2 * Copyright (C) 2018 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 "mesa/state_tracker/st_glsl_types.h"
36 #include "compiler/nir_types.h"
37 #include "main/imports.h"
38 #include "compiler/nir/nir_builder.h"
39 #include "util/half_float.h"
40 #include "util/register_allocate.h"
41 #include "util/u_debug.h"
42 #include "util/u_dynarray.h"
43 #include "util/list.h"
44 #include "main/mtypes.h"
47 #include "midgard_nir.h"
48 #include "midgard_compile.h"
51 #include "disassemble.h"
53 static const struct debug_named_value debug_options
[] = {
54 {"msgs", MIDGARD_DBG_MSGS
, "Print debug messages"},
55 {"shaders", MIDGARD_DBG_SHADERS
, "Dump shaders in NIR and MIR"},
59 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug
, "MIDGARD_MESA_DEBUG", debug_options
, 0)
61 int midgard_debug
= 0;
63 #define DBG(fmt, ...) \
64 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
65 fprintf(stderr, "%s:%d: "fmt, \
66 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
68 /* Instruction arguments represented as block-local SSA indices, rather than
69 * registers. Negative values mean unused. */
76 /* src1 is -not- SSA but instead a 16-bit inline constant to be smudged
77 * in. Only valid for ALU ops. */
81 /* Forward declare so midgard_branch can reference */
84 /* Target types. Defaults to TARGET_GOTO (the type corresponding directly to
85 * the hardware), hence why that must be zero. TARGET_DISCARD signals this
86 * instruction is actually a discard op. */
89 #define TARGET_BREAK 1
90 #define TARGET_CONTINUE 2
91 #define TARGET_DISCARD 3
93 typedef struct midgard_branch
{
94 /* If conditional, the condition is specified in r31.w */
97 /* For conditionals, if this is true, we branch on FALSE. If false, we branch on TRUE. */
98 bool invert_conditional
;
100 /* Branch targets: the start of a block, the start of a loop (continue), the end of a loop (break). Value is one of TARGET_ */
101 unsigned target_type
;
103 /* The actual target */
111 /* Generic in-memory data type repesenting a single logical instruction, rather
112 * than a single instruction group. This is the preferred form for code gen.
113 * Multiple midgard_insturctions will later be combined during scheduling,
114 * though this is not represented in this structure. Its format bridges
115 * the low-level binary representation with the higher level semantic meaning.
117 * Notably, it allows registers to be specified as block local SSA, for code
118 * emitted before the register allocation pass.
121 typedef struct midgard_instruction
{
122 /* Must be first for casting */
123 struct list_head link
;
125 unsigned type
; /* ALU, load/store, texture */
127 /* If the register allocator has not run yet... */
130 /* Special fields for an ALU instruction */
131 midgard_reg_info registers
;
133 /* I.e. (1 << alu_bit) */
138 uint16_t inline_constant
;
139 bool has_blend_constant
;
143 bool prepacked_branch
;
146 midgard_load_store_word load_store
;
147 midgard_vector_alu alu
;
148 midgard_texture_word texture
;
149 midgard_branch_extended branch_extended
;
152 /* General branch, rather than packed br_compact. Higher level
153 * than the other components */
154 midgard_branch branch
;
156 } midgard_instruction
;
158 typedef struct midgard_block
{
159 /* Link to next block. Must be first for mir_get_block */
160 struct list_head link
;
162 /* List of midgard_instructions emitted for the current block */
163 struct list_head instructions
;
167 /* List of midgard_bundles emitted (after the scheduler has run) */
168 struct util_dynarray bundles
;
170 /* Number of quadwords _actually_ emitted, as determined after scheduling */
171 unsigned quadword_count
;
173 struct midgard_block
*next_fallthrough
;
176 /* Helpers to generate midgard_instruction's using macro magic, since every
177 * driver seems to do it that way */
179 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
180 #define SWIZZLE_XYZW SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W)
182 #define M_LOAD_STORE(name, rname, uname) \
183 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
184 midgard_instruction i = { \
185 .type = TAG_LOAD_STORE_4, \
192 .op = midgard_op_##name, \
194 .swizzle = SWIZZLE_XYZW, \
202 #define M_LOAD(name) M_LOAD_STORE(name, dest, src0)
203 #define M_STORE(name) M_LOAD_STORE(name, src0, dest)
205 const midgard_vector_alu_src blank_alu_src
= {
206 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_W
),
209 const midgard_vector_alu_src blank_alu_src_xxxx
= {
210 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_X
, COMPONENT_X
, COMPONENT_X
),
213 const midgard_scalar_alu_src blank_scalar_alu_src
= {
217 /* Used for encoding the unused source of 1-op instructions */
218 const midgard_vector_alu_src zero_alu_src
= { 0 };
220 /* Coerce structs to integer */
223 vector_alu_srco_unsigned(midgard_vector_alu_src src
)
226 memcpy(&u
, &src
, sizeof(src
));
230 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
231 * the corresponding Midgard source */
233 static midgard_vector_alu_src
234 vector_alu_modifiers(nir_alu_src
*src
)
236 if (!src
) return blank_alu_src
;
238 midgard_vector_alu_src alu_src
= {
240 .negate
= src
->negate
,
243 .half
= 0, /* TODO */
244 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
)
250 /* 'Intrinsic' move for misc aliasing uses independent of actual NIR ALU code */
252 static midgard_instruction
253 v_fmov(unsigned src
, midgard_vector_alu_src mod
, unsigned dest
)
255 midgard_instruction ins
= {
258 .src0
= SSA_UNUSED_1
,
263 .op
= midgard_alu_op_fmov
,
264 .reg_mode
= midgard_reg_mode_full
,
265 .dest_override
= midgard_dest_override_none
,
267 .src1
= vector_alu_srco_unsigned(zero_alu_src
),
268 .src2
= vector_alu_srco_unsigned(mod
)
275 /* load/store instructions have both 32-bit and 16-bit variants, depending on
276 * whether we are using vectors composed of highp or mediump. At the moment, we
277 * don't support half-floats -- this requires changes in other parts of the
278 * compiler -- therefore the 16-bit versions are commented out. */
280 //M_LOAD(load_attr_16);
281 M_LOAD(load_attr_32
);
282 //M_LOAD(load_vary_16);
283 M_LOAD(load_vary_32
);
284 //M_LOAD(load_uniform_16);
285 M_LOAD(load_uniform_32
);
286 M_LOAD(load_color_buffer_8
);
287 //M_STORE(store_vary_16);
288 M_STORE(store_vary_32
);
289 M_STORE(store_cubemap_coords
);
291 static midgard_instruction
292 v_alu_br_compact_cond(midgard_jmp_writeout_op op
, unsigned tag
, signed offset
, unsigned cond
)
294 midgard_branch_cond branch
= {
302 memcpy(&compact
, &branch
, sizeof(branch
));
304 midgard_instruction ins
= {
306 .unit
= ALU_ENAB_BR_COMPACT
,
307 .prepacked_branch
= true,
308 .compact_branch
= true,
309 .br_compact
= compact
312 if (op
== midgard_jmp_writeout_op_writeout
)
318 static midgard_instruction
319 v_branch(bool conditional
, bool invert
)
321 midgard_instruction ins
= {
323 .unit
= ALU_ENAB_BRANCH
,
324 .compact_branch
= true,
326 .conditional
= conditional
,
327 .invert_conditional
= invert
334 static midgard_branch_extended
335 midgard_create_branch_extended( midgard_condition cond
,
336 midgard_jmp_writeout_op op
,
338 signed quadword_offset
)
340 /* For unclear reasons, the condition code is repeated 8 times */
341 uint16_t duplicated_cond
=
351 midgard_branch_extended branch
= {
353 .dest_tag
= dest_tag
,
354 .offset
= quadword_offset
,
355 .cond
= duplicated_cond
361 typedef struct midgard_bundle
{
362 /* Tag for the overall bundle */
365 /* Instructions contained by the bundle */
366 int instruction_count
;
367 midgard_instruction instructions
[5];
369 /* Bundle-wide ALU configuration */
372 bool has_embedded_constants
;
374 bool has_blend_constant
;
376 uint16_t register_words
[8];
377 int register_words_count
;
379 uint64_t body_words
[8];
381 int body_words_count
;
384 typedef struct compiler_context
{
386 gl_shader_stage stage
;
388 /* Is internally a blend shader? Depends on stage == FRAGMENT */
391 /* Tracking for blend constant patching */
392 int blend_constant_number
;
393 int blend_constant_offset
;
395 /* Current NIR function */
398 /* Unordered list of midgard_blocks */
400 struct list_head blocks
;
402 midgard_block
*initial_block
;
403 midgard_block
*previous_source_block
;
404 midgard_block
*final_block
;
406 /* List of midgard_instructions emitted for the current block */
407 midgard_block
*current_block
;
409 /* The index corresponding to the current loop, e.g. for breaks/contineus */
412 /* Constants which have been loaded, for later inlining */
413 struct hash_table_u64
*ssa_constants
;
415 /* SSA indices to be outputted to corresponding varying offset */
416 struct hash_table_u64
*ssa_varyings
;
418 /* SSA values / registers which have been aliased. Naively, these
419 * demand a fmov output; instead, we alias them in a later pass to
420 * avoid the wasted op.
422 * A note on encoding: to avoid dynamic memory management here, rather
423 * than ampping to a pointer, we map to the source index; the key
424 * itself is just the destination index. */
426 struct hash_table_u64
*ssa_to_alias
;
427 struct set
*leftover_ssa_to_alias
;
429 /* Actual SSA-to-register for RA */
430 struct hash_table_u64
*ssa_to_register
;
432 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
433 struct hash_table_u64
*hash_to_temp
;
437 /* Just the count of the max register used. Higher count => higher
438 * register pressure */
441 /* Used for cont/last hinting. Increase when a tex op is added.
