2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 #include <sys/types.h>
33 #include "main/mtypes.h"
34 #include "compiler/glsl/glsl_to_nir.h"
35 #include "compiler/nir_types.h"
36 #include "compiler/nir/nir_builder.h"
37 #include "util/half_float.h"
38 #include "util/u_math.h"
39 #include "util/u_debug.h"
40 #include "util/u_dynarray.h"
41 #include "util/list.h"
42 #include "main/mtypes.h"
45 #include "midgard_nir.h"
46 #include "midgard_compile.h"
47 #include "midgard_ops.h"
50 #include "midgard_quirks.h"
52 #include "disassemble.h"
54 static const struct debug_named_value debug_options
[] = {
55 {"msgs", MIDGARD_DBG_MSGS
, "Print debug messages"},
56 {"shaders", MIDGARD_DBG_SHADERS
, "Dump shaders in NIR and MIR"},
57 {"shaderdb", MIDGARD_DBG_SHADERDB
, "Prints shader-db statistics"},
61 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug
, "MIDGARD_MESA_DEBUG", debug_options
, 0)
63 unsigned SHADER_DB_COUNT
= 0;
65 int midgard_debug
= 0;
67 #define DBG(fmt, ...) \
68 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
69 fprintf(stderr, "%s:%d: "fmt, \
70 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
71 static midgard_block
*
72 create_empty_block(compiler_context
*ctx
)
74 midgard_block
*blk
= rzalloc(ctx
, midgard_block
);
76 blk
->base
.predecessors
= _mesa_set_create(blk
,
78 _mesa_key_pointer_equal
);
80 blk
->base
.name
= ctx
->block_source_count
++;
86 schedule_barrier(compiler_context
*ctx
)
88 midgard_block
*temp
= ctx
->after_block
;
89 ctx
->after_block
= create_empty_block(ctx
);
91 list_addtail(&ctx
->after_block
->base
.link
, &ctx
->blocks
);
92 list_inithead(&ctx
->after_block
->base
.instructions
);
93 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->after_block
->base
);
94 ctx
->current_block
= ctx
->after_block
;
95 ctx
->after_block
= temp
;
98 /* Helpers to generate midgard_instruction's using macro magic, since every
99 * driver seems to do it that way */
101 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
103 #define M_LOAD_STORE(name, store, T) \
104 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
105 midgard_instruction i = { \
106 .type = TAG_LOAD_STORE_4, \
109 .src = { ~0, ~0, ~0, ~0 }, \
110 .swizzle = SWIZZLE_IDENTITY_4, \
112 .op = midgard_op_##name, \
119 i.src_types[0] = T; \
128 #define M_LOAD(name, T) M_LOAD_STORE(name, false, T)
129 #define M_STORE(name, T) M_LOAD_STORE(name, true, T)
131 M_LOAD(ld_attr_32
, nir_type_uint32
);
132 M_LOAD(ld_vary_32
, nir_type_uint32
);
133 M_LOAD(ld_ubo_int4
, nir_type_uint32
);
134 M_LOAD(ld_int4
, nir_type_uint32
);
135 M_STORE(st_int4
, nir_type_uint32
);
136 M_LOAD(ld_color_buffer_32u
, nir_type_uint32
);
137 M_LOAD(ld_color_buffer_as_fp16
, nir_type_float16
);
138 M_STORE(st_vary_32
, nir_type_uint32
);
139 M_LOAD(ld_cubemap_coords
, nir_type_uint32
);
140 M_LOAD(ld_compute_id
, nir_type_uint32
);
142 static midgard_instruction
143 v_branch(bool conditional
, bool invert
)
145 midgard_instruction ins
= {
147 .unit
= ALU_ENAB_BRANCH
,
148 .compact_branch
= true,
150 .conditional
= conditional
,
151 .invert_conditional
= invert
154 .src
= { ~0, ~0, ~0, ~0 },
160 static midgard_branch_extended
161 midgard_create_branch_extended( midgard_condition cond
,
162 midgard_jmp_writeout_op op
,
164 signed quadword_offset
)
166 /* The condition code is actually a LUT describing a function to
167 * combine multiple condition codes. However, we only support a single
168 * condition code at the moment, so we just duplicate over a bunch of
171 uint16_t duplicated_cond
=
181 midgard_branch_extended branch
= {
183 .dest_tag
= dest_tag
,
184 .offset
= quadword_offset
,
185 .cond
= duplicated_cond
192 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
194 ins
->has_constants
= true;
195 memcpy(&ins
->constants
, constants
, 16);
199 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
201 return glsl_count_attribute_slots(type
, false);
204 /* Lower fdot2 to a vector multiplication followed by channel addition */
206 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
208 if (alu
->op
!= nir_op_fdot2
)
211 b
->cursor
= nir_before_instr(&alu
->instr
);
213 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
214 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
216 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
218 nir_ssa_def
*sum
= nir_fadd(b
,
219 nir_channel(b
, product
, 0),
220 nir_channel(b
, product
, 1));
222 /* Replace the fdot2 with this sum */
223 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
227 midgard_nir_lower_fdot2(nir_shader
*shader
)
229 bool progress
= false;
231 nir_foreach_function(function
, shader
) {
232 if (!function
->impl
) continue;
235 nir_builder
*b
= &_b
;
236 nir_builder_init(b
, function
->impl
);
238 nir_foreach_block(block
, function
->impl
) {
239 nir_foreach_instr_safe(instr
, block
) {
240 if (instr
->type
!= nir_instr_type_alu
) continue;
242 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
243 midgard_nir_lower_fdot2_body(b
, alu
);
249 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
256 static const nir_variable
*
257 search_var(struct exec_list
*vars
, unsigned driver_loc
)
259 nir_foreach_variable(var
, vars
) {
260 if (var
->data
.driver_location
== driver_loc
)
267 /* Midgard can write all of color, depth and stencil in a single writeout
268 * operation, so we merge depth/stencil stores with color stores.
269 * If there are no color stores, we add a write to the "depth RT".
272 midgard_nir_lower_zs_store(nir_shader
*nir
)
274 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
277 nir_variable
*z_var
= NULL
, *s_var
= NULL
;
279 nir_foreach_variable(var
, &nir
->outputs
) {
280 if (var
->data
.location
== FRAG_RESULT_DEPTH
)
282 else if (var
->data
.location
== FRAG_RESULT_STENCIL
)
286 if (!z_var
&& !s_var
)
289 bool progress
= false;
291 nir_foreach_function(function
, nir
) {
292 if (!function
->impl
) continue;
294 nir_intrinsic_instr
*z_store
= NULL
, *s_store
= NULL
;
296 nir_foreach_block(block
, function
->impl
) {
297 nir_foreach_instr_safe(instr
, block
) {
298 if (instr
->type
!= nir_instr_type_intrinsic
)
301 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
302 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
305 if (z_var
&& nir_intrinsic_base(intr
) == z_var
->data
.driver_location
) {
310 if (s_var
&& nir_intrinsic_base(intr
) == s_var
->data
.driver_location
) {
317 if (!z_store
&& !s_store
) continue;
319 bool replaced
= false;
321 nir_foreach_block(block
, function
->impl
) {
322 nir_foreach_instr_safe(instr
, block
) {
323 if (instr
->type
!= nir_instr_type_intrinsic
)
326 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
327 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
330 const nir_variable
*var
= search_var(&nir
->outputs
, nir_intrinsic_base(intr
));
333 if (var
->data
.location
!= FRAG_RESULT_COLOR
&&
334 var
->data
.location
< FRAG_RESULT_DATA0
)
337 assert(nir_src_is_const(intr
->src
[1]) && "no indirect outputs");
340 nir_builder_init(&b
, function
->impl
);
342 assert(!z_store
|| z_store
->instr
.block
== instr
->block
);
343 assert(!s_store
|| s_store
->instr
.block
== instr
->block
);
344 b
.cursor
= nir_after_block_before_jump(instr
->block
);
346 nir_intrinsic_instr
*combined_store
;
347 combined_store
= nir_intrinsic_instr_create(b
.shader
, nir_intrinsic_store_combined_output_pan
);
349 combined_store
->num_components
= intr
->src
[0].ssa
->num_components
;
351 nir_intrinsic_set_base(combined_store
, nir_intrinsic_base(intr
));
353 unsigned writeout
= PAN_WRITEOUT_C
;
355 writeout
|= PAN_WRITEOUT_Z
;
357 writeout
|= PAN_WRITEOUT_S
;
359 nir_intrinsic_set_component(combined_store
, writeout
);
361 struct nir_ssa_def
*zero
= nir_imm_int(&b
, 0);
363 struct nir_ssa_def
*src
[4] = {
366 z_store
? z_store
->src
[0].ssa
: zero
,
367 s_store
? s_store
->src
[0].ssa
: zero
,
370 for (int i
= 0; i
< 4; ++i
)
371 combined_store
->src
[i
] = nir_src_for_ssa(src
[i
]);
373 nir_builder_instr_insert(&b
, &combined_store
->instr
);
375 nir_instr_remove(instr
);
381 /* Insert a store to the depth RT (0xff) if needed */
384 nir_builder_init(&b
, function
->impl
);
386 nir_block
*block
= NULL
;
387 if (z_store
&& s_store
)
388 assert(z_store
->instr
.block
== s_store
->instr
.block
);
391 block
= z_store
->instr
.block
;
393 block
= s_store
->instr
.block
;
395 b
.cursor
= nir_after_block_before_jump(block
);
397 nir_intrinsic_instr
*combined_store
;
398 combined_store
= nir_intrinsic_instr_create(b
.shader
, nir_intrinsic_store_combined_output_pan
);
400 combined_store
->num_components
= 4;
402 nir_intrinsic_set_base(combined_store
, 0);
404 unsigned writeout
= 0;
406 writeout
|= PAN_WRITEOUT_Z
;
408 writeout
|= PAN_WRITEOUT_S
;
410 nir_intrinsic_set_component(combined_store
, writeout
);
412 struct nir_ssa_def
*zero
= nir_imm_int(&b
, 0);
414 struct nir_ssa_def
*src
[4] = {
415 nir_imm_vec4(&b
, 0, 0, 0, 0),
417 z_store
? z_store
->src
[0].ssa
: zero
,
418 s_store
? s_store
->src
[0].ssa
: zero
,
421 for (int i
= 0; i
< 4; ++i
)
422 combined_store
->src
[i
] = nir_src_for_ssa(src
[i
]);
424 nir_builder_instr_insert(&b
, &combined_store
->instr
);
428 nir_instr_remove(&z_store
->instr
);
431 nir_instr_remove(&s_store
->instr
);
433 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
440 /* Flushes undefined values to zero */
443 optimise_nir(nir_shader
*nir
, unsigned quirks
, bool is_blend
)
446 unsigned lower_flrp
=
447 (nir
->options
->lower_flrp16
? 16 : 0) |
448 (nir
->options
->lower_flrp32
? 32 : 0) |
449 (nir
->options
->lower_flrp64
? 64 : 0);
451 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
452 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
454 nir_lower_tex_options lower_tex_options
= {
455 .lower_txs_lod
= true,
457 .lower_tex_without_implicit_lod
=
458 (quirks
& MIDGARD_EXPLICIT_LOD
),
460 /* TODO: we have native gradient.. */
464 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
466 /* Must lower fdot2 after tex is lowered */
467 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
469 /* T720 is broken. */
471 if (quirks
& MIDGARD_BROKEN_LOD
)
472 NIR_PASS_V(nir
, midgard_nir_lod_errata
);
474 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_early
);
479 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
480 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
482 NIR_PASS(progress
, nir
, nir_copy_prop
);
483 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
484 NIR_PASS(progress
, nir
, nir_opt_dce
);
485 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
486 NIR_PASS(progress
, nir
, nir_opt_cse
);
487 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
488 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
489 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
491 if (lower_flrp
!= 0) {
492 bool lower_flrp_progress
= false;
493 NIR_PASS(lower_flrp_progress
,
497 false /* always_precise */,
498 nir
->options
->lower_ffma
);
499 if (lower_flrp_progress
) {
500 NIR_PASS(progress
, nir
,
501 nir_opt_constant_folding
);
505 /* Nothing should rematerialize any flrps, so we only
506 * need to do this lowering once.
