2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 #include <sys/types.h>
33 #include "main/mtypes.h"
34 #include "compiler/glsl/glsl_to_nir.h"
35 #include "compiler/nir_types.h"
36 #include "main/imports.h"
37 #include "compiler/nir/nir_builder.h"
38 #include "util/half_float.h"
39 #include "util/u_math.h"
40 #include "util/u_debug.h"
41 #include "util/u_dynarray.h"
42 #include "util/list.h"
43 #include "main/mtypes.h"
46 #include "midgard_nir.h"
47 #include "midgard_compile.h"
48 #include "midgard_ops.h"
51 #include "midgard_quirks.h"
53 #include "disassemble.h"
55 static const struct debug_named_value debug_options
[] = {
56 {"msgs", MIDGARD_DBG_MSGS
, "Print debug messages"},
57 {"shaders", MIDGARD_DBG_SHADERS
, "Dump shaders in NIR and MIR"},
58 {"shaderdb", MIDGARD_DBG_SHADERDB
, "Prints shader-db statistics"},
62 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug
, "MIDGARD_MESA_DEBUG", debug_options
, 0)
64 unsigned SHADER_DB_COUNT
= 0;
66 int midgard_debug
= 0;
68 #define DBG(fmt, ...) \
69 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
70 fprintf(stderr, "%s:%d: "fmt, \
71 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
72 static midgard_block
*
73 create_empty_block(compiler_context
*ctx
)
75 midgard_block
*blk
= rzalloc(ctx
, midgard_block
);
77 blk
->predecessors
= _mesa_set_create(blk
,
79 _mesa_key_pointer_equal
);
81 blk
->name
= ctx
->block_source_count
++;
87 midgard_block_add_successor(midgard_block
*block
, midgard_block
*successor
)
93 for (unsigned i
= 0; i
< block
->nr_successors
; ++i
) {
94 if (block
->successors
[i
] == successor
)
98 block
->successors
[block
->nr_successors
++] = successor
;
99 assert(block
->nr_successors
<= ARRAY_SIZE(block
->successors
));
101 /* Note the predecessor in the other direction */
102 _mesa_set_add(successor
->predecessors
, block
);
106 schedule_barrier(compiler_context
*ctx
)
108 midgard_block
*temp
= ctx
->after_block
;
109 ctx
->after_block
= create_empty_block(ctx
);
111 list_addtail(&ctx
->after_block
->link
, &ctx
->blocks
);
112 list_inithead(&ctx
->after_block
->instructions
);
113 midgard_block_add_successor(ctx
->current_block
, ctx
->after_block
);
114 ctx
->current_block
= ctx
->after_block
;
115 ctx
->after_block
= temp
;
118 /* Helpers to generate midgard_instruction's using macro magic, since every
119 * driver seems to do it that way */
121 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
123 #define M_LOAD_STORE(name, store) \
124 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
125 midgard_instruction i = { \
126 .type = TAG_LOAD_STORE_4, \
129 .src = { ~0, ~0, ~0, ~0 }, \
130 .swizzle = SWIZZLE_IDENTITY_4, \
132 .op = midgard_op_##name, \
145 #define M_LOAD(name) M_LOAD_STORE(name, false)
146 #define M_STORE(name) M_LOAD_STORE(name, true)
148 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
149 * the corresponding Midgard source */
151 static midgard_vector_alu_src
152 vector_alu_modifiers(nir_alu_src
*src
, bool is_int
, unsigned broadcast_count
,
153 bool half
, bool sext
)
155 /* Figure out how many components there are so we can adjust.