442 * Decrease when a tex op is removed. */
443 int texture_op_count
;
445 /* Mapping of texture register -> SSA index for unaliasing */
446 int texture_index
[2];
448 /* If any path hits a discard instruction */
451 /* The number of uniforms allowable for the fast path */
454 /* Count of instructions emitted from NIR overall, across all blocks */
455 int instruction_count
;
457 /* Alpha ref value passed in */
460 /* The index corresponding to the fragment output */
461 unsigned fragment_output
;
463 /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
464 unsigned sysvals
[MAX_SYSVAL_COUNT
];
465 unsigned sysval_count
;
466 struct hash_table_u64
*sysval_to_id
;
469 /* Append instruction to end of current block */
471 static midgard_instruction
*
472 mir_upload_ins(struct midgard_instruction ins
)
474 midgard_instruction
*heap
= malloc(sizeof(ins
));
475 memcpy(heap
, &ins
, sizeof(ins
));
480 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
482 list_addtail(&(mir_upload_ins(ins
))->link
, &ctx
->current_block
->instructions
);
486 mir_insert_instruction_before(struct midgard_instruction
*tag
, struct midgard_instruction ins
)
488 list_addtail(&(mir_upload_ins(ins
))->link
, &tag
->link
);
492 mir_remove_instruction(struct midgard_instruction
*ins
)
494 list_del(&ins
->link
);
497 static midgard_instruction
*
498 mir_prev_op(struct midgard_instruction
*ins
)
500 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
503 static midgard_instruction
*
504 mir_next_op(struct midgard_instruction
*ins
)
506 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
509 static midgard_block
*
510 mir_next_block(struct midgard_block
*blk
)
512 return list_first_entry(&(blk
->link
), midgard_block
, link
);
516 #define mir_foreach_block(ctx, v) list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
517 #define mir_foreach_block_from(ctx, from, v) list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
519 #define mir_foreach_instr(ctx, v) list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
520 #define mir_foreach_instr_safe(ctx, v) list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
521 #define mir_foreach_instr_in_block(block, v) list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
522 #define mir_foreach_instr_in_block_safe(block, v) list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
523 #define mir_foreach_instr_in_block_safe_rev(block, v) list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
524 #define mir_foreach_instr_in_block_from(block, v, from) list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
527 static midgard_instruction
*
528 mir_last_in_block(struct midgard_block
*block
)
530 return list_last_entry(&block
->instructions
, struct midgard_instruction
, link
);
533 static midgard_block
*
534 mir_get_block(compiler_context
*ctx
, int idx
)
536 struct list_head
*lst
= &ctx
->blocks
;
541 return (struct midgard_block
*) lst
;
544 /* Pretty printer for internal Midgard IR */
547 print_mir_source(int source
)
549 if (source
>= SSA_FIXED_MINIMUM
) {
550 /* Specific register */
551 int reg
= SSA_REG_FROM_FIXED(source
);
553 /* TODO: Moving threshold */
554 if (reg
> 16 && reg
< 24)
555 printf("u%d", 23 - reg
);
559 printf("%d", source
);
564 print_mir_instruction(midgard_instruction
*ins
)
570 midgard_alu_op op
= ins
->alu
.op
;
571 const char *name
= alu_opcode_names
[op
];
574 printf("%d.", ins
->unit
);
576 printf("%s", name
? name
: "??");
580 case TAG_LOAD_STORE_4
: {
581 midgard_load_store_op op
= ins
->load_store
.op
;
582 const char *name
= load_store_opcode_names
[op
];
589 case TAG_TEXTURE_4
: {
598 ssa_args
*args
= &ins
->ssa_args
;
600 printf(" %d, ", args
->dest
);
602 print_mir_source(args
->src0
);
605 if (args
->inline_constant
)
606 printf("#%d", ins
->inline_constant
);
608 print_mir_source(args
->src1
);
610 if (ins
->has_constants
)
611 printf(" <%f, %f, %f, %f>", ins
->constants
[0], ins
->constants
[1], ins
->constants
[2], ins
->constants
[3]);
617 print_mir_block(midgard_block
*block
)
621 mir_foreach_instr_in_block(block
, ins
) {
622 print_mir_instruction(ins
);
631 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
633 ins
->has_constants
= true;
634 memcpy(&ins
->constants
, constants
, 16);
636 /* If this is the special blend constant, mark this instruction */
638 if (ctx
->is_blend
&& ctx
->blend_constant_number
== name
)
639 ins
->has_blend_constant
= true;
643 glsl_type_size(const struct glsl_type
*type
)
645 return glsl_count_attribute_slots(type
, false);
649 uniform_type_size(const struct glsl_type
*type
)
651 return st_glsl_storage_type_size(type
, false);
654 /* Lower fdot2 to a vector multiplication followed by channel addition */
656 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
658 if (alu
->op
!= nir_op_fdot2
)
661 b
->cursor
= nir_before_instr(&alu
->instr
);
663 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
664 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
666 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
668 nir_ssa_def
*sum
= nir_fadd(b
,
669 nir_channel(b
, product
, 0),
670 nir_channel(b
, product
, 1));
672 /* Replace the fdot2 with this sum */
673 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
677 midgard_nir_sysval_for_intrinsic(nir_intrinsic_instr
*instr
)
679 switch (instr
->intrinsic
) {
680 case nir_intrinsic_load_viewport_scale
:
681 return PAN_SYSVAL_VIEWPORT_SCALE
;
682 case nir_intrinsic_load_viewport_offset
:
683 return PAN_SYSVAL_VIEWPORT_OFFSET
;
690 midgard_nir_assign_sysval_body(compiler_context
*ctx
, nir_instr
*instr
)
694 if (instr
->type
== nir_instr_type_intrinsic
) {
695 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
696 sysval
= midgard_nir_sysval_for_intrinsic(intr
);
702 /* We have a sysval load; check if it's already been assigned */
704 if (_mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
))
707 /* It hasn't -- so assign it now! */
709 unsigned id
= ctx
->sysval_count
++;
710 _mesa_hash_table_u64_insert(ctx
->sysval_to_id
, sysval
, (void *) ((uintptr_t) id
+ 1));
711 ctx
->sysvals
[id
] = sysval
;
715 midgard_nir_assign_sysvals(compiler_context
*ctx
, nir_shader
*shader
)
717 ctx
->sysval_count
= 0;
719 nir_foreach_function(function
, shader
) {
720 if (!function
->impl
) continue;
722 nir_foreach_block(block
, function
->impl
) {
723 nir_foreach_instr_safe(instr
, block
) {
724 midgard_nir_assign_sysval_body(ctx
, instr
);
731 midgard_nir_lower_fdot2(nir_shader
*shader
)
733 bool progress
= false;
735 nir_foreach_function(function
, shader
) {
736 if (!function
->impl
) continue;
739 nir_builder
*b
= &_b
;
740 nir_builder_init(b
, function
->impl
);
742 nir_foreach_block(block
, function
->impl
) {
743 nir_foreach_instr_safe(instr
, block
) {
744 if (instr
->type
!= nir_instr_type_alu
) continue;
746 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
747 midgard_nir_lower_fdot2_body(b
, alu
);
753 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
761 optimise_nir(nir_shader
*nir
)
765 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
766 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
768 nir_lower_tex_options lower_tex_options
= {
772 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
777 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic
);
778 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
779 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
781 NIR_PASS(progress
, nir
, nir_copy_prop
);
782 NIR_PASS(progress
, nir
, nir_opt_dce
);
783 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
784 NIR_PASS(progress
, nir
, nir_opt_cse
);
785 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
786 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
787 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
788 NIR_PASS(progress
, nir
, nir_opt_undef
);
789 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
792 nir_var_function_temp
);
794 /* TODO: Enable vectorize when merged upstream */
795 // NIR_PASS(progress, nir, nir_opt_vectorize);
798 /* Must be run at the end to prevent creation of fsin/fcos ops */
799 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
804 NIR_PASS(progress
, nir
, nir_opt_dce
);
805 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
806 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
807 NIR_PASS(progress
, nir
, nir_copy_prop
);
810 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
811 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
813 /* Lower mods for float ops only. Integer ops don't support modifiers
814 * (saturate doesn't make sense on integers, neg/abs require dedicated
817 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
818 NIR_PASS(progress
, nir
, nir_copy_prop
);
819 NIR_PASS(progress
, nir
, nir_opt_dce
);
821 /* We implement booleans as 32-bit 0/~0 */
822 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
824 /* Take us out of SSA */
825 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
826 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
828 /* We are a vector architecture; write combine where possible */
829 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
830 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
832 NIR_PASS(progress
, nir
, nir_opt_dce
);
835 /* Front-half of aliasing the SSA slots, merely by inserting the flag in the
836 * appropriate hash table. Intentional off-by-one to avoid confusing NULL with
837 * r0. See the comments in compiler_context */
840 alias_ssa(compiler_context
*ctx
, int dest
, int src
)
842 _mesa_hash_table_u64_insert(ctx
->ssa_to_alias
, dest
+ 1, (void *) ((uintptr_t) src
+ 1));
843 _mesa_set_add(ctx
->leftover_ssa_to_alias
, (void *) (uintptr_t) (dest
+ 1));
846 /* ...or undo it, after which the original index will be used (dummy move should be emitted alongside this) */
849 unalias_ssa(compiler_context
*ctx
, int dest
)
851 _mesa_hash_table_u64_remove(ctx
->ssa_to_alias
, dest
+ 1);
852 /* TODO: Remove from leftover or no? */
856 midgard_pin_output(compiler_context
*ctx
, int index
, int reg
)
858 _mesa_hash_table_u64_insert(ctx
->ssa_to_register
, index
+ 1, (void *) ((uintptr_t) reg
+ 1));
862 midgard_is_pinned(compiler_context
*ctx
, int index
)
864 return _mesa_hash_table_u64_search(ctx
->ssa_to_register
, index
+ 1) != NULL
;
867 /* Do not actually emit a load; instead, cache the constant for inlining */
870 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
872 nir_ssa_def def
= instr
->def
;
874 float *v
= ralloc_array(NULL
, float, 4);
875 memcpy(v
, &instr
->value
.f32
, 4 * sizeof(float));
876 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, def
.index
+ 1, v
);
879 /* Duplicate bits to convert sane 4-bit writemask to obscure 8-bit format (or
883 expand_writemask(unsigned mask
)
887 for (int i
= 0; i
< 4; ++i
)
895 squeeze_writemask(unsigned mask
)
899 for (int i
= 0; i
< 4; ++i
)
900 if (mask
& (3 << (2 * i
)))
907 /* Determines effective writemask, taking quirks and expansion into account */
909 effective_writemask(midgard_vector_alu
*alu
)
911 /* Channel count is off-by-one to fit in two-bits (0 channel makes no
914 unsigned channel_count
= GET_CHANNEL_COUNT(alu_opcode_props
[alu
->op
]);
916 /* If there is a fixed channel count, construct the appropriate mask */
919 return (1 << channel_count
) - 1;
921 /* Otherwise, just squeeze the existing mask */
922 return squeeze_writemask(alu
->mask
);
926 find_or_allocate_temp(compiler_context
*ctx
, unsigned hash
)
928 if ((hash
< 0) || (hash
>= SSA_FIXED_MINIMUM
))
931 unsigned temp
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->hash_to_temp
, hash
+ 1);
936 /* If no temp is find, allocate one */
937 temp
= ctx
->temp_count
++;
938 ctx
->max_hash
= MAX2(ctx
->max_hash
, hash
);
940 _mesa_hash_table_u64_insert(ctx
->hash_to_temp
, hash
+ 1, (void *) ((uintptr_t) temp
+ 1));
946 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
949 return src
->ssa
->index
;
951 return ctx
->func
->impl
->ssa_alloc
+ src
->reg
.reg
->index
;
955 nir_dest_index(compiler_context
*ctx
, nir_dest
*dst
)
958 return dst
->ssa
.index
;
960 return ctx
->func
->impl
->ssa_alloc
+ dst
->reg
.reg
->index
;
964 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
966 return nir_src_index(ctx
, &src
->src
);
969 /* Midgard puts conditionals in r31.w; move an arbitrary source (the output of
970 * a conditional test) into that register */
973 emit_condition(compiler_context
*ctx
, nir_src
*src
, bool for_branch
)
975 /* XXX: Force component correct */
976 int condition
= nir_src_index(ctx
, src
);
978 /* There is no boolean move instruction. Instead, we simulate a move by
979 * ANDing the condition with itself to get it into r31.w */
981 midgard_instruction ins
= {
983 .unit
= for_branch
? UNIT_SMUL
: UNIT_SADD
, /* TODO: DEDUCE THIS */
987 .dest
= SSA_FIXED_REGISTER(31),
990 .op
= midgard_alu_op_iand
,
991 .reg_mode
= midgard_reg_mode_full
,
992 .dest_override
= midgard_dest_override_none
,
993 .mask
= (0x3 << 6), /* w */
994 .src1
= vector_alu_srco_unsigned(blank_alu_src_xxxx
),
995 .src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
)
999 emit_mir_instruction(ctx
, ins
);
1002 #define ALU_CASE(nir, _op) \
1003 case nir_op_##nir: \
1004 op = midgard_alu_op_##_op; \
1008 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
1010 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
1012 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
.dest
);
1013 unsigned nr_components
= is_ssa
? instr
->dest
.dest
.ssa
.num_components
: instr
->dest
.dest
.reg
.reg
->num_components
;
1014 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
1016 /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
1017 * supported. A few do not and are commented for now. Also, there are a
1018 * number of NIR ops which Midgard does not support and need to be
1019 * lowered, also TODO. This switch block emits the opcode and calling
1020 * convention of the Midgard instruction; actual packing is done in
1025 switch (instr
->op
) {
1026 ALU_CASE(fadd
, fadd
);
1027 ALU_CASE(fmul
, fmul
);
1028 ALU_CASE(fmin
, fmin
);
1029 ALU_CASE(fmax
, fmax
);
1030 ALU_CASE(imin
, imin
);
1031 ALU_CASE(imax
, imax
);
1032 ALU_CASE(umin
, umin
);
1033 ALU_CASE(umax
, umax
);
1034 ALU_CASE(fmov
, fmov
);
1035 ALU_CASE(ffloor
, ffloor
);
1036 ALU_CASE(fround_even
, froundeven
);
1037 ALU_CASE(ftrunc
, ftrunc
);
1038 ALU_CASE(fceil
, fceil
);
1039 ALU_CASE(fdot3
, fdot3
);
1040 ALU_CASE(fdot4
, fdot4
);
1041 ALU_CASE(iadd
, iadd
);
1042 ALU_CASE(isub
, isub
);
1043 ALU_CASE(imul
, imul
);
1044 ALU_CASE(iabs
, iabs
);
1046 /* XXX: Use fmov, not imov, since imov was causing major
1047 * issues with texture precision? XXX research */
1048 ALU_CASE(imov
, fmov
);
1050 ALU_CASE(feq32
, feq
);
1051 ALU_CASE(fne32
, fne
);
1052 ALU_CASE(flt32
, flt
);
1053 ALU_CASE(ieq32
, ieq
);
1054 ALU_CASE(ine32
, ine
);
1055 ALU_CASE(ilt32
, ilt
);
1056 ALU_CASE(ult32
, ult
);
1058 /* We don't have a native b2f32 instruction. Instead, like many
1059 * GPUs, we exploit booleans as 0/~0 for false/true, and
1060 * correspondingly AND
1061 * by 1.0 to do the type conversion. For the moment, prime us
1064 * iand [whatever], #0
1066 * At the end of emit_alu (as MIR), we'll fix-up the constant
1069 ALU_CASE(b2f32
, iand
);
1070 ALU_CASE(b2i32
, iand
);
1072 /* Likewise, we don't have a dedicated f2b32 instruction, but
1073 * we can do a "not equal to 0.0" test. */
1075 ALU_CASE(f2b32
, fne
);
1076 ALU_CASE(i2b32
, ine
);
1078 ALU_CASE(frcp
, frcp
);
1079 ALU_CASE(frsq
, frsqrt
);
1080 ALU_CASE(fsqrt
, fsqrt
);
1081 ALU_CASE(fexp2
, fexp2
);
1082 ALU_CASE(flog2
, flog2
);
1084 ALU_CASE(f2i32
, f2i
);
1085 ALU_CASE(f2u32
, f2u
);
1086 ALU_CASE(i2f32
, i2f
);
1087 ALU_CASE(u2f32
, u2f
);
1089 ALU_CASE(fsin
, fsin
);
1090 ALU_CASE(fcos
, fcos
);
1092 ALU_CASE(iand
, iand
);
1094 ALU_CASE(ixor
, ixor
);
1095 ALU_CASE(inot
, inot
);
1096 ALU_CASE(ishl
, ishl
);
1097 ALU_CASE(ishr
, iasr
);
1098 ALU_CASE(ushr
, ilsr
);
1100 ALU_CASE(b32all_fequal2
, fball_eq
);
1101 ALU_CASE(b32all_fequal3
, fball_eq
);
1102 ALU_CASE(b32all_fequal4
, fball_eq
);
1104 ALU_CASE(b32any_fnequal2
, fbany_neq
);
1105 ALU_CASE(b32any_fnequal3
, fbany_neq
);
1106 ALU_CASE(b32any_fnequal4
, fbany_neq
);
1108 ALU_CASE(b32all_iequal2
, iball_eq
);
1109 ALU_CASE(b32all_iequal3
, iball_eq
);
1110 ALU_CASE(b32all_iequal4
, iball_eq
);
1112 ALU_CASE(b32any_inequal2
, ibany_neq
);
1113 ALU_CASE(b32any_inequal3
, ibany_neq
);
1114 ALU_CASE(b32any_inequal4
, ibany_neq
);
1116 /* For greater-or-equal, we lower to less-or-equal and flip the
1122 case nir_op_uge32
: {
1124 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
1125 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
1126 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
1127 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
1130 /* Swap via temporary */
1131 nir_alu_src temp
= instr
->src
[1];
1132 instr
->src
[1] = instr
->src
[0];
1133 instr
->src
[0] = temp
;
1138 case nir_op_b32csel
: {
1139 op
= midgard_alu_op_fcsel
;
1141 /* csel works as a two-arg in Midgard, since the condition is hardcoded in r31.w */
1144 emit_condition(ctx
, &instr
->src
[0].src
, false);
1146 /* The condition is the first argument; move the other
1147 * arguments up one to be a binary instruction for
1150 memmove(instr
->src
, instr
->src
+ 1, 2 * sizeof(nir_alu_src
));
1155 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
1160 /* Fetch unit, quirks, etc information */
1161 unsigned opcode_props
= alu_opcode_props
[op
];
1162 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
1164 /* Initialise fields common between scalar/vector instructions */
1165 midgard_outmod outmod
= instr
->dest
.saturate
? midgard_outmod_sat
: midgard_outmod_none
;
1167 /* src0 will always exist afaik, but src1 will not for 1-argument
1168 * instructions. The latter can only be fetched if the instruction
1169 * needs it, or else we may segfault. */
1171 unsigned src0
= nir_alu_src_index(ctx
, &instr
->src
[0]);
1172 unsigned src1
= nr_inputs
== 2 ? nir_alu_src_index(ctx
, &instr
->src
[1]) : SSA_UNUSED_0
;
1174 /* Rather than use the instruction generation helpers, we do it
1175 * ourselves here to avoid the mess */
1177 midgard_instruction ins
= {
1180 .src0
= quirk_flipped_r24
? SSA_UNUSED_1
: src0
,
1181 .src1
= quirk_flipped_r24
? src0
: src1
,
1186 nir_alu_src
*nirmods
[2] = { NULL
};
1188 if (nr_inputs
== 2) {
1189 nirmods
[0] = &instr
->src
[0];
1190 nirmods
[1] = &instr
->src
[1];
1191 } else if (nr_inputs
== 1) {
1192 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
1197 midgard_vector_alu alu
= {
1199 .reg_mode
= midgard_reg_mode_full
,
1200 .dest_override
= midgard_dest_override_none
,
1203 /* Writemask only valid for non-SSA NIR */
1204 .mask
= expand_writemask((1 << nr_components
) - 1),
1206 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0])),
1207 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1])),
1210 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
1213 alu
.mask
&= expand_writemask(instr
->dest
.write_mask
);
1217 /* Late fixup for emulated instructions */
1219 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1220 /* Presently, our second argument is an inline #0 constant.
