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; \
127 #define M_LOAD(name, T) M_LOAD_STORE(name, false, T)
128 #define M_STORE(name, T) M_LOAD_STORE(name, true, T)
130 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
131 * the corresponding Midgard source */
133 static midgard_vector_alu_src
134 vector_alu_modifiers(bool abs
, bool neg
, bool is_int
,
135 bool half
, bool sext
)
137 /* Figure out how many components there are so we can adjust.
138 * Specifically we want to broadcast the last channel so things like
142 midgard_vector_alu_src alu_src
= {
149 alu_src
.mod
= midgard_int_normal
;
151 /* Sign/zero-extend if needed */
155 midgard_int_sign_extend
156 : midgard_int_zero_extend
;
159 /* These should have been lowered away */
160 assert(!(abs
|| neg
));
162 alu_src
.mod
= (abs
<< 0) | (neg
<< 1);
168 M_LOAD(ld_attr_32
, nir_type_uint32
);
169 M_LOAD(ld_vary_32
, nir_type_uint32
);
170 M_LOAD(ld_ubo_int4
, nir_type_uint32
);
171 M_LOAD(ld_int4
, nir_type_uint32
);
172 M_STORE(st_int4
, nir_type_uint32
);
173 M_LOAD(ld_color_buffer_32u
, nir_type_uint32
);
174 M_STORE(st_vary_32
, nir_type_uint32
);
175 M_LOAD(ld_cubemap_coords
, nir_type_uint32
);
176 M_LOAD(ld_compute_id
, nir_type_uint32
);
178 static midgard_instruction
179 v_branch(bool conditional
, bool invert
)
181 midgard_instruction ins
= {
183 .unit
= ALU_ENAB_BRANCH
,
184 .compact_branch
= true,
186 .conditional
= conditional
,
187 .invert_conditional
= invert
190 .src
= { ~0, ~0, ~0, ~0 },
196 static midgard_branch_extended
197 midgard_create_branch_extended( midgard_condition cond
,
198 midgard_jmp_writeout_op op
,
200 signed quadword_offset
)
202 /* The condition code is actually a LUT describing a function to
203 * combine multiple condition codes. However, we only support a single
204 * condition code at the moment, so we just duplicate over a bunch of
207 uint16_t duplicated_cond
=
217 midgard_branch_extended branch
= {
219 .dest_tag
= dest_tag
,
220 .offset
= quadword_offset
,
221 .cond
= duplicated_cond
228 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
230 ins
->has_constants
= true;
231 memcpy(&ins
->constants
, constants
, 16);
235 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
237 return glsl_count_attribute_slots(type
, false);
240 /* Lower fdot2 to a vector multiplication followed by channel addition */
242 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
244 if (alu
->op
!= nir_op_fdot2
)
247 b
->cursor
= nir_before_instr(&alu
->instr
);
249 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
250 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
252 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
254 nir_ssa_def
*sum
= nir_fadd(b
,
255 nir_channel(b
, product
, 0),
256 nir_channel(b
, product
, 1));
258 /* Replace the fdot2 with this sum */
259 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
263 midgard_nir_lower_fdot2(nir_shader
*shader
)
265 bool progress
= false;
267 nir_foreach_function(function
, shader
) {
268 if (!function
->impl
) continue;
271 nir_builder
*b
= &_b
;
272 nir_builder_init(b
, function
->impl
);
274 nir_foreach_block(block
, function
->impl
) {
275 nir_foreach_instr_safe(instr
, block
) {
276 if (instr
->type
!= nir_instr_type_alu
) continue;
278 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
279 midgard_nir_lower_fdot2_body(b
, alu
);
285 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
292 /* Midgard can't write depth and stencil separately. It has to happen in a
293 * single store operation containing both. Let's add a panfrost specific
294 * intrinsic and turn all depth/stencil stores into a packed depth+stencil
298 midgard_nir_lower_zs_store(nir_shader
*nir
)
300 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
303 nir_variable
*z_var
= NULL
, *s_var
= NULL
;
305 nir_foreach_variable(var
, &nir
->outputs
) {
306 if (var
->data
.location
== FRAG_RESULT_DEPTH
)
308 else if (var
->data
.location
== FRAG_RESULT_STENCIL
)
312 if (!z_var
&& !s_var
)
315 bool progress
= false;
317 nir_foreach_function(function
, nir
) {
318 if (!function
->impl
) continue;
320 nir_intrinsic_instr
*z_store
= NULL
, *s_store
= NULL
, *last_store
= NULL
;
322 nir_foreach_block(block
, function
->impl
) {
323 nir_foreach_instr_safe(instr
, block
) {
324 if (instr
->type
!= nir_instr_type_intrinsic
)
327 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
328 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
331 if (z_var
&& nir_intrinsic_base(intr
) == z_var
->data
.driver_location
) {
337 if (s_var
&& nir_intrinsic_base(intr
) == s_var
->data
.driver_location
) {
345 if (!z_store
&& !s_store
) continue;
348 nir_builder_init(&b
, function
->impl
);
350 b
.cursor
= nir_before_instr(&last_store
->instr
);
352 nir_ssa_def
*zs_store_src
;
354 if (z_store
&& s_store
) {
355 nir_ssa_def
*srcs
[2] = {
356 nir_ssa_for_src(&b
, z_store
->src
[0], 1),
357 nir_ssa_for_src(&b
, s_store
->src
[0], 1),
360 zs_store_src
= nir_vec(&b
, srcs
, 2);
362 zs_store_src
= nir_ssa_for_src(&b
, last_store
->src
[0], 1);
365 nir_intrinsic_instr
*zs_store
;
367 zs_store
= nir_intrinsic_instr_create(b
.shader
,
368 nir_intrinsic_store_zs_output_pan
);
369 zs_store
->src
[0] = nir_src_for_ssa(zs_store_src
);
370 zs_store
->num_components
= z_store
&& s_store
? 2 : 1;
371 nir_intrinsic_set_component(zs_store
, z_store
? 0 : 1);
373 /* Replace the Z and S store by a ZS store */
374 nir_builder_instr_insert(&b
, &zs_store
->instr
);
377 nir_instr_remove(&z_store
->instr
);
380 nir_instr_remove(&s_store
->instr
);
382 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
389 /* Flushes undefined values to zero */
392 optimise_nir(nir_shader
*nir
, unsigned quirks
)
395 unsigned lower_flrp
=
396 (nir
->options
->lower_flrp16
? 16 : 0) |
397 (nir
->options
->lower_flrp32
? 32 : 0) |
398 (nir
->options
->lower_flrp64
? 64 : 0);
400 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
401 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
403 nir_lower_tex_options lower_tex_options
= {
404 .lower_txs_lod
= true,
406 .lower_tex_without_implicit_lod
=
407 (quirks
& MIDGARD_EXPLICIT_LOD
),
409 /* TODO: we have native gradient.. */
413 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
415 /* Must lower fdot2 after tex is lowered */
416 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
418 /* T720 is broken. */
420 if (quirks
& MIDGARD_BROKEN_LOD
)
421 NIR_PASS_V(nir
, midgard_nir_lod_errata
);
426 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
427 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
429 NIR_PASS(progress
, nir
, nir_copy_prop
);
430 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
431 NIR_PASS(progress
, nir
, nir_opt_dce
);
432 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
433 NIR_PASS(progress
, nir
, nir_opt_cse
);
434 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
435 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
436 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
438 if (lower_flrp
!= 0) {
439 bool lower_flrp_progress
= false;
440 NIR_PASS(lower_flrp_progress
,
444 false /* always_precise */,
445 nir
->options
->lower_ffma
);
446 if (lower_flrp_progress
) {
447 NIR_PASS(progress
, nir
,
448 nir_opt_constant_folding
);
452 /* Nothing should rematerialize any flrps, so we only
453 * need to do this lowering once.