156 * Specifically we want to broadcast the last channel so things like
160 if (broadcast_count
&& src
) {
161 uint8_t last_component
= src
->swizzle
[broadcast_count
- 1];
163 for (unsigned c
= broadcast_count
; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
164 src
->swizzle
[c
] = last_component
;
168 midgard_vector_alu_src alu_src
= {
175 alu_src
.mod
= midgard_int_normal
;
177 /* Sign/zero-extend if needed */
181 midgard_int_sign_extend
182 : midgard_int_zero_extend
;
185 /* These should have been lowered away */
187 assert(!(src
->abs
|| src
->negate
));
190 alu_src
.mod
= (src
->abs
<< 0) | (src
->negate
<< 1);
196 /* load/store instructions have both 32-bit and 16-bit variants, depending on
197 * whether we are using vectors composed of highp or mediump. At the moment, we
198 * don't support half-floats -- this requires changes in other parts of the
199 * compiler -- therefore the 16-bit versions are commented out. */
201 //M_LOAD(ld_attr_16);
203 //M_LOAD(ld_vary_16);
208 M_LOAD(ld_color_buffer_32u
);
209 //M_STORE(st_vary_16);
211 M_LOAD(ld_cubemap_coords
);
212 M_LOAD(ld_compute_id
);
214 static midgard_instruction
215 v_branch(bool conditional
, bool invert
)
217 midgard_instruction ins
= {
219 .unit
= ALU_ENAB_BRANCH
,
220 .compact_branch
= true,
222 .conditional
= conditional
,
223 .invert_conditional
= invert
226 .src
= { ~0, ~0, ~0, ~0 },
232 static midgard_branch_extended
233 midgard_create_branch_extended( midgard_condition cond
,
234 midgard_jmp_writeout_op op
,
236 signed quadword_offset
)
238 /* The condition code is actually a LUT describing a function to
239 * combine multiple condition codes. However, we only support a single
240 * condition code at the moment, so we just duplicate over a bunch of
243 uint16_t duplicated_cond
=
253 midgard_branch_extended branch
= {
255 .dest_tag
= dest_tag
,
256 .offset
= quadword_offset
,
257 .cond
= duplicated_cond
264 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
266 ins
->has_constants
= true;
267 memcpy(&ins
->constants
, constants
, 16);
271 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
273 return glsl_count_attribute_slots(type
, false);
276 /* Lower fdot2 to a vector multiplication followed by channel addition */
278 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
280 if (alu
->op
!= nir_op_fdot2
)
283 b
->cursor
= nir_before_instr(&alu
->instr
);
285 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
286 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
288 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
290 nir_ssa_def
*sum
= nir_fadd(b
,
291 nir_channel(b
, product
, 0),
292 nir_channel(b
, product
, 1));
294 /* Replace the fdot2 with this sum */
295 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
299 midgard_nir_lower_fdot2(nir_shader
*shader
)
301 bool progress
= false;
303 nir_foreach_function(function
, shader
) {
304 if (!function
->impl
) continue;
307 nir_builder
*b
= &_b
;
308 nir_builder_init(b
, function
->impl
);
310 nir_foreach_block(block
, function
->impl
) {
311 nir_foreach_instr_safe(instr
, block
) {
312 if (instr
->type
!= nir_instr_type_alu
) continue;
314 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
315 midgard_nir_lower_fdot2_body(b
, alu
);
321 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
328 /* Midgard can't write depth and stencil separately. It has to happen in a
329 * single store operation containing both. Let's add a panfrost specific
330 * intrinsic and turn all depth/stencil stores into a packed depth+stencil
334 midgard_nir_lower_zs_store(nir_shader
*nir
)
336 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
339 nir_variable
*z_var
= NULL
, *s_var
= NULL
;
341 nir_foreach_variable(var
, &nir
->outputs
) {
342 if (var
->data
.location
== FRAG_RESULT_DEPTH
)
344 else if (var
->data
.location
== FRAG_RESULT_STENCIL
)
348 if (!z_var
&& !s_var
)
351 bool progress
= false;
353 nir_foreach_function(function
, nir
) {
354 if (!function
->impl
) continue;
356 nir_intrinsic_instr
*z_store
= NULL
, *s_store
= NULL
, *last_store
= NULL
;
358 nir_foreach_block(block
, function
->impl
) {
359 nir_foreach_instr_safe(instr
, block
) {
360 if (instr
->type
!= nir_instr_type_intrinsic
)