511 NIR_PASS(progress
, nir
, nir_opt_undef
);
512 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
514 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
517 nir_var_function_temp
);
519 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
522 /* Run after opts so it can hit more */
524 NIR_PASS(progress
, nir
, nir_fuse_io_16
);
526 /* Must be run at the end to prevent creation of fsin/fcos ops */
527 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
532 NIR_PASS(progress
, nir
, nir_opt_dce
);
533 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
534 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
535 NIR_PASS(progress
, nir
, nir_copy_prop
);
538 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
539 NIR_PASS(progress
, nir
, nir_opt_algebraic_distribute_src_mods
);
541 /* We implement booleans as 32-bit 0/~0 */
542 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
544 /* Now that booleans are lowered, we can run out late opts */
545 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
546 NIR_PASS(progress
, nir
, midgard_nir_cancel_inot
);
548 NIR_PASS(progress
, nir
, nir_copy_prop
);
549 NIR_PASS(progress
, nir
, nir_opt_dce
);
551 /* Take us out of SSA */
552 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
553 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
555 /* We are a vector architecture; write combine where possible */
556 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
557 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
559 NIR_PASS(progress
, nir
, nir_opt_dce
);
562 /* Do not actually emit a load; instead, cache the constant for inlining */
565 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
567 nir_ssa_def def
= instr
->def
;
569 midgard_constants
*consts
= rzalloc(NULL
, midgard_constants
);
571 assert(instr
->def
.num_components
* instr
->def
.bit_size
<= sizeof(*consts
) * 8);
573 #define RAW_CONST_COPY(bits) \
574 nir_const_value_to_array(consts->u##bits, instr->value, \
575 instr->def.num_components, u##bits)
577 switch (instr
->def
.bit_size
) {
591 unreachable("Invalid bit_size for load_const instruction\n");
594 /* Shifted for SSA, +1 for off-by-one */
595 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, consts
);
598 /* Normally constants are embedded implicitly, but for I/O and such we have to
599 * explicitly emit a move with the constant source */
602 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
604 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
606 if (constant_value
) {
607 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), to
);
608 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
609 emit_mir_instruction(ctx
, ins
);
614 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
616 unsigned comp
= src
->swizzle
[0];
618 for (unsigned c
= 1; c
< nr_components
; ++c
) {
619 if (src
->swizzle
[c
] != comp
)
626 #define ALU_CASE(nir, _op) \
628 op = midgard_alu_op_##_op; \
629 assert(src_bitsize == dst_bitsize); \
632 #define ALU_CASE_RTZ(nir, _op) \
634 op = midgard_alu_op_##_op; \
635 roundmode = MIDGARD_RTZ; \
638 #define ALU_CHECK_CMP(sext) \
639 assert(src_bitsize == 16 || src_bitsize == 32); \
640 assert(dst_bitsize == 16 || dst_bitsize == 32); \
642 #define ALU_CASE_BCAST(nir, _op, count) \
644 op = midgard_alu_op_##_op; \
645 broadcast_swizzle = count; \
646 ALU_CHECK_CMP(true); \
649 #define ALU_CASE_CMP(nir, _op, sext) \
651 op = midgard_alu_op_##_op; \
652 ALU_CHECK_CMP(sext); \
655 /* Analyze the sizes of the dest and inputs to determine reg mode. */
657 static midgard_reg_mode
658 reg_mode_for_nir(nir_alu_instr
*instr
)
660 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
661 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
662 unsigned max_bitsize
= MAX2(src_bitsize
, dst_bitsize
);
664 /* We don't have fp16 LUTs, so we'll want to emit code like:
666 * vlut.fsinr hr0, hr0
668 * where both input and output are 16-bit but the operation is carried
680 max_bitsize
= MAX2(max_bitsize
, 32);
683 /* These get lowered to moves */
684 case nir_op_pack_32_4x8
:
687 case nir_op_pack_32_2x16
:
695 switch (max_bitsize
) {
696 /* Use 16 pipe for 8 since we don't support vec16 yet */
699 return midgard_reg_mode_16
;
701 return midgard_reg_mode_32
;
703 return midgard_reg_mode_64
;
705 unreachable("Invalid bit size");
709 /* Compare mir_lower_invert */
711 nir_accepts_inot(nir_op op
, unsigned src
)
715 case nir_op_iand
: /* TODO: b2f16 */
719 /* Only the condition */
727 mir_accept_dest_mod(compiler_context
*ctx
, nir_dest
**dest
, nir_op op
)
729 if (pan_has_dest_mod(dest
, op
)) {
730 assert((*dest
)->is_ssa
);
731 BITSET_SET(ctx
->already_emitted
, (*dest
)->ssa
.index
);
739 mir_copy_src(midgard_instruction
*ins
, nir_alu_instr
*instr
, unsigned i
, unsigned to
, bool *abs
, bool *neg
, bool *not, enum midgard_roundmode
*roundmode
, bool is_int
, unsigned bcast_count
)
741 nir_alu_src src
= instr
->src
[i
];
744 if (pan_has_source_mod(&src
, nir_op_fneg
))
747 if (pan_has_source_mod(&src
, nir_op_fabs
))
751 if (nir_accepts_inot(instr
->op
, i
) && pan_has_source_mod(&src
, nir_op_inot
))
755 if (pan_has_source_mod(&src
, nir_op_fround_even
))
756 *roundmode
= MIDGARD_RTE
;
758 if (pan_has_source_mod(&src
, nir_op_ftrunc
))
759 *roundmode
= MIDGARD_RTZ
;
761 if (pan_has_source_mod(&src
, nir_op_ffloor
))
762 *roundmode
= MIDGARD_RTN
;
764 if (pan_has_source_mod(&src
, nir_op_fceil
))
765 *roundmode
= MIDGARD_RTP
;
768 unsigned bits
= nir_src_bit_size(src
.src
);
770 ins
->src
[to
] = nir_src_index(NULL
, &src
.src
);
771 ins
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
] | bits
;
773 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
774 ins
->swizzle
[to
][c
] = src
.swizzle
[
775 (!bcast_count
|| c
< bcast_count
) ? c
:
780 /* Midgard features both fcsel and icsel, depending on whether you want int or
781 * float modifiers. NIR's csel is typeless, so we want a heuristic to guess if
782 * we should emit an int or float csel depending on what modifiers could be
783 * placed. In the absense of modifiers, this is probably arbitrary. */
786 mir_is_bcsel_float(nir_alu_instr
*instr
)
789 nir_op_i2i8
, nir_op_i2i16
,
790 nir_op_i2i32
, nir_op_i2i64
793 nir_op floatmods
[] = {
794 nir_op_fabs
, nir_op_fneg
,
795 nir_op_f2f16
, nir_op_f2f32
,
799 nir_op floatdestmods
[] = {
800 nir_op_fsat
, nir_op_fsat_signed
, nir_op_fclamp_pos
,
801 nir_op_f2f16
, nir_op_f2f32
806 for (unsigned i
= 1; i
< 3; ++i
) {
807 nir_alu_src s
= instr
->src
[i
];
808 for (unsigned q
= 0; q
< ARRAY_SIZE(intmods
); ++q
) {
809 if (pan_has_source_mod(&s
, intmods
[q
]))
814 for (unsigned i
= 1; i
< 3; ++i
) {
815 nir_alu_src s
= instr
->src
[i
];
816 for (unsigned q
= 0; q
< ARRAY_SIZE(floatmods
); ++q
) {
817 if (pan_has_source_mod(&s
, floatmods
[q
]))
822 for (unsigned q
= 0; q
< ARRAY_SIZE(floatdestmods
); ++q
) {
823 nir_dest
*dest
= &instr
->dest
.dest
;
824 if (pan_has_dest_mod(&dest
, floatdestmods
[q
]))
832 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
834 nir_dest
*dest
= &instr
->dest
.dest
;
836 if (dest
->is_ssa
&& BITSET_TEST(ctx
->already_emitted
, dest
->ssa
.index
))
839 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
840 * is handled elsewhere */
842 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
843 midgard_emit_derivatives(ctx
, instr
);
847 bool is_ssa
= dest
->is_ssa
;
849 unsigned nr_components
= nir_dest_num_components(*dest
);
850 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
853 /* Number of components valid to check for the instruction (the rest