1221 * Switch over to an embedded 1.0 constant (that can't fit
1222 * inline, since we're 32-bit, not 16-bit like the inline
1225 ins
.ssa_args
.inline_constant
= false;
1226 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1227 ins
.has_constants
= true;
1229 if (instr
->op
== nir_op_b2f32
) {
1230 ins
.constants
[0] = 1.0f
;
1232 /* Type pun it into place */
1234 memcpy(&ins
.constants
[0], &one
, sizeof(uint32_t));
1237 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1238 } else if (instr
->op
== nir_op_f2b32
|| instr
->op
== nir_op_i2b32
) {
1239 ins
.ssa_args
.inline_constant
= false;
1240 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1241 ins
.has_constants
= true;
1242 ins
.constants
[0] = 0.0f
;
1243 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1246 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1247 /* To avoid duplicating the lookup tables (probably), true LUT
1248 * instructions can only operate as if they were scalars. Lower
1249 * them here by changing the component. */
1251 uint8_t original_swizzle
[4];
1252 memcpy(original_swizzle
, nirmods
[0]->swizzle
, sizeof(nirmods
[0]->swizzle
));
1254 for (int i
= 0; i
< nr_components
; ++i
) {
1255 ins
.alu
.mask
= (0x3) << (2 * i
); /* Mask the associated component */
1257 for (int j
= 0; j
< 4; ++j
)
1258 nirmods
[0]->swizzle
[j
] = original_swizzle
[i
]; /* Pull from the correct component */
1260 ins
.alu
.src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0]));
1261 emit_mir_instruction(ctx
, ins
);
1264 emit_mir_instruction(ctx
, ins
);
1271 emit_uniform_read(compiler_context
*ctx
, unsigned dest
, unsigned offset
)
1273 /* TODO: half-floats */
1275 if (offset
< ctx
->uniform_cutoff
) {
1276 /* Fast path: For the first 16 uniform,
1277 * accesses are 0-cycle, since they're
1278 * just a register fetch in the usual
1279 * case. So, we alias the registers
1280 * while we're still in SSA-space */
1282 int reg_slot
= 23 - offset
;
1283 alias_ssa(ctx
, dest
, SSA_FIXED_REGISTER(reg_slot
));
1285 /* Otherwise, read from the 'special'
1286 * UBO to access higher-indexed
1287 * uniforms, at a performance cost */
1289 midgard_instruction ins
= m_load_uniform_32(dest
, offset
);
1291 /* TODO: Don't split */
1292 ins
.load_store
.varying_parameters
= (offset
& 7) << 7;
1293 ins
.load_store
.address
= offset
>> 3;
1295 ins
.load_store
.unknown
= 0x1E00; /* xxx: what is this? */
1296 emit_mir_instruction(ctx
, ins
);
1301 emit_sysval_read(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1303 /* First, pull out the destination */
1304 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
);
1306 /* Now, figure out which uniform this is */
1307 int sysval
= midgard_nir_sysval_for_intrinsic(instr
);
1308 void *val
= _mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
);
1310 /* Sysvals are prefix uniforms */
1311 unsigned uniform
= ((uintptr_t) val
) - 1;
1313 emit_uniform_read(ctx
, dest
, uniform
);
1317 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1319 unsigned offset
, reg
;
1321 switch (instr
->intrinsic
) {
1322 case nir_intrinsic_discard_if
:
1323 emit_condition(ctx
, &instr
->src
[0], true);
1327 case nir_intrinsic_discard
: {
1328 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1329 struct midgard_instruction discard
= v_branch(conditional
, false);
1330 discard
.branch
.target_type
= TARGET_DISCARD
;
1331 emit_mir_instruction(ctx
, discard
);
1333 ctx
->can_discard
= true;
1337 case nir_intrinsic_load_uniform
:
1338 case nir_intrinsic_load_input
:
1339 assert(nir_src_is_const(instr
->src
[0]) && "no indirect inputs");
1341 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[0]);
1343 reg
= nir_dest_index(ctx
, &instr
->dest
);
1345 if (instr
->intrinsic
== nir_intrinsic_load_uniform
&& !ctx
->is_blend
) {
1346 emit_uniform_read(ctx
, reg
, ctx
->sysval_count
+ offset
);
1347 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1348 /* XXX: Half-floats? */
1349 /* TODO: swizzle, mask */
1351 midgard_instruction ins
= m_load_vary_32(reg
, offset
);
1353 midgard_varying_parameter p
= {
1355 .interpolation
= midgard_interp_default
,
1356 .flat
= /*var->data.interpolation == INTERP_MODE_FLAT*/ 0
1360 memcpy(&u
, &p
, sizeof(p
));
1361 ins
.load_store
.varying_parameters
= u
;
1363 ins
.load_store
.unknown
= 0x1e9e; /* xxx: what is this? */
1364 emit_mir_instruction(ctx
, ins
);
1365 } else if (ctx
->is_blend
&& instr
->intrinsic
== nir_intrinsic_load_uniform
) {
1366 /* Constant encoded as a pinned constant */
1368 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, reg
);
1369 ins
.has_constants
= true;
1370 ins
.has_blend_constant
= true;
1371 emit_mir_instruction(ctx
, ins
);
1372 } else if (ctx
->is_blend
) {
1373 /* For blend shaders, a load might be
1374 * translated various ways depending on what
1375 * we're loading. Figure out how this is used */
1377 nir_variable
*out
= NULL
;
1379 nir_foreach_variable(var
, &ctx
->nir
->inputs
) {
1380 int drvloc
= var
->data
.driver_location
;
1382 if (nir_intrinsic_base(instr
) == drvloc
) {
1390 if (out
->data
.location
== VARYING_SLOT_COL0
) {
1391 /* Source color preloaded to r0 */
1393 midgard_pin_output(ctx
, reg
, 0);
1394 } else if (out
->data
.location
== VARYING_SLOT_COL1
) {
1395 /* Destination color must be read from framebuffer */
1397 midgard_instruction ins
= m_load_color_buffer_8(reg
, 0);
1398 ins
.load_store
.swizzle
= 0; /* xxxx */
1400 /* Read each component sequentially */
1402 for (int c
= 0; c
< 4; ++c
) {
1403 ins
.load_store
.mask
= (1 << c
);
1404 ins
.load_store
.unknown
= c
;
1405 emit_mir_instruction(ctx
, ins
);
1408 /* vadd.u2f hr2, abs(hr2), #0 */
1410 midgard_vector_alu_src alu_src
= blank_alu_src
;
1412 alu_src
.half
= true;
1414 midgard_instruction u2f
= {
1418 .src1
= SSA_UNUSED_0
,
1420 .inline_constant
= true
1423 .op
= midgard_alu_op_u2f
,
1424 .reg_mode
= midgard_reg_mode_half
,
1425 .dest_override
= midgard_dest_override_none
,
1427 .src1
= vector_alu_srco_unsigned(alu_src
),
1428 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
1432 emit_mir_instruction(ctx
, u2f
);
1434 /* vmul.fmul.sat r1, hr2, #0.00392151 */
1436 alu_src
.abs
= false;
1438 midgard_instruction fmul
= {
1440 .inline_constant
= _mesa_float_to_half(1.0 / 255.0),
1444 .src1
= SSA_UNUSED_0
,
1445 .inline_constant
= true
1448 .op
= midgard_alu_op_fmul
,
1449 .reg_mode
= midgard_reg_mode_full
,
1450 .dest_override
= midgard_dest_override_none
,
1451 .outmod
= midgard_outmod_sat
,
1453 .src1
= vector_alu_srco_unsigned(alu_src
),
1454 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
1458 emit_mir_instruction(ctx
, fmul
);
1460 DBG("Unknown input in blend shader\n");
1463 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1464 midgard_instruction ins
= m_load_attr_32(reg
, offset
);
1465 ins
.load_store
.unknown
= 0x1E1E; /* XXX: What is this? */
1466 ins
.load_store
.mask
= (1 << instr
->num_components
) - 1;
1467 emit_mir_instruction(ctx
, ins
);
1469 DBG("Unknown load\n");
1475 case nir_intrinsic_store_output
:
1476 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1478 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1480 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1482 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1483 /* gl_FragColor is not emitted with load/store
1484 * instructions. Instead, it gets plonked into
1485 * r0 at the end of the shader and we do the
1486 * framebuffer writeout dance. TODO: Defer
1489 midgard_pin_output(ctx
, reg
, 0);
1491 /* Save the index we're writing to for later reference
1492 * in the epilogue */
1494 ctx
->fragment_output
= reg
;
1495 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1496 /* Varyings are written into one of two special
1497 * varying register, r26 or r27. The register itself is selected as the register
1498 * in the st_vary instruction, minus the base of 26. E.g. write into r27 and then call st_vary(1)
1500 * Normally emitting fmov's is frowned upon,
1501 * but due to unique constraints of
1502 * REGISTER_VARYING, fmov emission + a
1503 * dedicated cleanup pass is the only way to
1504 * guarantee correctness when considering some
1505 * (common) edge cases XXX: FIXME */
1507 /* If this varying corresponds to a constant (why?!),
1508 * emit that now since it won't get picked up by
1509 * hoisting (since there is no corresponding move
1510 * emitted otherwise) */
1512 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, reg
+ 1);
1514 if (constant_value
) {
1515 /* Special case: emit the varying write
1516 * directly to r26 (looks funny in asm but it's
1517 * fine) and emit the store _now_. Possibly
1518 * slightly slower, but this is a really stupid
1519 * special case anyway (why on earth would you
1520 * have a constant varying? Your own fault for
1521 * slightly worse perf :P) */
1523 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, SSA_FIXED_REGISTER(26));
1524 attach_constants(ctx
, &ins
, constant_value
, reg
+ 1);
1525 emit_mir_instruction(ctx
, ins
);
1527 midgard_instruction st
= m_store_vary_32(SSA_FIXED_REGISTER(0), offset
);
1528 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
1529 emit_mir_instruction(ctx
, st
);
1531 /* Do not emit the varying yet -- instead, just mark down that we need to later */
1533 _mesa_hash_table_u64_insert(ctx
->ssa_varyings
, reg
+ 1, (void *) ((uintptr_t) (offset
+ 1)));
1536 DBG("Unknown store\n");
1542 case nir_intrinsic_load_alpha_ref_float
:
1543 assert(instr
->dest
.is_ssa
);
1545 float ref_value
= ctx
->alpha_ref
;
1547 float *v
= ralloc_array(NULL
, float, 4);
1548 memcpy(v
, &ref_value
, sizeof(float));
1549 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, instr
->dest
.ssa
.index
+ 1, v
);
1552 case nir_intrinsic_load_viewport_scale
:
1553 case nir_intrinsic_load_viewport_offset
:
1554 emit_sysval_read(ctx
, instr
);
1558 printf ("Unhandled intrinsic\n");
1565 midgard_tex_format(enum glsl_sampler_dim dim
)
1568 case GLSL_SAMPLER_DIM_2D
:
1569 case GLSL_SAMPLER_DIM_EXTERNAL
:
1572 case GLSL_SAMPLER_DIM_3D
:
1575 case GLSL_SAMPLER_DIM_CUBE
:
1576 return TEXTURE_CUBE
;
1579 DBG("Unknown sampler dim type\n");
1586 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1589 //assert (!instr->sampler);
1590 //assert (!instr->texture_array_size);
1591 assert (instr
->op
== nir_texop_tex
);
1593 /* Allocate registers via a round robin scheme to alternate between the two registers */
1594 int reg
= ctx
->texture_op_count
& 1;
1595 int in_reg
= reg
, out_reg
= reg
;
1597 /* Make room for the reg */
1599 if (ctx
->texture_index
[reg
] > -1)
1600 unalias_ssa(ctx
, ctx
->texture_index
[reg
]);
1602 int texture_index
= instr
->texture_index
;
1603 int sampler_index
= texture_index
;
1605 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1606 switch (instr
->src
[i
].src_type
) {
1607 case nir_tex_src_coord
: {
1608 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1610 midgard_vector_alu_src alu_src
= blank_alu_src
;
1612 int reg
= SSA_FIXED_REGISTER(REGISTER_TEXTURE_BASE
+ in_reg
);
