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 /* Flushes undefined values to zero */
259 optimise_nir(nir_shader
*nir
, unsigned quirks
, bool is_blend
)
262 unsigned lower_flrp
=
263 (nir
->options
->lower_flrp16
? 16 : 0) |
264 (nir
->options
->lower_flrp32
? 32 : 0) |
265 (nir
->options
->lower_flrp64
? 64 : 0);
267 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
268 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
270 nir_lower_tex_options lower_tex_options
= {
271 .lower_txs_lod
= true,
273 .lower_tex_without_implicit_lod
=
274 (quirks
& MIDGARD_EXPLICIT_LOD
),
276 /* TODO: we have native gradient.. */
280 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
282 /* Must lower fdot2 after tex is lowered */
283 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
285 /* T720 is broken. */
287 if (quirks
& MIDGARD_BROKEN_LOD
)
288 NIR_PASS_V(nir
, midgard_nir_lod_errata
);
290 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_early
);
293 NIR_PASS(progress
, nir
, nir_fuse_io_16
);
298 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
299 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
301 NIR_PASS(progress
, nir
, nir_copy_prop
);
302 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
303 NIR_PASS(progress
, nir
, nir_opt_dce
);
304 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
305 NIR_PASS(progress
, nir
, nir_opt_cse
);
306 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
307 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
308 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
310 if (lower_flrp
!= 0) {
311 bool lower_flrp_progress
= false;
312 NIR_PASS(lower_flrp_progress
,
316 false /* always_precise */,
317 nir
->options
->lower_ffma
);
318 if (lower_flrp_progress
) {
319 NIR_PASS(progress
, nir
,
320 nir_opt_constant_folding
);
324 /* Nothing should rematerialize any flrps, so we only
325 * need to do this lowering once.
330 NIR_PASS(progress
, nir
, nir_opt_undef
);
331 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
333 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
336 nir_var_function_temp
);
338 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
341 /* Must be run at the end to prevent creation of fsin/fcos ops */
342 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
347 NIR_PASS(progress
, nir
, nir_opt_dce
);
348 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
349 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
350 NIR_PASS(progress
, nir
, nir_copy_prop
);
353 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
354 NIR_PASS(progress
, nir
, nir_opt_algebraic_distribute_src_mods
);
356 /* We implement booleans as 32-bit 0/~0 */
357 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
359 /* Now that booleans are lowered, we can run out late opts */
360 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
361 NIR_PASS(progress
, nir
, midgard_nir_cancel_inot
);
363 NIR_PASS(progress
, nir
, nir_copy_prop
);
364 NIR_PASS(progress
, nir
, nir_opt_dce
);
366 /* Take us out of SSA */
367 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
368 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
370 /* We are a vector architecture; write combine where possible */
371 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
372 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
374 NIR_PASS(progress
, nir
, nir_opt_dce
);
377 /* Do not actually emit a load; instead, cache the constant for inlining */
380 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
382 nir_ssa_def def
= instr
->def
;
384 midgard_constants
*consts
= rzalloc(NULL
, midgard_constants
);
386 assert(instr
->def
.num_components
* instr
->def
.bit_size
<= sizeof(*consts
) * 8);
388 #define RAW_CONST_COPY(bits) \
389 nir_const_value_to_array(consts->u##bits, instr->value, \
390 instr->def.num_components, u##bits)
392 switch (instr
->def
.bit_size
) {
406 unreachable("Invalid bit_size for load_const instruction\n");
409 /* Shifted for SSA, +1 for off-by-one */
410 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, consts
);
413 /* Normally constants are embedded implicitly, but for I/O and such we have to
414 * explicitly emit a move with the constant source */
417 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
419 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
421 if (constant_value
) {
422 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), to
);
423 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
424 emit_mir_instruction(ctx
, ins
);
429 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
431 unsigned comp
= src
->swizzle
[0];
433 for (unsigned c
= 1; c
< nr_components
; ++c
) {
434 if (src
->swizzle
[c
] != comp
)
441 #define ALU_CASE(nir, _op) \
443 op = midgard_alu_op_##_op; \
444 assert(src_bitsize == dst_bitsize); \
447 #define ALU_CASE_RTZ(nir, _op) \
449 op = midgard_alu_op_##_op; \
450 roundmode = MIDGARD_RTZ; \
453 #define ALU_CHECK_CMP(sext) \
454 assert(src_bitsize == 16 || src_bitsize == 32); \
455 assert(dst_bitsize == 16 || dst_bitsize == 32); \
457 #define ALU_CASE_BCAST(nir, _op, count) \
459 op = midgard_alu_op_##_op; \
460 broadcast_swizzle = count; \
461 ALU_CHECK_CMP(true); \
464 #define ALU_CASE_CMP(nir, _op, sext) \
466 op = midgard_alu_op_##_op; \
467 ALU_CHECK_CMP(sext); \
470 /* Analyze the sizes of the dest and inputs to determine reg mode. */
472 static midgard_reg_mode
473 reg_mode_for_nir(nir_alu_instr
*instr
)
475 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
476 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
477 unsigned max_bitsize
= MAX2(src_bitsize
, dst_bitsize
);
479 /* We don't have fp16 LUTs, so we'll want to emit code like:
481 * vlut.fsinr hr0, hr0
483 * where both input and output are 16-bit but the operation is carried
495 max_bitsize
= MAX2(max_bitsize
, 32);
498 /* These get lowered to moves */
499 case nir_op_pack_32_4x8
:
502 case nir_op_pack_32_2x16
:
510 switch (max_bitsize
) {
511 /* Use 16 pipe for 8 since we don't support vec16 yet */
514 return midgard_reg_mode_16
;
516 return midgard_reg_mode_32
;
518 return midgard_reg_mode_64
;
520 unreachable("Invalid bit size");
524 /* Compare mir_lower_invert */
526 nir_accepts_inot(nir_op op
, unsigned src
)
530 case nir_op_iand
: /* TODO: b2f16 */
534 /* Only the condition */
542 mir_accept_dest_mod(compiler_context
*ctx
, nir_dest
**dest
, nir_op op
)
544 if (pan_has_dest_mod(dest
, op
)) {
545 assert((*dest
)->is_ssa
);
546 BITSET_SET(ctx
->already_emitted
, (*dest
)->ssa
.index
);
554 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
)
556 nir_alu_src src
= instr
->src
[i
];
559 if (pan_has_source_mod(&src
, nir_op_fneg
))
562 if (pan_has_source_mod(&src
, nir_op_fabs
))
566 if (nir_accepts_inot(instr
->op
, i
) && pan_has_source_mod(&src
, nir_op_inot
))
570 if (pan_has_source_mod(&src
, nir_op_fround_even
))
571 *roundmode
= MIDGARD_RTE
;
573 if (pan_has_source_mod(&src
, nir_op_ftrunc
))
574 *roundmode
= MIDGARD_RTZ
;
576 if (pan_has_source_mod(&src
, nir_op_ffloor
))
577 *roundmode
= MIDGARD_RTN
;
579 if (pan_has_source_mod(&src
, nir_op_fceil
))
580 *roundmode
= MIDGARD_RTP
;
583 unsigned bits
= nir_src_bit_size(src
.src
);
585 ins
->src
[to
] = nir_src_index(NULL
, &src
.src
);
586 ins
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
] | bits
;
588 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
589 ins
->swizzle
[to
][c
] = src
.swizzle
[
590 (!bcast_count
|| c
< bcast_count
) ? c
:
595 /* Midgard features both fcsel and icsel, depending on whether you want int or
596 * float modifiers. NIR's csel is typeless, so we want a heuristic to guess if
597 * we should emit an int or float csel depending on what modifiers could be
598 * placed. In the absense of modifiers, this is probably arbitrary. */
601 mir_is_bcsel_float(nir_alu_instr
*instr
)
604 nir_op_i2i8
, nir_op_i2i16
,
605 nir_op_i2i32
, nir_op_i2i64
608 nir_op floatmods
[] = {
609 nir_op_fabs
, nir_op_fneg
,
610 nir_op_f2f16
, nir_op_f2f32
,
614 nir_op floatdestmods
[] = {
615 nir_op_fsat
, nir_op_fsat_signed
, nir_op_fclamp_pos
,
616 nir_op_f2f16
, nir_op_f2f32
621 for (unsigned i
= 1; i
< 3; ++i
) {
622 nir_alu_src s
= instr
->src
[i
];
623 for (unsigned q
= 0; q
< ARRAY_SIZE(intmods
); ++q
) {
624 if (pan_has_source_mod(&s
, intmods
[q
]))
