remove_instruction(then_exit
);
2363 then_branch
->branch
.target_block
= after_else_idx
;
2365 then_branch
->branch
.target_block
= else_idx
;
2366 then_exit
->branch
.target_block
= after_else_idx
;
2369 /* Wire up the successors */
2371 ctx
->after_block
= create_empty_block(ctx
);
2373 midgard_block_add_successor(before_block
, then_block
);
2374 midgard_block_add_successor(before_block
, else_block
);
2376 midgard_block_add_successor(end_then_block
, ctx
->after_block
);
2377 midgard_block_add_successor(end_else_block
, ctx
->after_block
);
2381 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2383 /* Remember where we are */
2384 midgard_block
*start_block
= ctx
->current_block
;
2386 /* Allocate a loop number, growing the current inner loop depth */
2387 int loop_idx
= ++ctx
->current_loop_depth
;
2389 /* Get index from before the body so we can loop back later */
2390 int start_idx
= ctx
->block_count
;
2392 /* Emit the body itself */
2393 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2395 /* Branch back to loop back */
2396 struct midgard_instruction br_back
= v_branch(false, false);
2397 br_back
.branch
.target_block
= start_idx
;
2398 emit_mir_instruction(ctx
, br_back
);
2400 /* Mark down that branch in the graph. */
2401 midgard_block_add_successor(start_block
, loop_block
);
2402 midgard_block_add_successor(ctx
->current_block
, loop_block
);
2404 /* Find the index of the block about to follow us (note: we don't add
2405 * one; blocks are 0-indexed so we get a fencepost problem) */
2406 int break_block_idx
= ctx
->block_count
;
2408 /* Fix up the break statements we emitted to point to the right place,
2409 * now that we can allocate a block number for them */
2410 ctx
->after_block
= create_empty_block(ctx
);
2412 list_for_each_entry_from(struct midgard_block
, block
, start_block
, &ctx
->blocks
, link
) {
2413 mir_foreach_instr_in_block(block
, ins
) {
2414 if (ins
->type
!= TAG_ALU_4
) continue;
2415 if (!ins
->compact_branch
) continue;
2416 if (ins
->prepacked_branch
) continue;
2418 /* We found a branch -- check the type to see if we need to do anything */
2419 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2421 /* It's a break! Check if it's our break */
2422 if (ins
->branch
.target_break
!= loop_idx
) continue;
2424 /* Okay, cool, we're breaking out of this loop.
2425 * Rewrite from a break to a goto */
2427 ins
->branch
.target_type
= TARGET_GOTO
;
2428 ins
->branch
.target_block
= break_block_idx
;
2430 midgard_block_add_successor(block
, ctx
->after_block
);
2434 /* Now that we've finished emitting the loop, free up the depth again
2435 * so we play nice with recursion amid nested loops */
2436 --ctx
->current_loop_depth
;
2438 /* Dump loop stats */
2442 static midgard_block
*
2443 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2445 midgard_block
*start_block
= NULL
;
2447 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2448 switch (node
->type
) {
2449 case nir_cf_node_block
: {
2450 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2453 start_block
= block
;
2458 case nir_cf_node_if
:
2459 emit_if(ctx
, nir_cf_node_as_if(node
));
2462 case nir_cf_node_loop
:
2463 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2466 case nir_cf_node_function
:
2475 /* Due to lookahead, we need to report the first tag executed in the command
2476 * stream and in branch targets. An initial block might be empty, so iterate
2477 * until we find one that 'works' */
2480 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2482 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2484 unsigned first_tag
= 0;
2486 mir_foreach_block_from(ctx
, initial_block
, v
) {
2487 midgard_bundle
*initial_bundle
=
2488 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2490 if (initial_bundle
) {
2491 first_tag
= initial_bundle
->tag
;
2500 midgard_compile_shader_nir(struct midgard_screen
*screen
, nir_shader
*nir
, midgard_program
*program
, bool is_blend
)
2502 struct util_dynarray
*compiled
= &program
->compiled
;
2504 midgard_debug
= debug_get_option_midgard_debug();
2506 /* TODO: Bound against what? */
2507 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2510 ctx
->screen
= screen
;
2511 ctx
->stage
= nir
->info
.stage
;
2512 ctx
->is_blend
= is_blend
;
2513 ctx
->alpha_ref
= program
->alpha_ref
;
2515 /* Start off with a safe cutoff, allowing usage of all 16 work
2516 * registers. Later, we'll promote uniform reads to uniform registers
2517 * if we determine it is beneficial to do so */
2518 ctx
->uniform_cutoff
= 8;
2520 /* Initialize at a global (not block) level hash tables */