1614 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1615 /* For cubemaps, we need to load coords into
1616 * special r27, and then use a special ld/st op
1617 * to copy into the texture register */
1619 alu_src
.swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_X
);
1621 midgard_instruction move
= v_fmov(index
, alu_src
, SSA_FIXED_REGISTER(27));
1622 emit_mir_instruction(ctx
, move
);
1624 midgard_instruction st
= m_store_cubemap_coords(reg
, 0);
1625 st
.load_store
.unknown
= 0x24; /* XXX: What is this? */
1626 st
.load_store
.mask
= 0x3; /* xy? */
1627 st
.load_store
.swizzle
= alu_src
.swizzle
;
1628 emit_mir_instruction(ctx
, st
);
1631 alu_src
.swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_X
, COMPONENT_X
);
1633 midgard_instruction ins
= v_fmov(index
, alu_src
, reg
);
1634 emit_mir_instruction(ctx
, ins
);
1637 //midgard_pin_output(ctx, index, REGISTER_TEXTURE_BASE + in_reg);
1643 DBG("Unknown source type\n");
1650 /* No helper to build texture words -- we do it all here */
1651 midgard_instruction ins
= {
1652 .type
= TAG_TEXTURE_4
,
1654 .op
= TEXTURE_OP_NORMAL
,
1655 .format
= midgard_tex_format(instr
->sampler_dim
),
1656 .texture_handle
= texture_index
,
1657 .sampler_handle
= sampler_index
,
1659 /* TODO: Don't force xyzw */
1660 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_W
),
1672 /* Assume we can continue; hint it out later */
1677 /* Set registers to read and write from the same place */
1678 ins
.texture
.in_reg_select
= in_reg
;
1679 ins
.texture
.out_reg_select
= out_reg
;
1681 /* TODO: Dynamic swizzle input selection, half-swizzles? */
1682 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_3D
) {
1683 ins
.texture
.in_reg_swizzle_right
= COMPONENT_X
;
1684 ins
.texture
.in_reg_swizzle_left
= COMPONENT_Y
;
1685 //ins.texture.in_reg_swizzle_third = COMPONENT_Z;
1687 ins
.texture
.in_reg_swizzle_left
= COMPONENT_X
;
1688 ins
.texture
.in_reg_swizzle_right
= COMPONENT_Y
;
1689 //ins.texture.in_reg_swizzle_third = COMPONENT_X;
1692 emit_mir_instruction(ctx
, ins
);
1694 /* Simultaneously alias the destination and emit a move for it. The move will be eliminated if possible */
1696 int o_reg
= REGISTER_TEXTURE_BASE
+ out_reg
, o_index
= nir_dest_index(ctx
, &instr
->dest
);
1697 alias_ssa(ctx
, o_index
, SSA_FIXED_REGISTER(o_reg
));
1698 ctx
->texture_index
[reg
] = o_index
;
1700 midgard_instruction ins2
= v_fmov(SSA_FIXED_REGISTER(o_reg
), blank_alu_src
, o_index
);
1701 emit_mir_instruction(ctx
, ins2
);
1703 /* Used for .cont and .last hinting */
1704 ctx
->texture_op_count
++;
1708 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1710 switch (instr
->type
) {
1711 case nir_jump_break
: {
1712 /* Emit a branch out of the loop */
1713 struct midgard_instruction br
= v_branch(false, false);
1714 br
.branch
.target_type
= TARGET_BREAK
;
1715 br
.branch
.target_break
= ctx
->current_loop
;
1716 emit_mir_instruction(ctx
, br
);
1723 DBG("Unknown jump type %d\n", instr
->type
);
1729 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1731 switch (instr
->type
) {
1732 case nir_instr_type_load_const
:
1733 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1736 case nir_instr_type_intrinsic
:
1737 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1740 case nir_instr_type_alu
:
1741 emit_alu(ctx
, nir_instr_as_alu(instr
));
1744 case nir_instr_type_tex
:
1745 emit_tex(ctx
, nir_instr_as_tex(instr
));
1748 case nir_instr_type_jump
:
1749 emit_jump(ctx
, nir_instr_as_jump(instr
));
1752 case nir_instr_type_ssa_undef
:
1757 DBG("Unhandled instruction type\n");
1762 /* Determine the actual hardware from the index based on the RA results or special values */
1765 dealias_register(compiler_context
*ctx
, struct ra_graph
*g
, int reg
, int maxreg
)
1767 if (reg
>= SSA_FIXED_MINIMUM
)
1768 return SSA_REG_FROM_FIXED(reg
);
1771 assert(reg
< maxreg
);
1772 int r
= ra_get_node_reg(g
, reg
);
1773 ctx
->work_registers
= MAX2(ctx
->work_registers
, r
);
1778 /* fmov style unused */
1780 return REGISTER_UNUSED
;
1782 /* lut style unused */
1784 return REGISTER_UNUSED
;
1787 DBG("Unknown SSA register alias %d\n", reg
);
1794 midgard_ra_select_callback(struct ra_graph
*g
, BITSET_WORD
*regs
, void *data
)
1796 /* Choose the first available register to minimise reported register pressure */
1798 for (int i
= 0; i
< 16; ++i
) {
1799 if (BITSET_TEST(regs
, i
)) {
1809 midgard_is_live_in_instr(midgard_instruction
*ins
, int src
)
1811 if (ins
->ssa_args
.src0
== src
) return true;
1812 if (ins
->ssa_args
.src1
== src
) return true;
1818 is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
)
1820 /* Check the rest of the block for liveness */
1821 mir_foreach_instr_in_block_from(block
, ins
, mir_next_op(start
)) {
1822 if (midgard_is_live_in_instr(ins
, src
))
1826 /* Check the rest of the blocks for liveness */
1827 mir_foreach_block_from(ctx
, mir_next_block(block
), b
) {
1828 mir_foreach_instr_in_block(b
, ins
) {
1829 if (midgard_is_live_in_instr(ins
, src
))
1834 /* TODO: How does control flow interact in complex shaders? */
1840 allocate_registers(compiler_context
*ctx
)
1842 /* First, initialize the RA */
1843 struct ra_regs
*regs
= ra_alloc_reg_set(NULL
, 32, true);
1845 /* Create a primary (general purpose) class, as well as special purpose
1846 * pipeline register classes */
1848 int primary_class
= ra_alloc_reg_class(regs
);
1849 int varying_class
= ra_alloc_reg_class(regs
);
1851 /* Add the full set of work registers */
1852 int work_count
= 16 - MAX2((ctx
->uniform_cutoff
- 8), 0);
1853 for (int i
= 0; i
< work_count
; ++i
)
1854 ra_class_add_reg(regs
, primary_class
, i
);
1856 /* Add special registers */
1857 ra_class_add_reg(regs
, varying_class
, REGISTER_VARYING_BASE
);
1858 ra_class_add_reg(regs
, varying_class
, REGISTER_VARYING_BASE
+ 1);
1860 /* We're done setting up */
1861 ra_set_finalize(regs
, NULL
);
1863 /* Transform the MIR into squeezed index form */
1864 mir_foreach_block(ctx
, block
) {
1865 mir_foreach_instr_in_block(block
, ins
) {
1866 if (ins
->compact_branch
) continue;
1868 ins
->ssa_args
.src0
= find_or_allocate_temp(ctx
, ins
->ssa_args
.src0
);
1869 ins
->ssa_args
.src1
= find_or_allocate_temp(ctx
, ins
->ssa_args
.src1
);
1870 ins
->ssa_args
.dest
= find_or_allocate_temp(ctx
, ins
->ssa_args
.dest
);
1872 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
1873 print_mir_block(block
);
1876 /* Let's actually do register allocation */
1877 int nodes
= ctx
->temp_count
;
1878 struct ra_graph
*g
= ra_alloc_interference_graph(regs
, nodes
);
1880 /* Set everything to the work register class, unless it has somewhere
1883 mir_foreach_block(ctx
, block
) {
1884 mir_foreach_instr_in_block(block
, ins
) {
1885 if (ins
->compact_branch
) continue;
1887 if (ins
->ssa_args
.dest
< 0) continue;
1889 if (ins
->ssa_args
.dest
>= SSA_FIXED_MINIMUM
) continue;
1891 int class = primary_class
;
1893 ra_set_node_class(g
, ins
->ssa_args
.dest
, class);
1897 for (int index
= 0; index
<= ctx
->max_hash
; ++index
) {
1898 unsigned temp
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_to_register
, index
+ 1);
1901 unsigned reg
= temp
- 1;
1902 int t
= find_or_allocate_temp(ctx
, index
);
1903 ra_set_node_reg(g
, t
, reg
);
1907 /* Determine liveness */
1909 int *live_start
= malloc(nodes
* sizeof(int));
1910 int *live_end
= malloc(nodes
* sizeof(int));
1912 /* Initialize as non-existent */
1914 for (int i
= 0; i
< nodes
; ++i
) {
1915 live_start
[i
] = live_end
[i
] = -1;
1920 mir_foreach_block(ctx
, block
) {
1921 mir_foreach_instr_in_block(block
, ins
) {
1922 if (ins
->compact_branch
) continue;
1924 if (ins
->ssa_args
.dest
< SSA_FIXED_MINIMUM
) {
1925 /* If this destination is not yet live, it is now since we just wrote it */
1927 int dest
= ins
->ssa_args
.dest
;
1929 if (live_start
[dest
] == -1)
1930 live_start
[dest
] = d
;
1933 /* Since we just used a source, the source might be
1934 * dead now. Scan the rest of the block for
1935 * invocations, and if there are none, the source dies
1938 int sources
[2] = { ins
->ssa_args
.src0
, ins
->ssa_args
.src1
};
1940 for (int src
= 0; src
< 2; ++src
) {
1941 int s
= sources
[src
];
1943 if (s
< 0) continue;
1945 if (s
>= SSA_FIXED_MINIMUM
) continue;
1947 if (!is_live_after(ctx
, block
, ins
, s
)) {
1956 /* If a node still hasn't been killed, kill it now */
1958 for (int i
= 0; i
< nodes
; ++i
) {
1959 /* live_start == -1 most likely indicates a pinned output */
1961 if (live_end
[i
] == -1)
1965 /* Setup interference between nodes that are live at the same time */
1967 for (int i
= 0; i
< nodes
; ++i
) {
1968 for (int j
= i
+ 1; j
< nodes
; ++j
) {
1969 if (!(live_start
[i
] >= live_end
[j
] || live_start
[j
] >= live_end
[i
]))
1970 ra_add_node_interference(g
, i
, j
);
1974 ra_set_select_reg_callback(g
, midgard_ra_select_callback
, NULL
);
1976 if (!ra_allocate(g
)) {
1977 DBG("Error allocating registers\n");
1985 mir_foreach_block(ctx
, block
) {
1986 mir_foreach_instr_in_block(block
, ins
) {
1987 if (ins
->compact_branch
) continue;
1989 ssa_args args
= ins
->ssa_args
;
1991 switch (ins
->type
) {
1993 ins
->registers
.src1_reg
= dealias_register(ctx
, g
, args
.src0
, nodes
);
1995 ins
->registers
.src2_imm
= args
.inline_constant
;
1997 if (args
.inline_constant
) {
1998 /* Encode inline 16-bit constant as a vector by default */
2000 ins
->registers
.src2_reg
= ins
->inline_constant
>> 11;
2002 int lower_11
= ins
->inline_constant
& ((1 << 12) - 1);
2004 uint16_t imm
= ((lower_11
>> 8) & 0x7) | ((lower_11
& 0xFF) << 3);
2005 ins
->alu
.src2
= imm
<< 2;
2007 ins
->registers
.src2_reg
= dealias_register(ctx
, g
, args
.src1
, nodes
);
2010 ins
->registers
.out_reg
= dealias_register(ctx
, g
, args
.dest
, nodes
);
2014 case TAG_LOAD_STORE_4
: {
2015 if (OP_IS_STORE_VARY(ins
->load_store
.op
)) {
2016 /* TODO: use ssa_args for store_vary */
2017 ins
->load_store
.reg
= 0;
2019 bool has_dest
= args
.dest
>= 0;
2020 int ssa_arg
= has_dest
? args
.dest
: args
.src0
;
2022 ins
->load_store
.reg
= dealias_register(ctx
, g
, ssa_arg
, nodes
);
2035 /* Midgard IR only knows vector ALU types, but we sometimes need to actually
2036 * use scalar ALU instructions, for functional or performance reasons. To do
2037 * this, we just demote vector ALU payloads to scalar. */
2040 component_from_mask(unsigned mask
)
2042 for (int c
= 0; c
< 4; ++c
) {
2043 if (mask
& (3 << (2 * c
)))
2052 is_single_component_mask(unsigned mask
)
2056 for (int c
= 0; c
< 4; ++c
)
2057 if (mask
& (3 << (2 * c
)))
2060 return components
== 1;
2063 /* Create a mask of accessed components from a swizzle to figure out vector
2067 swizzle_to_access_mask(unsigned swizzle
)
2069 unsigned component_mask
= 0;
2071 for (int i
= 0; i
< 4; ++i
) {
2072 unsigned c
= (swizzle
>> (2 * i
)) & 3;
2073 component_mask
|= (1 << c
);
2076 return component_mask
;
2080 vector_to_scalar_source(unsigned u
)
2082 midgard_vector_alu_src v
;
2083 memcpy(&v
, &u
, sizeof(v
));
2085 midgard_scalar_alu_src s
= {
2089 .component
= (v
.swizzle
& 3) << 1
2093 memcpy(&o
, &s
, sizeof(s
));
2095 return o
& ((1 << 6) - 1);