629 for (unsigned i
= 1; i
< 3; ++i
) {
630 nir_alu_src s
= instr
->src
[i
];
631 for (unsigned q
= 0; q
< ARRAY_SIZE(floatmods
); ++q
) {
632 if (pan_has_source_mod(&s
, floatmods
[q
]))
637 for (unsigned q
= 0; q
< ARRAY_SIZE(floatdestmods
); ++q
) {
638 nir_dest
*dest
= &instr
->dest
.dest
;
639 if (pan_has_dest_mod(&dest
, floatdestmods
[q
]))
647 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
649 nir_dest
*dest
= &instr
->dest
.dest
;
651 if (dest
->is_ssa
&& BITSET_TEST(ctx
->already_emitted
, dest
->ssa
.index
))
654 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
655 * is handled elsewhere */
657 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
658 midgard_emit_derivatives(ctx
, instr
);
662 bool is_ssa
= dest
->is_ssa
;
664 unsigned nr_components
= nir_dest_num_components(*dest
);
665 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
668 /* Number of components valid to check for the instruction (the rest
669 * will be forced to the last), or 0 to use as-is. Relevant as
670 * ball-type instructions have a channel count in NIR but are all vec4
673 unsigned broadcast_swizzle
= 0;
675 /* What register mode should we operate in? */
676 midgard_reg_mode reg_mode
=
677 reg_mode_for_nir(instr
);
679 /* Should we swap arguments? */
680 bool flip_src12
= false;
682 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
683 unsigned dst_bitsize
= nir_dest_bit_size(*dest
);
685 enum midgard_roundmode roundmode
= MIDGARD_RTE
;
688 ALU_CASE(fadd
, fadd
);
689 ALU_CASE(fmul
, fmul
);
690 ALU_CASE(fmin
, fmin
);
691 ALU_CASE(fmax
, fmax
);
692 ALU_CASE(imin
, imin
);
693 ALU_CASE(imax
, imax
);
694 ALU_CASE(umin
, umin
);
695 ALU_CASE(umax
, umax
);
696 ALU_CASE(ffloor
, ffloor
);
697 ALU_CASE(fround_even
, froundeven
);
698 ALU_CASE(ftrunc
, ftrunc
);
699 ALU_CASE(fceil
, fceil
);
700 ALU_CASE(fdot3
, fdot3
);
701 ALU_CASE(fdot4
, fdot4
);
702 ALU_CASE(iadd
, iadd
);
703 ALU_CASE(isub
, isub
);
704 ALU_CASE(imul
, imul
);
706 /* Zero shoved as second-arg */
707 ALU_CASE(iabs
, iabsdiff
);
711 ALU_CASE_CMP(feq32
, feq
, false);
712 ALU_CASE_CMP(fne32
, fne
, false);
713 ALU_CASE_CMP(flt32
, flt
, false);
714 ALU_CASE_CMP(ieq32
, ieq
, true);
715 ALU_CASE_CMP(ine32
, ine
, true);
716 ALU_CASE_CMP(ilt32
, ilt
, true);
717 ALU_CASE_CMP(ult32
, ult
, false);
719 /* We don't have a native b2f32 instruction. Instead, like many
720 * GPUs, we exploit booleans as 0/~0 for false/true, and
721 * correspondingly AND
722 * by 1.0 to do the type conversion. For the moment, prime us
725 * iand [whatever], #0
727 * At the end of emit_alu (as MIR), we'll fix-up the constant
730 ALU_CASE_CMP(b2f32
, iand
, true);
731 ALU_CASE_CMP(b2f16
, iand
, true);
732 ALU_CASE_CMP(b2i32
, iand
, true);
734 /* Likewise, we don't have a dedicated f2b32 instruction, but
735 * we can do a "not equal to 0.0" test. */
737 ALU_CASE_CMP(f2b32
, fne
, false);
738 ALU_CASE_CMP(i2b32
, ine
, true);
740 ALU_CASE(frcp
, frcp
);
741 ALU_CASE(frsq
, frsqrt
);
742 ALU_CASE(fsqrt
, fsqrt
);
743 ALU_CASE(fexp2
, fexp2
);
744 ALU_CASE(flog2
, flog2
);
746 ALU_CASE_RTZ(f2i64
, f2i_rte
);
747 ALU_CASE_RTZ(f2u64
, f2u_rte
);
748 ALU_CASE_RTZ(i2f64
, i2f_rte
);
749 ALU_CASE_RTZ(u2f64
, u2f_rte
);
751 ALU_CASE_RTZ(f2i32
, f2i_rte
);
752 ALU_CASE_RTZ(f2u32
, f2u_rte
);
753 ALU_CASE_RTZ(i2f32
, i2f_rte
);
754 ALU_CASE_RTZ(u2f32
, u2f_rte
);
756 ALU_CASE_RTZ(f2i8
, f2i_rte
);
757 ALU_CASE_RTZ(f2u8
, f2u_rte
);
759 ALU_CASE_RTZ(f2i16
, f2i_rte
);
760 ALU_CASE_RTZ(f2u16
, f2u_rte
);
761 ALU_CASE_RTZ(i2f16
, i2f_rte
);
762 ALU_CASE_RTZ(u2f16
, u2f_rte
);
764 ALU_CASE(fsin
, fsin
);
765 ALU_CASE(fcos
, fcos
);
767 /* We'll get 0 in the second arg, so:
768 * ~a = ~(a | 0) = nor(a, 0) */
769 ALU_CASE(inot
, inor
);
770 ALU_CASE(iand
, iand
);
772 ALU_CASE(ixor
, ixor
);
773 ALU_CASE(ishl
, ishl
);
774 ALU_CASE(ishr
, iasr
);
775 ALU_CASE(ushr
, ilsr
);
777 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
778 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
779 ALU_CASE_CMP(b32all_fequal4
, fball_eq
, true);
781 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
782 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
783 ALU_CASE_CMP(b32any_fnequal4
, fbany_neq
, true);
785 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
786 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
787 ALU_CASE_CMP(b32all_iequal4
, iball_eq
, true);
789 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
790 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
791 ALU_CASE_CMP(b32any_inequal4
, ibany_neq
, true);
793 /* Source mods will be shoved in later */
794 ALU_CASE(fabs
, fmov
);
795 ALU_CASE(fneg
, fmov
);
796 ALU_CASE(fsat
, fmov
);
797 ALU_CASE(fsat_signed
, fmov
);
798 ALU_CASE(fclamp_pos
, fmov
);
800 /* For size conversion, we use a move. Ideally though we would squash
801 * these ops together; maybe that has to happen after in NIR as part of
802 * propagation...? An earlier algebraic pass ensured we step down by
803 * only / exactly one size. If stepping down, we use a dest override to
804 * reduce the size; if stepping up, we use a larger-sized move with a
805 * half source and a sign/zero-extension modifier */
818 if (instr
->op
== nir_op_f2f16
|| instr
->op
== nir_op_f2f32
||
819 instr
->op
== nir_op_f2f64
)
820 op
= midgard_alu_op_fmov
;
822 op
= midgard_alu_op_imov
;
827 /* For greater-or-equal, we lower to less-or-equal and flip the
835 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
836 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
837 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
838 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
842 ALU_CHECK_CMP(false);
846 case nir_op_b32csel
: {
847 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
848 bool is_float
= mir_is_bcsel_float(instr
);
850 (mixed
? midgard_alu_op_fcsel_v
: midgard_alu_op_fcsel
) :
851 (mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
);
856 case nir_op_unpack_32_2x16
:
857 case nir_op_unpack_32_4x8
:
858 case nir_op_pack_32_2x16
:
859 case nir_op_pack_32_4x8
: {
860 op
= midgard_alu_op_imov
;
865 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
870 /* Promote imov to fmov if it might help inline a constant */
871 if (op
== midgard_alu_op_imov
&& nir_src_is_const(instr
->src
[0].src
)
872 && nir_src_bit_size(instr
->src
[0].src
) == 32
873 && nir_is_same_comp_swizzle(instr
->src
[0].swizzle
,
874 nir_src_num_components(instr
->src
[0].src
))) {
875 op
= midgard_alu_op_fmov
;
878 /* Midgard can perform certain modifiers on output of an ALU op */
881 bool is_int
= midgard_is_integer_op(op
);
883 if (midgard_is_integer_out_op(op
)) {
884 outmod
= midgard_outmod_int_wrap
;
885 } else if (instr
->op
== nir_op_fsat
) {
886 outmod
= midgard_outmod_sat
;
887 } else if (instr
->op
== nir_op_fsat_signed
) {
888 outmod
= midgard_outmod_sat_signed
;
889 } else if (instr
->op
== nir_op_fclamp_pos
) {
890 outmod
= midgard_outmod_pos
;
893 /* Fetch unit, quirks, etc information */
894 unsigned opcode_props
= alu_opcode_props
[op
].props
;
895 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
897 /* Look for floating point mods. We have the mods fsat, fsat_signed,
898 * and fpos. We also have the relations (note 3 * 2 = 6 cases):
900 * fsat_signed(fpos(x)) = fsat(x)
901 * fsat_signed(fsat(x)) = fsat(x)
902 * fpos(fsat_signed(x)) = fsat(x)
903 * fpos(fsat(x)) = fsat(x)
904 * fsat(fsat_signed(x)) = fsat(x)
905 * fsat(fpos(x)) = fsat(x)
907 * So by cases any composition of output modifiers is equivalent to
911 if (!is_int
&& !(opcode_props
& OP_TYPE_CONVERT
)) {
912 bool fpos
= mir_accept_dest_mod(ctx
, &dest
, nir_op_fclamp_pos
);
913 bool fsat
= mir_accept_dest_mod(ctx
, &dest
, nir_op_fsat
);
914 bool ssat
= mir_accept_dest_mod(ctx
, &dest
, nir_op_fsat_signed
);
915 bool prior
= (outmod
!= midgard_outmod_none
);
916 int count
= (int) prior
+ (int) fpos
+ (int) ssat
+ (int) fsat
;
918 outmod
= ((count
> 1) || fsat
) ? midgard_outmod_sat
:
919 fpos
? midgard_outmod_pos
:
920 ssat
? midgard_outmod_sat_signed
:
924 midgard_instruction ins
= {
926 .dest
= nir_dest_index(dest
),
927 .dest_type
= nir_op_infos
[instr
->op
].output_type
928 | nir_dest_bit_size(*dest
),
929 .roundmode
= roundmode
,
932 enum midgard_roundmode
*roundptr
= (opcode_props
& MIDGARD_ROUNDS
) ?