854 * will be forced to the last), or 0 to use as-is. Relevant as
855 * ball-type instructions have a channel count in NIR but are all vec4
858 unsigned broadcast_swizzle
= 0;
860 /* What register mode should we operate in? */
861 midgard_reg_mode reg_mode
=
862 reg_mode_for_nir(instr
);
864 /* Should we swap arguments? */
865 bool flip_src12
= false;
867 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
868 unsigned dst_bitsize
= nir_dest_bit_size(*dest
);
870 enum midgard_roundmode roundmode
= MIDGARD_RTE
;
873 ALU_CASE(fadd
, fadd
);
874 ALU_CASE(fmul
, fmul
);
875 ALU_CASE(fmin
, fmin
);
876 ALU_CASE(fmax
, fmax
);
877 ALU_CASE(imin
, imin
);
878 ALU_CASE(imax
, imax
);
879 ALU_CASE(umin
, umin
);
880 ALU_CASE(umax
, umax
);
881 ALU_CASE(ffloor
, ffloor
);
882 ALU_CASE(fround_even
, froundeven
);
883 ALU_CASE(ftrunc
, ftrunc
);
884 ALU_CASE(fceil
, fceil
);
885 ALU_CASE(fdot3
, fdot3
);
886 ALU_CASE(fdot4
, fdot4
);
887 ALU_CASE(iadd
, iadd
);
888 ALU_CASE(isub
, isub
);
889 ALU_CASE(imul
, imul
);
891 /* Zero shoved as second-arg */
892 ALU_CASE(iabs
, iabsdiff
);
896 ALU_CASE_CMP(feq32
, feq
, false);
897 ALU_CASE_CMP(fne32
, fne
, false);
898 ALU_CASE_CMP(flt32
, flt
, false);
899 ALU_CASE_CMP(ieq32
, ieq
, true);
900 ALU_CASE_CMP(ine32
, ine
, true);
901 ALU_CASE_CMP(ilt32
, ilt
, true);
902 ALU_CASE_CMP(ult32
, ult
, false);
904 /* We don't have a native b2f32 instruction. Instead, like many
905 * GPUs, we exploit booleans as 0/~0 for false/true, and
906 * correspondingly AND
907 * by 1.0 to do the type conversion. For the moment, prime us
910 * iand [whatever], #0
912 * At the end of emit_alu (as MIR), we'll fix-up the constant
915 ALU_CASE_CMP(b2f32
, iand
, true);
916 ALU_CASE_CMP(b2f16
, iand
, true);
917 ALU_CASE_CMP(b2i32
, iand
, true);
919 /* Likewise, we don't have a dedicated f2b32 instruction, but
920 * we can do a "not equal to 0.0" test. */
922 ALU_CASE_CMP(f2b32
, fne
, false);
923 ALU_CASE_CMP(i2b32
, ine
, true);
925 ALU_CASE(frcp
, frcp
);
926 ALU_CASE(frsq
, frsqrt
);
927 ALU_CASE(fsqrt
, fsqrt
);
928 ALU_CASE(fexp2
, fexp2
);
929 ALU_CASE(flog2
, flog2
);
931 ALU_CASE_RTZ(f2i64
, f2i_rte
);
932 ALU_CASE_RTZ(f2u64
, f2u_rte
);
933 ALU_CASE_RTZ(i2f64
, i2f_rte
);
934 ALU_CASE_RTZ(u2f64
, u2f_rte
);
936 ALU_CASE_RTZ(f2i32
, f2i_rte
);
937 ALU_CASE_RTZ(f2u32
, f2u_rte
);
938 ALU_CASE_RTZ(i2f32
, i2f_rte
);
939 ALU_CASE_RTZ(u2f32
, u2f_rte
);
941 ALU_CASE_RTZ(f2i8
, f2i_rte
);
942 ALU_CASE_RTZ(f2u8
, f2u_rte
);
944 ALU_CASE_RTZ(f2i16
, f2i_rte
);
945 ALU_CASE_RTZ(f2u16
, f2u_rte
);
946 ALU_CASE_RTZ(i2f16
, i2f_rte
);
947 ALU_CASE_RTZ(u2f16
, u2f_rte
);
949 ALU_CASE(fsin
, fsin
);
950 ALU_CASE(fcos
, fcos
);
952 /* We'll get 0 in the second arg, so:
953 * ~a = ~(a | 0) = nor(a, 0) */
954 ALU_CASE(inot
, inor
);
955 ALU_CASE(iand
, iand
);
957 ALU_CASE(ixor
, ixor
);
958 ALU_CASE(ishl
, ishl
);
959 ALU_CASE(ishr
, iasr
);
960 ALU_CASE(ushr
, ilsr
);
962 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
963 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
964 ALU_CASE_CMP(b32all_fequal4
, fball_eq
, true);
966 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
967 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
968 ALU_CASE_CMP(b32any_fnequal4
, fbany_neq
, true);
970 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
971 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
972 ALU_CASE_CMP(b32all_iequal4
, iball_eq
, true);
974 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
975 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
976 ALU_CASE_CMP(b32any_inequal4
, ibany_neq
, true);
978 /* Source mods will be shoved in later */
979 ALU_CASE(fabs
, fmov
);
980 ALU_CASE(fneg
, fmov
);
981 ALU_CASE(fsat
, fmov
);
982 ALU_CASE(fsat_signed
, fmov
);
983 ALU_CASE(fclamp_pos
, fmov
);
985 /* For size conversion, we use a move. Ideally though we would squash
986 * these ops together; maybe that has to happen after in NIR as part of
987 * propagation...? An earlier algebraic pass ensured we step down by
988 * only / exactly one size. If stepping down, we use a dest override to
989 * reduce the size; if stepping up, we use a larger-sized move with a
990 * half source and a sign/zero-extension modifier */
1002 case nir_op_f2f64
: {
1003 if (instr
->op
== nir_op_f2f16
|| instr
->op
== nir_op_f2f32
||
1004 instr
->op
== nir_op_f2f64
)
1005 op
= midgard_alu_op_fmov
;
1007 op
= midgard_alu_op_imov
;
1012 /* For greater-or-equal, we lower to less-or-equal and flip the
1018 case nir_op_uge32
: {
1020 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
1021 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
1022 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
1023 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
1027 ALU_CHECK_CMP(false);
1031 case nir_op_b32csel
: {
1032 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
1033 bool is_float
= mir_is_bcsel_float(instr
);
1035 (mixed
? midgard_alu_op_fcsel_v
: midgard_alu_op_fcsel
) :
1036 (mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
);
1041 case nir_op_unpack_32_2x16
:
1042 case nir_op_unpack_32_4x8
:
1043 case nir_op_pack_32_2x16
:
1044 case nir_op_pack_32_4x8
: {
1045 op
= midgard_alu_op_imov
;
1050 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
1055 /* Promote imov to fmov if it might help inline a constant */
1056 if (op
== midgard_alu_op_imov
&& nir_src_is_const(instr
->src
[0].src
)
1057 && nir_src_bit_size(instr
->src
[0].src
) == 32
1058 && nir_is_same_comp_swizzle(instr
->src
[0].swizzle
,
1059 nir_src_num_components(instr
->src
[0].src
))) {
1060 op
= midgard_alu_op_fmov
;
1063 /* Midgard can perform certain modifiers on output of an ALU op */
1065 unsigned outmod
= 0;
1066 bool is_int
= midgard_is_integer_op(op
);
1068 if (midgard_is_integer_out_op(op
)) {
1069 outmod
= midgard_outmod_int_wrap
;
1070 } else if (instr
->op
== nir_op_fsat
) {
1071 outmod
= midgard_outmod_sat
;
1072 } else if (instr
->op
== nir_op_fsat_signed
) {
1073 outmod
= midgard_outmod_sat_signed
;
1074 } else if (instr
->op
== nir_op_fclamp_pos
) {
1075 outmod
= midgard_outmod_pos
;
1078 /* Fetch unit, quirks, etc information */
1079 unsigned opcode_props
= alu_opcode_props
[op
].props
;
1080 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
1082 /* Look for floating point mods. We have the mods fsat, fsat_signed,
1083 * and fpos. We also have the relations (note 3 * 2 = 6 cases):
1085 * fsat_signed(fpos(x)) = fsat(x)
1086 * fsat_signed(fsat(x)) = fsat(x)
1087 * fpos(fsat_signed(x)) = fsat(x)
1088 * fpos(fsat(x)) = fsat(x)
1089 * fsat(fsat_signed(x)) = fsat(x)
1090 * fsat(fpos(x)) = fsat(x)
1092 * So by cases any composition of output modifiers is equivalent to
1096 if (!is_int
&& !(opcode_props
& OP_TYPE_CONVERT
)) {
1097 bool fpos
= mir_accept_dest_mod(ctx
, &dest
, nir_op_fclamp_pos
);
1098 bool fsat
= mir_accept_dest_mod(ctx
, &dest
, nir_op_fsat
);
1099 bool ssat
= mir_accept_dest_mod(ctx
, &dest
, nir_op_fsat_signed
);
1100 bool prior
= (outmod
!= midgard_outmod_none
);
1101 int count
= (int) prior
+ (int) fpos
+ (int) ssat
+ (int) fsat
;
1103 outmod
= ((count
> 1) || fsat
) ? midgard_outmod_sat
:
1104 fpos
? midgard_outmod_pos
:
1105 ssat
? midgard_outmod_sat_signed
:
1109 midgard_instruction ins
= {
1111 .dest
= nir_dest_index(dest
),
1112 .dest_type
= nir_op_infos
[instr
->op
].output_type
1113 | nir_dest_bit_size(*dest
),
1114 .roundmode
= roundmode
,
1117 enum midgard_roundmode
*roundptr
= (opcode_props
& MIDGARD_ROUNDS
) ?