458 NIR_PASS(progress
, nir
, nir_opt_undef
);
459 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
461 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
464 nir_var_function_temp
);
466 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
469 /* Must be run at the end to prevent creation of fsin/fcos ops */
470 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
475 NIR_PASS(progress
, nir
, nir_opt_dce
);
476 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
477 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
478 NIR_PASS(progress
, nir
, nir_copy_prop
);
481 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
482 NIR_PASS(progress
, nir
, nir_opt_algebraic_distribute_src_mods
);
484 /* We implement booleans as 32-bit 0/~0 */
485 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
487 /* Now that booleans are lowered, we can run out late opts */
488 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
489 NIR_PASS(progress
, nir
, midgard_nir_cancel_inot
);
491 NIR_PASS(progress
, nir
, nir_copy_prop
);
492 NIR_PASS(progress
, nir
, nir_opt_dce
);
494 /* Take us out of SSA */
495 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
496 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
498 /* We are a vector architecture; write combine where possible */
499 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
500 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
502 NIR_PASS(progress
, nir
, nir_opt_dce
);
505 /* Do not actually emit a load; instead, cache the constant for inlining */
508 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
510 nir_ssa_def def
= instr
->def
;
512 midgard_constants
*consts
= rzalloc(NULL
, midgard_constants
);
514 assert(instr
->def
.num_components
* instr
->def
.bit_size
<= sizeof(*consts
) * 8);
516 #define RAW_CONST_COPY(bits) \
517 nir_const_value_to_array(consts->u##bits, instr->value, \
518 instr->def.num_components, u##bits)
520 switch (instr
->def
.bit_size
) {
534 unreachable("Invalid bit_size for load_const instruction\n");
537 /* Shifted for SSA, +1 for off-by-one */
538 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, consts
);
541 /* Normally constants are embedded implicitly, but for I/O and such we have to
542 * explicitly emit a move with the constant source */
545 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
547 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
549 if (constant_value
) {
550 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), to
);
551 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
552 emit_mir_instruction(ctx
, ins
);
557 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
559 unsigned comp
= src
->swizzle
[0];
561 for (unsigned c
= 1; c
< nr_components
; ++c
) {
562 if (src
->swizzle
[c
] != comp
)
569 #define ALU_CASE(nir, _op) \
571 op = midgard_alu_op_##_op; \
572 assert(src_bitsize == dst_bitsize); \
575 #define ALU_CASE_BCAST(nir, _op, count) \
577 op = midgard_alu_op_##_op; \
578 broadcast_swizzle = count; \
579 assert(src_bitsize == dst_bitsize); \
581 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
582 * special treatment override this anyway. */
584 static midgard_reg_mode
585 reg_mode_for_nir(nir_alu_instr
*instr
)
587 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
589 switch (src_bitsize
) {
591 return midgard_reg_mode_8
;
593 return midgard_reg_mode_16
;
595 return midgard_reg_mode_32
;
597 return midgard_reg_mode_64
;
599 unreachable("Invalid bit size");
603 /* Compare mir_lower_invert */
605 nir_accepts_inot(nir_op op
, unsigned src
)
613 /* Only the condition */
621 mir_copy_src(midgard_instruction
*ins
, nir_alu_instr
*instr
, unsigned i
, unsigned to
, bool *abs
, bool *neg
, bool *not, bool is_int
, unsigned bcast_count
)
623 nir_alu_src src
= instr
->src
[i
];
626 if (pan_has_source_mod(&src
, nir_op_fneg
))
629 if (pan_has_source_mod(&src
, nir_op_fabs
))
633 if (nir_accepts_inot(instr
->op
, i
) && pan_has_source_mod(&src
, nir_op_inot
))
636 unsigned bits
= nir_src_bit_size(src
.src
);
638 ins
->src
[to
] = nir_src_index(NULL
, &src
.src
);
639 ins
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
] | bits
;
641 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
642 ins
->swizzle
[to
][c
] = src
.swizzle
[
643 (!bcast_count
|| c
< bcast_count
) ? c
:
649 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
651 nir_dest
*dest
= &instr
->dest
.dest
;
653 if (dest
->is_ssa
&& BITSET_TEST(ctx
->already_emitted
, dest
->ssa
.index
))
656 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
657 * is handled elsewhere */
659 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
660 midgard_emit_derivatives(ctx
, instr
);
664 bool is_ssa
= dest
->is_ssa
;
666 unsigned nr_components
= nir_dest_num_components(*dest
);
667 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
670 /* Number of components valid to check for the instruction (the rest
671 * will be forced to the last), or 0 to use as-is. Relevant as
672 * ball-type instructions have a channel count in NIR but are all vec4
675 unsigned broadcast_swizzle
= 0;
677 /* What register mode should we operate in? */
678 midgard_reg_mode reg_mode
=
679 reg_mode_for_nir(instr
);
681 /* Do we need a destination override? Used for inline
684 midgard_dest_override dest_override
=
685 midgard_dest_override_none
;
687 /* Should we use a smaller respective source and sign-extend? */
689 bool half_1
= false, sext_1
= false;
690 bool half_2
= false, sext_2
= false;
692 /* Should we swap arguments? */
693 bool flip_src12
= false;
695 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
696 unsigned dst_bitsize
= nir_dest_bit_size(*dest
);
699 ALU_CASE(fadd
, fadd
);
700 ALU_CASE(fmul
, fmul
);
701 ALU_CASE(fmin
, fmin
);
702 ALU_CASE(fmax
, fmax
);
703 ALU_CASE(imin
, imin
);
704 ALU_CASE(imax
, imax
);
705 ALU_CASE(umin
, umin
);
706 ALU_CASE(umax
, umax
);
707 ALU_CASE(ffloor
, ffloor
);
708 ALU_CASE(fround_even
, froundeven
);
709 ALU_CASE(ftrunc
, ftrunc
);
710 ALU_CASE(fceil
, fceil
);
711 ALU_CASE(fdot3
, fdot3
);
712 ALU_CASE(fdot4
, fdot4
);
713 ALU_CASE(iadd
, iadd
);
714 ALU_CASE(isub
, isub
);
715 ALU_CASE(imul
, imul
);
717 /* Zero shoved as second-arg */
718 ALU_CASE(iabs
, iabsdiff
);
722 ALU_CASE(feq32
, feq
);
723 ALU_CASE(fne32
, fne
);
724 ALU_CASE(flt32
, flt
);
725 ALU_CASE(ieq32
, ieq
);
726 ALU_CASE(ine32
, ine
);
727 ALU_CASE(ilt32
, ilt
);
728 ALU_CASE(ult32
, ult
);
730 /* We don't have a native b2f32 instruction. Instead, like many
731 * GPUs, we exploit booleans as 0/~0 for false/true, and
732 * correspondingly AND
733 * by 1.0 to do the type conversion. For the moment, prime us
736 * iand [whatever], #0
738 * At the end of emit_alu (as MIR), we'll fix-up the constant
741 ALU_CASE(b2f32
, iand
);
742 ALU_CASE(b2i32
, iand
);
744 /* Likewise, we don't have a dedicated f2b32 instruction, but
745 * we can do a "not equal to 0.0" test. */
747 ALU_CASE(f2b32
, fne
);
748 ALU_CASE(i2b32
, ine
);
750 ALU_CASE(frcp
, frcp
);
751 ALU_CASE(frsq
, frsqrt
);
752 ALU_CASE(fsqrt
, fsqrt
);
753 ALU_CASE(fexp2
, fexp2
);
754 ALU_CASE(flog2
, flog2
);
756 ALU_CASE(f2i64
, f2i_rtz
);
757 ALU_CASE(f2u64
, f2u_rtz
);
758 ALU_CASE(i2f64
, i2f_rtz
);
759 ALU_CASE(u2f64
, u2f_rtz
);
761 ALU_CASE(f2i32
, f2i_rtz
);
762 ALU_CASE(f2u32
, f2u_rtz
);
763 ALU_CASE(i2f32
, i2f_rtz
);
764 ALU_CASE(u2f32
, u2f_rtz
);
766 ALU_CASE(f2i16
, f2i_rtz
);
767 ALU_CASE(f2u16
, f2u_rtz
);
768 ALU_CASE(i2f16
, i2f_rtz
);
769 ALU_CASE(u2f16
, u2f_rtz
);
771 ALU_CASE(fsin
, fsin
);
772 ALU_CASE(fcos
, fcos
);
774 /* We'll get 0 in the second arg, so:
775 * ~a = ~(a | 0) = nor(a, 0) */
776 ALU_CASE(inot
, inor
);
777 ALU_CASE(iand
, iand
);
779 ALU_CASE(ixor
, ixor
);
780 ALU_CASE(ishl
, ishl
);
781 ALU_CASE(ishr
, iasr
);
782 ALU_CASE(ushr
, ilsr
);
784 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
785 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
786 ALU_CASE(b32all_fequal4
, fball_eq
);
788 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
789 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
790 ALU_CASE(b32any_fnequal4
, fbany_neq
);
792 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
793 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