933 &ins
.roundmode
: NULL
;
935 for (unsigned i
= nr_inputs
; i
< ARRAY_SIZE(ins
.src
); ++i
)
938 if (quirk_flipped_r24
) {
940 mir_copy_src(&ins
, instr
, 0, 1, &ins
.src_abs
[1], &ins
.src_neg
[1], &ins
.src_invert
[1], roundptr
, is_int
, broadcast_swizzle
);
942 for (unsigned i
= 0; i
< nr_inputs
; ++i
) {
945 if (instr
->op
== nir_op_b32csel
) {
946 /* The condition is the first argument; move
947 * the other arguments up one to be a binary
948 * instruction for Midgard with the condition
957 } else if (flip_src12
) {
961 mir_copy_src(&ins
, instr
, i
, to
, &ins
.src_abs
[to
], &ins
.src_neg
[to
], &ins
.src_invert
[to
], roundptr
, is_int
, broadcast_swizzle
);
963 /* (!c) ? a : b = c ? b : a */
964 if (instr
->op
== nir_op_b32csel
&& ins
.src_invert
[2]) {
965 ins
.src_invert
[2] = false;
971 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
972 /* Lowered to move */
973 if (instr
->op
== nir_op_fneg
)
974 ins
.src_neg
[1] ^= true;
976 if (instr
->op
== nir_op_fabs
)
977 ins
.src_abs
[1] = true;
980 ins
.mask
= mask_of(nr_components
);
982 midgard_vector_alu alu
= {
984 .reg_mode
= reg_mode
,
988 /* Apply writemask if non-SSA, keeping in mind that we can't write to
989 * components that don't exist. Note modifier => SSA => !reg => no
990 * writemask, so we don't have to worry about writemasks here.*/
993 ins
.mask
&= instr
->dest
.write_mask
;
997 /* Late fixup for emulated instructions */
999 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1000 /* Presently, our second argument is an inline #0 constant.
1001 * Switch over to an embedded 1.0 constant (that can't fit
1002 * inline, since we're 32-bit, not 16-bit like the inline
1005 ins
.has_inline_constant
= false;
1006 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1007 ins
.src_types
[1] = nir_type_float32
;
1008 ins
.has_constants
= true;
1010 if (instr
->op
== nir_op_b2f32
)
1011 ins
.constants
.f32
[0] = 1.0f
;
1013 ins
.constants
.i32
[0] = 1;
1015 for (unsigned c
= 0; c
< 16; ++c
)
1016 ins
.swizzle
[1][c
] = 0;
1017 } else if (instr
->op
== nir_op_b2f16
) {
1018 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1019 ins
.src_types
[1] = nir_type_float16
;
1020 ins
.has_constants
= true;
1021 ins
.constants
.i16
[0] = _mesa_float_to_half(1.0);
1023 for (unsigned c
= 0; c
< 16; ++c
)
1024 ins
.swizzle
[1][c
] = 0;
1025 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1026 /* Lots of instructions need a 0 plonked in */
1027 ins
.has_inline_constant
= false;
1028 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1029 ins
.src_types
[1] = nir_type_uint32
;
1030 ins
.has_constants
= true;
1031 ins
.constants
.u32
[0] = 0;
1033 for (unsigned c
= 0; c
< 16; ++c
)
1034 ins
.swizzle
[1][c
] = 0;
1035 } else if (instr
->op
== nir_op_pack_32_2x16
) {
1036 ins
.dest_type
= nir_type_uint16
;
1037 ins
.mask
= mask_of(nr_components
* 2);
1039 } else if (instr
->op
== nir_op_pack_32_4x8
) {
1040 ins
.dest_type
= nir_type_uint8
;
1041 ins
.mask
= mask_of(nr_components
* 4);
1043 } else if (instr
->op
== nir_op_unpack_32_2x16
) {
1044 ins
.dest_type
= nir_type_uint32
;
1045 ins
.mask
= mask_of(nr_components
>> 1);
1047 } else if (instr
->op
== nir_op_unpack_32_4x8
) {
1048 ins
.dest_type
= nir_type_uint32
;
1049 ins
.mask
= mask_of(nr_components
>> 2);
1053 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1054 /* To avoid duplicating the lookup tables (probably), true LUT
1055 * instructions can only operate as if they were scalars. Lower
1056 * them here by changing the component. */
1058 unsigned orig_mask
= ins
.mask
;
1060 unsigned swizzle_back
[MIR_VEC_COMPONENTS
];
1061 memcpy(&swizzle_back
, ins
.swizzle
[0], sizeof(swizzle_back
));
1063 for (int i
= 0; i
< nr_components
; ++i
) {
1064 /* Mask the associated component, dropping the
1065 * instruction if needed */
1068 ins
.mask
&= orig_mask
;
1073 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1074 ins
.swizzle
[0][j
] = swizzle_back
[i
]; /* Pull from the correct component */
1076 emit_mir_instruction(ctx
, ins
);
1079 emit_mir_instruction(ctx
, ins
);
1086 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1088 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1089 unsigned nir_mask
= 0;
1093 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1094 dsize
= nir_dest_bit_size(intr
->dest
);
1096 nir_mask
= nir_intrinsic_write_mask(intr
);
1100 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1101 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1102 mir_set_bytemask(ins
, bytemask
);
1103 ins
->dest_type
= nir_type_uint
| dsize
;
1106 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1107 * optimized) versions of UBO #0 */
1109 static midgard_instruction
*
1111 compiler_context
*ctx
,
1115 nir_src
*indirect_offset
,
1116 unsigned indirect_shift
,
1119 /* TODO: half-floats */
1121 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1122 ins
.constants
.u32
[0] = offset
;
1124 if (instr
->type
== nir_instr_type_intrinsic
)
1125 mir_set_intr_mask(instr
, &ins
, true);
1127 if (indirect_offset
) {
1128 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1129 ins
.src_types
[2] = nir_type_uint32
;
1130 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1132 ins
.load_store
.arg_2
= 0x1E;
1135 ins
.load_store
.arg_1
= index
;
1137 return emit_mir_instruction(ctx
, ins
);
1140 /* Globals are like UBOs if you squint. And shared memory is like globals if
1141 * you squint even harder */
1145 compiler_context
*ctx
,
1154 midgard_instruction ins
;
1157 ins
= m_ld_int4(srcdest
, 0);
1159 ins
= m_st_int4(srcdest
, 0);
1161 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1162 mir_set_intr_mask(instr
, &ins
, is_read
);
1164 emit_mir_instruction(ctx
, ins
);
1169 compiler_context
*ctx
,
1170 unsigned dest
, unsigned offset
,
1171 unsigned nr_comp
, unsigned component
,
1172 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1174 /* XXX: Half-floats? */
1175 /* TODO: swizzle, mask */
1177 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1178 ins
.mask
= mask_of(nr_comp
);
1179 ins
.dest_type
= type
;
1181 if (type
== nir_type_float16
) {
1182 /* Ensure we are aligned so we can pack it later */
1183 ins
.mask
= mask_of(ALIGN_POT(nr_comp
, 2));
1186 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1187 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1189 midgard_varying_parameter p
= {
1191 .interpolation
= midgard_interp_default
,
1196 memcpy(&u
, &p
, sizeof(p
));
1197 ins
.load_store