363 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
364 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
367 if (z_var
&& nir_intrinsic_base(intr
) == z_var
->data
.driver_location
) {
373 if (s_var
&& nir_intrinsic_base(intr
) == s_var
->data
.driver_location
) {
381 if (!z_store
&& !s_store
) continue;
384 nir_builder_init(&b
, function
->impl
);
386 b
.cursor
= nir_before_instr(&last_store
->instr
);
388 nir_ssa_def
*zs_store_src
;
390 if (z_store
&& s_store
) {
391 nir_ssa_def
*srcs
[2] = {
392 nir_ssa_for_src(&b
, z_store
->src
[0], 1),
393 nir_ssa_for_src(&b
, s_store
->src
[0], 1),
396 zs_store_src
= nir_vec(&b
, srcs
, 2);
398 zs_store_src
= nir_ssa_for_src(&b
, last_store
->src
[0], 1);
401 nir_intrinsic_instr
*zs_store
;
403 zs_store
= nir_intrinsic_instr_create(b
.shader
,
404 nir_intrinsic_store_zs_output_pan
);
405 zs_store
->src
[0] = nir_src_for_ssa(zs_store_src
);
406 zs_store
->num_components
= z_store
&& s_store
? 2 : 1;
407 nir_intrinsic_set_component(zs_store
, z_store
? 0 : 1);
409 /* Replace the Z and S store by a ZS store */
410 nir_builder_instr_insert(&b
, &zs_store
->instr
);
413 nir_instr_remove(&z_store
->instr
);
416 nir_instr_remove(&s_store
->instr
);
418 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
425 /* Flushes undefined values to zero */
428 optimise_nir(nir_shader
*nir
, unsigned quirks
)
431 unsigned lower_flrp
=
432 (nir
->options
->lower_flrp16
? 16 : 0) |
433 (nir
->options
->lower_flrp32
? 32 : 0) |
434 (nir
->options
->lower_flrp64
? 64 : 0);
436 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
437 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
439 nir_lower_tex_options lower_tex_options
= {
440 .lower_txs_lod
= true,
442 .lower_tex_without_implicit_lod
=
443 (quirks
& MIDGARD_EXPLICIT_LOD
),
445 /* TODO: we have native gradient.. */
449 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
451 /* Must lower fdot2 after tex is lowered */
452 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
454 /* T720 is broken. */
456 if (quirks
& MIDGARD_BROKEN_LOD
)
457 NIR_PASS_V(nir
, midgard_nir_lod_errata
);
462 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
463 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
465 NIR_PASS(progress
, nir
, nir_copy_prop
);
466 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
467 NIR_PASS(progress
, nir
, nir_opt_dce
);
468 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
469 NIR_PASS(progress
, nir
, nir_opt_cse
);
470 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
471 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
472 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
474 if (lower_flrp
!= 0) {
475 bool lower_flrp_progress
= false;
476 NIR_PASS(lower_flrp_progress
,
480 false /* always_precise */,
481 nir
->options
->lower_ffma
);
482 if (lower_flrp_progress
) {
483 NIR_PASS(progress
, nir
,
484 nir_opt_constant_folding
);
488 /* Nothing should rematerialize any flrps, so we only
489 * need to do this lowering once.
494 NIR_PASS(progress
, nir
, nir_opt_undef
);
495 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
497 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
500 nir_var_function_temp
);
502 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
505 /* Must be run at the end to prevent creation of fsin/fcos ops */
506 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
511 NIR_PASS(progress
, nir
, nir_opt_dce
);
512 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
513 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
514 NIR_PASS(progress
, nir
, nir_copy_prop
);
517 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
519 /* We implement booleans as 32-bit 0/~0 */
520 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
522 /* Now that booleans are lowered, we can run out late opts */
523 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
525 /* Lower mods for float ops only. Integer ops don't support modifiers
526 * (saturate doesn't make sense on integers, neg/abs require dedicated
529 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
530 NIR_PASS(progress
, nir
, nir_copy_prop
);
531 NIR_PASS(progress
, nir
, nir_opt_dce
);
533 /* Take us out of SSA */
534 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