794 ALU_CASE(b32all_iequal4
, iball_eq
);
796 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
797 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
798 ALU_CASE(b32any_inequal4
, ibany_neq
);
800 /* Source mods will be shoved in later */
801 ALU_CASE(fabs
, fmov
);
802 ALU_CASE(fneg
, fmov
);
803 ALU_CASE(fsat
, fmov
);
804 ALU_CASE(fsat_signed
, fmov
);
805 ALU_CASE(fclamp_pos
, fmov
);
807 /* For size conversion, we use a move. Ideally though we would squash
808 * these ops together; maybe that has to happen after in NIR as part of
809 * propagation...? An earlier algebraic pass ensured we step down by
810 * only / exactly one size. If stepping down, we use a dest override to
811 * reduce the size; if stepping up, we use a larger-sized move with a
812 * half source and a sign/zero-extension modifier */
818 /* If we end up upscale, we'll need a sign-extend on the
819 * operand (the second argument) */
830 if (instr
->op
== nir_op_f2f16
|| instr
->op
== nir_op_f2f32
||
831 instr
->op
== nir_op_f2f64
)
832 op
= midgard_alu_op_fmov
;
834 op
= midgard_alu_op_imov
;
836 if (dst_bitsize
== (src_bitsize
* 2)) {
840 /* Use a greater register mode */
842 } else if (src_bitsize
== (dst_bitsize
* 2)) {
843 /* Converting down */
844 dest_override
= midgard_dest_override_lower
;
850 /* For greater-or-equal, we lower to less-or-equal and flip the
858 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
859 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
860 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
861 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
868 case nir_op_b32csel
: {
869 /* Midgard features both fcsel and icsel, depending on
870 * the type of the arguments/output. However, as long
871 * as we're careful we can _always_ use icsel and
872 * _never_ need fcsel, since the latter does additional
873 * floating-point-specific processing whereas the
874 * former just moves bits on the wire. It's not obvious
875 * why these are separate opcodes, save for the ability
876 * to do things like sat/pos/abs/neg for free */
878 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
879 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
885 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
890 /* Midgard can perform certain modifiers on output of an ALU op */
894 bool abs
[4] = { false };
895 bool neg
[4] = { false };
896 bool is_int
= midgard_is_integer_op(op
);
898 if (midgard_is_integer_out_op(op
)) {
899 outmod
= midgard_outmod_int_wrap
;
900 } else if (instr
->op
== nir_op_fsat
) {
901 outmod
= midgard_outmod_sat
;
902 } else if (instr
->op
== nir_op_fsat_signed
) {
903 outmod
= midgard_outmod_sat_signed
;
904 } else if (instr
->op
== nir_op_fclamp_pos
) {
905 outmod
= midgard_outmod_pos
;
908 /* Fetch unit, quirks, etc information */
909 unsigned opcode_props
= alu_opcode_props
[op
].props
;
910 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
912 midgard_instruction ins
= {
914 .dest
= nir_dest_index(dest
),
915 .dest_type
= nir_op_infos
[instr
->op
].output_type
916 | nir_dest_bit_size(*dest
),
919 for (unsigned i
= nr_inputs
; i
< ARRAY_SIZE(ins
.src
); ++i
)
922 if (quirk_flipped_r24
) {
924 mir_copy_src(&ins
, instr
, 0, 1, &abs
[1], &neg
[1], &ins
.src_invert
[1], is_int
, broadcast_swizzle
);
926 for (unsigned i
= 0; i
< nr_inputs
; ++i
) {
929 if (instr
->op
== nir_op_b32csel
) {
930 /* The condition is the first argument; move
931 * the other arguments up one to be a binary
932 * instruction for Midgard with the condition
941 } else if (flip_src12
) {
945 mir_copy_src(&ins
, instr
, i
, to
, &abs
[to
], &neg
[to
], &ins
.src_invert
[to
], is_int
, broadcast_swizzle
);
947 /* (!c) ? a : b = c ? b : a */
948 if (instr
->op
== nir_op_b32csel
&& ins
.src_invert
[2]) {
949 ins
.src_invert
[2] = false;
955 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
956 /* Lowered to move */
957 if (instr
->op
== nir_op_fneg
)
960 if (instr
->op
== nir_op_fabs
)
964 ins
.mask
= mask_of(nr_components
);
966 midgard_vector_alu alu
= {
968 .reg_mode
= reg_mode
,
969 .dest_override
= dest_override
,
972 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(abs
[0], neg
[0], is_int
, half_1
, sext_1
)),
973 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(abs
[1], neg
[1], is_int
, half_2
, sext_2
)),
976 /* Apply writemask if non-SSA, keeping in mind that we can't write to
977 * components that don't exist. Note modifier => SSA => !reg => no
978 * writemask, so we don't have to worry about writemasks here.*/
981 ins
.mask
&= instr
->dest
.write_mask
;
985 /* Arrange for creation of iandnot/iornot */
986 if (ins
.src_invert
[0] && !ins
.src_invert
[1]) {
988 ins
.src_invert
[0] = false;
989 ins
.src_invert
[1] = true;
992 /* Late fixup for emulated instructions */
994 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
995 /* Presently, our second argument is an inline #0 constant.
996 * Switch over to an embedded 1.0 constant (that can't fit
997 * inline, since we're 32-bit, not 16-bit like the inline
1000 ins
.has_inline_constant
= false;
1001 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1002 ins
.src_types
[1] = nir_type_float32
;
1003 ins
.has_constants
= true;
1005 if (instr
->op
== nir_op_b2f32
)
1006 ins
.constants
.f32
[0] = 1.0f
;
1008 ins
.constants
.i32
[0] = 1;
1010 for (unsigned c
= 0; c
< 16; ++c
)
1011 ins
.swizzle
[1][c
] = 0;
1012 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1013 /* Lots of instructions need a 0 plonked in */
1014 ins
.has_inline_constant
= false;
1015 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1016 ins
.src_types
[1] = nir_type_uint32
;
1017 ins
.has_constants
= true;
1018 ins
.constants
.u32
[0] = 0;
1020 for (unsigned c
= 0; c
< 16; ++c
)
1021 ins
.swizzle
[1][c
] = 0;
1024 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1025 /* To avoid duplicating the lookup tables (probably), true LUT
1026 * instructions can only operate as if they were scalars. Lower
1027 * them here by changing the component. */
1029 unsigned orig_mask
= ins
.mask
;
1031 unsigned swizzle_back
[MIR_VEC_COMPONENTS
];
1032 memcpy(&swizzle_back
, ins
.swizzle
[0], sizeof(swizzle_back
));
1034 for (int i
= 0; i
< nr_components
; ++i
) {
1035 /* Mask the associated component, dropping the
1036 * instruction if needed */
1039 ins
.mask
&= orig_mask
;
1044 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1045 ins
.swizzle
[0][j
] = swizzle_back
[i
]; /* Pull from the correct component */
1047 emit_mir_instruction(ctx
, ins
);
1050 emit_mir_instruction(ctx
, ins
);
1057 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1059 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1060 unsigned nir_mask
= 0;
1064 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1065 dsize
= nir_dest_bit_size(intr
->dest
);
1067 nir_mask
= nir_intrinsic_write_mask(intr
);
1071 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1072 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1073 mir_set_bytemask(ins
, bytemask
);
1076 ins
->load_64
= true;
1079 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1080 * optimized) versions of UBO #0 */
1082 static midgard_instruction
*
1084 compiler_context
*ctx
,
1088 nir_src
*indirect_offset
,
1089 unsigned indirect_shift
,
1092 /* TODO: half-floats */
1094 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1095 ins
.constants
.u32
[0] = offset
;
1097 if (instr
->type
== nir_instr_type_intrinsic
)
1098 mir_set_intr_mask(instr
, &ins
, true);
1100 if (indirect_offset
) {
1101 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1102 ins
.src_types
[2] = nir_type_uint32
;
1103 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1105 ins
.load_store
.arg_2
= 0x1E;
1108 ins
.load_store
.arg_1
= index
;
1110 return emit_mir_instruction(ctx
, ins
);
1113 /* Globals are like UBOs if you squint. And shared memory is like globals if
1114 * you squint even harder */
1118 compiler_context
*ctx
,
1127 midgard_instruction ins
;
1130 ins
= m_ld_int4(srcdest
, 0);
1132 ins
= m_st_int4(srcdest
, 0);
1134 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1135 mir_set_intr_mask(instr
, &ins
, is_read
);
1137 emit_mir_instruction(ctx
, ins
);
1142 compiler_context
*ctx
,
1143 unsigned dest