.varying_parameters
= u
;
1199 if (indirect_offset
) {
1200 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1201 ins
.src_types
[2] = nir_type_uint32
;
1203 ins
.load_store
.arg_2
= 0x1E;
1205 ins
.load_store
.arg_1
= 0x9E;
1207 /* Use the type appropriate load */
1209 case nir_type_uint32
:
1210 case nir_type_bool32
:
1211 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1213 case nir_type_int32
:
1214 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1216 case nir_type_float32
:
1217 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1219 case nir_type_float16
:
1220 ins
.load_store
.op
= midgard_op_ld_vary_16
;
1223 unreachable("Attempted to load unknown type");
1227 emit_mir_instruction(ctx
, ins
);
1232 compiler_context
*ctx
,
1233 unsigned dest
, unsigned offset
,
1234 unsigned nr_comp
, nir_alu_type t
)
1236 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1237 ins
.load_store
.arg_1
= 0x1E;
1238 ins
.load_store
.arg_2
= 0x1E;
1239 ins
.mask
= mask_of(nr_comp
);
1241 /* Use the type appropriate load */
1245 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1248 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1250 case nir_type_float
:
1251 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1254 unreachable("Attempted to load unknown type");
1258 emit_mir_instruction(ctx
, ins
);
1262 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1263 unsigned nr_components
, unsigned offset
)
1267 /* Figure out which uniform this is */
1268 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1269 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1271 unsigned dest
= nir_dest_index(&nir_dest
);
1273 /* Sysvals are prefix uniforms */
1274 unsigned uniform
= ((uintptr_t) val
) - 1;
1276 /* Emit the read itself -- this is never indirect */
1277 midgard_instruction
*ins
=
1278 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1280 ins
->mask
= mask_of(nr_components
);
1284 compute_builtin_arg(nir_op op
)
1287 case nir_intrinsic_load_work_group_id
:
1289 case nir_intrinsic_load_local_invocation_id
:
1292 unreachable("Invalid compute paramater loaded");
1297 emit_fragment_store(compiler_context
*ctx
, unsigned src
, enum midgard_rt_id rt
)
1299 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1301 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1305 emit_explicit_constant(ctx
, src
, src
);
1307 struct midgard_instruction ins
=
1308 v_branch(false, false);
1310 bool depth_only
= (rt
== MIDGARD_ZS_RT
);
1312 ins
.writeout
= depth_only
? PAN_WRITEOUT_Z
: PAN_WRITEOUT_C
;
1314 /* Add dependencies */
1316 ins
.src_types
[0] = nir_type_uint32
;
1317 ins
.constants
.u32
[0] = depth_only
? 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1318 for (int i
= 0; i
< 4; ++i
)
1319 ins
.swizzle
[0][i
] = i
;
1321 /* Emit the branch */
1322 br
= emit_mir_instruction(ctx
, ins
);
1323 schedule_barrier(ctx
);
1324 ctx
->writeout_branch
[rt
] = br
;
1326 /* Push our current location = current block count - 1 = where we'll
1327 * jump to. Maybe a bit too clever for my own good */
1329 br
->branch
.target_block
= ctx
->block_count
- 1;
1333 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1335 unsigned reg
= nir_dest_index(&instr
->dest
);
1336 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1337 ins
.mask
= mask_of(3);
1338 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1339 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1340 emit_mir_instruction(ctx
, ins
);
1344 vertex_builtin_arg(nir_op op
)
1347 case nir_intrinsic_load_vertex_id
:
1348 return PAN_VERTEX_ID
;
1349 case nir_intrinsic_load_instance_id
:
1350 return PAN_INSTANCE_ID
;
1352 unreachable("Invalid vertex builtin");
1357 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1359 unsigned reg
= nir_dest_index(&instr
->dest
);
1360 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1364 emit_control_barrier(compiler_context
*ctx
)
1366 midgard_instruction ins
= {
1367 .type
= TAG_TEXTURE_4
,
1369 .src
= { ~0, ~0, ~0, ~0 },
1371 .op
= TEXTURE_OP_BARRIER
,
1373 /* TODO: optimize */
1374 .out_of_order
= MIDGARD_BARRIER_BUFFER
|
1375 MIDGARD_BARRIER_SHARED
,
1379 emit_mir_instruction(ctx
, ins
);
1382 static const nir_variable
*
1383 search_var(struct exec_list
*vars
, unsigned driver_loc
)
1385 nir_foreach_variable(var
, vars
) {
1386 if (var
->data
.driver_location
== driver_loc
)
1394 mir_get_branch_cond(nir_src
*src
, bool *invert
)
1396 /* Wrap it. No swizzle since it's a scalar */
1402 *invert
= pan_has_source_mod(&alu
, nir_op_inot
);
1403 return nir_src_index(NULL
, &alu
.src
);
1407 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1409 unsigned offset
= 0, reg
;
1411 switch (instr
->intrinsic
) {
1412 case nir_intrinsic_discard_if
:
1413 case nir_intrinsic_discard
: {
1414 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1415 struct midgard_instruction discard
= v_branch(conditional
, false);
1416 discard
.branch
.target_type
= TARGET_DISCARD
;
1419 discard
.src
[0] = mir_get_branch_cond(&instr
->src
[0],
1420 &discard
.branch
.invert_conditional
);
1421 discard
.src_types
[0] = nir_type_uint32
;
1424 emit_mir_instruction(ctx
, discard
);
1425 schedule_barrier(ctx
);
1430 case nir_intrinsic_load_uniform
:
1431 case nir_intrinsic_load_ubo
:
1432 case nir_intrinsic_load_global
:
1433 case nir_intrinsic_load_shared
:
1434 case nir_intrinsic_load_input
:
1435 case nir_intrinsic_load_interpolated_input
: {
1436 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1437 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1438 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1439 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1440 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1441 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1443 /* Get the base type of the intrinsic */
1444 /* TODO: Infer type? Does it matter? */
1446 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1447 (is_interp
) ? nir_type_float
:
1448 nir_intrinsic_type(instr
);
1450 t
= nir_alu_type_get_base_type(t
);
1452 if (!(is_ubo
|| is_global
)) {
1453 offset
= nir_intrinsic_base(instr
);
1456 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1458 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1460 bool direct
= nir_src_is_const(*src_offset
);
1461 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1464 offset
+= nir_src_as_uint(*src_offset
);
1466 /* We may need to apply a fractional offset */
1467 int component
= (is_flat
|| is_interp
) ?