535 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
537 /* We are a vector architecture; write combine where possible */
538 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
539 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
541 NIR_PASS(progress
, nir
, nir_opt_dce
);
544 /* Do not actually emit a load; instead, cache the constant for inlining */
547 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
549 nir_ssa_def def
= instr
->def
;
551 midgard_constants
*consts
= rzalloc(NULL
, midgard_constants
);
553 assert(instr
->def
.num_components
* instr
->def
.bit_size
<= sizeof(*consts
) * 8);
555 #define RAW_CONST_COPY(bits) \
556 nir_const_value_to_array(consts->u##bits, instr->value, \
557 instr->def.num_components, u##bits)
559 switch (instr
->def
.bit_size
) {
573 unreachable("Invalid bit_size for load_const instruction\n");
576 /* Shifted for SSA, +1 for off-by-one */
577 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, consts
);
580 /* Normally constants are embedded implicitly, but for I/O and such we have to
581 * explicitly emit a move with the constant source */
584 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
586 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
588 if (constant_value
) {
589 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), to
);
590 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
591 emit_mir_instruction(ctx
, ins
);
596 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
598 unsigned comp
= src
->swizzle
[0];
600 for (unsigned c
= 1; c
< nr_components
; ++c
) {
601 if (src
->swizzle
[c
] != comp
)
608 #define ALU_CASE(nir, _op) \
610 op = midgard_alu_op_##_op; \
611 assert(src_bitsize == dst_bitsize); \
614 #define ALU_CASE_BCAST(nir, _op, count) \
616 op = midgard_alu_op_##_op; \
617 broadcast_swizzle = count; \
618 assert(src_bitsize == dst_bitsize); \
621 nir_is_fzero_constant(nir_src src
)
623 if (!nir_src_is_const(src
))
626 for (unsigned c
= 0; c
< nir_src_num_components(src
); ++c
) {
627 if (nir_src_comp_as_float(src
, c
) != 0.0)
634 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
635 * special treatment override this anyway. */
637 static midgard_reg_mode
638 reg_mode_for_nir(nir_alu_instr
*instr
)
640 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
642 switch (src_bitsize
) {
644 return midgard_reg_mode_8
;
646 return midgard_reg_mode_16
;
648 return midgard_reg_mode_32
;
650 return midgard_reg_mode_64
;
652 unreachable("Invalid bit size");
657 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
659 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
660 * is handled elsewhere */
662 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
663 midgard_emit_derivatives(ctx
, instr
);
667 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
669 unsigned dest
= nir_dest_index(&instr
->dest
.dest
);
670 unsigned nr_components
= nir_dest_num_components(instr
->dest
.dest
);
671 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
673 /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
674 * supported. A few do not and are commented for now. Also, there are a
675 * number of NIR ops which Midgard does not support and need to be
676 * lowered, also TODO. This switch block emits the opcode and calling
677 * convention of the Midgard instruction; actual packing is done in
682 /* Number of components valid to check for the instruction (the rest
683 * will be forced to the last), or 0 to use as-is. Relevant as
684 * ball-type instructions have a channel count in NIR but are all vec4
687 unsigned broadcast_swizzle
= 0;
689 /* What register mode should we operate in? */
690 midgard_reg_mode reg_mode
=
691 reg_mode_for_nir(instr
);
693 /* Do we need a destination override? Used for inline
696 midgard_dest_override dest_override
=
697 midgard_dest_override_none
;
699 /* Should we use a smaller respective source and sign-extend? */
701 bool half_1
= false, sext_1
= false;
702 bool half_2
= false, sext_2
= false;
704 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
705 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
708 ALU_CASE(fadd
, fadd
);
709 ALU_CASE(fmul
, fmul
);
710 ALU_CASE(fmin
, fmin
);
711 ALU_CASE(fmax
, fmax
);
712 ALU_CASE(imin
, imin
);