2098 static midgard_scalar_alu
2099 vector_to_scalar_alu(midgard_vector_alu v
, midgard_instruction
*ins
)
2101 /* The output component is from the mask */
2102 midgard_scalar_alu s
= {
2104 .src1
= vector_to_scalar_source(v
.src1
),
2105 .src2
= vector_to_scalar_source(v
.src2
),
2108 .output_full
= 1, /* TODO: Half */
2109 .output_component
= component_from_mask(v
.mask
) << 1,
2112 /* Inline constant is passed along rather than trying to extract it
2115 if (ins
->ssa_args
.inline_constant
) {
2117 int lower_11
= ins
->inline_constant
& ((1 << 12) - 1);
2118 imm
|= (lower_11
>> 9) & 3;
2119 imm
|= (lower_11
>> 6) & 4;
2120 imm
|= (lower_11
>> 2) & 0x38;
2121 imm
|= (lower_11
& 63) << 6;
2129 /* Midgard prefetches instruction types, so during emission we need to
2130 * lookahead too. Unless this is the last instruction, in which we return 1. Or
2131 * if this is the second to last and the last is an ALU, then it's also 1... */
2133 #define IS_ALU(tag) (tag == TAG_ALU_4 || tag == TAG_ALU_8 || \
2134 tag == TAG_ALU_12 || tag == TAG_ALU_16)
2136 #define EMIT_AND_COUNT(type, val) util_dynarray_append(emission, type, val); \
2137 bytes_emitted += sizeof(type)
2140 emit_binary_vector_instruction(midgard_instruction
*ains
,
2141 uint16_t *register_words
, int *register_words_count
,
2142 uint64_t *body_words
, size_t *body_size
, int *body_words_count
,
2143 size_t *bytes_emitted
)
2145 memcpy(®ister_words
[(*register_words_count
)++], &ains
->registers
, sizeof(ains
->registers
));
2146 *bytes_emitted
+= sizeof(midgard_reg_info
);
2148 body_size
[*body_words_count
] = sizeof(midgard_vector_alu
);
2149 memcpy(&body_words
[(*body_words_count
)++], &ains
->alu
, sizeof(ains
->alu
));
2150 *bytes_emitted
+= sizeof(midgard_vector_alu
);
2153 /* Checks for an SSA data hazard between two adjacent instructions, keeping in
2154 * mind that we are a vector architecture and we can write to different
2155 * components simultaneously */
2158 can_run_concurrent_ssa(midgard_instruction
*first
, midgard_instruction
*second
)
2160 /* Each instruction reads some registers and writes to a register. See
2161 * where the first writes */
2163 /* Figure out where exactly we wrote to */
2164 int source
= first
->ssa_args
.dest
;
2165 int source_mask
= first
->type
== TAG_ALU_4
? squeeze_writemask(first
->alu
.mask
) : 0xF;
2167 /* As long as the second doesn't read from the first, we're okay */
2168 if (second
->ssa_args
.src0
== source
) {
2169 if (first
->type
== TAG_ALU_4
) {
2170 /* Figure out which components we just read from */
2172 int q
= second
->alu
.src1
;
2173 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2175 /* Check if there are components in common, and fail if so */
2176 if (swizzle_to_access_mask(m
->swizzle
) & source_mask
)
2183 if (second
->ssa_args
.src1
== source
)
2186 /* Otherwise, it's safe in that regard. Another data hazard is both
2187 * writing to the same place, of course */
2189 if (second
->ssa_args
.dest
== source
) {
2190 /* ...but only if the components overlap */
2191 int dest_mask
= second
->type
== TAG_ALU_4
? squeeze_writemask(second
->alu
.mask
) : 0xF;
2193 if (dest_mask
& source_mask
)
2203 midgard_instruction
**segment
, unsigned segment_size
,
2204 midgard_instruction
*ains
)
2206 for (int s
= 0; s
< segment_size
; ++s
)
2207 if (!can_run_concurrent_ssa(segment
[s
], ains
))
2215 /* Schedules, but does not emit, a single basic block. After scheduling, the
2216 * final tag and size of the block are known, which are necessary for branching
2219 static midgard_bundle
2220 schedule_bundle(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*ins
, int *skip
)
2222 int instructions_emitted
= 0, instructions_consumed
= -1;
2223 midgard_bundle bundle
= { 0 };
2225 uint8_t tag
= ins
->type
;
2227 /* Default to the instruction's tag */
2230 switch (ins
->type
) {
2232 uint32_t control
= 0;
2233 size_t bytes_emitted
= sizeof(control
);
2235 /* TODO: Constant combining */
2236 int index
= 0, last_unit
= 0;
2238 /* Previous instructions, for the purpose of parallelism */
2239 midgard_instruction
*segment
[4] = {0};
2240 int segment_size
= 0;
2242 instructions_emitted
= -1;
2243 midgard_instruction
*pins
= ins
;
2246 midgard_instruction
*ains
= pins
;
2248 /* Advance instruction pointer */
2250 ains
= mir_next_op(pins
);
2254 /* Out-of-work condition */
2255 if ((struct list_head
*) ains
== &block
->instructions
)
2258 /* Ensure that the chain can continue */
2259 if (ains
->type
!= TAG_ALU_4
) break;
2261 /* According to the presentation "The ARM
2262 * Mali-T880 Mobile GPU" from HotChips 27,
2263 * there are two pipeline stages. Branching
2264 * position determined experimentally. Lines
2265 * are executed in parallel:
2268 * [ VADD ] [ SMUL ] [ LUT ] [ BRANCH ]
2270 * Verify that there are no ordering dependencies here.
2272 * TODO: Allow for parallelism!!!
2275 /* Pick a unit for it if it doesn't force a particular unit */
2277 int unit
= ains
->unit
;
2280 int op
= ains
->alu
.op
;
2281 int units
= alu_opcode_props
[op
];
2283 /* TODO: Promotion of scalars to vectors */
2284 int vector
= ((!is_single_component_mask(ains
->alu
.mask
)) || ((units
& UNITS_SCALAR
) == 0)) && (units
& UNITS_ANY_VECTOR
);
2287 assert(units
& UNITS_SCALAR
);
2290 if (last_unit
>= UNIT_VADD
) {
2291 if (units
& UNIT_VLUT
)
2296 if ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
))
2298 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
2300 else if (units
& UNIT_VLUT
)
2306 if (last_unit
>= UNIT_VADD
) {
2307 if ((units
& UNIT_SMUL
) && !(control
& UNIT_SMUL
))
2309 else if (units
& UNIT_VLUT
)
2314 if ((units
& UNIT_SADD
) && !(control
& UNIT_SADD
) && !midgard_has_hazard(segment
, segment_size
, ains
))
2316 else if (units
& UNIT_SMUL
)
2317 unit
= ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
)) ? UNIT_VMUL
: UNIT_SMUL
;
2318 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
2325 assert(unit
& units
);
2328 /* Late unit check, this time for encoding (not parallelism) */
2329 if (unit
<= last_unit
) break;
2331 /* Clear the segment */
2332 if (last_unit
< UNIT_VADD
&& unit
>= UNIT_VADD
)
2335 if (midgard_has_hazard(segment
, segment_size
, ains
))
2338 /* We're good to go -- emit the instruction */
2341 segment
[segment_size
++] = ains
;
2343 /* Only one set of embedded constants per
2344 * bundle possible; if we have more, we must
2345 * break the chain early, unfortunately */
2347 if (ains
->has_constants
) {
2348 if (bundle
.has_embedded_constants
) {
2349 /* ...but if there are already
2350 * constants but these are the
2351 * *same* constants, we let it
2354 if (memcmp(bundle
.constants
, ains
->constants
, sizeof(bundle
.constants
)))
2357 bundle
.has_embedded_constants
= true;
2358 memcpy(bundle
.constants
, ains
->constants
, sizeof(bundle
.constants
));
2360 /* If this is a blend shader special constant, track it for patching */
2361 if (ains
->has_blend_constant
)
2362 bundle
.has_blend_constant
= true;
2366 if (ains
->unit
& UNITS_ANY_VECTOR
) {
2367 emit_binary_vector_instruction(ains
, bundle
.register_words
,
2368 &bundle
.register_words_count
, bundle
.body_words
,
2369 bundle
.body_size
, &bundle
.body_words_count
, &bytes_emitted
);
2370 } else if (ains
->compact_branch
) {
2371 /* All of r0 has to be written out
2372 * along with the branch writeout.
2375 if (ains
->writeout
) {
2377 midgard_instruction ins
= v_fmov(0, blank_alu_src
, SSA_FIXED_REGISTER(0));
2378 ins
.unit
= UNIT_VMUL
;
2380 control
|= ins
.unit
;
2382 emit_binary_vector_instruction(&ins
, bundle
.register_words
,
2383 &bundle
.register_words_count
, bundle
.body_words
,
2384 bundle
.body_size
, &bundle
.body_words_count
, &bytes_emitted
);
2386 /* Analyse the group to see if r0 is written in full, on-time, without hanging dependencies*/
2387 bool written_late
= false;
2388 bool components
[4] = { 0 };
2389 uint16_t register_dep_mask
= 0;
2390 uint16_t written_mask
= 0;
2392 midgard_instruction
*qins
= ins
;
2393 for (int t
= 0; t
< index
; ++t
) {
2394 if (qins
->registers
.out_reg
!= 0) {
2395 /* Mark down writes */
2397 written_mask
|= (1 << qins
->registers
.out_reg
);
2399 /* Mark down the register dependencies for errata check */
2401 if (qins
->registers
.src1_reg
< 16)
2402 register_dep_mask
|= (1 << qins
->registers
.src1_reg
);
2404 if (qins
->registers
.src2_reg
< 16)
2405 register_dep_mask
|= (1 << qins
->registers
.src2_reg
);
2407 int mask
= qins
->alu
.mask
;
2409 for (int c
= 0; c
< 4; ++c
)
2410 if (mask
& (0x3 << (2 * c
)))
2411 components
[c
] = true;
2413 /* ..but if the writeout is too late, we have to break up anyway... for some reason */
2415 if (qins
->unit
== UNIT_VLUT
)
2416 written_late
= true;
2419 /* Advance instruction pointer */
2420 qins
= mir_next_op(qins
);
2424 /* ERRATA (?): In a bundle ending in a fragment writeout, the register dependencies of r0 cannot be written within this bundle (discovered in -bshading:shading=phong) */
2425 if (register_dep_mask
& written_mask
) {
2426 DBG("ERRATA WORKAROUND: Breakup for writeout dependency masks %X vs %X (common %X)\n", register_dep_mask
, written_mask
, register_dep_mask
& written_mask
);
2433 /* If even a single component is not written, break it up (conservative check). */
2434 bool breakup
= false;
2436 for (int c
= 0; c
< 4; ++c
)
2443 /* Otherwise, we're free to proceed */
2447 if (ains
->unit
== ALU_ENAB_BRANCH
) {
2448 bundle
.body_size
[bundle
.body_words_count
] = sizeof(midgard_branch_extended
);
2449 memcpy(&bundle
.body_words
[bundle
.body_words_count
++], &ains
->branch_extended
, sizeof(midgard_branch_extended
));
2450 bytes_emitted
+= sizeof(midgard_branch_extended
);
2452 bundle
.body_size
[bundle
.body_words_count
] = sizeof(ains
->br_compact
);
2453 memcpy(&bundle
.body_words
[bundle
.body_words_count
++], &ains
->br_compact
, sizeof(ains
->br_compact
));
2454 bytes_emitted
+= sizeof(ains
->br_compact
);
2457 memcpy(&bundle
.register_words
[bundle
.register_words_count
++], &ains
->registers
, sizeof(ains
->registers
));
2458 bytes_emitted
+= sizeof(midgard_reg_info
);
2460 bundle
.body_size
[bundle
.body_words_count
] = sizeof(midgard_scalar_alu
);
2461 bundle
.body_words_count
++;
2462 bytes_emitted
+= sizeof(midgard_scalar_alu
);
2465 /* Defer marking until after writing to allow for break */
2466 control
|= ains
->unit
;
2467 last_unit
= ains
->unit
;
2468 ++instructions_emitted
;
2472 /* Bubble up the number of instructions for skipping */
2473 instructions_consumed
= index
- 1;
2477 /* Pad ALU op to nearest word */
2479 if (bytes_emitted
& 15) {
2480 padding
= 16 - (bytes_emitted
& 15);
2481 bytes_emitted
+= padding
;
2484 /* Constants must always be quadwords */
2485 if (bundle
.has_embedded_constants
)
2486 bytes_emitted
+= 16;
2488 /* Size ALU instruction for tag */
2489 bundle
.tag
= (TAG_ALU_4
) + (bytes_emitted
/ 16) - 1;
2490 bundle
.padding
= padding
;
2491 bundle
.control
= bundle
.tag
| control
;
2496 case TAG_LOAD_STORE_4
: {
2497 /* Load store instructions have two words at once. If
2498 * we only have one queued up, we need to NOP pad.