1118 &ins
.roundmode
: NULL
;
1120 for (unsigned i
= nr_inputs
; i
< ARRAY_SIZE(ins
.src
); ++i
)
1123 if (quirk_flipped_r24
) {
1125 mir_copy_src(&ins
, instr
, 0, 1, &ins
.src_abs
[1], &ins
.src_neg
[1], &ins
.src_invert
[1], roundptr
, is_int
, broadcast_swizzle
);
1127 for (unsigned i
= 0; i
< nr_inputs
; ++i
) {
1130 if (instr
->op
== nir_op_b32csel
) {
1131 /* The condition is the first argument; move
1132 * the other arguments up one to be a binary
1133 * instruction for Midgard with the condition
1138 else if (flip_src12
)
1142 } else if (flip_src12
) {
1146 mir_copy_src(&ins
, instr
, i
, to
, &ins
.src_abs
[to
], &ins
.src_neg
[to
], &ins
.src_invert
[to
], roundptr
, is_int
, broadcast_swizzle
);
1148 /* (!c) ? a : b = c ? b : a */
1149 if (instr
->op
== nir_op_b32csel
&& ins
.src_invert
[2]) {
1150 ins
.src_invert
[2] = false;
1156 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
1157 /* Lowered to move */
1158 if (instr
->op
== nir_op_fneg
)
1159 ins
.src_neg
[1] ^= true;
1161 if (instr
->op
== nir_op_fabs
)
1162 ins
.src_abs
[1] = true;
1165 ins
.mask
= mask_of(nr_components
);
1167 midgard_vector_alu alu
= {
1169 .reg_mode
= reg_mode
,
1173 /* Apply writemask if non-SSA, keeping in mind that we can't write to
1174 * components that don't exist. Note modifier => SSA => !reg => no
1175 * writemask, so we don't have to worry about writemasks here.*/
1178 ins
.mask
&= instr
->dest
.write_mask
;
1182 /* Late fixup for emulated instructions */
1184 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1185 /* Presently, our second argument is an inline #0 constant.
1186 * Switch over to an embedded 1.0 constant (that can't fit
1187 * inline, since we're 32-bit, not 16-bit like the inline
1190 ins
.has_inline_constant
= false;
1191 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1192 ins
.src_types
[1] = nir_type_float32
;
1193 ins
.has_constants
= true;
1195 if (instr
->op
== nir_op_b2f32
)
1196 ins
.constants
.f32
[0] = 1.0f
;
1198 ins
.constants
.i32
[0] = 1;
1200 for (unsigned c
= 0; c
< 16; ++c
)
1201 ins
.swizzle
[1][c
] = 0;
1202 } else if (instr
->op
== nir_op_b2f16
) {
1203 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1204 ins
.src_types
[1] = nir_type_float16
;
1205 ins
.has_constants
= true;
1206 ins
.constants
.i16
[0] = _mesa_float_to_half(1.0);
1208 for (unsigned c
= 0; c
< 16; ++c
)
1209 ins
.swizzle
[1][c
] = 0;
1210 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1211 /* Lots of instructions need a 0 plonked in */
1212 ins
.has_inline_constant
= false;
1213 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1214 ins
.src_types
[1] = nir_type_uint32
;
1215 ins
.has_constants
= true;
1216 ins
.constants
.u32
[0] = 0;
1218 for (unsigned c
= 0; c
< 16; ++c
)
1219 ins
.swizzle
[1][c
] = 0;
1220 } else if (instr
->op
== nir_op_pack_32_2x16
) {
1221 ins
.dest_type
= nir_type_uint16
;
1222 ins
.mask
= mask_of(nr_components
* 2);
1224 } else if (instr
->op
== nir_op_pack_32_4x8
) {
1225 ins
.dest_type
= nir_type_uint8
;
1226 ins
.mask
= mask_of(nr_components
* 4);
1228 } else if (instr
->op
== nir_op_unpack_32_2x16
) {
1229 ins
.dest_type
= nir_type_uint32
;
1230 ins
.mask
= mask_of(nr_components
>> 1);
1232 } else if (instr
->op
== nir_op_unpack_32_4x8
) {
1233 ins
.dest_type
= nir_type_uint32
;
1234 ins
.mask
= mask_of(nr_components
>> 2);
1238 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1239 /* To avoid duplicating the lookup tables (probably), true LUT
1240 * instructions can only operate as if they were scalars. Lower
1241 * them here by changing the component. */
1243 unsigned orig_mask
= ins
.mask
;
1245 unsigned swizzle_back
[MIR_VEC_COMPONENTS
];
1246 memcpy(&swizzle_back
, ins
.swizzle
[0], sizeof(swizzle_back
));
1248 midgard_instruction ins_split
[MIR_VEC_COMPONENTS
];
1249 unsigned ins_count
= 0;
1251 for (int i
= 0; i
< nr_components
; ++i
) {
1252 /* Mask the associated component, dropping the
1253 * instruction if needed */
1256 ins
.mask
&= orig_mask
;
1258 for (unsigned j
= 0; j
< ins_count
; ++j
) {
1259 if (swizzle_back
[i
] == ins_split
[j
].swizzle
[0][0]) {
1260 ins_split
[j
].mask
|= ins
.mask
;
1269 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1270 ins
.swizzle
[0][j
] = swizzle_back
[i
]; /* Pull from the correct component */
1272 ins_split
[ins_count
] = ins
;
1277 for (unsigned i
= 0; i
< ins_count
; ++i
) {
1278 emit_mir_instruction(ctx
, ins_split
[i
]);
1281 emit_mir_instruction(ctx
, ins
);
1288 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1290 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1291 unsigned nir_mask
= 0;
1295 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1296 dsize
= nir_dest_bit_size(intr
->dest
);
1298 nir_mask
= nir_intrinsic_write_mask(intr
);
1302 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1303 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1304 mir_set_bytemask(ins
, bytemask
);
1305 ins
->dest_type
= nir_type_uint
| dsize
;
1308 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1309 * optimized) versions of UBO #0 */
1311 static midgard_instruction
*
1313 compiler_context
*ctx
,
1317 nir_src
*indirect_offset
,
1318 unsigned indirect_shift
,
1321 /* TODO: half-floats */
1323 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1324 ins
.constants
.u32
[0] = offset
;
1326 if (instr
->type
== nir_instr_type_intrinsic
)
1327 mir_set_intr_mask(instr
, &ins
, true);
1329 if (indirect_offset
) {
1330 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1331 ins
.src_types
[2] = nir_type_uint32
;
1332 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1334 /* X component for the whole swizzle to prevent register
1335 * pressure from ballooning from the extra components */
1336 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[2]); ++i
)
1337 ins
.swizzle
[2][i
] = 0;
1339 ins
.load_store
.arg_2
= 0x1E;
1342 ins
.load_store
.arg_1
= index
;
1344 return emit_mir_instruction(ctx
, ins
);
1347 /* Globals are like UBOs if you squint. And shared memory is like globals if
1348 * you squint even harder */
1352 compiler_context
*ctx
,
1361 midgard_instruction ins
;
1364 ins
= m_ld_int4(srcdest
, 0);
1366 ins
= m_st_int4(srcdest
, 0);
1368 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1369 mir_set_intr_mask(instr
, &ins
, is_read
);
1371 emit_mir_instruction(ctx
, ins
);
1376 compiler_context
*ctx
,
1377 unsigned dest
, unsigned offset
,
1378 unsigned nr_comp
, unsigned component
,
1379 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1381 /* XXX: Half-floats? */
1382 /* TODO: swizzle, mask */
1384 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1385 ins
.mask
= mask_of(nr_comp
);
1386 ins
.dest_type
= type
;
1388 if (type
== nir_type_float16
) {
1389 /* Ensure we are aligned so we can pack it later */
1390 ins
.mask
= mask_of(ALIGN_POT(nr_comp
, 2));
1393 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1394 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1396 midgard_varying_parameter p
= {
1398 .interpolation
= midgard_interp_default
,
1403 memcpy(&u
, &p
, sizeof(p
));
1404 ins
.load_store
.varying_parameters
= u
;
1406 if (indirect_offset
) {
1407 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1408 ins
.src_types
[2] = nir_type_uint32
;
1410 ins
.load_store
.arg_2
= 0x1E;
1412 ins
.load_store
.arg_1
= 0x9E;
1414 /* Use the type appropriate load */
1416 case nir_type_uint32
:
1417 case nir_type_bool32
:
1418 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1420 case nir_type_int32
:
1421 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1423 case nir_type_float32
:
1424 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1426 case nir_type_float16
:
1427 ins
.load_store
.op
= midgard_op_ld_vary_16
;
1430 unreachable("Attempted to load unknown type");
1434 emit_mir_instruction(ctx
, ins
);
1439 compiler_context
*ctx
,
1440 unsigned dest
, unsigned offset
,
1441 unsigned nr_comp
, nir_alu_type t
)
1443 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1444 ins
.load_store
.arg_1
= 0x1E;
1445 ins
.load_store
.arg_2
= 0x1E;
1446 ins
.mask
= mask_of(nr_comp
);
1448 /* Use the type appropriate load */
1452 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1455 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1457 case nir_type_float
:
1458 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1461 unreachable("Attempted to load unknown type");
1465 emit_mir_instruction(ctx
, ins
);
1469 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1470 unsigned nr_components
, unsigned offset
)
1474 /* Figure out which uniform this is */
1475 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1476 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1478 unsigned dest
= nir_dest_index(&nir_dest
);
1480 /* Sysvals are prefix uniforms */
1481 unsigned uniform
= ((uintptr_t) val
) - 1;
1483 /* Emit the read itself -- this is never indirect */
1484 midgard_instruction
*ins
=
1485 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1487 ins
->mask
= mask_of(nr_components
);
1491 compute_builtin_arg(nir_op op
)
1494 case nir_intrinsic_load_work_group_id
:
1496 case nir_intrinsic_load_local_invocation_id
:
1499 unreachable("Invalid compute paramater loaded");
1504 emit_fragment_store(compiler_context
*ctx
, unsigned src
, unsigned src_z
, unsigned src_s
, enum midgard_rt_id rt
)
1506 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1508 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1512 emit_explicit_constant(ctx
, src
, src
);
1514 struct midgard_instruction ins
=
1515 v_branch(false, false);
1517 bool depth_only
= (rt
== MIDGARD_ZS_RT
);
1519 ins
.writeout
= depth_only
? 0 : PAN_WRITEOUT_C
;
1521 /* Add dependencies */
1523 ins
.src_types
[0] = nir_type_uint32
;
1524 ins
.constants
.u32
[0] = depth_only
? 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1525 for (int i
= 0; i
< 4; ++i
)
1526 ins
.swizzle
[0][i
] = i
;
1529 emit_explicit_constant(ctx
, src_z
, src_z
);
1531 ins
.src_types
[2] = nir_type_uint32
;
1532 ins
.writeout
|= PAN_WRITEOUT_Z
;
1535 emit_explicit_constant(ctx
, src_s
, src_s
);
1537 ins
.src_types
[3] = nir_type_uint32
;
1538 ins
.writeout
|= PAN_WRITEOUT_S
;
1541 /* Emit the branch */
1542 br
= emit_mir_instruction(ctx
, ins
);
1543 schedule_barrier(ctx
);
1544 ctx
->writeout_branch
[rt
] = br
;
1546 /* Push our current location = current block count - 1 = where we'll
1547 * jump to. Maybe a bit too clever for my own good */