713 ALU_CASE(imax
, imax
);
714 ALU_CASE(umin
, umin
);
715 ALU_CASE(umax
, umax
);
716 ALU_CASE(ffloor
, ffloor
);
717 ALU_CASE(fround_even
, froundeven
);
718 ALU_CASE(ftrunc
, ftrunc
);
719 ALU_CASE(fceil
, fceil
);
720 ALU_CASE(fdot3
, fdot3
);
721 ALU_CASE(fdot4
, fdot4
);
722 ALU_CASE(iadd
, iadd
);
723 ALU_CASE(isub
, isub
);
724 ALU_CASE(imul
, imul
);
726 /* Zero shoved as second-arg */
727 ALU_CASE(iabs
, iabsdiff
);
731 ALU_CASE(feq32
, feq
);
732 ALU_CASE(fne32
, fne
);
733 ALU_CASE(flt32
, flt
);
734 ALU_CASE(ieq32
, ieq
);
735 ALU_CASE(ine32
, ine
);
736 ALU_CASE(ilt32
, ilt
);
737 ALU_CASE(ult32
, ult
);
739 /* We don't have a native b2f32 instruction. Instead, like many
740 * GPUs, we exploit booleans as 0/~0 for false/true, and
741 * correspondingly AND
742 * by 1.0 to do the type conversion. For the moment, prime us
745 * iand [whatever], #0
747 * At the end of emit_alu (as MIR), we'll fix-up the constant
750 ALU_CASE(b2f32
, iand
);
751 ALU_CASE(b2i32
, iand
);
753 /* Likewise, we don't have a dedicated f2b32 instruction, but
754 * we can do a "not equal to 0.0" test. */
756 ALU_CASE(f2b32
, fne
);
757 ALU_CASE(i2b32
, ine
);
759 ALU_CASE(frcp
, frcp
);
760 ALU_CASE(frsq
, frsqrt
);
761 ALU_CASE(fsqrt
, fsqrt
);
762 ALU_CASE(fexp2
, fexp2
);
763 ALU_CASE(flog2
, flog2
);
765 ALU_CASE(f2i64
, f2i_rtz
);
766 ALU_CASE(f2u64
, f2u_rtz
);
767 ALU_CASE(i2f64
, i2f_rtz
);
768 ALU_CASE(u2f64
, u2f_rtz
);
770 ALU_CASE(f2i32
, f2i_rtz
);
771 ALU_CASE(f2u32
, f2u_rtz
);
772 ALU_CASE(i2f32
, i2f_rtz
);
773 ALU_CASE(u2f32
, u2f_rtz
);
775 ALU_CASE(f2i16
, f2i_rtz
);
776 ALU_CASE(f2u16
, f2u_rtz
);
777 ALU_CASE(i2f16
, i2f_rtz
);
778 ALU_CASE(u2f16
, u2f_rtz
);
780 ALU_CASE(fsin
, fsin
);
781 ALU_CASE(fcos
, fcos
);
783 /* We'll set invert */
784 ALU_CASE(inot
, imov
);
785 ALU_CASE(iand
, iand
);
787 ALU_CASE(ixor
, ixor
);
788 ALU_CASE(ishl
, ishl
);
789 ALU_CASE(ishr
, iasr
);
790 ALU_CASE(ushr
, ilsr
);
792 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
793 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
794 ALU_CASE(b32all_fequal4
, fball_eq
);
796 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
797 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
798 ALU_CASE(b32any_fnequal4
, fbany_neq
);
800 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
801 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
802 ALU_CASE(b32all_iequal4
, iball_eq
);
804 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
805 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
806 ALU_CASE(b32any_inequal4
, ibany_neq
);
808 /* Source mods will be shoved in later */
809 ALU_CASE(fabs
, fmov
);
810 ALU_CASE(fneg
, fmov
);
811 ALU_CASE(fsat
, fmov
);
813 /* For size conversion, we use a move. Ideally though we would squash
814 * these ops together; maybe that has to happen after in NIR as part of
815 * propagation...? An earlier algebraic pass ensured we step down by
816 * only / exactly one size. If stepping down, we use a dest override to
817 * reduce the size; if stepping up, we use a larger-sized move with a
818 * half source and a sign/zero-extension modifier */
824 /* If we end up upscale, we'll need a sign-extend on the
825 * operand (the second argument) */
836 if (instr
->op
== nir_op_f2f16
|| instr
->op
== nir_op_f2f32
||
837 instr
->op
== nir_op_f2f64
)
838 op
= midgard_alu_op_fmov
;
840 op
= midgard_alu_op_imov
;
842 if (dst_bitsize
== (src_bitsize
* 2)) {
846 /* Use a greater register mode */
848 } else if (src_bitsize
== (dst_bitsize
* 2)) {
849 /* Converting down */
850 dest_override
= midgard_dest_override_lower
;
856 /* For greater-or-equal, we lower to less-or-equal and flip the
864 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
865 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
866 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
867 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
870 /* Swap via temporary */
871 nir_alu_src temp
= instr
->src
[1];
872 instr
->src
[1] = instr
->src
[0];
873 instr
->src
[0] = temp
;
878 case nir_op_b32csel
: {
879 /* Midgard features both fcsel and icsel, depending on
880 * the type of the arguments/output. However, as long
881 * as we're careful we can _always_ use icsel and