2499 * Otherwise, we store both in succession to save space
2500 * and cycles -- letting them go in parallel -- skip
2501 * the next. The usefulness of this optimisation is
2502 * greatly dependent on the quality of the instruction
2506 midgard_instruction
*next_op
= mir_next_op(ins
);
2508 if ((struct list_head
*) next_op
!= &block
->instructions
&& next_op
->type
== TAG_LOAD_STORE_4
) {
2509 /* As the two operate concurrently, make sure
2510 * they are not dependent */
2512 if (can_run_concurrent_ssa(ins
, next_op
) || true) {
2513 /* Skip ahead, since it's redundant with the pair */
2514 instructions_consumed
= 1 + (instructions_emitted
++);
2522 /* Texture ops default to single-op-per-bundle scheduling */
2526 /* Copy the instructions into the bundle */
2527 bundle
.instruction_count
= instructions_emitted
+ 1;
2531 midgard_instruction
*uins
= ins
;
2532 for (int i
= 0; used_idx
< bundle
.instruction_count
; ++i
) {
2533 bundle
.instructions
[used_idx
++] = *uins
;
2534 uins
= mir_next_op(uins
);
2537 *skip
= (instructions_consumed
== -1) ? instructions_emitted
: instructions_consumed
;
2543 quadword_size(int tag
)
2558 case TAG_LOAD_STORE_4
:
2570 /* Schedule a single block by iterating its instruction to create bundles.
2571 * While we go, tally about the bundle sizes to compute the block size. */
2574 schedule_block(compiler_context
*ctx
, midgard_block
*block
)
2576 util_dynarray_init(&block
->bundles
, NULL
);
2578 block
->quadword_count
= 0;
2580 mir_foreach_instr_in_block(block
, ins
) {
2582 midgard_bundle bundle
= schedule_bundle(ctx
, block
, ins
, &skip
);
2583 util_dynarray_append(&block
->bundles
, midgard_bundle
, bundle
);
2585 if (bundle
.has_blend_constant
) {
2586 /* TODO: Multiblock? */
2587 int quadwords_within_block
= block
->quadword_count
+ quadword_size(bundle
.tag
) - 1;
2588 ctx
->blend_constant_offset
= quadwords_within_block
* 0x10;
2592 ins
= mir_next_op(ins
);
2594 block
->quadword_count
+= quadword_size(bundle
.tag
);
2597 block
->is_scheduled
= true;
2601 schedule_program(compiler_context
*ctx
)
2603 allocate_registers(ctx
);
2605 mir_foreach_block(ctx
, block
) {
2606 schedule_block(ctx
, block
);
2610 /* After everything is scheduled, emit whole bundles at a time */
2613 emit_binary_bundle(compiler_context
*ctx
, midgard_bundle
*bundle
, struct util_dynarray
*emission
, int next_tag
)
2615 int lookahead
= next_tag
<< 4;
2617 switch (bundle
->tag
) {
2622 /* Actually emit each component */
2623 util_dynarray_append(emission
, uint32_t, bundle
->control
| lookahead
);
2625 for (int i
= 0; i
< bundle
->register_words_count
; ++i
)
2626 util_dynarray_append(emission
, uint16_t, bundle
->register_words
[i
]);
2628 /* Emit body words based on the instructions bundled */
2629 for (int i
= 0; i
< bundle
->instruction_count
; ++i
) {
2630 midgard_instruction
*ins
= &bundle
->instructions
[i
];
2632 if (ins
->unit
& UNITS_ANY_VECTOR
) {
2633 memcpy(util_dynarray_grow(emission
, sizeof(midgard_vector_alu
)), &ins
->alu
, sizeof(midgard_vector_alu
));
2634 } else if (ins
->compact_branch
) {
2635 /* Dummy move, XXX DRY */
2636 if ((i
== 0) && ins
->writeout
) {
2637 midgard_instruction ins
= v_fmov(0, blank_alu_src
, SSA_FIXED_REGISTER(0));
2638 memcpy(util_dynarray_grow(emission
, sizeof(midgard_vector_alu
)), &ins
.alu
, sizeof(midgard_vector_alu
));
2641 if (ins
->unit
== ALU_ENAB_BR_COMPACT
) {
2642 memcpy(util_dynarray_grow(emission
, sizeof(ins
->br_compact
)), &ins
->br_compact
, sizeof(ins
->br_compact
));
2644 memcpy(util_dynarray_grow(emission
, sizeof(ins
->branch_extended
)), &ins
->branch_extended
, sizeof(ins
->branch_extended
));
2648 midgard_scalar_alu scalarised
= vector_to_scalar_alu(ins
->alu
, ins
);
2649 memcpy(util_dynarray_grow(emission
, sizeof(scalarised
)), &scalarised
, sizeof(scalarised
));
2653 /* Emit padding (all zero) */
2654 memset(util_dynarray_grow(emission
, bundle
->padding
), 0, bundle
->padding
);
2656 /* Tack on constants */
2658 if (bundle
->has_embedded_constants
) {
2659 util_dynarray_append(emission
, float, bundle
->constants
[0]);
2660 util_dynarray_append(emission
, float, bundle
->constants
[1]);
2661 util_dynarray_append(emission
, float, bundle
->constants
[2]);
2662 util_dynarray_append(emission
, float, bundle
->constants
[3]);
2668 case TAG_LOAD_STORE_4
: {
2669 /* One or two composing instructions */
2671 uint64_t current64
, next64
= LDST_NOP
;
2673 memcpy(¤t64
, &bundle
->instructions
[0].load_store
, sizeof(current64
));
2675 if (bundle
->instruction_count
== 2)
2676 memcpy(&next64
, &bundle
->instructions
[1].load_store
, sizeof(next64
));
2678 midgard_load_store instruction
= {
2679 .type
= bundle
->tag
,
2680 .next_type
= next_tag
,
2685 util_dynarray_append(emission
, midgard_load_store
, instruction
);
2690 case TAG_TEXTURE_4
: {
2691 /* Texture instructions are easy, since there is no
2692 * pipelining nor VLIW to worry about. We may need to set the .last flag */
2694 midgard_instruction
*ins
= &bundle
->instructions
[0];
2696 ins
->texture
.type
= TAG_TEXTURE_4
;
2697 ins
->texture
.next_type
= next_tag
;
2699 ctx
->texture_op_count
--;
2701 if (!ctx
->texture_op_count
) {
2702 ins
->texture
.cont
= 0;
2703 ins
->texture
.last
= 1;
2706 util_dynarray_append(emission
, midgard_texture_word
, ins
->texture
);
2711 DBG("Unknown midgard instruction type\n");
2718 /* ALU instructions can inline or embed constants, which decreases register
2719 * pressure and saves space. */
2721 #define CONDITIONAL_ATTACH(src) { \
2722 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->ssa_args.src + 1); \
2725 attach_constants(ctx, alu, entry, alu->ssa_args.src + 1); \
2726 alu->ssa_args.src = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2731 inline_alu_constants(compiler_context
*ctx
)
2733 mir_foreach_instr(ctx
, alu
) {
2734 /* Other instructions cannot inline constants */
2735 if (alu
->type
!= TAG_ALU_4
) continue;
2737 /* If there is already a constant here, we can do nothing */
2738 if (alu
->has_constants
) continue;
2740 CONDITIONAL_ATTACH(src0
);
2742 if (!alu
->has_constants
) {
2743 CONDITIONAL_ATTACH(src1
)
2744 } else if (!alu
->inline_constant
) {
2745 /* Corner case: _two_ vec4 constants, for instance with a
2746 * csel. For this case, we can only use a constant
2747 * register for one, we'll have to emit a move for the
2748 * other. Note, if both arguments are constants, then
2749 * necessarily neither argument depends on the value of
2750 * any particular register. As the destination register
2751 * will be wiped, that means we can spill the constant
2752 * to the destination register.
2755 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->ssa_args
.src1
+ 1);
2756 unsigned scratch
= alu
->ssa_args
.dest
;
2759 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, scratch
);
2760 attach_constants(ctx
, &ins
, entry
, alu
->ssa_args
.src1
+ 1);
2762 /* Force a break XXX Defer r31 writes */
2763 ins
.unit
= UNIT_VLUT
;
2765 /* Set the source */
2766 alu
->ssa_args
.src1
= scratch
;
2768 /* Inject us -before- the last instruction which set r31 */
2769 mir_insert_instruction_before(mir_prev_op(alu
), ins
);
2775 /* Midgard supports two types of constants, embedded constants (128-bit) and
2776 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2777 * constants can be demoted to inline constants, for space savings and
2778 * sometimes a performance boost */
2781 embedded_to_inline_constant(compiler_context
*ctx
)
2783 mir_foreach_instr(ctx
, ins
) {
2784 if (!ins
->has_constants
) continue;
2786 if (ins
->ssa_args
.inline_constant
) continue;
2788 /* Blend constants must not be inlined by definition */
2789 if (ins
->has_blend_constant
) continue;
2791 /* src1 cannot be an inline constant due to encoding
2792 * restrictions. So, if possible we try to flip the arguments
2795 int op
= ins
->alu
.op
;
2797 if (ins
->ssa_args
.src0
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2798 /* Flip based on op. Fallthrough intentional */
2801 /* These ops require an operational change to flip their arguments TODO */
2802 case midgard_alu_op_flt
:
2803 case midgard_alu_op_fle
:
2804 case midgard_alu_op_ilt
:
2805 case midgard_alu_op_ile
:
2806 case midgard_alu_op_fcsel
:
2807 case midgard_alu_op_icsel
:
2808 case midgard_alu_op_isub
:
2809 DBG("Missed non-commutative flip (%s)\n", alu_opcode_names
[op
]);
2812 /* These ops are commutative and Just Flip */
2813 case midgard_alu_op_fne
:
2814 case midgard_alu_op_fadd
:
2815 case midgard_alu_op_fmul
:
2816 case midgard_alu_op_fmin
:
2817 case midgard_alu_op_fmax
:
2818 case midgard_alu_op_iadd
:
2819 case midgard_alu_op_imul
:
2820 case midgard_alu_op_feq
:
2821 case midgard_alu_op_ieq
:
2822 case midgard_alu_op_ine
:
2823 case midgard_alu_op_iand
:
2824 case midgard_alu_op_ior
:
2825 case midgard_alu_op_ixor
:
2826 /* Flip the SSA numbers */
2827 ins
->ssa_args
.src0
= ins
->ssa_args
.src1
;
2828 ins
->ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
2830 /* And flip the modifiers */
2834 src_temp
= ins
->alu
.src2
;
2835 ins
->alu
.src2
= ins
->alu
.src1
;
2836 ins
->alu
.src1
= src_temp
;
2843 if (ins
->ssa_args
.src1
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2844 /* Extract the source information */
2846 midgard_vector_alu_src
*src
;
2847 int q
= ins
->alu
.src2
;
2848 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2851 /* Component is from the swizzle, e.g. r26.w -> w component. TODO: What if x is masked out? */
2852 int component
= src
->swizzle
& 3;
2854 /* Scale constant appropriately, if we can legally */
2855 uint16_t scaled_constant
= 0;
2857 /* XXX: Check legality */
2858 if (midgard_is_integer_op(op
)) {
2859 /* TODO: Inline integer */
2862 unsigned int *iconstants
= (unsigned int *) ins
->constants
;
2863 scaled_constant
= (uint16_t) iconstants
[component
];
2865 /* Constant overflow after resize */
2866 if (scaled_constant
!= iconstants
[component
])
2869 scaled_constant
= _mesa_float_to_half((float) ins
->constants
[component
]);
2872 /* We don't know how to handle these with a constant */
2874 if (src
->abs
|| src
->negate
|| src
->half
|| src
->rep_low
|| src
->rep_high
) {
2875 DBG("Bailing inline constant...\n");
2879 /* Make sure that the constant is not itself a
2880 * vector by checking if all accessed values
2881 * (by the swizzle) are the same. */
2883 uint32_t *cons
= (uint32_t *) ins
->constants
;
2884 uint32_t value
= cons
[component
];
2886 bool is_vector
= false;
2887 unsigned mask
= effective_writemask(&ins
->alu
);
2889 for (int c
= 1; c
< 4; ++c
) {
2890 /* We only care if this component is actually used */
2891 if (!(mask
& (1 << c
)))
2894 uint32_t test
= cons
[(src
->swizzle
>> (2 * c
)) & 3];
2896 if (test
!= value
) {
2905 /* Get rid of the embedded constant */
2906 ins
->has_constants
= false;
2907 ins
->ssa_args
.src1
= SSA_UNUSED_0
;
2908 ins
->ssa_args
.inline_constant
= true;
2909 ins
->inline_constant
= scaled_constant
;
2914 /* Map normal SSA sources to other SSA sources / fixed registers (like
2918 map_ssa_to_alias(compiler_context
*ctx
, int *ref
)
2920 unsigned int alias
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_to_alias
, *ref
+ 1);
2923 /* Remove entry in leftovers to avoid a redunant fmov */
2925 struct set_entry
*leftover
= _mesa_set_search(ctx
->leftover_ssa_to_alias
, ((void *) (uintptr_t) (*ref
+ 1)));
2928 _mesa_set_remove(ctx
->leftover_ssa_to_alias
, leftover
);
2930 /* Assign the alias map */
2936 #define AS_SRC(to, u) \
2937 int q##to = ins->alu.src2; \
2938 midgard_vector_alu_src *to = (midgard_vector_alu_src *) &q##to;
2940 /* Removing unused moves is necessary to clean up the texture pipeline results.