1549 br
->branch
.target_block
= ctx
->block_count
- 1;
1553 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1555 unsigned reg
= nir_dest_index(&instr
->dest
);
1556 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1557 ins
.mask
= mask_of(3);
1558 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1559 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1560 emit_mir_instruction(ctx
, ins
);
1564 vertex_builtin_arg(nir_op op
)
1567 case nir_intrinsic_load_vertex_id
:
1568 return PAN_VERTEX_ID
;
1569 case nir_intrinsic_load_instance_id
:
1570 return PAN_INSTANCE_ID
;
1572 unreachable("Invalid vertex builtin");
1577 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1579 unsigned reg
= nir_dest_index(&instr
->dest
);
1580 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1584 emit_control_barrier(compiler_context
*ctx
)
1586 midgard_instruction ins
= {
1587 .type
= TAG_TEXTURE_4
,
1589 .src
= { ~0, ~0, ~0, ~0 },
1591 .op
= TEXTURE_OP_BARRIER
,
1593 /* TODO: optimize */
1594 .out_of_order
= MIDGARD_BARRIER_BUFFER
|
1595 MIDGARD_BARRIER_SHARED
,
1599 emit_mir_instruction(ctx
, ins
);
1603 mir_get_branch_cond(nir_src
*src
, bool *invert
)
1605 /* Wrap it. No swizzle since it's a scalar */
1611 *invert
= pan_has_source_mod(&alu
, nir_op_inot
);
1612 return nir_src_index(NULL
, &alu
.src
);
1616 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1618 unsigned offset
= 0, reg
;
1620 switch (instr
->intrinsic
) {
1621 case nir_intrinsic_discard_if
:
1622 case nir_intrinsic_discard
: {
1623 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1624 struct midgard_instruction discard
= v_branch(conditional
, false);
1625 discard
.branch
.target_type
= TARGET_DISCARD
;
1628 discard
.src
[0] = mir_get_branch_cond(&instr
->src
[0],
1629 &discard
.branch
.invert_conditional
);
1630 discard
.src_types
[0] = nir_type_uint32
;
1633 emit_mir_instruction(ctx
, discard
);
1634 schedule_barrier(ctx
);
1639 case nir_intrinsic_load_uniform
:
1640 case nir_intrinsic_load_ubo
:
1641 case nir_intrinsic_load_global
:
1642 case nir_intrinsic_load_shared
:
1643 case nir_intrinsic_load_input
:
1644 case nir_intrinsic_load_interpolated_input
: {
1645 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1646 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1647 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1648 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1649 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1650 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1652 /* Get the base type of the intrinsic */
1653 /* TODO: Infer type? Does it matter? */
1655 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1656 (is_interp
) ? nir_type_float
:
1657 nir_intrinsic_type(instr
);
1659 t
= nir_alu_type_get_base_type(t
);
1661 if (!(is_ubo
|| is_global
)) {
1662 offset
= nir_intrinsic_base(instr
);
1665 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1667 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1669 bool direct
= nir_src_is_const(*src_offset
);
1670 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1673 offset
+= nir_src_as_uint(*src_offset
);
1675 /* We may need to apply a fractional offset */
1676 int component
= (is_flat
|| is_interp
) ?
1677 nir_intrinsic_component(instr
) : 0;
1678 reg
= nir_dest_index(&instr
->dest
);
1680 if (is_uniform
&& !ctx
->is_blend
) {
1681 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1682 } else if (is_ubo
) {
1683 nir_src index
= instr
->src
[0];
1685 /* TODO: Is indirect block number possible? */
1686 assert(nir_src_is_const(index
));
1688 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1689 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1690 } else if (is_global
|| is_shared
) {
1691 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1692 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1693 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
| nir_dest_bit_size(instr
->dest
), is_flat
);
1694 } else if (ctx
->is_blend
) {
1695 /* ctx->blend_input will be precoloured to r0, where
1696 * the input is preloaded */
1698 if (ctx
->blend_input
== ~0)
1699 ctx
->blend_input
= reg
;
1701 emit_mir_instruction(ctx
, v_mov(ctx
->blend_input
, reg
));
1702 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1703 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1705 DBG("Unknown load\n");
1712 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1713 case nir_intrinsic_load_barycentric_pixel
:
1714 case nir_intrinsic_load_barycentric_centroid
:
1717 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1719 case nir_intrinsic_load_raw_output_pan
: {
1720 reg
= nir_dest_index(&instr
->dest
);
1721 assert(ctx
->is_blend
);
1723 /* T720 and below use different blend opcodes with slightly
1724 * different semantics than T760 and up */
1726 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1728 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1729 ld
.load_store
.op
= midgard_op_ld_color_buffer_32u_old
;
1730 ld
.load_store
.address
= 16;
1731 ld
.load_store
.arg_2
= 0x1E;
1734 emit_mir_instruction(ctx
, ld
);
1738 case nir_intrinsic_load_output
: {
1739 reg
= nir_dest_index(&instr
->dest
);
1740 assert(ctx
->is_blend
);
1742 midgard_instruction ld
= m_ld_color_buffer_as_fp16(reg
, 0);
1744 for (unsigned c
= 4; c
< 16; ++c
)
1745 ld
.swizzle
[0][c
] = 0;
1747 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1748 ld
.load_store
.op
= midgard_op_ld_color_buffer_as_fp16_old
;
1749 ld
.load_store
.address
= 1;
1750 ld
.load_store
.arg_2
= 0x1E;
1753 emit_mir_instruction(ctx
, ld
);
1757 case nir_intrinsic_load_blend_const_color_rgba
: {
1758 assert(ctx
->is_blend
);
1759 reg
= nir_dest_index(&instr
->dest
);
1761 /* Blend constants are embedded directly in the shader and
1762 * patched in, so we use some magic routing */
1764 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1765 ins
.has_constants
= true;
1766 ins
.has_blend_constant
= true;
1767 emit_mir_instruction(ctx
, ins
);
1771 case nir_intrinsic_store_output
:
1772 case nir_intrinsic_store_combined_output_pan
:
1773 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1775 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1777 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1779 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1780 bool combined
= instr
->intrinsic
==
1781 nir_intrinsic_store_combined_output_pan
;
1783 const nir_variable
*var
;
1784 enum midgard_rt_id rt
;
1786 var
= search_var(&ctx
->nir
->outputs
,
1787 nir_intrinsic_base(instr
));
1789 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1790 rt
= MIDGARD_COLOR_RT0
;
1791 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1792 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1799 unsigned reg_z
= ~0, reg_s
= ~0;
1801 unsigned writeout
= nir_intrinsic_component(instr
);
1802 if (writeout
& PAN_WRITEOUT_Z
)
1803 reg_z
= nir_src_index(ctx
, &instr
->src
[2]);
1804 if (writeout
& PAN_WRITEOUT_S
)
1805 reg_s
= nir_src_index(ctx
, &instr
->src
[3]);
1808 emit_fragment_store(ctx
, reg
, reg_z
, reg_s
, rt
);
1809 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1810 assert(instr
->intrinsic
== nir_intrinsic_store_output
);
1812 /* We should have been vectorized, though we don't
1813 * currently check that st_vary is emitted only once
1814 * per slot (this is relevant, since there's not a mask
1815 * parameter available on the store [set to 0 by the
1816 * blob]). We do respect the component by adjusting the
1817 * swizzle. If this is a constant source, we'll need to
1818 * emit that explicitly. */
1820 emit_explicit_constant(ctx
, reg
, reg
);
1822 unsigned dst_component
= nir_intrinsic_component(instr
);
1823 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1825 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1826 st
.load_store
.arg_1
= 0x9E;
1827 st
.load_store
.arg_2
= 0x1E;
1829 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1832 st
.load_store
.op
= midgard_op_st_vary_32u
;
1835 st
.load_store
.op
= midgard_op_st_vary_32i
;
1837 case nir_type_float
:
1838 st
.load_store
.op
= midgard_op_st_vary_32
;
1841 unreachable("Attempted to store unknown type");
1845 /* nir_intrinsic_component(store_intr) encodes the
1846 * destination component start. Source component offset
1847 * adjustment is taken care of in
1848 * install_registers_instr(), when offset_swizzle() is
1851 unsigned src_component
= COMPONENT_X
;
1853 assert(nr_comp
> 0);
1854 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1855 st
.swizzle
[0][i
] = src_component
;
1856 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1860 emit_mir_instruction(ctx
, st
);
1862 DBG("Unknown store\n");
1868 /* Special case of store_output for lowered blend shaders */
1869 case nir_intrinsic_store_raw_output_pan
:
1870 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1871 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1872 emit_fragment_store(ctx
, reg
, ~0, ~0, ctx
->blend_rt
);
1875 case nir_intrinsic_store_global
:
1876 case nir_intrinsic_store_shared
:
1877 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1878 emit_explicit_constant(ctx
, reg
, reg
);
1880 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1883 case nir_intrinsic_load_ssbo_address
:
1884 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1887 case nir_intrinsic_get_buffer_size
:
1888 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1891 case nir_intrinsic_load_viewport_scale
:
1892 case nir_intrinsic_load_viewport_offset
:
1893 case nir_intrinsic_load_num_work_groups
:
1894 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1895 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1898 case nir_intrinsic_load_work_group_id
:
1899 case nir_intrinsic_load_local_invocation_id
:
1900 emit_compute_builtin(ctx
, instr
);
1903 case nir_intrinsic_load_vertex_id
:
1904 case nir_intrinsic_load_instance_id
:
1905 emit_vertex_builtin(ctx
, instr
);
1908 case nir_intrinsic_memory_barrier_buffer
:
1909 case nir_intrinsic_memory_barrier_shared
:
1912 case nir_intrinsic_control_barrier
:
1913 schedule_barrier(ctx
);
1914 emit_control_barrier(ctx
);
1915 schedule_barrier(ctx
);
1919 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1926 midgard_tex_format(enum glsl_sampler_dim dim
)
1929 case GLSL_SAMPLER_DIM_1D
:
1930 case GLSL_SAMPLER_DIM_BUF
:
1933 case GLSL_SAMPLER_DIM_2D
:
1934 case GLSL_SAMPLER_DIM_EXTERNAL