2522 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2523 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2524 ctx
->sysval_to_id
= _mesa_hash_table_u64_create(NULL
);
2526 /* Record the varying mapping for the command stream's bookkeeping */
2528 struct exec_list
*varyings
=
2529 ctx
->stage
== MESA_SHADER_VERTEX
? &nir
->outputs
: &nir
->inputs
;
2531 unsigned max_varying
= 0;
2532 nir_foreach_variable(var
, varyings
) {
2533 unsigned loc
= var
->data
.driver_location
;
2534 unsigned sz
= glsl_type_size(var
->type
, FALSE
);
2536 for (int c
= 0; c
< sz
; ++c
) {
2537 program
->varyings
[loc
+ c
] = var
->data
.location
+ c
;
2538 max_varying
= MAX2(max_varying
, loc
+ c
);
2542 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2543 * (so we don't accidentally duplicate the epilogue since mesa/st has
2544 * messed with our I/O quite a bit already) */
2546 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2548 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2549 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2550 NIR_PASS_V(nir
, nir_clamp_psiz
, 1.0, 1024.0);
2553 NIR_PASS_V(nir
, nir_lower_var_copies
);
2554 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2555 NIR_PASS_V(nir
, nir_split_var_copies
);
2556 NIR_PASS_V(nir
, nir_lower_var_copies
);
2557 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2558 NIR_PASS_V(nir
, nir_lower_var_copies
);
2559 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2561 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2563 /* Optimisation passes */
2567 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2568 nir_print_shader(nir
, stdout
);
2571 /* Assign sysvals and counts, now that we're sure
2572 * (post-optimisation) */
2574 midgard_nir_assign_sysvals(ctx
, nir
);
2576 program
->uniform_count
= nir
->num_uniforms
;
2577 program
->sysval_count
= ctx
->sysval_count
;
2578 memcpy(program
->sysvals
, ctx
->sysvals
, sizeof(ctx
->sysvals
[0]) * ctx
->sysval_count
);
2580 nir_foreach_function(func
, nir
) {
2584 list_inithead(&ctx
->blocks
);
2585 ctx
->block_count
= 0;
2588 emit_cf_list(ctx
, &func
->impl
->body
);
2590 /* Emit empty exit block with successor */
2592 struct midgard_block
*semi_end
= ctx
->current_block
;
2594 struct midgard_block
*end
=
2595 emit_block(ctx
, func
->impl
->end_block
);
2597 midgard_block_add_successor(semi_end
, end
);
2599 break; /* TODO: Multi-function shaders */
2602 util_dynarray_init(compiled
, NULL
);
2604 /* MIR-level optimizations */
2606 bool progress
= false;
2611 mir_foreach_block(ctx
, block
) {
2612 progress
|= midgard_opt_pos_propagate(ctx
, block
);
2613 progress
|= midgard_opt_copy_prop(ctx
, block
);
2614 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2615 progress
|= midgard_opt_combine_projection(ctx
, block
);
2616 progress
|= midgard_opt_varying_projection(ctx
, block
);
2617 progress
|= midgard_opt_not_propagate(ctx
, block
);
2618 progress
|= midgard_opt_fuse_src_invert(ctx
, block
);
2619 progress
|= midgard_opt_fuse_dest_invert(ctx
, block
);
2623 mir_foreach_block(ctx
, block
) {
2624 midgard_lower_invert(ctx
, block
);
2625 midgard_lower_derivatives(ctx
, block
);
2628 /* Nested control-flow can result in dead branches at the end of the
2629 * block. This messes with our analysis and is just dead code, so cull
2631 mir_foreach_block(ctx
, block
) {
2632 midgard_opt_cull_dead_branch(ctx
, block
);
2635 /* Ensure we were lowered */
2636 mir_foreach_instr_global(ctx
, ins
) {
2637 assert(!ins
->invert
);
2641 schedule_program(ctx
);
2643 /* Now that all the bundles are scheduled and we can calculate block
2644 * sizes, emit actual branch instructions rather than placeholders */
2646 int br_block_idx
= 0;
2648 mir_foreach_block(ctx
, block
) {
2649 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2650 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2651 midgard_instruction
*ins
= bundle
->instructions
[c
];
2653 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2655 if (ins
->prepacked_branch
) continue;
2657 /* Parse some basic branch info */
2658 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2659 bool is_conditional
= ins
->branch
.conditional
;
2660 bool is_inverted
= ins
->branch
.invert_conditional
;
2661 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2663 /* Determine the block we're jumping to */
2664 int target_number
= ins
->branch
.target_block
;
2666 /* Report the destination tag */
2667 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2669 /* Count up the number of quadwords we're