1468 nir_intrinsic_component(instr
) : 0;
1469 reg
= nir_dest_index(&instr
->dest
);
1471 if (is_uniform
&& !ctx
->is_blend
) {
1472 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1473 } else if (is_ubo
) {
1474 nir_src index
= instr
->src
[0];
1476 /* TODO: Is indirect block number possible? */
1477 assert(nir_src_is_const(index
));
1479 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1480 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1481 } else if (is_global
|| is_shared
) {
1482 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1483 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1484 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
| nir_dest_bit_size(instr
->dest
), is_flat
);
1485 } else if (ctx
->is_blend
) {
1486 /* For blend shaders, load the input color, which is
1487 * preloaded to r0 */
1489 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), reg
);
1490 emit_mir_instruction(ctx
, move
);
1491 schedule_barrier(ctx
);
1492 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1493 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1495 DBG("Unknown load\n");
1502 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1503 case nir_intrinsic_load_barycentric_pixel
:
1504 case nir_intrinsic_load_barycentric_centroid
:
1507 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1509 case nir_intrinsic_load_raw_output_pan
: {
1510 reg
= nir_dest_index(&instr
->dest
);
1511 assert(ctx
->is_blend
);
1513 /* T720 and below use different blend opcodes with slightly
1514 * different semantics than T760 and up */
1516 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1518 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1519 ld
.load_store
.op
= midgard_op_ld_color_buffer_32u_old
;
1520 ld
.load_store
.address
= 16;
1521 ld
.load_store
.arg_2
= 0x1E;
1524 emit_mir_instruction(ctx
, ld
);
1528 case nir_intrinsic_load_output
: {
1529 reg
= nir_dest_index(&instr
->dest
);
1530 assert(ctx
->is_blend
);
1532 midgard_instruction ld
= m_ld_color_buffer_as_fp16(reg
, 0);
1534 for (unsigned c
= 4; c
< 16; ++c
)
1535 ld
.swizzle
[0][c
] = 0;
1537 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1538 ld
.load_store
.op
= midgard_op_ld_color_buffer_as_fp16_old
;
1539 ld
.load_store
.address
= 1;
1540 ld
.load_store
.arg_2
= 0x1E;
1543 emit_mir_instruction(ctx
, ld
);
1547 case nir_intrinsic_load_blend_const_color_rgba
: {
1548 assert(ctx
->is_blend
);
1549 reg
= nir_dest_index(&instr
->dest
);
1551 /* Blend constants are embedded directly in the shader and
1552 * patched in, so we use some magic routing */
1554 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1555 ins
.has_constants
= true;
1556 ins
.has_blend_constant
= true;
1557 emit_mir_instruction(ctx
, ins
);
1561 case nir_intrinsic_store_output
:
1562 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1564 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1566 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1568 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1569 const nir_variable
*var
;
1570 enum midgard_rt_id rt
;
1572 var
= search_var(&ctx
->nir
->outputs
,
1573 nir_intrinsic_base(instr
));
1575 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1576 rt
= MIDGARD_COLOR_RT0
;
1577 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1578 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1583 emit_fragment_store(ctx
, reg
, rt
);
1584 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1585 /* We should have been vectorized, though we don't
1586 * currently check that st_vary is emitted only once
1587 * per slot (this is relevant, since there's not a mask
1588 * parameter available on the store [set to 0 by the
1589 * blob]). We do respect the component by adjusting the
1590 * swizzle. If this is a constant source, we'll need to
1591 * emit that explicitly. */
1593 emit_explicit_constant(ctx
, reg
, reg
);
1595 unsigned dst_component
= nir_intrinsic_component(instr
);
1596 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1598 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1599 st
.load_store
.arg_1
= 0x9E;
1600 st
.load_store
.arg_2
= 0x1E;
1602 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1605 st
.load_store
.op
= midgard_op_st_vary_32u
;
1608 st
.load_store
.op
= midgard_op_st_vary_32i
;
1610 case nir_type_float
:
1611 st
.load_store
.op
= midgard_op_st_vary_32
;
1614 unreachable("Attempted to store unknown type");
1618 /* nir_intrinsic_component(store_intr) encodes the
1619 * destination component start. Source component offset
1620 * adjustment is taken care of in
1621 * install_registers_instr(), when offset_swizzle() is
1624 unsigned src_component
= COMPONENT_X
;
1626 assert(nr_comp
> 0);
1627 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1628 st
.swizzle
[0][i
] = src_component
;
1629 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1633 emit_mir_instruction(ctx
, st
);
1635 DBG("Unknown store\n");
1641 /* Special case of store_output for lowered blend shaders */
1642 case nir_intrinsic_store_raw_output_pan
:
1643 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1644 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1645 emit_fragment_store(ctx
, reg
, ctx
->blend_rt
);
1648 case nir_intrinsic_store_global
:
1649 case nir_intrinsic_store_shared
:
1650 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1651 emit_explicit_constant(ctx
, reg
, reg
);
1653 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1656 case nir_intrinsic_load_ssbo_address
:
1657 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1660 case nir_intrinsic_get_buffer_size
:
1661 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1664 case nir_intrinsic_load_viewport_scale
:
1665 case nir_intrinsic_load_viewport_offset
:
1666 case nir_intrinsic_load_num_work_groups
:
1667 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1668 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1671 case nir_intrinsic_load_work_group_id
:
1672 case nir_intrinsic_load_local_invocation_id
:
1673 emit_compute_builtin(ctx
, instr
);
1676 case nir_intrinsic_load_vertex_id
:
1677 case nir_intrinsic_load_instance_id
:
1678 emit_vertex_builtin(ctx
, instr
);
1681 case nir_intrinsic_memory_barrier_buffer
:
1682 case nir_intrinsic_memory_barrier_shared
:
1685 case nir_intrinsic_control_barrier
:
1686 schedule_barrier(ctx
);
1687 emit_control_barrier(ctx
);
1688 schedule_barrier(ctx
);
1692 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1699 midgard_tex_format(enum glsl_sampler_dim dim
)
1702 case GLSL_SAMPLER_DIM_1D
:
1703 case GLSL_SAMPLER_DIM_BUF
:
1706 case GLSL_SAMPLER_DIM_2D
:
1707 case GLSL_SAMPLER_DIM_EXTERNAL
:
1708 case GLSL_SAMPLER_DIM_RECT
:
1711 case GLSL_SAMPLER_DIM_3D
:
1714 case GLSL_SAMPLER_DIM_CUBE
:
1715 return MALI_TEX_CUBE
;
1718 DBG("Unknown sampler dim type\n");
1724 /* Tries to attach an explicit LOD or bias as a constant. Returns whether this
1728 pan_attach_constant_bias(
1729 compiler_context
*ctx
,
1731 midgard_texture_word
*word
)
1733 /* To attach as constant, it has to *be* constant */
1735 if (!nir_src_is_const(lod
))
1738 float f
= nir_src_as_float(lod
);
1740 /* Break into fixed-point */
1742 float lod_frac
= f
- lod_int
;
1744 /* Carry over negative fractions */
1745 if (lod_frac
< 0.0) {
1751 word
->bias
= float_to_ubyte(lod_frac
);
1752 word
->bias_int
= lod_int
;
1758 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1759 unsigned midgard_texop
)
1762 //assert (!instr->sampler);
1764 int texture_index
= instr
->texture_index
;
1765 int sampler_index
= texture_index
;
1767 nir_alu_type dest_base
= nir_alu_type_get_base_type(instr
->dest_type
);
1768 nir_alu_type dest_type
= dest_base
| nir_dest_bit_size(instr
->dest
);
1770 midgard_instruction ins
= {
1771 .type
= TAG_TEXTURE_4
,
1773 .dest
= nir_dest_index(&instr
->dest
),
1774 .src
= { ~0, ~0, ~0, ~0 },
1775 .dest_type
= dest_type
,
1776 .swizzle
= SWIZZLE_IDENTITY_4
,
1778 .op
= midgard_texop
,
1779 .format
= midgard_tex_format(instr
->sampler_dim
),
1780 .texture_handle
= texture_index
,
1781 .sampler_handle
= sampler_index
,
1782 .shadow
= instr
->is_shadow
,
1786 if (instr
->is_shadow
&& !instr
->is_new_style_shadow
)
1787 for (int i
= 0; i
< 4; ++i
)
1788 ins
.swizzle
[0][i
] = COMPONENT_X
;
1790 /* We may need a temporary for the coordinate */
1792 bool needs_temp_coord
=
1793 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
1794 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
1797 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
1799 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1800 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1801 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1802 unsigned sz
= nir_src_bit_size(instr
->src
[i
].src
);
1803 nir_alu_type T
= nir_tex_instr_src_type(instr
, i
) | sz
;
1805 switch (instr
->src
[i
].src_type
) {
1806 case nir_tex_src_coord
: {