:
1935 case GLSL_SAMPLER_DIM_RECT
:
1938 case GLSL_SAMPLER_DIM_3D
:
1941 case GLSL_SAMPLER_DIM_CUBE
:
1942 return MALI_TEX_CUBE
;
1945 DBG("Unknown sampler dim type\n");
1951 /* Tries to attach an explicit LOD or bias as a constant. Returns whether this
1955 pan_attach_constant_bias(
1956 compiler_context
*ctx
,
1958 midgard_texture_word
*word
)
1960 /* To attach as constant, it has to *be* constant */
1962 if (!nir_src_is_const(lod
))
1965 float f
= nir_src_as_float(lod
);
1967 /* Break into fixed-point */
1969 float lod_frac
= f
- lod_int
;
1971 /* Carry over negative fractions */
1972 if (lod_frac
< 0.0) {
1978 word
->bias
= float_to_ubyte(lod_frac
);
1979 word
->bias_int
= lod_int
;
1985 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1986 unsigned midgard_texop
)
1989 //assert (!instr->sampler);
1991 int texture_index
= instr
->texture_index
;
1992 int sampler_index
= texture_index
;
1994 nir_alu_type dest_base
= nir_alu_type_get_base_type(instr
->dest_type
);
1995 nir_alu_type dest_type
= dest_base
| nir_dest_bit_size(instr
->dest
);
1997 midgard_instruction ins
= {
1998 .type
= TAG_TEXTURE_4
,
2000 .dest
= nir_dest_index(&instr
->dest
),
2001 .src
= { ~0, ~0, ~0, ~0 },
2002 .dest_type
= dest_type
,
2003 .swizzle
= SWIZZLE_IDENTITY_4
,
2005 .op
= midgard_texop
,
2006 .format
= midgard_tex_format(instr
->sampler_dim
),
2007 .texture_handle
= texture_index
,
2008 .sampler_handle
= sampler_index
,
2009 .shadow
= instr
->is_shadow
,
2013 if (instr
->is_shadow
&& !instr
->is_new_style_shadow
)
2014 for (int i
= 0; i
< 4; ++i
)
2015 ins
.swizzle
[0][i
] = COMPONENT_X
;
2017 /* We may need a temporary for the coordinate */
2019 bool needs_temp_coord
=
2020 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
2021 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
2024 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
2026 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
2027 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
2028 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
2029 unsigned sz
= nir_src_bit_size(instr
->src
[i
].src
);
2030 nir_alu_type T
= nir_tex_instr_src_type(instr
, i
) | sz
;
2032 switch (instr
->src
[i
].src_type
) {
2033 case nir_tex_src_coord
: {
2034 emit_explicit_constant(ctx
, index
, index
);
2036 unsigned coord_mask
= mask_of(instr
->coord_components
);
2038 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
2041 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
2043 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
2044 /* texelFetch is undefined on samplerCube */
2045 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
2047 /* For cubemaps, we use a special ld/st op to
2048 * select the face and copy the xy into the
2049 * texture register */
2051 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
2053 ld
.src_types
[1] = T
;
2054 ld
.mask
= 0x3; /* xy */
2055 ld
.load_store
.arg_1
= 0x20;
2056 ld
.swizzle
[1][3] = COMPONENT_X
;
2057 emit_mir_instruction(ctx
, ld
);
2060 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
2061 ins
.swizzle
[1][3] = COMPONENT_X
;
2062 } else if (needs_temp_coord
) {
2063 /* mov coord_temp, coords */
2064 midgard_instruction mov
= v_mov(index
, coords
);
2065 mov
.mask
= coord_mask
;
2068 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
2070 emit_mir_instruction(ctx
, mov
);
2075 ins
.src
[1] = coords
;
2076 ins
.src_types
[1] = T
;
2078 /* Texelfetch coordinates uses all four elements
2079 * (xyz/index) regardless of texture dimensionality,
2080 * which means it's necessary to zero the unused
2081 * components to keep everything happy */
2083 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
2084 /* mov index.zw, #0, or generalized */
2085 midgard_instruction mov
=
2086 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
2087 mov
.has_constants
= true;
2088 mov
.mask
= coord_mask
^ 0xF;
2089 emit_mir_instruction(ctx
, mov
);
2092 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
2093 /* Array component in w but NIR wants it in z,
2094 * but if we have a temp coord we already fixed
2097 if (nr_components
== 3) {
2098 ins
.swizzle
[1][2] = COMPONENT_Z
;
2099 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
2100 } else if (nr_components
== 2) {
2102 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
2103 ins
.swizzle
[1][3] = COMPONENT_X
;
2105 unreachable("Invalid texture 2D components");
2108 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
2110 ins
.swizzle
[1][2] = COMPONENT_Z
;
2111 ins
.swizzle
[1][3] = COMPONENT_W
;
2117 case nir_tex_src_bias
:
2118 case nir_tex_src_lod
: {
2119 /* Try as a constant if we can */
2121 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
2122 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
2125 ins
.texture
.lod_register
= true;
2127 ins
.src_types
[2] = T
;
2129 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
2130 ins
.swizzle
[2][c
] = COMPONENT_X
;
2132 emit_explicit_constant(ctx
, index
, index
);
2137 case nir_tex_src_offset
: {
2138 ins
.texture
.offset_register
= true;
2140 ins
.src_types
[3] = T
;
2142 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
2143 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
2145 emit_explicit_constant(ctx
, index
, index
);
2149 case nir_tex_src_comparator
: {
2150 unsigned comp
= COMPONENT_Z
;
2152 /* mov coord_temp.foo, coords */
2153 midgard_instruction mov
= v_mov(index
, coords
);
2154 mov
.mask
= 1 << comp
;
2156 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
2157 mov
.swizzle
[1][i
] = COMPONENT_X
;
2159 emit_mir_instruction(ctx
, mov
);
2164 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
2170 emit_mir_instruction(ctx
, ins
);
2174 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
2176 switch (instr
->op
) {
2179 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
2182 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
2185 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
2188 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
2191 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
2198 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
2200 switch (instr
->type
) {
2201 case nir_jump_break
: {
2202 /* Emit a branch out of the loop */
2203 struct midgard_instruction br
= v_branch(false, false);
2204 br
.branch
.target_type
= TARGET_BREAK
;
2205 br
.branch
.target_break
= ctx
->current_loop_depth
;
2206 emit_mir_instruction(ctx
, br
);
2211 DBG("Unknown jump type %d\n", instr
->type
);
2217 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
2219 switch (instr
->type
) {
2220 case nir_instr_type_load_const
:
2221 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2224 case nir_instr_type_intrinsic
:
2225 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2228 case nir_instr_type_alu
:
2229 emit_alu(ctx
, nir_instr_as_alu(instr
));
2232 case nir_instr_type_tex
:
2233 emit_tex(ctx
, nir_instr_as_tex(instr
));
2236 case nir_instr_type_jump
:
2237 emit_jump(ctx
, nir_instr_as_jump(instr
));
2240 case nir_instr_type_ssa_undef
:
2245 DBG("Unhandled instruction type\n");
2251 /* ALU instructions can inline or embed constants, which decreases register
2252 * pressure and saves space. */
2254 #define CONDITIONAL_ATTACH(idx) { \
2255 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
2258 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
2259 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2264 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
2266 mir_foreach_instr_in_block(block
, alu
) {
2267 /* Other instructions cannot inline constants */
2268 if (alu
->type
!= TAG_ALU_4
) continue;
2269 if (alu
->compact_branch
) continue;
2271 /* If there is already a constant here, we can do nothing */
2272 if (alu
->has_constants
) continue;
2274 CONDITIONAL_ATTACH(0);
2276 if (!alu
->has_constants
) {
2277 CONDITIONAL_ATTACH(1)
2278 } else if (!alu
->inline_constant
) {
2279 /* Corner case: _two_ vec4 constants, for instance with a
2280 * csel. For this case, we can only use a constant
2281 * register for one, we'll have to emit a move for the
2284 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2285 unsigned scratch
= make_compiler_temp(ctx
);
2288 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2289 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2291 /* Set the source */
2292 alu
->src
[1] = scratch
;
2294 /* Inject us -before- the last instruction which set r31 */
2295 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2301 /* Midgard supports two types of constants, embedded constants (128-bit) and
2302 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2303 * constants can be demoted to inline constants, for space savings and
2304 * sometimes a performance boost */
2307 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2309 mir_foreach_instr_in_block(block
, ins
) {
2310 if (!ins
->has_constants
) continue;
2311 if (ins
->has_inline_constant
) continue;
2313 /* Blend constants must not be inlined by definition */
2314 if (ins
->has_blend_constant
) continue;
2316 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2317 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2318 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2320 if (!(is_16
|| is_32
))
2323 /* src1 cannot be an inline constant due to encoding
2324 * restrictions. So, if possible we try to flip the arguments
2327 int op
= ins
->alu
.op
;
2329 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
) &&
2330 alu_opcode_props
[op
].props
& OP_COMMUTES
) {
2334 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2335 /* Component is from the swizzle. Take a nonzero component */
2337 unsigned first_comp
= ffs(ins
->mask
) - 1;
2338 unsigned component
= ins
->swizzle
[1][first_comp
];
2340 /* Scale constant appropriately, if we can legally */
2341 int16_t scaled_constant
= 0;
2344 scaled_constant
= ins
->constants
.u16
[component
];
2345 } else if (midgard_is_integer_op(op
)) {
2346 scaled_constant
= ins
->constants
.u32
[component
];
2348 /* Constant overflow after resize */
2349 if (scaled_constant
!= ins
->constants
.u32
[component
])
2352 float original
= ins
->constants
.f32
[component
];
2353 scaled_constant
= _mesa_float_to_half(original
);
2355 /* Check for loss of precision. If this is
2356 * mediump, we don't care, but for a highp
2357 * shader, we need to pay attention. NIR
2358 * doesn't yet tell us which mode we're in!