, unsigned offset
,
1144 unsigned nr_comp
, unsigned component
,
1145 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1147 /* XXX: Half-floats? */
1148 /* TODO: swizzle, mask */
1150 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1151 ins
.mask
= mask_of(nr_comp
);
1153 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1154 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1156 midgard_varying_parameter p
= {
1158 .interpolation
= midgard_interp_default
,
1163 memcpy(&u
, &p
, sizeof(p
));
1164 ins
.load_store
.varying_parameters
= u
;
1166 if (indirect_offset
) {
1167 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1168 ins
.src_types
[2] = nir_type_uint32
;
1170 ins
.load_store
.arg_2
= 0x1E;
1172 ins
.load_store
.arg_1
= 0x9E;
1174 /* Use the type appropriate load */
1178 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1181 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1183 case nir_type_float
:
1184 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1187 unreachable("Attempted to load unknown type");
1191 emit_mir_instruction(ctx
, ins
);
1196 compiler_context
*ctx
,
1197 unsigned dest
, unsigned offset
,
1198 unsigned nr_comp
, nir_alu_type t
)
1200 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1201 ins
.load_store
.arg_1
= 0x1E;
1202 ins
.load_store
.arg_2
= 0x1E;
1203 ins
.mask
= mask_of(nr_comp
);
1205 /* Use the type appropriate load */
1209 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1212 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1214 case nir_type_float
:
1215 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1218 unreachable("Attempted to load unknown type");
1222 emit_mir_instruction(ctx
, ins
);
1226 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1227 unsigned nr_components
, unsigned offset
)
1231 /* Figure out which uniform this is */
1232 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1233 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1235 unsigned dest
= nir_dest_index(&nir_dest
);
1237 /* Sysvals are prefix uniforms */
1238 unsigned uniform
= ((uintptr_t) val
) - 1;
1240 /* Emit the read itself -- this is never indirect */
1241 midgard_instruction
*ins
=
1242 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1244 ins
->mask
= mask_of(nr_components
);
1248 compute_builtin_arg(nir_op op
)
1251 case nir_intrinsic_load_work_group_id
:
1253 case nir_intrinsic_load_local_invocation_id
:
1256 unreachable("Invalid compute paramater loaded");
1261 emit_fragment_store(compiler_context
*ctx
, unsigned src
, enum midgard_rt_id rt
)
1263 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1265 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1269 emit_explicit_constant(ctx
, src
, src
);
1271 struct midgard_instruction ins
=
1272 v_branch(false, false);
1274 ins
.writeout
= true;
1276 /* Add dependencies */
1278 ins
.src_types
[0] = nir_type_uint32
;
1279 ins
.constants
.u32
[0] = rt
== MIDGARD_ZS_RT
?
1280 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1282 /* Emit the branch */
1283 br
= emit_mir_instruction(ctx
, ins
);
1284 schedule_barrier(ctx
);
1285 ctx
->writeout_branch
[rt
] = br
;
1287 /* Push our current location = current block count - 1 = where we'll
1288 * jump to. Maybe a bit too clever for my own good */
1290 br
->branch
.target_block
= ctx
->block_count
- 1;
1294 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1296 unsigned reg
= nir_dest_index(&instr
->dest
);
1297 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1298 ins
.mask
= mask_of(3);
1299 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1300 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1301 emit_mir_instruction(ctx
, ins
);
1305 vertex_builtin_arg(nir_op op
)
1308 case nir_intrinsic_load_vertex_id
:
1309 return PAN_VERTEX_ID
;
1310 case nir_intrinsic_load_instance_id
:
1311 return PAN_INSTANCE_ID
;
1313 unreachable("Invalid vertex builtin");
1318 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1320 unsigned reg
= nir_dest_index(&instr
->dest
);
1321 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1325 emit_control_barrier(compiler_context
*ctx
)
1327 midgard_instruction ins
= {
1328 .type
= TAG_TEXTURE_4
,
1329 .src
= { ~0, ~0, ~0, ~0 },
1331 .op
= TEXTURE_OP_BARRIER
,
1333 /* TODO: optimize */
1334 .barrier_buffer
= 1,
1339 emit_mir_instruction(ctx
, ins
);
1342 static const nir_variable
*
1343 search_var(struct exec_list
*vars
, unsigned driver_loc
)
1345 nir_foreach_variable(var
, vars
) {
1346 if (var
->data
.driver_location
== driver_loc
)
1354 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1356 unsigned offset
= 0, reg
;
1358 switch (instr
->intrinsic
) {
1359 case nir_intrinsic_discard_if
:
1360 case nir_intrinsic_discard
: {
1361 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1362 struct midgard_instruction discard
= v_branch(conditional
, false);
1363 discard
.branch
.target_type
= TARGET_DISCARD
;
1366 discard
.src
[0] = nir_src_index(ctx
, &instr
->src
[0]);
1367 discard
.src_types
[0] = nir_type_uint32
;
1370 emit_mir_instruction(ctx
, discard
);
1371 schedule_barrier(ctx
);
1376 case nir_intrinsic_load_uniform
:
1377 case nir_intrinsic_load_ubo
:
1378 case nir_intrinsic_load_global
:
1379 case nir_intrinsic_load_shared
:
1380 case nir_intrinsic_load_input
:
1381 case nir_intrinsic_load_interpolated_input
: {
1382 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1383 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1384 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1385 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1386 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1387 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1389 /* Get the base type of the intrinsic */
1390 /* TODO: Infer type? Does it matter? */
1392 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1393 (is_interp
) ? nir_type_float
:
1394 nir_intrinsic_type(instr
);
1396 t
= nir_alu_type_get_base_type(t
);
1398 if (!(is_ubo
|| is_global
)) {
1399 offset
= nir_intrinsic_base(instr
);
1402 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1404 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1406 bool direct
= nir_src_is_const(*src_offset
);
1407 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1410 offset
+= nir_src_as_uint(*src_offset
);
1412 /* We may need to apply a fractional offset */
1413 int component
= (is_flat
|| is_interp
) ?
1414 nir_intrinsic_component(instr
) : 0;
1415 reg
= nir_dest_index(&instr
->dest
);
1417 if (is_uniform
&& !ctx
->is_blend
) {
1418 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1419 } else if (is_ubo
) {
1420 nir_src index
= instr
->src
[0];
1422 /* TODO: Is indirect block number possible? */
1423 assert(nir_src_is_const(index
));
1425 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1426 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1427 } else if (is_global
|| is_shared
) {
1428 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1429 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1430 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
, is_flat
);
1431 } else if (ctx
->is_blend
) {
1432 /* For blend shaders, load the input color, which is
1433 * preloaded to r0 */
1435 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), reg
);
1436 emit_mir_instruction(ctx
, move
);
1437 schedule_barrier(ctx
);
1438 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1439 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1441 DBG("Unknown load\n");
1448 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1449 case nir_intrinsic_load_barycentric_pixel
:
1450 case nir_intrinsic_load_barycentric_centroid
:
1453 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1455 case nir_intrinsic_load_raw_output_pan
:
1456 case nir_intrinsic_load_output_u8_as_fp16_pan
:
1457 reg
= nir_dest_index(&instr
->dest
);
1458 assert(ctx
->is_blend
);
1460 /* T720 and below use different blend opcodes with slightly
1461 * different semantics than T760 and up */
1463 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1464 bool old_blend
= ctx
->quirks
& MIDGARD_OLD_BLEND
;
1466 if (instr
->intrinsic
== nir_intrinsic_load_output_u8_as_fp16_pan
) {
1467 ld
.load_store
.op
= old_blend
?