1807 emit_explicit_constant(ctx
, index
, index
);
1809 unsigned coord_mask
= mask_of(instr
->coord_components
);
1811 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
1814 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
1816 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1817 /* texelFetch is undefined on samplerCube */
1818 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1820 /* For cubemaps, we use a special ld/st op to
1821 * select the face and copy the xy into the
1822 * texture register */
1824 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
1826 ld
.src_types
[1] = T
;
1827 ld
.mask
= 0x3; /* xy */
1828 ld
.load_store
.arg_1
= 0x20;
1829 ld
.swizzle
[1][3] = COMPONENT_X
;
1830 emit_mir_instruction(ctx
, ld
);
1833 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1834 ins
.swizzle
[1][3] = COMPONENT_X
;
1835 } else if (needs_temp_coord
) {
1836 /* mov coord_temp, coords */
1837 midgard_instruction mov
= v_mov(index
, coords
);
1838 mov
.mask
= coord_mask
;
1841 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
1843 emit_mir_instruction(ctx
, mov
);
1848 ins
.src
[1] = coords
;
1849 ins
.src_types
[1] = T
;
1851 /* Texelfetch coordinates uses all four elements
1852 * (xyz/index) regardless of texture dimensionality,
1853 * which means it's necessary to zero the unused
1854 * components to keep everything happy */
1856 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1857 /* mov index.zw, #0, or generalized */
1858 midgard_instruction mov
=
1859 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
1860 mov
.has_constants
= true;
1861 mov
.mask
= coord_mask
^ 0xF;
1862 emit_mir_instruction(ctx
, mov
);
1865 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1866 /* Array component in w but NIR wants it in z,
1867 * but if we have a temp coord we already fixed
1870 if (nr_components
== 3) {
1871 ins
.swizzle
[1][2] = COMPONENT_Z
;
1872 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
1873 } else if (nr_components
== 2) {
1875 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1876 ins
.swizzle
[1][3] = COMPONENT_X
;
1878 unreachable("Invalid texture 2D components");
1881 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1883 ins
.swizzle
[1][2] = COMPONENT_Z
;
1884 ins
.swizzle
[1][3] = COMPONENT_W
;
1890 case nir_tex_src_bias
:
1891 case nir_tex_src_lod
: {
1892 /* Try as a constant if we can */
1894 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1895 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1898 ins
.texture
.lod_register
= true;
1900 ins
.src_types
[2] = T
;
1902 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1903 ins
.swizzle
[2][c
] = COMPONENT_X
;
1905 emit_explicit_constant(ctx
, index
, index
);
1910 case nir_tex_src_offset
: {
1911 ins
.texture
.offset_register
= true;
1913 ins
.src_types
[3] = T
;
1915 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1916 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
1918 emit_explicit_constant(ctx
, index
, index
);
1922 case nir_tex_src_comparator
: {
1923 unsigned comp
= COMPONENT_Z
;
1925 /* mov coord_temp.foo, coords */
1926 midgard_instruction mov
= v_mov(index
, coords
);
1927 mov
.mask
= 1 << comp
;
1929 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
1930 mov
.swizzle
[1][i
] = COMPONENT_X
;
1932 emit_mir_instruction(ctx
, mov
);
1937 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
1943 emit_mir_instruction(ctx
, ins
);
1947 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1949 switch (instr
->op
) {
1952 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1955 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1958 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1961 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
1964 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
1971 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1973 switch (instr
->type
) {
1974 case nir_jump_break
: {
1975 /* Emit a branch out of the loop */
1976 struct midgard_instruction br
= v_branch(false, false);
1977 br
.branch
.target_type
= TARGET_BREAK
;
1978 br
.branch
.target_break
= ctx
->current_loop_depth
;
1979 emit_mir_instruction(ctx
, br
);
1984 DBG("Unknown jump type %d\n", instr
->type
);
1990 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1992 switch (instr
->type
) {
1993 case nir_instr_type_load_const
:
1994 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1997 case nir_instr_type_intrinsic
:
1998 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2001 case nir_instr_type_alu
:
2002 emit_alu(ctx
, nir_instr_as_alu(instr
));
2005 case nir_instr_type_tex
:
2006 emit_tex(ctx
, nir_instr_as_tex(instr
));
2009 case nir_instr_type_jump
:
2010 emit_jump(ctx
, nir_instr_as_jump(instr
));
2013 case nir_instr_type_ssa_undef
:
2018 DBG("Unhandled instruction type\n");
2024 /* ALU instructions can inline or embed constants, which decreases register
2025 * pressure and saves space. */
2027 #define CONDITIONAL_ATTACH(idx) { \
2028 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
2031 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
2032 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2037 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
2039 mir_foreach_instr_in_block(block
, alu
) {
2040 /* Other instructions cannot inline constants */
2041 if (alu
->type
!= TAG_ALU_4
) continue;
2042 if (alu
->compact_branch
) continue;
2044 /* If there is already a constant here, we can do nothing */
2045 if (alu
->has_constants
) continue;
2047 CONDITIONAL_ATTACH(0);
2049 if (!alu
->has_constants
) {
2050 CONDITIONAL_ATTACH(1)
2051 } else if (!alu
->inline_constant
) {
2052 /* Corner case: _two_ vec4 constants, for instance with a
2053 * csel. For this case, we can only use a constant
2054 * register for one, we'll have to emit a move for the
2057 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2058 unsigned scratch
= make_compiler_temp(ctx
);
2061 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2062 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2064 /* Set the source */
2065 alu
->src
[1] = scratch
;
2067 /* Inject us -before- the last instruction which set r31 */
2068 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2074 /* Midgard supports two types of constants, embedded constants (128-bit) and
2075 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2076 * constants can be demoted to inline constants, for space savings and
2077 * sometimes a performance boost */
2080 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2082 mir_foreach_instr_in_block(block
, ins
) {
2083 if (!ins
->has_constants
) continue;
2084 if (ins
->has_inline_constant
) continue;
2086 /* Blend constants must not be inlined by definition */
2087 if (ins
->has_blend_constant
) continue;
2089 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2090 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2091 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2093 if (!(is_16
|| is_32
))
2096 /* src1 cannot be an inline constant due to encoding
2097 * restrictions. So, if possible we try to flip the arguments
2100 int op
= ins
->alu
.op
;
2102 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
) &&
2103 alu_opcode_props
[op
].props
& OP_COMMUTES
) {
2107 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2108 /* Component is from the swizzle. Take a nonzero component */
2110 unsigned first_comp
= ffs(ins
->mask
) - 1;
2111 unsigned component
= ins
->swizzle
[1][first_comp
];
2113 /* Scale constant appropriately, if we can legally */
2114 int16_t scaled_constant
= 0;
2117 scaled_constant
= ins
->constants
.u16
[component
];
2118 } else if (midgard_is_integer_op(op
)) {
2119 scaled_constant
= ins
->constants
.u32
[component
];
2121 /* Constant overflow after resize */
2122 if (scaled_constant
!= ins
->constants
.u32
[component
])
2125 float original
= ins
->constants
.f32
[component
];
2126 scaled_constant
= _mesa_float_to_half(original
);
2128 /* Check for loss of precision. If this is
2129 * mediump, we don't care, but for a highp
2130 * shader, we need to pay attention. NIR
2131 * doesn't yet tell us which mode we're in!
2132 * Practically this prevents most constants
2133 * from being inlined, sadly. */
2135 float fp32
= _mesa_half_to_float(scaled_constant
);
2137 if (fp32
!= original
)
2141 /* Should've been const folded */
2142 if (ins
->src_abs
[1] || ins
->src_neg
[1])
2145 /* Make sure that the constant is not itself a vector
2146 * by checking if all accessed values are the same. */
2148 const midgard_constants
*cons
= &ins
->constants
;
2149 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2151 bool is_vector
= false;
2152 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2154 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2155 /* We only care if this component is actually used */
2156 if (!(mask
& (1 << c
)))
2159 uint32_t test
= is_16
?