2942 * To do so, we find moves in the MIR. We check if their destination is live later. If it's not, the move is redundant. */
2945 midgard_eliminate_orphan_moves(compiler_context
*ctx
, midgard_block
*block
)
2947 mir_foreach_instr_in_block_safe(block
, ins
) {
2948 if (ins
->type
!= TAG_ALU_4
) continue;
2950 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2952 if (ins
->ssa_args
.dest
>= SSA_FIXED_MINIMUM
) continue;
2954 if (midgard_is_pinned(ctx
, ins
->ssa_args
.dest
)) continue;
2956 if (is_live_after(ctx
, block
, ins
, ins
->ssa_args
.dest
)) continue;
2958 mir_remove_instruction(ins
);
2962 /* The following passes reorder MIR instructions to enable better scheduling */
2965 midgard_pair_load_store(compiler_context
*ctx
, midgard_block
*block
)
2967 mir_foreach_instr_in_block_safe(block
, ins
) {
2968 if (ins
->type
!= TAG_LOAD_STORE_4
) continue;
2970 /* We've found a load/store op. Check if next is also load/store. */
2971 midgard_instruction
*next_op
= mir_next_op(ins
);
2972 if (&next_op
->link
!= &block
->instructions
) {
2973 if (next_op
->type
== TAG_LOAD_STORE_4
) {
2974 /* If so, we're done since we're a pair */
2975 ins
= mir_next_op(ins
);
2979 /* Maximum search distance to pair, to avoid register pressure disasters */
2980 int search_distance
= 8;
2982 /* Otherwise, we have an orphaned load/store -- search for another load */
2983 mir_foreach_instr_in_block_from(block
, c
, mir_next_op(ins
)) {
2984 /* Terminate search if necessary */
2985 if (!(search_distance
--)) break;
2987 if (c
->type
!= TAG_LOAD_STORE_4
) continue;
2989 if (OP_IS_STORE(c
->load_store
.op
)) continue;
2991 /* We found one! Move it up to pair and remove it from the old location */
2993 mir_insert_instruction_before(ins
, *c
);
2994 mir_remove_instruction(c
);
3002 /* Emit varying stores late */
3005 midgard_emit_store(compiler_context
*ctx
, midgard_block
*block
) {
3006 /* Iterate in reverse to get the final write, rather than the first */
3008 mir_foreach_instr_in_block_safe_rev(block
, ins
) {
3009 /* Check if what we just wrote needs a store */
3010 int idx
= ins
->ssa_args
.dest
;
3011 uintptr_t varying
= ((uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_varyings
, idx
+ 1));
3013 if (!varying
) continue;
3017 /* We need to store to the appropriate varying, so emit the
3020 /* TODO: Integrate with special purpose RA (and scheduler?) */
3021 bool high_varying_register
= false;
3023 midgard_instruction mov
= v_fmov(idx
, blank_alu_src
, SSA_FIXED_REGISTER(REGISTER_VARYING_BASE
+ high_varying_register
));
3025 midgard_instruction st
= m_store_vary_32(SSA_FIXED_REGISTER(high_varying_register
), varying
);
3026 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
3028 mir_insert_instruction_before(mir_next_op(ins
), st
);
3029 mir_insert_instruction_before(mir_next_op(ins
), mov
);
3031 /* We no longer need to store this varying */
3032 _mesa_hash_table_u64_remove(ctx
->ssa_varyings
, idx
+ 1);
3036 /* If there are leftovers after the below pass, emit actual fmov
3037 * instructions for the slow-but-correct path */
3040 emit_leftover_move(compiler_context
*ctx
)
3042 set_foreach(ctx
->leftover_ssa_to_alias
, leftover
) {
3043 int base
= ((uintptr_t) leftover
->key
) - 1;
3046 map_ssa_to_alias(ctx
, &mapped
);
3047 EMIT(fmov
, mapped
, blank_alu_src
, base
);
3052 actualise_ssa_to_alias(compiler_context
*ctx
)
3054 mir_foreach_instr(ctx
, ins
) {
3055 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src0
);
3056 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src1
);
3059 emit_leftover_move(ctx
);
3062 /* Vertex shaders do not write gl_Position as is; instead, they write a
3063 * transformed screen space position as a varying. See section 12.5 "Coordinate
3064 * Transformation" of the ES 3.2 full specification for details.
3066 * This transformation occurs early on, as NIR and prior to optimisation, in
3067 * order to take advantage of NIR optimisation passes of the transform itself.
3071 write_transformed_position(nir_builder
*b
, nir_src input_point_src
)
3073 nir_ssa_def
*input_point
= nir_ssa_for_src(b
, input_point_src
, 4);
3074 nir_ssa_def
*scale
= nir_load_viewport_scale(b
);
3075 nir_ssa_def
*offset
= nir_load_viewport_offset(b
);
3077 /* World space to normalised device coordinates to screen space */
3079 nir_ssa_def
*w_recip
= nir_frcp(b
, nir_channel(b
, input_point
, 3));
3080 nir_ssa_def
*ndc_point
= nir_fmul(b
, nir_channels(b
, input_point
, 0x7), w_recip
);
3081 nir_ssa_def
*screen
= nir_fadd(b
, nir_fmul(b
, ndc_point
, scale
), offset
);
3083 /* gl_Position will be written out in screenspace xyz, with w set to
3084 * the reciprocal we computed earlier. The transformed w component is
3085 * then used for perspective-correct varying interpolation. The
3086 * transformed w component must preserve its original sign; this is
3087 * used in depth clipping computations */
3089 nir_ssa_def
*screen_space
= nir_vec4(b
,
3090 nir_channel(b
, screen
, 0),
3091 nir_channel(b
, screen
, 1),
3092 nir_channel(b
, screen
, 2),
3095 /* Finally, write out the transformed values to the varying */
3097 nir_intrinsic_instr
*store
;
3098 store
= nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_store_output
);
3099 store
->num_components
= 4;
3100 nir_intrinsic_set_base(store
, 0);
3101 nir_intrinsic_set_write_mask(store
, 0xf);
3102 store
->src
[0].ssa
= screen_space
;
3103 store
->src
[0].is_ssa
= true;
3104 store
->src
[1] = nir_src_for_ssa(nir_imm_int(b
, 0));
3105 nir_builder_instr_insert(b
, &store
->instr
);
3109 transform_position_writes(nir_shader
*shader
)
3111 nir_foreach_function(func
, shader
) {
3112 nir_foreach_block(block
, func
->impl
) {
3113 nir_foreach_instr_safe(instr
, block
) {
3114 if (instr
->type
!= nir_instr_type_intrinsic
) continue;
3116 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
3117 nir_variable
*out
= NULL
;
3119 switch (intr
->intrinsic
) {
3120 case nir_intrinsic_store_output
:
3121 /* already had i/o lowered.. lookup the matching output var: */
3122 nir_foreach_variable(var
, &shader
->outputs
) {
3123 int drvloc
= var
->data
.driver_location
;
3125 if (nir_intrinsic_base(intr
) == drvloc
) {
3139 if (out
->data
.mode
!= nir_var_shader_out
)
3142 if (out
->data
.location
!= VARYING_SLOT_POS
)
3146 nir_builder_init(&b
, func
->impl
);
3147 b
.cursor
= nir_before_instr(instr
);
3149 write_transformed_position(&b
, intr
->src
[0]);
3150 nir_instr_remove(instr
);
3157 emit_fragment_epilogue(compiler_context
*ctx
)
3159 /* Special case: writing out constants requires us to include the move
3160 * explicitly now, so shove it into r0 */
3162 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, ctx
->fragment_output
+ 1);
3164 if (constant_value
) {
3165 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, SSA_FIXED_REGISTER(0));
3166 attach_constants(ctx
, &ins
, constant_value
, ctx
->fragment_output
+ 1);
3167 emit_mir_instruction(ctx
, ins
);
3170 /* Perform the actual fragment writeout. We have two writeout/branch
3171 * instructions, forming a loop until writeout is successful as per the
3172 * docs. TODO: gl_FragDepth */
3174 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
3175 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
3178 /* For the blend epilogue, we need to convert the blended fragment vec4 (stored
3179 * in r0) to a RGBA8888 value by scaling and type converting. We then output it
3180 * with the int8 analogue to the fragment epilogue */
3183 emit_blend_epilogue(compiler_context
*ctx
)
3185 /* vmul.fmul.none.fulllow hr48, r0, #255 */
3187 midgard_instruction scale
= {
3190 .inline_constant
= _mesa_float_to_half(255.0),
3192 .src0
= SSA_FIXED_REGISTER(0),
3193 .src1
= SSA_UNUSED_0
,
3194 .dest
= SSA_FIXED_REGISTER(24),
3195 .inline_constant
= true
3198 .op
= midgard_alu_op_fmul
,
3199 .reg_mode
= midgard_reg_mode_full
,
3200 .dest_override
= midgard_dest_override_lower
,
3202 .src1
= vector_alu_srco_unsigned(blank_alu_src
),
3203 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3207 emit_mir_instruction(ctx
, scale
);
3209 /* vadd.f2u8.pos.low hr0, hr48, #0 */
3211 midgard_vector_alu_src alu_src
= blank_alu_src
;
3212 alu_src
.half
= true;
3214 midgard_instruction f2u8
= {
3217 .src0
= SSA_FIXED_REGISTER(24),
3218 .src1
= SSA_UNUSED_0
,
3219 .dest
= SSA_FIXED_REGISTER(0),
3220 .inline_constant
= true
3223 .op
= midgard_alu_op_f2u8
,
3224 .reg_mode
= midgard_reg_mode_half
,
3225 .dest_override
= midgard_dest_override_lower
,
3226 .outmod
= midgard_outmod_pos
,
3228 .src1
= vector_alu_srco_unsigned(alu_src
),
3229 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3233 emit_mir_instruction(ctx
, f2u8
);
3235 /* vmul.imov.quarter r0, r0, r0 */
3237 midgard_instruction imov_8
= {
3240 .src0
= SSA_UNUSED_1
,
3241 .src1
= SSA_FIXED_REGISTER(0),
3242 .dest
= SSA_FIXED_REGISTER(0),
3245 .op
= midgard_alu_op_imov
,
3246 .reg_mode
= midgard_reg_mode_quarter
,
3247 .dest_override
= midgard_dest_override_none
,
3249 .src1
= vector_alu_srco_unsigned(blank_alu_src
),
3250 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3254 /* Emit branch epilogue with the 8-bit move as the source */
3256 emit_mir_instruction(ctx
, imov_8
);
3257 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
3259 emit_mir_instruction(ctx
, imov_8
);
3260 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
3263 static midgard_block
*
3264 emit_block(compiler_context
*ctx
, nir_block
*block
)
3266 midgard_block
*this_block
= malloc(sizeof(midgard_block
));
3267 list_addtail(&this_block
->link
, &ctx
->blocks
);
3269 this_block
->is_scheduled
= false;
3272 ctx
->texture_index
[0] = -1;
3273 ctx
->texture_index
[1] = -1;
3275 /* Set up current block */
3276 list_inithead(&this_block
->instructions
);
3277 ctx
->current_block
= this_block
;
3279 nir_foreach_instr(instr
, block
) {
3280 emit_instr(ctx
, instr
);
3281 ++ctx
->instruction_count
;
3284 inline_alu_constants(ctx
);
3285 embedded_to_inline_constant(ctx
);
3287 /* Perform heavylifting for aliasing */
3288 actualise_ssa_to_alias(ctx
);
3290 midgard_emit_store(ctx
, this_block
);
3291 midgard_eliminate_orphan_moves(ctx
, this_block
);
3292 midgard_pair_load_store(ctx
, this_block
);
3294 /* Append fragment shader epilogue (value writeout) */
3295 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
3296 if (block
== nir_impl_last_block(ctx
->func
->impl
)) {
3298 emit_blend_epilogue(ctx
);
3300 emit_fragment_epilogue(ctx
);
3304 /* Fallthrough save */
3305 this_block
->next_fallthrough
= ctx
->previous_source_block
;
3307 if (block
== nir_start_block(ctx
->func
->impl
))
3308 ctx
->initial_block
= this_block
;
3310 if (block
== nir_impl_last_block(ctx
->func
->impl
))
3311 ctx
->final_block
= this_block
;
3313 /* Allow the next control flow to access us retroactively, for
3315 ctx
->current_block
= this_block
;
3317 /* Document the fallthrough chain */
3318 ctx
->previous_source_block
= this_block
;
3323 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
3326 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
3328 /* Conditional branches expect the condition in r31.w; emit a move for
3329 * that in the _previous_ block (which is the current block). */
3330 emit_condition(ctx
, &nif
->condition
, true);
3332 /* Speculatively emit the branch, but we can't fill it in until later */
3333 EMIT(branch
, true, true);
3334 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
3336 /* Emit the two subblocks */
3337 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
3339 /* Emit a jump from the end of the then block to the end of the else */
3340 EMIT(branch
, false, false);
3341 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
3343 /* Emit second block, and check if it's empty */
3345 int else_idx
= ctx
->block_count
;
3346 int count_in
= ctx
->instruction_count
;
3347 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
3348 int after_else_idx
= ctx
->block_count
;
3350 /* Now that we have the subblocks emitted, fix up the branches */
3355 if (ctx
->instruction_count
== count_in
) {
3356 /* The else block is empty, so don't emit an exit jump */
3357 mir_remove_instruction(then_exit
);
3358 then_branch
->branch
.target_block
= after_else_idx
;
3360 then_branch
->branch
.target_block
= else_idx
;
3361 then_exit
->branch
.target_block
= after_else_idx
;
3366 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
3368 /* Remember where we are */
3369 midgard_block
*start_block
= ctx
->current_block
;
3371 /* Allocate a loop number for this. TODO: Nested loops. Instead of a
3372 * single current_loop variable, maybe we need a stack */
3374 int loop_idx
= ++ctx
->current_loop
;
3376 /* Get index from before the body so we can loop back later */
3377 int start_idx
= ctx
->block_count
;
3379 /* Emit the body itself */
3380 emit_cf_list(ctx
, &nloop
->body
);
3382 /* Branch back to loop back */
3383 struct midgard_instruction br_back
= v_branch(false, false);
3384 br_back
.branch
.target_block
= start_idx
;
3385 emit_mir_instruction(ctx
, br_back
);
3387 /* Find the index of the block about to follow us (note: we don't add
3388 * one; blocks are 0-indexed so we get a fencepost problem) */
3389 int break_block_idx
= ctx
->block_count
;
3391 /* Fix up the break statements we emitted to point to the right place,
3392 * now that we can allocate a block number for them */
3394 list_for_each_entry_from(struct midgard_block
, block
, start_block
, &ctx
->blocks
, link
) {
3395 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
3396 print_mir_block(block
);
3397 mir_foreach_instr_in_block(block
, ins
) {
3398 if (ins
->type
!= TAG_ALU_4
) continue;
3399 if (!ins
->compact_branch
) continue;
3400 if (ins
->prepacked_branch
) continue;
3402 /* We found a branch -- check the type to see if we need to do anything */
3403 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
3405 /* It's a break! Check if it's our break */
3406 if (ins
->branch
.target_break
!= loop_idx
) continue;
3408 /* Okay, cool, we're breaking out of this loop.