1468 midgard_op_ld_color_buffer_u8_as_fp16_old
:
1469 midgard_op_ld_color_buffer_u8_as_fp16
;
1472 ld
.load_store
.address
= 1;
1473 ld
.load_store
.arg_2
= 0x1E;
1476 for (unsigned c
= 2; c
< 16; ++c
)
1477 ld
.swizzle
[0][c
] = 0;
1480 emit_mir_instruction(ctx
, ld
);
1483 case nir_intrinsic_load_blend_const_color_rgba
: {
1484 assert(ctx
->is_blend
);
1485 reg
= nir_dest_index(&instr
->dest
);
1487 /* Blend constants are embedded directly in the shader and
1488 * patched in, so we use some magic routing */
1490 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1491 ins
.has_constants
= true;
1492 ins
.has_blend_constant
= true;
1493 emit_mir_instruction(ctx
, ins
);
1497 case nir_intrinsic_store_zs_output_pan
: {
1498 assert(ctx
->stage
== MESA_SHADER_FRAGMENT
);
1499 emit_fragment_store(ctx
, nir_src_index(ctx
, &instr
->src
[0]),
1502 midgard_instruction
*br
= ctx
->writeout_branch
[MIDGARD_ZS_RT
];
1504 if (!nir_intrinsic_component(instr
))
1505 br
->writeout_depth
= true;
1506 if (nir_intrinsic_component(instr
) ||
1507 instr
->num_components
)
1508 br
->writeout_stencil
= true;
1509 assert(br
->writeout_depth
| br
->writeout_stencil
);
1513 case nir_intrinsic_store_output
:
1514 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1516 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1518 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1520 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1521 const nir_variable
*var
;
1522 enum midgard_rt_id rt
;
1524 var
= search_var(&ctx
->nir
->outputs
,
1525 nir_intrinsic_base(instr
));
1527 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1528 rt
= MIDGARD_COLOR_RT0
;
1529 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1530 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1535 emit_fragment_store(ctx
, reg
, rt
);
1536 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1537 /* We should have been vectorized, though we don't
1538 * currently check that st_vary is emitted only once
1539 * per slot (this is relevant, since there's not a mask
1540 * parameter available on the store [set to 0 by the
1541 * blob]). We do respect the component by adjusting the
1542 * swizzle. If this is a constant source, we'll need to
1543 * emit that explicitly. */
1545 emit_explicit_constant(ctx
, reg
, reg
);
1547 unsigned dst_component
= nir_intrinsic_component(instr
);
1548 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1550 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1551 st
.load_store
.arg_1
= 0x9E;
1552 st
.load_store
.arg_2
= 0x1E;
1554 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1557 st
.load_store
.op
= midgard_op_st_vary_32u
;
1560 st
.load_store
.op
= midgard_op_st_vary_32i
;
1562 case nir_type_float
:
1563 st
.load_store
.op
= midgard_op_st_vary_32
;
1566 unreachable("Attempted to store unknown type");
1570 /* nir_intrinsic_component(store_intr) encodes the
1571 * destination component start. Source component offset
1572 * adjustment is taken care of in
1573 * install_registers_instr(), when offset_swizzle() is
1576 unsigned src_component
= COMPONENT_X
;
1578 assert(nr_comp
> 0);
1579 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1580 st
.swizzle
[0][i
] = src_component
;
1581 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1585 emit_mir_instruction(ctx
, st
);
1587 DBG("Unknown store\n");
1593 /* Special case of store_output for lowered blend shaders */
1594 case nir_intrinsic_store_raw_output_pan
:
1595 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1596 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1598 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1599 /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
1600 * reg = r0.x. We want to splatter. So we can do a 32-bit move
1603 * imov r0.xyzw, r0.xxxx
1606 unsigned expanded
= make_compiler_temp(ctx
);
1608 midgard_instruction splatter
= v_mov(reg
, expanded
);
1610 for (unsigned c
= 0; c
< 16; ++c
)
1611 splatter
.swizzle
[1][c
] = 0;
1613 emit_mir_instruction(ctx
, splatter
);
1614 emit_fragment_store(ctx
, expanded
, ctx
->blend_rt
);
1616 emit_fragment_store(ctx
, reg
, ctx
->blend_rt
);
1620 case nir_intrinsic_store_global
:
1621 case nir_intrinsic_store_shared
:
1622 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1623 emit_explicit_constant(ctx
, reg
, reg
);
1625 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1628 case nir_intrinsic_load_ssbo_address
:
1629 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1632 case nir_intrinsic_get_buffer_size
:
1633 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1636 case nir_intrinsic_load_viewport_scale
:
1637 case nir_intrinsic_load_viewport_offset
:
1638 case nir_intrinsic_load_num_work_groups
:
1639 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1640 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1643 case nir_intrinsic_load_work_group_id
:
1644 case nir_intrinsic_load_local_invocation_id
:
1645 emit_compute_builtin(ctx
, instr
);
1648 case nir_intrinsic_load_vertex_id
:
1649 case nir_intrinsic_load_instance_id
:
1650 emit_vertex_builtin(ctx
, instr
);
1653 case nir_intrinsic_memory_barrier_buffer
:
1654 case nir_intrinsic_memory_barrier_shared
:
1657 case nir_intrinsic_control_barrier
:
1658 schedule_barrier(ctx
);
1659 emit_control_barrier(ctx
);
1660 schedule_barrier(ctx
);
1664 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1671 midgard_tex_format(enum glsl_sampler_dim dim
)
1674 case GLSL_SAMPLER_DIM_1D
:
1675 case GLSL_SAMPLER_DIM_BUF
:
1678 case GLSL_SAMPLER_DIM_2D
:
1679 case GLSL_SAMPLER_DIM_EXTERNAL
:
1680 case GLSL_SAMPLER_DIM_RECT
:
1683 case GLSL_SAMPLER_DIM_3D
:
1686 case GLSL_SAMPLER_DIM_CUBE
:
1687 return MALI_TEX_CUBE
;
1690 DBG("Unknown sampler dim type\n");
1696 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1700 pan_attach_constant_bias(
1701 compiler_context
*ctx
,
1703 midgard_texture_word
*word
)
1705 /* To attach as constant, it has to *be* constant */
1707 if (!nir_src_is_const(lod
))
1710 float f
= nir_src_as_float(lod
);
1712 /* Break into fixed-point */
1714 float lod_frac
= f
- lod_int
;
1716 /* Carry over negative fractions */
1717 if (lod_frac
< 0.0) {
1723 word
->bias
= float_to_ubyte(lod_frac
);
1724 word
->bias_int
= lod_int
;
1730 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1731 unsigned midgard_texop
)
1734 //assert (!instr->sampler);
1736 int texture_index
= instr
->texture_index
;
1737 int sampler_index
= texture_index
;
1739 nir_alu_type dest_base
= nir_alu_type_get_base_type(instr
->dest_type
);
1740 nir_alu_type dest_type
= dest_base
| nir_dest_bit_size(instr
->dest
);
1742 midgard_instruction ins
= {
1743 .type
= TAG_TEXTURE_4
,
1745 .dest
= nir_dest_index(&instr
->dest
),
1746 .src
= { ~0, ~0, ~0, ~0 },
1747 .dest_type
= dest_type
,
1748 .swizzle
= SWIZZLE_IDENTITY_4
,
1750 .op
= midgard_texop
,
1751 .format
= midgard_tex_format(instr
->sampler_dim
),
1752 .texture_handle
= texture_index
,
1753 .sampler_handle
= sampler_index
,
1754 .shadow
= instr
->is_shadow
,
1758 if (instr
->is_shadow
&& !instr
->is_new_style_shadow
)
1759 for (int i
= 0; i
< 4; ++i
)
1760 ins
.swizzle
[0][i
] = COMPONENT_X
;
1762 /* We may need a temporary for the coordinate */
1764 bool needs_temp_coord
=
1765 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
1766 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
1769 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
1771 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1772 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1773 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1774 unsigned sz
= nir_src_bit_size(instr
->src
[i
].src
);
1775 nir_alu_type T
= nir_tex_instr_src_type(instr
, i
) | sz
;
1777 switch (instr
->src
[i
].src_type
) {
1778 case nir_tex_src_coord
: {
1779 emit_explicit_constant(ctx
, index
, index
);
1781 unsigned coord_mask
= mask_of(instr
->coord_components
);
1783 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
1786 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
1788 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1789 /* texelFetch is undefined on samplerCube */
1790 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1792 /* For cubemaps, we use a special ld/st op to
1793 * select the face and copy the xy into the
1794 * texture register */
1796 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
1798 ld
.src_types
[1] = T
;
1799 ld
.mask
= 0x3; /* xy */
1800 ld
.load_store
.arg_1
= 0x20;
1801 ld
.swizzle