2160 cons
->u16
[ins
->swizzle
[1][c
]] :
2161 cons
->u32
[ins
->swizzle
[1][c
]];
2163 if (test
!= value
) {
2172 /* Get rid of the embedded constant */
2173 ins
->has_constants
= false;
2175 ins
->has_inline_constant
= true;
2176 ins
->inline_constant
= scaled_constant
;
2181 /* Dead code elimination for branches at the end of a block - only one branch
2182 * per block is legal semantically */
2185 midgard_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2187 bool branched
= false;
2189 mir_foreach_instr_in_block_safe(block
, ins
) {
2190 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2193 mir_remove_instruction(ins
);
2199 /* We want to force the invert on AND/OR to the second slot to legalize into
2200 * iandnot/iornot. The relevant patterns are for AND (and OR respectively)
2202 * ~a & #b = ~a & ~(#~b)
2207 midgard_legalize_invert(compiler_context
*ctx
, midgard_block
*block
)
2209 mir_foreach_instr_in_block(block
, ins
) {
2210 if (ins
->type
!= TAG_ALU_4
) continue;
2212 if (ins
->alu
.op
!= midgard_alu_op_iand
&&
2213 ins
->alu
.op
!= midgard_alu_op_ior
) continue;
2215 if (ins
->src_invert
[1] || !ins
->src_invert
[0]) continue;
2217 if (ins
->has_inline_constant
) {
2218 /* ~(#~a) = ~(~#a) = a, so valid, and forces both
2220 ins
->inline_constant
= ~ins
->inline_constant
;
2221 ins
->src_invert
[1] = true;
2223 /* Flip to the right invert order. Note
2224 * has_inline_constant false by assumption on the
2225 * branch, so flipping makes sense. */
2232 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2234 /* Loop to ourselves */
2235 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2236 struct midgard_instruction ins
= v_branch(false, false);
2237 ins
.writeout
= br
->writeout
;
2238 ins
.branch
.target_block
= ctx
->block_count
- 1;
2239 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2240 memcpy(&ins
.src_types
, &br
->src_types
, sizeof(ins
.src_types
));
2241 emit_mir_instruction(ctx
, ins
);
2243 ctx
->current_block
->epilogue
= true;
2244 schedule_barrier(ctx
);
2245 return ins
.branch
.target_block
;
2248 static midgard_block
*
2249 emit_block(compiler_context
*ctx
, nir_block
*block
)
2251 midgard_block
*this_block
= ctx
->after_block
;
2252 ctx
->after_block
= NULL
;
2255 this_block
= create_empty_block(ctx
);
2257 list_addtail(&this_block
->base
.link
, &ctx
->blocks
);
2259 this_block
->scheduled
= false;
2262 /* Set up current block */
2263 list_inithead(&this_block
->base
.instructions
);
2264 ctx
->current_block
= this_block
;
2266 nir_foreach_instr(instr
, block
) {
2267 emit_instr(ctx
, instr
);
2268 ++ctx
->instruction_count
;
2274 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2277 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2279 midgard_block
*before_block
= ctx
->current_block
;
2281 /* Speculatively emit the branch, but we can't fill it in until later */
2283 EMIT(branch
, true, true);
2284 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2285 then_branch
->src
[0] = mir_get_branch_cond(&nif
->condition
, &inv
);
2286 then_branch
->src_types
[0] = nir_type_uint32
;
2287 then_branch
->branch
.invert_conditional
= !inv
;
2289 /* Emit the two subblocks. */
2290 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2291 midgard_block
*end_then_block
= ctx
->current_block
;
2293 /* Emit a jump from the end of the then block to the end of the else */
2294 EMIT(branch
, false, false);
2295 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2297 /* Emit second block, and check if it's empty */
2299 int else_idx
= ctx
->block_count
;
2300 int count_in
= ctx
->instruction_count
;
2301 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2302 midgard_block
*end_else_block
= ctx
->current_block
;
2303 int after_else_idx
= ctx
->block_count
;
2305 /* Now that we have the subblocks emitted, fix up the branches */
2310 if (ctx
->instruction_count
== count_in
) {
2311 /* The else block is empty, so don't emit an exit jump */
2312 mir_remove_instruction(then_exit
);
2313 then_branch
->branch
.target_block
= after_else_idx
;
2315 then_branch
->branch
.target_block
= else_idx
;
2316 then_exit
->branch
.target_block
= after_else_idx
;
2319 /* Wire up the successors */
2321 ctx
->after_block
= create_empty_block(ctx
);
2323 pan_block_add_successor(&before_block
->base
, &then_block
->base
);
2324 pan_block_add_successor(&before_block
->base
, &else_block
->base
);
2326 pan_block_add_successor(&end_then_block
->base
, &ctx
->after_block
->base
);
2327 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
);
2331 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2333 /* Remember where we are */
2334 midgard_block
*start_block
= ctx
->current_block
;
2336 /* Allocate a loop number, growing the current inner loop depth */
2337 int loop_idx
= ++ctx
->current_loop_depth
;
2339 /* Get index from before the body so we can loop back later */
2340 int start_idx
= ctx
->block_count
;
2342 /* Emit the body itself */
2343 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2345 /* Branch back to loop back */
2346 struct midgard_instruction br_back
= v_branch(false, false);
2347 br_back
.branch
.target_block
= start_idx
;
2348 emit_mir_instruction(ctx
, br_back
);
2350 /* Mark down that branch in the graph. */
2351 pan_block_add_successor(&start_block
->base
, &loop_block
->base
);
2352 pan_block_add_successor(&ctx
->current_block
->base
, &loop_block
->base
);
2354 /* Find the index of the block about to follow us (note: we don't add
2355 * one; blocks are 0-indexed so we get a fencepost problem) */
2356 int break_block_idx
= ctx
->block_count
;
2358 /* Fix up the break statements we emitted to point to the right place,
2359 * now that we can allocate a block number for them */
2360 ctx
->after_block
= create_empty_block(ctx
);
2362 mir_foreach_block_from(ctx
, start_block
, _block
) {
2363 mir_foreach_instr_in_block(((midgard_block
*) _block
), ins
) {
2364 if (ins
->type
!= TAG_ALU_4
) continue;
2365 if (!ins
->compact_branch
) continue;
2367 /* We found a branch -- check the type to see if we need to do anything */
2368 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2370 /* It's a break! Check if it's our break */
2371 if (ins
->branch
.target_break
!= loop_idx
) continue;
2373 /* Okay, cool, we're breaking out of this loop.
2374 * Rewrite from a break to a goto */
2376 ins
->branch
.target_type
= TARGET_GOTO
;
2377 ins
->branch
.target_block
= break_block_idx
;
2379 pan_block_add_successor(_block
, &ctx
->after_block
->base
);
2383 /* Now that we've finished emitting the loop, free up the depth again
2384 * so we play nice with recursion amid nested loops */
2385 --ctx
->current_loop_depth
;
2387 /* Dump loop stats */
2391 static midgard_block
*
2392 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2394 midgard_block
*start_block
= NULL
;
2396 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2397 switch (node
->type
) {
2398 case nir_cf_node_block
: {
2399 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2402 start_block
= block
;
2407 case nir_cf_node_if
:
2408 emit_if(ctx
, nir_cf_node_as_if(node
));
2411 case nir_cf_node_loop
:
2412 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2415 case nir_cf_node_function
:
2424 /* Due to lookahead, we need to report the first tag executed in the command
2425 * stream and in branch targets. An initial block might be empty, so iterate
2426 * until we find one that 'works' */
2429 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2431 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2433 mir_foreach_block_from(ctx
, initial_block
, _v
) {
2434 midgard_block
*v
= (midgard_block
*) _v
;
2435 if (v
->quadword_count
) {
2436 midgard_bundle
*initial_bundle
=
2437 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2439 return initial_bundle
->tag
;
2443 /* Default to a tag 1 which will break from the shader, in case we jump
2444 * to the exit block (i.e. `return` in a compute shader) */
2449 /* For each fragment writeout instruction, generate a writeout loop to
2450 * associate with it */
2453 mir_add_writeout_loops(compiler_context
*ctx
)
2455 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2456 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2459 unsigned popped
= br
->branch
.target_block
;
2460 pan_block_add_successor(&(mir_get_block(ctx
, popped
- 1)->base
), &ctx
->current_block
->base
);
2461 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2462 br
->branch
.target_type
= TARGET_GOTO
;
2464 /* If we have more RTs, we'll need to restore back after our
2465 * loop terminates */
2467 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2468 midgard_instruction uncond
= v_branch(false, false);
2469 uncond
.branch
.target_block
= popped
;
2470 uncond
.branch
.target_type
= TARGET_GOTO
;
2471 emit_mir_instruction(ctx
, uncond
);
2472 pan_block_add_successor(&ctx
->current_block
->base
, &(mir_get_block(ctx
, popped
)->base
));
2473 schedule_barrier(ctx
);
2475 /* We're last, so we can terminate here */
2476 br
->last_writeout
= true;
2482 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2484 struct util_dynarray
*compiled
= &program
->compiled
;
2486 midgard_debug
= debug_get_option_midgard_debug();
2488 /* TODO: Bound against what? */
2489 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2492 ctx