2359 * Practically this prevents most constants
2360 * from being inlined, sadly. */
2362 float fp32
= _mesa_half_to_float(scaled_constant
);
2364 if (fp32
!= original
)
2368 /* Should've been const folded */
2369 if (ins
->src_abs
[1] || ins
->src_neg
[1])
2372 /* Make sure that the constant is not itself a vector
2373 * by checking if all accessed values are the same. */
2375 const midgard_constants
*cons
= &ins
->constants
;
2376 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2378 bool is_vector
= false;
2379 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2381 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2382 /* We only care if this component is actually used */
2383 if (!(mask
& (1 << c
)))
2386 uint32_t test
= is_16
?
2387 cons
->u16
[ins
->swizzle
[1][c
]] :
2388 cons
->u32
[ins
->swizzle
[1][c
]];
2390 if (test
!= value
) {
2399 /* Get rid of the embedded constant */
2400 ins
->has_constants
= false;
2402 ins
->has_inline_constant
= true;
2403 ins
->inline_constant
= scaled_constant
;
2408 /* Dead code elimination for branches at the end of a block - only one branch
2409 * per block is legal semantically */
2412 midgard_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2414 bool branched
= false;
2416 mir_foreach_instr_in_block_safe(block
, ins
) {
2417 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2420 mir_remove_instruction(ins
);
2426 /* We want to force the invert on AND/OR to the second slot to legalize into
2427 * iandnot/iornot. The relevant patterns are for AND (and OR respectively)
2429 * ~a & #b = ~a & ~(#~b)
2434 midgard_legalize_invert(compiler_context
*ctx
, midgard_block
*block
)
2436 mir_foreach_instr_in_block(block
, ins
) {
2437 if (ins
->type
!= TAG_ALU_4
) continue;
2439 if (ins
->alu
.op
!= midgard_alu_op_iand
&&
2440 ins
->alu
.op
!= midgard_alu_op_ior
) continue;
2442 if (ins
->src_invert
[1] || !ins
->src_invert
[0]) continue;
2444 if (ins
->has_inline_constant
) {
2445 /* ~(#~a) = ~(~#a) = a, so valid, and forces both
2447 ins
->inline_constant
= ~ins
->inline_constant
;
2448 ins
->src_invert
[1] = true;
2450 /* Flip to the right invert order. Note
2451 * has_inline_constant false by assumption on the
2452 * branch, so flipping makes sense. */
2459 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2461 /* Loop to ourselves */
2462 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2463 struct midgard_instruction ins
= v_branch(false, false);
2464 ins
.writeout
= br
->writeout
;
2465 ins
.branch
.target_block
= ctx
->block_count
- 1;
2466 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2467 memcpy(&ins
.src_types
, &br
->src_types
, sizeof(ins
.src_types
));
2468 emit_mir_instruction(ctx
, ins
);
2470 ctx
->current_block
->epilogue
= true;
2471 schedule_barrier(ctx
);
2472 return ins
.branch
.target_block
;
2475 static midgard_block
*
2476 emit_block(compiler_context
*ctx
, nir_block
*block
)
2478 midgard_block
*this_block
= ctx
->after_block
;
2479 ctx
->after_block
= NULL
;
2482 this_block
= create_empty_block(ctx
);
2484 list_addtail(&this_block
->base
.link
, &ctx
->blocks
);
2486 this_block
->scheduled
= false;
2489 /* Set up current block */
2490 list_inithead(&this_block
->base
.instructions
);
2491 ctx
->current_block
= this_block
;
2493 nir_foreach_instr(instr
, block
) {
2494 emit_instr(ctx
, instr
);
2495 ++ctx
->instruction_count
;
2501 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2504 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2506 midgard_block
*before_block
= ctx
->current_block
;
2508 /* Speculatively emit the branch, but we can't fill it in until later */
2510 EMIT(branch
, true, true);
2511 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2512 then_branch
->src
[0] = mir_get_branch_cond(&nif
->condition
, &inv
);
2513 then_branch
->src_types
[0] = nir_type_uint32
;
2514 then_branch
->branch
.invert_conditional
= !inv
;
2516 /* Emit the two subblocks. */
2517 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2518 midgard_block
*end_then_block
= ctx
->current_block
;
2520 /* Emit a jump from the end of the then block to the end of the else */
2521 EMIT(branch
, false, false);
2522 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2524 /* Emit second block, and check if it's empty */
2526 int else_idx
= ctx
->block_count
;
2527 int count_in
= ctx
->instruction_count
;
2528 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2529 midgard_block
*end_else_block
= ctx
->current_block
;
2530 int after_else_idx
= ctx
->block_count
;
2532 /* Now that we have the subblocks emitted, fix up the branches */
2537 if (ctx
->instruction_count
== count_in
) {
2538 /* The else block is empty, so don't emit an exit jump */
2539 mir_remove_instruction(then_exit
);
2540 then_branch
->branch
.target_block
= after_else_idx
;
2542 then_branch
->branch
.target_block
= else_idx
;
2543 then_exit
->branch
.target_block
= after_else_idx
;
2546 /* Wire up the successors */
2548 ctx
->after_block
= create_empty_block(ctx
);
2550 pan_block_add_successor(&before_block
->base
, &then_block
->base
);
2551 pan_block_add_successor(&before_block
->base
, &else_block
->base
);
2553 pan_block_add_successor(&end_then_block
->base
, &ctx
->after_block
->base
);
2554 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
);
2558 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2560 /* Remember where we are */
2561 midgard_block
*start_block
= ctx
->current_block
;
2563 /* Allocate a loop number, growing the current inner loop depth */
2564 int loop_idx
= ++ctx
->current_loop_depth
;
2566 /* Get index from before the body so we can loop back later */
2567 int start_idx
= ctx
->block_count
;
2569 /* Emit the body itself */
2570 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2572 /* Branch back to loop back */
2573 struct midgard_instruction br_back
= v_branch(false, false);
2574 br_back
.branch
.target_block
= start_idx
;
2575 emit_mir_instruction(ctx
, br_back
);
2577 /* Mark down that branch in the graph. */
2578 pan_block_add_successor(&start_block
->base
, &loop_block
->base
);
2579 pan_block_add_successor(&ctx
->current_block
->base
, &loop_block
->base
);
2581 /* Find the index of the block about to follow us (note: we don't add
2582 * one; blocks are 0-indexed so we get a fencepost problem) */
2583 int break_block_idx
= ctx
->block_count
;
2585 /* Fix up the break statements we emitted to point to the right place,
2586 * now that we can allocate a block number for them */
2587 ctx
->after_block
= create_empty_block(ctx
);
2589 mir_foreach_block_from(ctx
, start_block
, _block
) {
2590 mir_foreach_instr_in_block(((midgard_block
*) _block
), ins
) {
2591 if (ins
->type
!= TAG_ALU_4
) continue;
2592 if (!ins
->compact_branch
) continue;
2594 /* We found a branch -- check the type to see if we need to do anything */
2595 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2597 /* It's a break! Check if it's our break */
2598 if (ins
->branch
.target_break
!= loop_idx
) continue;
2600 /* Okay, cool, we're breaking out of this loop.