[1][3] = COMPONENT_X
;
1802 emit_mir_instruction(ctx
, ld
);
1805 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1806 ins
.swizzle
[1][3] = COMPONENT_X
;
1807 } else if (needs_temp_coord
) {
1808 /* mov coord_temp, coords */
1809 midgard_instruction mov
= v_mov(index
, coords
);
1810 mov
.mask
= coord_mask
;
1813 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
1815 emit_mir_instruction(ctx
, mov
);
1820 ins
.src
[1] = coords
;
1821 ins
.src_types
[1] = T
;
1823 /* Texelfetch coordinates uses all four elements
1824 * (xyz/index) regardless of texture dimensionality,
1825 * which means it's necessary to zero the unused
1826 * components to keep everything happy */
1828 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1829 /* mov index.zw, #0, or generalized */
1830 midgard_instruction mov
=
1831 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
1832 mov
.has_constants
= true;
1833 mov
.mask
= coord_mask
^ 0xF;
1834 emit_mir_instruction(ctx
, mov
);
1837 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1838 /* Array component in w but NIR wants it in z,
1839 * but if we have a temp coord we already fixed
1842 if (nr_components
== 3) {
1843 ins
.swizzle
[1][2] = COMPONENT_Z
;
1844 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
1845 } else if (nr_components
== 2) {
1847 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1848 ins
.swizzle
[1][3] = COMPONENT_X
;
1850 unreachable("Invalid texture 2D components");
1853 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1855 ins
.swizzle
[1][2] = COMPONENT_Z
;
1856 ins
.swizzle
[1][3] = COMPONENT_W
;
1862 case nir_tex_src_bias
:
1863 case nir_tex_src_lod
: {
1864 /* Try as a constant if we can */
1866 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1867 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1870 ins
.texture
.lod_register
= true;
1872 ins
.src_types
[2] = T
;
1874 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1875 ins
.swizzle
[2][c
] = COMPONENT_X
;
1877 emit_explicit_constant(ctx
, index
, index
);
1882 case nir_tex_src_offset
: {
1883 ins
.texture
.offset_register
= true;
1885 ins
.src_types
[3] = T
;
1887 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1888 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
1890 emit_explicit_constant(ctx
, index
, index
);
1894 case nir_tex_src_comparator
: {
1895 unsigned comp
= COMPONENT_Z
;
1897 /* mov coord_temp.foo, coords */
1898 midgard_instruction mov
= v_mov(index
, coords
);
1899 mov
.mask
= 1 << comp
;
1901 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
1902 mov
.swizzle
[1][i
] = COMPONENT_X
;
1904 emit_mir_instruction(ctx
, mov
);
1909 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
1915 emit_mir_instruction(ctx
, ins
);
1919 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1921 switch (instr
->op
) {
1924 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1927 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1930 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1933 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
1936 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
1943 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1945 switch (instr
->type
) {
1946 case nir_jump_break
: {
1947 /* Emit a branch out of the loop */
1948 struct midgard_instruction br
= v_branch(false, false);
1949 br
.branch
.target_type
= TARGET_BREAK
;
1950 br
.branch
.target_break
= ctx
->current_loop_depth
;
1951 emit_mir_instruction(ctx
, br
);
1956 DBG("Unknown jump type %d\n", instr
->type
);
1962 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1964 switch (instr
->type
) {
1965 case nir_instr_type_load_const
:
1966 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1969 case nir_instr_type_intrinsic
:
1970 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1973 case nir_instr_type_alu
:
1974 emit_alu(ctx
, nir_instr_as_alu(instr
));
1977 case nir_instr_type_tex
:
1978 emit_tex(ctx
, nir_instr_as_tex(instr
));
1981 case nir_instr_type_jump
:
1982 emit_jump(ctx
, nir_instr_as_jump(instr
));
1985 case nir_instr_type_ssa_undef
:
1990 DBG("Unhandled instruction type\n");
1996 /* ALU instructions can inline or embed constants, which decreases register
1997 * pressure and saves space. */
1999 #define CONDITIONAL_ATTACH(idx) { \
2000 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
2003 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
2004 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2009 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
2011 mir_foreach_instr_in_block(block
, alu
) {
2012 /* Other instructions cannot inline constants */
2013 if (alu
->type
!= TAG_ALU_4
) continue;
2014 if (alu
->compact_branch
) continue;
2016 /* If there is already a constant here, we can do nothing */
2017 if (alu
->has_constants
) continue;
2019 CONDITIONAL_ATTACH(0);
2021 if (!alu
->has_constants
) {
2022 CONDITIONAL_ATTACH(1)
2023 } else if (!alu
->inline_constant
) {
2024 /* Corner case: _two_ vec4 constants, for instance with a
2025 * csel. For this case, we can only use a constant
2026 * register for one, we'll have to emit a move for the
2029 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2030 unsigned scratch
= make_compiler_temp(ctx
);
2033 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2034 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2036 /* Set the source */
2037 alu
->src
[1] = scratch
;
2039 /* Inject us -before- the last instruction which set r31 */
2040 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2046 /* Midgard supports two types of constants, embedded constants (128-bit) and
2047 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2048 * constants can be demoted to inline constants, for space savings and
2049 * sometimes a performance boost */
2052 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2054 mir_foreach_instr_in_block(block
, ins
) {
2055 if (!ins
->has_constants
) continue;
2056 if (ins
->has_inline_constant
) continue;
2058 /* Blend constants must not be inlined by definition */
2059 if (ins
->has_blend_constant
) continue;
2061 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2062 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2063 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2065 if (!(is_16
|| is_32
))
2068 /* src1 cannot be an inline constant due to encoding
2069 * restrictions. So, if possible we try to flip the arguments
2072 int op
= ins
->alu
.op
;
2074 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
) &&
2075 alu_opcode_props
[op
].props
& OP_COMMUTES
) {
2079 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2080 /* Extract the source information */
2082 midgard_vector_alu_src
*src
;
2083 int q
= ins
->alu
.src2
;
2084 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2087 /* Component is from the swizzle. Take a nonzero component */
2089 unsigned first_comp
= ffs(ins
->mask
) - 1;
2090 unsigned component
= ins
->swizzle
[1][first_comp
];
2092 /* Scale constant appropriately, if we can legally */
2093 uint16_t scaled_constant
= 0;
2096 scaled_constant
= ins
->constants
.u16
[component
];
2097 } else if (midgard_is_integer_op(op
)) {
2098 scaled_constant
= ins
->constants
.u32
[component
];
2100 /* Constant overflow after resize */
2101 if (scaled_constant
!= ins
->constants
.u32
[component
])
2104 float original
= ins
->constants
.f32
[component
];
2105 scaled_constant
= _mesa_float_to_half(original
);
2107 /* Check for loss of precision. If this is
2108 * mediump, we don't care, but for a highp
2109 * shader, we need to pay attention. NIR
2110 * doesn't yet tell us which mode we're in!
2111 * Practically this prevents most constants
2112 * from being inlined, sadly. */
2114 float fp32
= _mesa_half_to_float(scaled_constant
);
2116 if (fp32
!= original
)
2120 /* We don't know how to handle these with a constant */
2122 if (mir_nontrivial_source2_mod_simple(ins
) || src
->rep_low
|| src
->rep_high
) {
2123 DBG("Bailing inline constant...\n");
2127 /* Make sure that the constant is not itself a vector
2128 * by checking if all accessed values are the same. */
2130 const midgard_constants
*cons
= &ins
->constants
;
2131 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2133 bool is_vector
= false;
2134 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2136 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2137 /* We only care if this component is actually used */
2138 if (!(mask
& (1 << c
)))
2141 uint32_t test
= is_16
?