->stage
= nir
->info
.stage
;
2493 ctx
->is_blend
= is_blend
;
2494 ctx
->alpha_ref
= program
->alpha_ref
;
2495 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2496 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2498 /* Start off with a safe cutoff, allowing usage of all 16 work
2499 * registers. Later, we'll promote uniform reads to uniform registers
2500 * if we determine it is beneficial to do so */
2501 ctx
->uniform_cutoff
= 8;
2503 /* Initialize at a global (not block) level hash tables */
2505 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2506 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2508 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2509 * (so we don't accidentally duplicate the epilogue since mesa/st has
2510 * messed with our I/O quite a bit already) */
2512 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2514 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2515 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2516 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2519 NIR_PASS_V(nir
, nir_lower_var_copies
);
2520 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2521 NIR_PASS_V(nir
, nir_split_var_copies
);
2522 NIR_PASS_V(nir
, nir_lower_var_copies
);
2523 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2524 NIR_PASS_V(nir
, nir_lower_var_copies
);
2525 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2527 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2528 NIR_PASS_V(nir
, nir_lower_ssbo
);
2530 /* Optimisation passes */
2532 optimise_nir(nir
, ctx
->quirks
, is_blend
);
2534 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2535 nir_print_shader(nir
, stdout
);
2538 /* Assign sysvals and counts, now that we're sure
2539 * (post-optimisation) */
2541 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2542 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2543 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2545 nir_foreach_function(func
, nir
) {
2549 list_inithead(&ctx
->blocks
);
2550 ctx
->block_count
= 0;
2552 ctx
->already_emitted
= calloc(BITSET_WORDS(func
->impl
->ssa_alloc
), sizeof(BITSET_WORD
));
2554 emit_cf_list(ctx
, &func
->impl
->body
);
2555 free(ctx
->already_emitted
);
2556 break; /* TODO: Multi-function shaders */
2559 util_dynarray_init(compiled
, NULL
);
2561 /* Per-block lowering before opts */
2563 mir_foreach_block(ctx
, _block
) {
2564 midgard_block
*block
= (midgard_block
*) _block
;
2565 inline_alu_constants(ctx
, block
);
2566 embedded_to_inline_constant(ctx
, block
);
2568 /* MIR-level optimizations */
2570 bool progress
= false;
2574 progress
|= midgard_opt_dead_code_eliminate(ctx
);
2576 mir_foreach_block(ctx
, _block
) {
2577 midgard_block
*block
= (midgard_block
*) _block
;
2578 progress
|= midgard_opt_copy_prop(ctx
, block
);
2579 progress
|= midgard_opt_combine_projection(ctx
, block
);
2580 progress
|= midgard_opt_varying_projection(ctx
, block
);
2584 mir_foreach_block(ctx
, _block
) {
2585 midgard_block
*block
= (midgard_block
*) _block
;
2586 midgard_lower_derivatives(ctx
, block
);
2587 midgard_legalize_invert(ctx
, block
);
2588 midgard_cull_dead_branch(ctx
, block
);
2591 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2592 mir_add_writeout_loops(ctx
);
2594 /* Analyze now that the code is known but before scheduling creates
2595 * pipeline registers which are harder to track */
2596 mir_analyze_helper_terminate(ctx
);
2597 mir_analyze_helper_requirements(ctx
);
2600 midgard_schedule_program(ctx
);
2603 /* Now that all the bundles are scheduled and we can calculate block
2604 * sizes, emit actual branch instructions rather than placeholders */
2606 int br_block_idx
= 0;
2608 mir_foreach_block(ctx
, _block
) {
2609 midgard_block
*block
= (midgard_block
*) _block
;
2610 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2611 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2612 midgard_instruction
*ins
= bundle
->instructions
[c
];
2614 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2616 /* Parse some basic branch info */
2617 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2618 bool is_conditional
= ins
->branch
.conditional
;
2619 bool is_inverted
= ins
->branch
.invert_conditional
;
2620 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2621 bool is_writeout
= ins
->writeout
;
2623 /* Determine the block we're jumping to */
2624 int target_number
= ins
->branch
.target_block
;
2626 /* Report the destination tag */
2627 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2629 /* Count up the number of quadwords we're
2630 * jumping over = number of quadwords until
2631 * (br_block_idx, target_number) */
2633 int quadword_offset
= 0;
2637 } else if (target_number
> br_block_idx
) {
2640 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2641 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2644 quadword_offset
+= blk
->quadword_count
;
2647 /* Jump backwards */
2649 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2650 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2653 quadword_offset
-= blk
->quadword_count
;
2657 /* Unconditional extended branches (far jumps)
2658 * have issues, so we always use a conditional
2659 * branch, setting the condition to always for
2660 * unconditional. For compact unconditional
2661 * branches, cond isn't used so it doesn't
2662 * matter what we pick. */
2664 midgard_condition cond
=
2665 !is_conditional
? midgard_condition_always
:
2666 is_inverted
? midgard_condition_false
:
2667 midgard_condition_true
;
2669 midgard_jmp_writeout_op op
=
2670 is_discard
? midgard_jmp_writeout_op_discard
:
2671 is_writeout
? midgard_jmp_writeout_op_writeout
:
2672 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2673 midgard_jmp_writeout_op_branch_cond
;
2676 midgard_branch_extended branch
=
2677 midgard_create_branch_extended(
2682 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2683 } else if (is_conditional
|| is_discard
) {
2684 midgard_branch_cond branch
= {
2686 .dest_tag
= dest_tag
,
2687 .offset
= quadword_offset
,
2691 assert(branch
.offset
== quadword_offset
);
2693 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2695 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2697 midgard_branch_uncond branch
= {
2699 .dest_tag
= dest_tag
,
2700 .offset
= quadword_offset
,
2704 assert(branch
.offset
== quadword_offset
);
2706 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2714 /* Emit flat binary from the instruction arrays. Iterate each block in
2715 * sequence. Save instruction boundaries such that lookahead tags can
2716 * be assigned easily */
2718 /* Cache _all_ bundles in source order for lookahead across failed branches */
2720 int bundle_count
= 0;
2721 mir_foreach_block(ctx
, _block
) {
2722 midgard_block
*block
= (midgard_block
*) _block
;
2723 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2725 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2727 mir_foreach_block(ctx
, _block
) {
2728 midgard_block
*block
= (midgard_block
*) _block
;
2729 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2730 source_order_bundles
[bundle_idx
++] = bundle
;
2734 int current_bundle
= 0;
2736 /* Midgard prefetches instruction types, so during emission we
2737 * need to lookahead. Unless this is the last instruction, in
2738 * which we return 1. */
2740 mir_foreach_block(ctx
, _block
) {
2741 midgard_block
*block
= (midgard_block
*) _block
;
2742 mir_foreach_bundle_in_block(block
, bundle
) {
2745 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2746 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2748 emit_binary_bundle(ctx
, block
, bundle
, compiled
, lookahead
);
2752 /* TODO: Free deeper */
2753 //util_dynarray_fini(&block->instructions);
2756 free(source_order_bundles
);
2758 /* Report the very first tag executed */
2759 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2761 /* Deal with off-by-one related to the fencepost problem */
2762 program
->work_register_count
= ctx
->work_registers
+ 1;
2763 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2765 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2766 program
->tls_size
= ctx
->tls_size
;
2768 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2769 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
2771 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
2772 unsigned nr_bundles
= 0, nr_ins
= 0;
2774 /* Count instructions and bundles */
2776 mir_foreach_block(ctx
, _block
) {
2777 midgard_block
*block
= (midgard_block
*) _block
;
2778 nr_bundles
+= util_dynarray_num_elements(
2779 &block
->bundles
, midgard_bundle
);
2781 mir_foreach_bundle_in_block(block
, bun
)
2782 nr_ins
+= bun
->instruction_count
;
2785 /* Calculate thread count. There are certain cutoffs by
2786 * register count for thread count */
2788 unsigned nr_registers
= program
->work_register_count
;
2790 unsigned nr_threads
=
2791 (nr_registers
<= 4) ? 4 :
2792 (nr_registers
<= 8) ? 2 :
2797 fprintf(stderr
, "shader%d - %s shader: "
2798 "%u inst, %u bundles, %u quadwords, "
2799 "%u registers, %u threads, %u loops, "
2800 "%u:%u spills:fills\n",
2802 ctx
->is_blend
? "PAN_SHADER_BLEND" :
2803 gl_shader_stage_name(ctx
->stage
),
2804 nr_ins
, nr_bundles
, ctx
->quadword_count
,
2805 nr_registers
, nr_threads
,
2807 ctx
->spills
, ctx
->fills
);