2601 * Rewrite from a break to a goto */
2603 ins
->branch
.target_type
= TARGET_GOTO
;
2604 ins
->branch
.target_block
= break_block_idx
;
2606 pan_block_add_successor(_block
, &ctx
->after_block
->base
);
2610 /* Now that we've finished emitting the loop, free up the depth again
2611 * so we play nice with recursion amid nested loops */
2612 --ctx
->current_loop_depth
;
2614 /* Dump loop stats */
2618 static midgard_block
*
2619 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2621 midgard_block
*start_block
= NULL
;
2623 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2624 switch (node
->type
) {
2625 case nir_cf_node_block
: {
2626 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2629 start_block
= block
;
2634 case nir_cf_node_if
:
2635 emit_if(ctx
, nir_cf_node_as_if(node
));
2638 case nir_cf_node_loop
:
2639 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2642 case nir_cf_node_function
:
2651 /* Due to lookahead, we need to report the first tag executed in the command
2652 * stream and in branch targets. An initial block might be empty, so iterate
2653 * until we find one that 'works' */
2656 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2658 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2660 mir_foreach_block_from(ctx
, initial_block
, _v
) {
2661 midgard_block
*v
= (midgard_block
*) _v
;
2662 if (v
->quadword_count
) {
2663 midgard_bundle
*initial_bundle
=
2664 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2666 return initial_bundle
->tag
;
2670 /* Default to a tag 1 which will break from the shader, in case we jump
2671 * to the exit block (i.e. `return` in a compute shader) */
2676 /* For each fragment writeout instruction, generate a writeout loop to
2677 * associate with it */
2680 mir_add_writeout_loops(compiler_context
*ctx
)
2682 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2683 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2686 unsigned popped
= br
->branch
.target_block
;
2687 pan_block_add_successor(&(mir_get_block(ctx
, popped
- 1)->base
), &ctx
->current_block
->base
);
2688 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2689 br
->branch
.target_type
= TARGET_GOTO
;
2691 /* If we have more RTs, we'll need to restore back after our
2692 * loop terminates */
2694 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2695 midgard_instruction uncond
= v_branch(false, false);
2696 uncond
.branch
.target_block
= popped
;
2697 uncond
.branch
.target_type
= TARGET_GOTO
;
2698 emit_mir_instruction(ctx
, uncond
);
2699 pan_block_add_successor(&ctx
->current_block
->base
, &(mir_get_block(ctx
, popped
)->base
));
2700 schedule_barrier(ctx
);
2702 /* We're last, so we can terminate here */
2703 br
->last_writeout
= true;
2709 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2711 struct util_dynarray
*compiled
= &program
->compiled
;
2713 midgard_debug
= debug_get_option_midgard_debug();
2715 /* TODO: Bound against what? */
2716 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2719 ctx
->stage
= nir
->info
.stage
;
2720 ctx
->is_blend
= is_blend
;
2721 ctx
->alpha_ref
= program
->alpha_ref
;
2722 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2723 ctx
->blend_input
= ~0;
2724 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2726 /* Start off with a safe cutoff, allowing usage of all 16 work
2727 * registers. Later, we'll promote uniform reads to uniform registers
2728 * if we determine it is beneficial to do so */
2729 ctx
->uniform_cutoff
= 8;
2731 /* Initialize at a global (not block) level hash tables */
2733 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2734 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2736 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2737 * (so we don't accidentally duplicate the epilogue since mesa/st has
2738 * messed with our I/O quite a bit already) */
2740 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2742 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2743 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2744 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2747 NIR_PASS_V(nir
, nir_lower_var_copies
);
2748 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2749 NIR_PASS_V(nir
, nir_split_var_copies
);
2750 NIR_PASS_V(nir
, nir_lower_var_copies
);
2751 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2752 NIR_PASS_V(nir
, nir_lower_var_copies
);
2753 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2755 NIR_PASS_V(nir
, nir_lower_io
, nir_var_shader_in
| nir_var_shader_out
,
2757 NIR_PASS_V(nir
, nir_lower_ssbo
);
2758 NIR_PASS_V(nir
, midgard_nir_lower_zs_store
);
2760 /* Optimisation passes */
2762 optimise_nir(nir
, ctx
->quirks
, is_blend
);
2764 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2765 nir_print_shader(nir
, stdout
);
2768 /* Assign sysvals and counts, now that we're sure
2769 * (post-optimisation) */
2771 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2772 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2773 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2775 nir_foreach_function(func
, nir
) {
2779 list_inithead(&ctx
->blocks
);
2780 ctx
->block_count
= 0;
2782 ctx
->already_emitted
= calloc(BITSET_WORDS(func
->impl
->ssa_alloc
), sizeof(BITSET_WORD
));
2784 emit_cf_list(ctx
, &func
->impl
->body
);
2785 free(ctx
->already_emitted
);
2786 break; /* TODO: Multi-function shaders */
2789 util_dynarray_init(compiled
, NULL
);
2791 /* Per-block lowering before opts */
2793 mir_foreach_block(ctx
, _block
) {
2794 midgard_block
*block
= (midgard_block
*) _block
;
2795 inline_alu_constants(ctx
, block
);
2796 embedded_to_inline_constant(ctx
, block
);
2798 /* MIR-level optimizations */
2800 bool progress
= false;
2804 progress
|= midgard_opt_dead_code_eliminate(ctx
);
2806 mir_foreach_block(ctx
, _block
) {
2807 midgard_block
*block
= (midgard_block
*) _block
;
2808 progress
|= midgard_opt_copy_prop(ctx
, block
);
2809 progress
|= midgard_opt_combine_projection(ctx
, block
);
2810 progress
|= midgard_opt_varying_projection(ctx
, block
);
2814 mir_foreach_block(ctx
, _block
) {
2815 midgard_block
*block
= (midgard_block
*) _block
;
2816 midgard_lower_derivatives(ctx
, block
);
2817 midgard_legalize_invert(ctx
, block
);
2818 midgard_cull_dead_branch(ctx
, block
);
2821 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2822 mir_add_writeout_loops(ctx
);
2824 /* Analyze now that the code is known but before scheduling creates
2825 * pipeline registers which are harder to track */
2826 mir_analyze_helper_terminate(ctx
);
2827 mir_analyze_helper_requirements(ctx
);
2830 midgard_schedule_program(ctx
);
2833 /* Now that all the bundles are scheduled and we can calculate block
2834 * sizes, emit actual branch instructions rather than placeholders */
2836 int br_block_idx
= 0;
2838 mir_foreach_block(ctx
, _block
) {
2839 midgard_block
*block
= (midgard_block
*) _block
;
2840 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2841 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2842 midgard_instruction
*ins
= bundle
->instructions
[c
];
2844 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2846 /* Parse some basic branch info */
2847 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2848 bool is_conditional
= ins
->branch
.conditional
;
2849 bool is_inverted
= ins
->branch
.invert_conditional
;
2850 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2851 bool is_writeout
= ins
->writeout
;
2853 /* Determine the block we're jumping to */
2854 int target_number
= ins
->branch
.target_block
;
2856 /* Report the destination tag */
2857 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2859 /* Count up the number of quadwords we're
2860 * jumping over = number of quadwords until
2861 * (br_block_idx, target_number) */
2863 int quadword_offset
= 0;
2867 } else if (target_number
> br_block_idx
) {
2870 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2871 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2874 quadword_offset
+= blk
->quadword_count
;
2877 /* Jump backwards */
2879 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2880 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2883 quadword_offset
-= blk
->quadword_count
;
2887 /* Unconditional extended branches (far jumps)
2888 * have issues, so we always use a conditional
2889 * branch, setting the condition to always for
2890 * unconditional. For compact unconditional
2891 * branches, cond isn't used so it doesn't
2892 * matter what we pick. */
2894 midgard_condition cond
=
2895 !is_conditional
? midgard_condition_always
:
2896 is_inverted
? midgard_condition_false
:
2897 midgard_condition_true
;
2899 midgard_jmp_writeout_op op
=
2900 is_discard
? midgard_jmp_writeout_op_discard
:
2901 is_writeout
? midgard_jmp_writeout_op_writeout
:
2902 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2903 midgard_jmp_writeout_op_branch_cond
;
2906 midgard_branch_extended branch
=
2907 midgard_create_branch_extended(
2912 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2913 } else if (is_conditional
|| is_discard
) {
2914 midgard_branch_cond branch
= {
2916 .dest_tag
= dest_tag
,
2917 .offset
= quadword_offset
,
2921 assert(branch
.offset
== quadword_offset
);
2923 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2925 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2927 midgard_branch_uncond branch
= {
2929 .dest_tag
= dest_tag
,
2930 .offset
= quadword_offset
,
2934 assert(branch
.offset
== quadword_offset
);
2936 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2944 /* Emit flat binary from the instruction arrays. Iterate each block in
2945 * sequence. Save instruction boundaries such that lookahead tags can
2946 * be assigned easily */
2948 /* Cache _all_ bundles in source order for lookahead across failed branches */
2950 int bundle_count
= 0;
2951 mir_foreach_block(ctx
, _block
) {
2952 midgard_block
*block
= (midgard_block
*) _block
;
2953 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2955 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2957 mir_foreach_block(ctx
, _block
) {
2958 midgard_block
*block
= (midgard_block
*) _block
;
2959 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2960 source_order_bundles
[bundle_idx
++] = bundle
;
2964 int current_bundle
= 0;
2966 /* Midgard prefetches instruction types, so during emission we
2967 * need to lookahead. Unless this is the last instruction, in
2968 * which we return 1. */
2970 mir_foreach_block(ctx
, _block
) {
2971 midgard_block
*block
= (midgard_block
*) _block
;
2972 mir_foreach_bundle_in_block(block
, bundle
) {
2975 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2976 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2978 emit_binary_bundle(ctx
, block
, bundle
, compiled
, lookahead
);
2982 /* TODO: Free deeper */
2983 //util_dynarray_fini(&block->instructions);
2986 free(source_order_bundles
);
2988 /* Report the very first tag executed */
2989 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2991 /* Deal with off-by-one related to the fencepost problem */
2992 program
->work_register_count
= ctx
->work_registers
+ 1;
2993 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2995 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2996 program
->tls_size
= ctx
->tls_size
;
2998 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2999 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
3001 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
3002 unsigned nr_bundles
= 0, nr_ins
= 0;
3004 /* Count instructions and bundles */
3006 mir_foreach_block(ctx
, _block
) {
3007 midgard_block
*block
= (midgard_block
*) _block
;
3008 nr_bundles
+= util_dynarray_num_elements(
3009 &block
->bundles
, midgard_bundle
);
3011 mir_foreach_bundle_in_block(block
, bun
)
3012 nr_ins
+= bun
->instruction_count
;
3015 /* Calculate thread count. There are certain cutoffs by
3016 * register count for thread count */
3018 unsigned nr_registers
= program
->work_register_count
;
3020 unsigned nr_threads
=
3021 (nr_registers
<= 4) ? 4 :
3022 (nr_registers
<= 8) ? 2 :
3027 fprintf(stderr
, "shader%d - %s shader: "
3028 "%u inst, %u bundles, %u quadwords, "
3029 "%u registers, %u threads, %u loops, "
3030 "%u:%u spills:fills\n",
3032 ctx
->is_blend
? "PAN_SHADER_BLEND" :
3033 gl_shader_stage_name(ctx
->stage
),
3034 nr_ins
, nr_bundles
, ctx
->quadword_count
,
3035 nr_registers
, nr_threads
,
3037 ctx
->spills
, ctx
->fills
);