3409 * Rewrite from a break to a goto */
3411 ins
->branch
.target_type
= TARGET_GOTO
;
3412 ins
->branch
.target_block
= break_block_idx
;
3417 static midgard_block
*
3418 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
3420 midgard_block
*start_block
= NULL
;
3422 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
3423 switch (node
->type
) {
3424 case nir_cf_node_block
: {
3425 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
3428 start_block
= block
;
3433 case nir_cf_node_if
:
3434 emit_if(ctx
, nir_cf_node_as_if(node
));
3437 case nir_cf_node_loop
:
3438 emit_loop(ctx
, nir_cf_node_as_loop(node
));
3441 case nir_cf_node_function
:
3450 /* Due to lookahead, we need to report the first tag executed in the command
3451 * stream and in branch targets. An initial block might be empty, so iterate
3452 * until we find one that 'works' */
3455 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
3457 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
3459 unsigned first_tag
= 0;
3462 midgard_bundle
*initial_bundle
= util_dynarray_element(&initial_block
->bundles
, midgard_bundle
, 0);
3464 if (initial_bundle
) {
3465 first_tag
= initial_bundle
->tag
;
3469 /* Initial block is empty, try the next block */
3470 initial_block
= list_first_entry(&(initial_block
->link
), midgard_block
, link
);
3471 } while(initial_block
!= NULL
);
3478 midgard_compile_shader_nir(nir_shader
*nir
, midgard_program
*program
, bool is_blend
)
3480 struct util_dynarray
*compiled
= &program
->compiled
;
3482 midgard_debug
= debug_get_option_midgard_debug();
3484 compiler_context ictx
= {
3486 .stage
= nir
->info
.stage
,
3488 .is_blend
= is_blend
,
3489 .blend_constant_offset
= -1,
3491 .alpha_ref
= program
->alpha_ref
3494 compiler_context
*ctx
= &ictx
;
3496 /* TODO: Decide this at runtime */
3497 ctx
->uniform_cutoff
= 8;
3499 /* Assign var locations early, so the epilogue can use them if necessary */
3501 nir_assign_var_locations(&nir
->outputs
, &nir
->num_outputs
, glsl_type_size
);
3502 nir_assign_var_locations(&nir
->inputs
, &nir
->num_inputs
, glsl_type_size
);
3503 nir_assign_var_locations(&nir
->uniforms
, &nir
->num_uniforms
, uniform_type_size
);
3505 /* Initialize at a global (not block) level hash tables */
3507 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
3508 ctx
->ssa_varyings
= _mesa_hash_table_u64_create(NULL
);
3509 ctx
->ssa_to_alias
= _mesa_hash_table_u64_create(NULL
);
3510 ctx
->ssa_to_register
= _mesa_hash_table_u64_create(NULL
);
3511 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
3512 ctx
->sysval_to_id
= _mesa_hash_table_u64_create(NULL
);
3513 ctx
->leftover_ssa_to_alias
= _mesa_set_create(NULL
, _mesa_hash_pointer
, _mesa_key_pointer_equal
);
3515 /* Record the varying mapping for the command stream's bookkeeping */
3517 struct exec_list
*varyings
=
3518 ctx
->stage
== MESA_SHADER_VERTEX
? &nir
->outputs
: &nir
->inputs
;
3520 nir_foreach_variable(var
, varyings
) {
3521 unsigned loc
= var
->data
.driver_location
;
3522 program
->varyings
[loc
] = var
->data
.location
;
3525 /* Lower vars -- not I/O -- before epilogue */
3527 NIR_PASS_V(nir
, nir_lower_var_copies
);
3528 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
3529 NIR_PASS_V(nir
, nir_split_var_copies
);
3530 NIR_PASS_V(nir
, nir_lower_var_copies
);
3531 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
3532 NIR_PASS_V(nir
, nir_lower_var_copies
);
3533 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
3535 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
3537 /* Append vertex epilogue before optimisation, so the epilogue itself
3540 if (ctx
->stage
== MESA_SHADER_VERTEX
)
3541 transform_position_writes(nir
);
3543 /* Optimisation passes */
3547 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
3548 nir_print_shader(nir
, stdout
);
3551 /* Assign sysvals and counts, now that we're sure
3552 * (post-optimisation) */
3554 midgard_nir_assign_sysvals(ctx
, nir
);
3556 program
->uniform_count
= nir
->num_uniforms
;
3557 program
->sysval_count
= ctx
->sysval_count
;
3558 memcpy(program
->sysvals
, ctx
->sysvals
, sizeof(ctx
->sysvals
[0]) * ctx
->sysval_count
);
3560 program
->attribute_count
= (ctx
->stage
== MESA_SHADER_VERTEX
) ? nir
->num_inputs
: 0;
3561 program
->varying_count
= (ctx
->stage
== MESA_SHADER_VERTEX
) ? nir
->num_outputs
: ((ctx
->stage
== MESA_SHADER_FRAGMENT
) ? nir
->num_inputs
: 0);
3563 nir_foreach_function(func
, nir
) {
3567 list_inithead(&ctx
->blocks
);
3568 ctx
->block_count
= 0;
3571 emit_cf_list(ctx
, &func
->impl
->body
);
3572 emit_block(ctx
, func
->impl
->end_block
);
3574 break; /* TODO: Multi-function shaders */
3577 util_dynarray_init(compiled
, NULL
);
3580 schedule_program(ctx
);
3582 /* Now that all the bundles are scheduled and we can calculate block
3583 * sizes, emit actual branch instructions rather than placeholders */
3585 int br_block_idx
= 0;
3587 mir_foreach_block(ctx
, block
) {
3588 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3589 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
3590 midgard_instruction
*ins
= &bundle
->instructions
[c
];
3592 if (!midgard_is_branch_unit(ins
->unit
)) continue;
3594 if (ins
->prepacked_branch
) continue;
3596 /* Parse some basic branch info */
3597 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
3598 bool is_conditional
= ins
->branch
.conditional
;
3599 bool is_inverted
= ins
->branch
.invert_conditional
;
3600 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
3602 /* Determine the block we're jumping to */
3603 int target_number
= ins
->branch
.target_block
;
3605 /* Report the destination tag. Discards don't need this */
3606 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
3608 /* Count up the number of quadwords we're jumping over. That is, the number of quadwords in each of the blocks between (br_block_idx, target_number) */
3609 int quadword_offset
= 0;
3612 /* Jump to the end of the shader. We
3613 * need to include not only the
3614 * following blocks, but also the
3615 * contents of our current block (since
3616 * discard can come in the middle of
3619 midgard_block
*blk
= mir_get_block(ctx
, br_block_idx
+ 1);
3621 for (midgard_bundle
*bun
= bundle
+ 1; bun
< (midgard_bundle
*)((char*) block
->bundles
.data
+ block
->bundles
.size
); ++bun
) {
3622 quadword_offset
+= quadword_size(bun
->tag
);
3625 mir_foreach_block_from(ctx
, blk
, b
) {
3626 quadword_offset
+= b
->quadword_count
;
3629 } else if (target_number
> br_block_idx
) {
3632 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
3633 midgard_block
*blk
= mir_get_block(ctx
, idx
);
3636 quadword_offset
+= blk
->quadword_count
;
3639 /* Jump backwards */
3641 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
3642 midgard_block
*blk
= mir_get_block(ctx
, idx
);
3645 quadword_offset
-= blk
->quadword_count
;
3649 /* Unconditional extended branches (far jumps)
3650 * have issues, so we always use a conditional
3651 * branch, setting the condition to always for
3652 * unconditional. For compact unconditional
3653 * branches, cond isn't used so it doesn't
3654 * matter what we pick. */
3656 midgard_condition cond
=
3657 !is_conditional
? midgard_condition_always
:
3658 is_inverted
? midgard_condition_false
:
3659 midgard_condition_true
;
3661 midgard_jmp_writeout_op op
=
3662 is_discard
? midgard_jmp_writeout_op_discard
:
3663 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
3664 midgard_jmp_writeout_op_branch_cond
;
3667 midgard_branch_extended branch
=
3668 midgard_create_branch_extended(
3673 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
3674 } else if (is_conditional
|| is_discard
) {
3675 midgard_branch_cond branch
= {
3677 .dest_tag
= dest_tag
,
3678 .offset
= quadword_offset
,
3682 assert(branch
.offset
== quadword_offset
);
3684 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
3686 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
3688 midgard_branch_uncond branch
= {
3690 .dest_tag
= dest_tag
,
3691 .offset
= quadword_offset
,
3695 assert(branch
.offset
== quadword_offset
);
3697 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
3705 /* Emit flat binary from the instruction arrays. Iterate each block in
3706 * sequence. Save instruction boundaries such that lookahead tags can
3707 * be assigned easily */
3709 /* Cache _all_ bundles in source order for lookahead across failed branches */
3711 int bundle_count
= 0;
3712 mir_foreach_block(ctx
, block
) {
3713 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
3715 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
3717 mir_foreach_block(ctx
, block
) {
3718 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3719 source_order_bundles
[bundle_idx
++] = bundle
;
3723 int current_bundle
= 0;
3725 mir_foreach_block(ctx
, block
) {
3726 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3729 if (current_bundle
+ 1 < bundle_count
) {
3730 uint8_t next
= source_order_bundles
[current_bundle
+ 1]->tag
;
3732 if (!(current_bundle
+ 2 < bundle_count
) && IS_ALU(next
)) {
3739 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
3743 /* TODO: Free deeper */
3744 //util_dynarray_fini(&block->instructions);
3747 free(source_order_bundles
);
3749 /* Report the very first tag executed */
3750 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
3752 /* Deal with off-by-one related to the fencepost problem */
3753 program
->work_register_count
= ctx
->work_registers
+ 1;
3755 program
->can_discard
= ctx
->can_discard
;
3756 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
3758 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
3760 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
3761 disassemble_midgard(program
->compiled
.data
, program
->compiled
.size
);