2142 cons
->u16
[ins
->swizzle
[1][c
]] :
2143 cons
->u32
[ins
->swizzle
[1][c
]];
2145 if (test
!= value
) {
2154 /* Get rid of the embedded constant */
2155 ins
->has_constants
= false;
2157 ins
->has_inline_constant
= true;
2158 ins
->inline_constant
= scaled_constant
;
2163 /* Dead code elimination for branches at the end of a block - only one branch
2164 * per block is legal semantically */
2167 midgard_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2169 bool branched
= false;
2171 mir_foreach_instr_in_block_safe(block
, ins
) {
2172 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2175 mir_remove_instruction(ins
);
2182 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2184 /* Loop to ourselves */
2185 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2186 struct midgard_instruction ins
= v_branch(false, false);
2187 ins
.writeout
= true;
2188 ins
.writeout_depth
= br
->writeout_depth
;
2189 ins
.writeout_stencil
= br
->writeout_stencil
;
2190 ins
.branch
.target_block
= ctx
->block_count
- 1;
2191 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2192 emit_mir_instruction(ctx
, ins
);
2194 ctx
->current_block
->epilogue
= true;
2195 schedule_barrier(ctx
);
2196 return ins
.branch
.target_block
;
2199 static midgard_block
*
2200 emit_block(compiler_context
*ctx
, nir_block
*block
)
2202 midgard_block
*this_block
= ctx
->after_block
;
2203 ctx
->after_block
= NULL
;
2206 this_block
= create_empty_block(ctx
);
2208 list_addtail(&this_block
->base
.link
, &ctx
->blocks
);
2210 this_block
->scheduled
= false;
2213 /* Set up current block */
2214 list_inithead(&this_block
->base
.instructions
);
2215 ctx
->current_block
= this_block
;
2217 nir_foreach_instr(instr
, block
) {
2218 emit_instr(ctx
, instr
);
2219 ++ctx
->instruction_count
;
2225 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2228 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2230 midgard_block
*before_block
= ctx
->current_block
;
2232 /* Speculatively emit the branch, but we can't fill it in until later */
2233 EMIT(branch
, true, true);
2234 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2235 then_branch
->src
[0] = nir_src_index(ctx
, &nif
->condition
);
2236 then_branch
->src_types
[0] = nir_type_uint32
;
2238 /* Emit the two subblocks. */
2239 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2240 midgard_block
*end_then_block
= ctx
->current_block
;
2242 /* Emit a jump from the end of the then block to the end of the else */
2243 EMIT(branch
, false, false);
2244 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2246 /* Emit second block, and check if it's empty */
2248 int else_idx
= ctx
->block_count
;
2249 int count_in
= ctx
->instruction_count
;
2250 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2251 midgard_block
*end_else_block
= ctx
->current_block
;
2252 int after_else_idx
= ctx
->block_count
;
2254 /* Now that we have the subblocks emitted, fix up the branches */
2259 if (ctx
->instruction_count
== count_in
) {
2260 /* The else block is empty, so don't emit an exit jump */
2261 mir_remove_instruction(then_exit
);
2262 then_branch
->branch
.target_block
= after_else_idx
;
2264 then_branch
->branch
.target_block
= else_idx
;
2265 then_exit
->branch
.target_block
= after_else_idx
;
2268 /* Wire up the successors */
2270 ctx
->after_block
= create_empty_block(ctx
);
2272 pan_block_add_successor(&before_block
->base
, &then_block
->base
);
2273 pan_block_add_successor(&before_block
->base
, &else_block
->base
);
2275 pan_block_add_successor(&end_then_block
->base
, &ctx
->after_block
->base
);
2276 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
);
2280 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2282 /* Remember where we are */
2283 midgard_block
*start_block
= ctx
->current_block
;
2285 /* Allocate a loop number, growing the current inner loop depth */
2286 int loop_idx
= ++ctx
->current_loop_depth
;
2288 /* Get index from before the body so we can loop back later */
2289 int start_idx
= ctx
->block_count
;
2291 /* Emit the body itself */
2292 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2294 /* Branch back to loop back */
2295 struct midgard_instruction br_back
= v_branch(false, false);
2296 br_back
.branch
.target_block
= start_idx
;
2297 emit_mir_instruction(ctx
, br_back
);
2299 /* Mark down that branch in the graph. */
2300 pan_block_add_successor(&start_block
->base
, &loop_block
->base
);
2301 pan_block_add_successor(&ctx
->current_block
->base
, &loop_block
->base
);
2303 /* Find the index of the block about to follow us (note: we don't add
2304 * one; blocks are 0-indexed so we get a fencepost problem) */
2305 int break_block_idx
= ctx
->block_count
;
2307 /* Fix up the break statements we emitted to point to the right place,
2308 * now that we can allocate a block number for them */
2309 ctx
->after_block
= create_empty_block(ctx
);
2311 mir_foreach_block_from(ctx
, start_block
, _block
) {
2312 mir_foreach_instr_in_block(((midgard_block
*) _block
), ins
) {
2313 if (ins
->type
!= TAG_ALU_4
) continue;
2314 if (!ins
->compact_branch
) continue;
2316 /* We found a branch -- check the type to see if we need to do anything */
2317 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2319 /* It's a break! Check if it's our break */
2320 if (ins
->branch
.target_break
!= loop_idx
) continue;
2322 /* Okay, cool, we're breaking out of this loop.
2323 * Rewrite from a break to a goto */
2325 ins
->branch
.target_type
= TARGET_GOTO
;
2326 ins
->branch
.target_block
= break_block_idx
;
2328 pan_block_add_successor(_block
, &ctx
->after_block
->base
);
2332 /* Now that we've finished emitting the loop, free up the depth again
2333 * so we play nice with recursion amid nested loops */
2334 --ctx
->current_loop_depth
;
2336 /* Dump loop stats */
2340 static midgard_block
*
2341 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2343 midgard_block
*start_block
= NULL
;
2345 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2346 switch (node
->type
) {
2347 case nir_cf_node_block
: {
2348 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2351 start_block
= block
;
2356 case nir_cf_node_if
:
2357 emit_if(ctx
, nir_cf_node_as_if(node
));
2360 case nir_cf_node_loop
:
2361 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2364 case nir_cf_node_function
:
2373 /* Due to lookahead, we need to report the first tag executed in the command
2374 * stream and in branch targets. An initial block might be empty, so iterate
2375 * until we find one that 'works' */
2378 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2380 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2382 mir_foreach_block_from(ctx
, initial_block
, _v
) {
2383 midgard_block
*v
= (midgard_block
*) _v
;
2384 if (v
->quadword_count
) {
2385 midgard_bundle
*initial_bundle
=
2386 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2388 return initial_bundle
->tag
;
2392 /* Default to a tag 1 which will break from the shader, in case we jump
2393 * to the exit block (i.e. `return` in a compute shader) */
2398 /* For each fragment writeout instruction, generate a writeout loop to
2399 * associate with it */
2402 mir_add_writeout_loops(compiler_context
*ctx
)
2404 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2405 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2408 unsigned popped
= br
->branch
.target_block
;
2409 pan_block_add_successor(&(mir_get_block(ctx
, popped
- 1)->base
), &ctx
->current_block
->base
);
2410 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2411 br
->branch
.target_type
= TARGET_GOTO
;
2413 /* If we have more RTs, we'll need to restore back after our
2414 * loop terminates */
2416 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2417 midgard_instruction uncond
= v_branch(false, false);
2418 uncond
.branch
.target_block
= popped
;
2419 uncond
.branch
.target_type
= TARGET_GOTO
;
2420 emit_mir_instruction(ctx
, uncond
);
2421 pan_block_add_successor(&ctx
->current_block
->base
, &(mir_get_block(ctx
, popped
)->base
));
2422 schedule_barrier(ctx
);
2424 /* We're last, so we can terminate here */
2425 br
->last_writeout
= true;
2431 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2433 struct util_dynarray
*compiled
= &program
->compiled
;
2435 midgard_debug
= debug_get_option_midgard_debug();
2437 /* TODO: Bound against what? */
2438 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2441 ctx
->stage
= nir
->info
.stage
;
2442 ctx
->is_blend
= is_blend
;
2443 ctx
->alpha_ref
= program
->alpha_ref
;
2444 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2445 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2447 /* Start off with a safe cutoff, allowing usage of all 16 work
2448 * registers. Later, we'll promote uniform reads to uniform registers
2449 * if we determine it is beneficial to do so */
2450 ctx