2670 * jumping over = number of quadwords until
2671 * (br_block_idx, target_number) */
2673 int quadword_offset
= 0;
2677 } else if (target_number
> br_block_idx
) {
2680 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2681 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2684 quadword_offset
+= blk
->quadword_count
;
2687 /* Jump backwards */
2689 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2690 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2693 quadword_offset
-= blk
->quadword_count
;
2697 /* Unconditional extended branches (far jumps)
2698 * have issues, so we always use a conditional
2699 * branch, setting the condition to always for
2700 * unconditional. For compact unconditional
2701 * branches, cond isn't used so it doesn't
2702 * matter what we pick. */
2704 midgard_condition cond
=
2705 !is_conditional
? midgard_condition_always
:
2706 is_inverted
? midgard_condition_false
:
2707 midgard_condition_true
;
2709 midgard_jmp_writeout_op op
=
2710 is_discard
? midgard_jmp_writeout_op_discard
:
2711 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2712 midgard_jmp_writeout_op_branch_cond
;
2715 midgard_branch_extended branch
=
2716 midgard_create_branch_extended(
2721 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2722 } else if (is_conditional
|| is_discard
) {
2723 midgard_branch_cond branch
= {
2725 .dest_tag
= dest_tag
,
2726 .offset
= quadword_offset
,
2730 assert(branch
.offset
== quadword_offset
);
2732 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2734 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2736 midgard_branch_uncond branch
= {
2738 .dest_tag
= dest_tag
,
2739 .offset
= quadword_offset
,
2743 assert(branch
.offset
== quadword_offset
);
2745 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2753 /* Emit flat binary from the instruction arrays. Iterate each block in
2754 * sequence. Save instruction boundaries such that lookahead tags can
2755 * be assigned easily */
2757 /* Cache _all_ bundles in source order for lookahead across failed branches */
2759 int bundle_count
= 0;
2760 mir_foreach_block(ctx
, block
) {
2761 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2763 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2765 mir_foreach_block(ctx
, block
) {
2766 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2767 source_order_bundles
[bundle_idx
++] = bundle
;
2771 int current_bundle
= 0;
2773 /* Midgard prefetches instruction types, so during emission we
2774 * need to lookahead. Unless this is the last instruction, in
2775 * which we return 1. Or if this is the second to last and the
2776 * last is an ALU, then it's also 1... */
2778 mir_foreach_block(ctx
, block
) {
2779 mir_foreach_bundle_in_block(block
, bundle
) {
2782 if (current_bundle
+ 1 < bundle_count
) {
2783 uint8_t next
= source_order_bundles
[current_bundle
+ 1]->tag
;
2785 if (!(current_bundle
+ 2 < bundle_count
) && IS_ALU(next
)) {
2792 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2796 /* TODO: Free deeper */
2797 //util_dynarray_fini(&block->instructions);
2800 free(source_order_bundles
);
2802 /* Report the very first tag executed */
2803 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2805 /* Deal with off-by-one related to the fencepost problem */
2806 program
->work_register_count
= ctx
->work_registers
+ 1;
2807 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2809 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2810 program
->tls_size
= ctx
->tls_size
;
2812 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2813 disassemble_midgard(program
->compiled
.data
, program
->compiled
.size
, false, 0, "");
2815 if (midgard_debug
& MIDGARD_DBG_SHADERDB
) {
2816 unsigned nr_bundles
= 0, nr_ins
= 0, nr_quadwords
= 0;
2818 /* Count instructions and bundles */
2820 mir_foreach_instr_global(ctx
, ins
) {
2824 mir_foreach_block(ctx
, block
) {
2825 nr_bundles
+= util_dynarray_num_elements(
2826 &block
->bundles
, midgard_bundle
);
2828 nr_quadwords
+= block
->quadword_count
;
2831 /* Calculate thread count. There are certain cutoffs by
2832 * register count for thread count */
2834 unsigned nr_registers
= program
->work_register_count
;
2836 unsigned nr_threads
=
2837 (nr_registers
<= 4) ? 4 :
2838 (nr_registers
<= 8) ? 2 :
2843 fprintf(stderr
, "shader%d - %s shader: "
2844 "%u inst, %u bundles, %u quadwords, "
2845 "%u registers, %u threads, %u loops, "
2846 "%d:%d spills:fills\n",
2848 gl_shader_stage_name(ctx
->stage
),
2849 nr_ins
, nr_bundles
, nr_quadwords
,
2850 nr_registers
, nr_threads
,
2852 ctx
->spills
, ctx
->fills
);