->uniform_cutoff
= 8;
2452 /* Initialize at a global (not block) level hash tables */
2454 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2455 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2457 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2458 * (so we don't accidentally duplicate the epilogue since mesa/st has
2459 * messed with our I/O quite a bit already) */
2461 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2463 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2464 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2465 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2468 NIR_PASS_V(nir
, nir_lower_var_copies
);
2469 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2470 NIR_PASS_V(nir
, nir_split_var_copies
);
2471 NIR_PASS_V(nir
, nir_lower_var_copies
);
2472 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2473 NIR_PASS_V(nir
, nir_lower_var_copies
);
2474 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2476 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2477 NIR_PASS_V(nir
, nir_lower_ssbo
);
2478 NIR_PASS_V(nir
, midgard_nir_lower_zs_store
);
2480 /* Optimisation passes */
2482 optimise_nir(nir
, ctx
->quirks
);
2484 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2485 nir_print_shader(nir
, stdout
);
2488 /* Assign sysvals and counts, now that we're sure
2489 * (post-optimisation) */
2491 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2492 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2493 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2495 nir_foreach_function(func
, nir
) {
2499 list_inithead(&ctx
->blocks
);
2500 ctx
->block_count
= 0;
2502 ctx
->already_emitted
= calloc(BITSET_WORDS(func
->impl
->ssa_alloc
), sizeof(BITSET_WORD
));
2504 emit_cf_list(ctx
, &func
->impl
->body
);
2505 free(ctx
->already_emitted
);
2506 break; /* TODO: Multi-function shaders */
2509 util_dynarray_init(compiled
, NULL
);
2511 /* Per-block lowering before opts */
2513 mir_foreach_block(ctx
, _block
) {
2514 midgard_block
*block
= (midgard_block
*) _block
;
2515 inline_alu_constants(ctx
, block
);
2516 midgard_opt_promote_fmov(ctx
, block
);
2517 embedded_to_inline_constant(ctx
, block
);
2519 /* MIR-level optimizations */
2521 bool progress
= false;
2526 mir_foreach_block(ctx
, _block
) {
2527 midgard_block
*block
= (midgard_block
*) _block
;
2528 progress
|= midgard_opt_copy_prop(ctx
, block
);
2529 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2530 progress
|= midgard_opt_combine_projection(ctx
, block
);
2531 progress
|= midgard_opt_varying_projection(ctx
, block
);
2535 mir_foreach_block(ctx
, _block
) {
2536 midgard_block
*block
= (midgard_block
*) _block
;
2537 midgard_lower_derivatives(ctx
, block
);
2538 midgard_cull_dead_branch(ctx
, block
);
2541 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2542 mir_add_writeout_loops(ctx
);
2544 /* Analyze now that the code is known but before scheduling creates
2545 * pipeline registers which are harder to track */
2546 mir_analyze_helper_terminate(ctx
);
2547 mir_analyze_helper_requirements(ctx
);
2550 midgard_schedule_program(ctx
);
2553 /* Now that all the bundles are scheduled and we can calculate block
2554 * sizes, emit actual branch instructions rather than placeholders */
2556 int br_block_idx
= 0;
2558 mir_foreach_block(ctx
, _block
) {
2559 midgard_block
*block
= (midgard_block
*) _block
;
2560 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2561 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2562 midgard_instruction
*ins
= bundle
->instructions
[c
];
2564 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2566 /* Parse some basic branch info */
2567 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2568 bool is_conditional
= ins
->branch
.conditional
;
2569 bool is_inverted
= ins
->branch
.invert_conditional
;
2570 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2571 bool is_writeout
= ins
->writeout
;
2573 /* Determine the block we're jumping to */
2574 int target_number
= ins
->branch
.target_block
;
2576 /* Report the destination tag */
2577 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2579 /* Count up the number of quadwords we're
2580 * jumping over = number of quadwords until
2581 * (br_block_idx, target_number) */
2583 int quadword_offset
= 0;
2587 } else if (target_number
> br_block_idx
) {
2590 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2591 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2594 quadword_offset
+= blk
->quadword_count
;
2597 /* Jump backwards */
2599 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2600 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2603 quadword_offset
-= blk
->quadword_count
;
2607 /* Unconditional extended branches (far jumps)
2608 * have issues, so we always use a conditional
2609 * branch, setting the condition to always for
2610 * unconditional. For compact unconditional
2611 * branches, cond isn't used so it doesn't
2612 * matter what we pick. */
2614 midgard_condition cond
=
2615 !is_conditional
? midgard_condition_always
:
2616 is_inverted
? midgard_condition_false
:
2617 midgard_condition_true
;
2619 midgard_jmp_writeout_op op
=
2620 is_discard
? midgard_jmp_writeout_op_discard
:
2621 is_writeout
? midgard_jmp_writeout_op_writeout
:
2622 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2623 midgard_jmp_writeout_op_branch_cond
;
2626 midgard_branch_extended branch
=
2627 midgard_create_branch_extended(
2632 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2633 } else if (is_conditional
|| is_discard
) {
2634 midgard_branch_cond branch
= {
2636 .dest_tag
= dest_tag
,
2637 .offset
= quadword_offset
,
2641 assert(branch
.offset
== quadword_offset
);
2643 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2645 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2647 midgard_branch_uncond branch
= {
2649 .dest_tag
= dest_tag
,
2650 .offset
= quadword_offset
,
2654 assert(branch
.offset
== quadword_offset
);
2656 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2664 /* Emit flat binary from the instruction arrays. Iterate each block in
2665 * sequence. Save instruction boundaries such that lookahead tags can
2666 * be assigned easily */
2668 /* Cache _all_ bundles in source order for lookahead across failed branches */
2670 int bundle_count
= 0;
2671 mir_foreach_block(ctx
, _block
) {
2672 midgard_block
*block
= (midgard_block
*) _block
;
2673 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2675 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2677 mir_foreach_block(ctx
, _block
) {
2678 midgard_block
*block
= (midgard_block
*) _block
;
2679 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2680 source_order_bundles
[bundle_idx
++] = bundle
;
2684 int current_bundle
= 0;
2686 /* Midgard prefetches instruction types, so during emission we
2687 * need to lookahead. Unless this is the last instruction, in
2688 * which we return 1. */
2690 mir_foreach_block(ctx
, _block
) {
2691 midgard_block
*block
= (midgard_block
*) _block
;
2692 mir_foreach_bundle_in_block(block
, bundle
) {
2695 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2696 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2698 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2702 /* TODO: Free deeper */
2703 //util_dynarray_fini(&block->instructions);
2706 free(source_order_bundles
);
2708 /* Report the very first tag executed */
2709 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2711 /* Deal with off-by-one related to the fencepost problem */
2712 program
->work_register_count
= ctx
->work_registers
+ 1;
2713 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2715 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2716 program
->tls_size
= ctx
->tls_size
;
2718 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2719 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
2721 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
2722 unsigned nr_bundles
= 0, nr_ins
= 0;
2724 /* Count instructions and bundles */
2726 mir_foreach_block(ctx
, _block
) {
2727 midgard_block
*block
= (midgard_block
*) _block
;
2728 nr_bundles
+= util_dynarray_num_elements(
2729 &block
->bundles
, midgard_bundle
);
2731 mir_foreach_bundle_in_block(block
, bun
)
2732 nr_ins
+= bun
->instruction_count
;
2735 /* Calculate thread count. There are certain cutoffs by
2736 * register count for thread count */
2738 unsigned nr_registers
= program
->work_register_count
;
2740 unsigned nr_threads
=
2741 (nr_registers
<= 4) ? 4 :
2742 (nr_registers
<= 8) ? 2 :
2747 fprintf(stderr
, "shader%d - %s shader: "
2748 "%u inst, %u bundles, %u quadwords, "
2749 "%u registers, %u threads, %u loops, "
2750 "%u:%u spills:fills\n",
2752 gl_shader_stage_name(ctx
->stage
),
2753 nr_ins
, nr_bundles
, ctx
->quadword_count
,
2754 nr_registers
, nr_threads
,
2756 ctx
->spills
, ctx
->fills
);