882 * _never_ need fcsel, since the latter does additional
883 * floating-point-specific processing whereas the
884 * former just moves bits on the wire. It's not obvious
885 * why these are separate opcodes, save for the ability
886 * to do things like sat/pos/abs/neg for free */
888 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
889 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
891 /* The condition is the first argument; move the other
892 * arguments up one to be a binary instruction for
893 * Midgard with the condition last */
895 nir_alu_src temp
= instr
->src
[2];
897 instr
->src
[2] = instr
->src
[0];
898 instr
->src
[0] = instr
->src
[1];
899 instr
->src
[1] = temp
;
905 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
910 /* Midgard can perform certain modifiers on output of an ALU op */
913 if (midgard_is_integer_out_op(op
)) {
914 outmod
= midgard_outmod_int_wrap
;
916 bool sat
= instr
->dest
.saturate
|| instr
->op
== nir_op_fsat
;
917 outmod
= sat
? midgard_outmod_sat
: midgard_outmod_none
;
920 /* fmax(a, 0.0) can turn into a .pos modifier as an optimization */
922 if (instr
->op
== nir_op_fmax
) {
923 if (nir_is_fzero_constant(instr
->src
[0].src
)) {
924 op
= midgard_alu_op_fmov
;
926 outmod
= midgard_outmod_pos
;
927 instr
->src
[0] = instr
->src
[1];
928 } else if (nir_is_fzero_constant(instr
->src
[1].src
)) {
929 op
= midgard_alu_op_fmov
;
931 outmod
= midgard_outmod_pos
;
935 /* Fetch unit, quirks, etc information */
936 unsigned opcode_props
= alu_opcode_props
[op
].props
;
937 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
939 /* src0 will always exist afaik, but src1 will not for 1-argument
940 * instructions. The latter can only be fetched if the instruction
941 * needs it, or else we may segfault. */
943 unsigned src0
= nir_alu_src_index(ctx
, &instr
->src
[0]);
944 unsigned src1
= nr_inputs
>= 2 ? nir_alu_src_index(ctx
, &instr
->src
[1]) : ~0;
945 unsigned src2
= nr_inputs
== 3 ? nir_alu_src_index(ctx
, &instr
->src
[2]) : ~0;
946 assert(nr_inputs
<= 3);
948 /* Rather than use the instruction generation helpers, we do it
949 * ourselves here to avoid the mess */
951 midgard_instruction ins
= {
954 quirk_flipped_r24
? ~0 : src0
,
955 quirk_flipped_r24
? src0
: src1
,
962 nir_alu_src
*nirmods
[3] = { NULL
};
964 if (nr_inputs
>= 2) {
965 nirmods
[0] = &instr
->src
[0];
966 nirmods
[1] = &instr
->src
[1];
967 } else if (nr_inputs
== 1) {
968 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
974 nirmods
[2] = &instr
->src
[2];
976 /* These were lowered to a move, so apply the corresponding mod */
978 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
979 nir_alu_src
*s
= nirmods
[quirk_flipped_r24
];
981 if (instr
->op
== nir_op_fneg
)
982 s
->negate
= !s
->negate
;
984 if (instr
->op
== nir_op_fabs
)
988 bool is_int
= midgard_is_integer_op(op
);
990 ins
.mask
= mask_of(nr_components
);
992 midgard_vector_alu alu
= {
994 .reg_mode
= reg_mode
,
995 .dest_override
= dest_override
,
998 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, sext_1
)),
999 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
, broadcast_swizzle
, half_2
, sext_2
)),
1002 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
1005 ins
.mask
&= instr
->dest
.write_mask
;
1007 for (unsigned m
= 0; m
< 3; ++m
) {
1011 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; ++c
)
1012 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[c
];
1014 /* Replicate. TODO: remove when vec16 lands */
1015 for (unsigned c
= NIR_MAX_VEC_COMPONENTS
; c
< MIR_VEC_COMPONENTS
; ++c
)
1016 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[NIR_MAX_VEC_COMPONENTS
- 1];
1019 if (nr_inputs
== 3) {
1020 /* Conditions can't have mods */
1021 assert(!nirmods
[2]->abs
);
1022 assert(!nirmods
[2]->negate
);
1027 /* Late fixup for emulated instructions */
1029 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1030 /* Presently, our second argument is an inline #0 constant.
1031 * Switch over to an embedded 1.0 constant (that can't fit
1032 * inline, since we're 32-bit, not 16-bit like the inline
1035 ins
.has_inline_constant
= false;
1036 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1037 ins
.has_constants
= true;
1039 if (instr
->op
== nir_op_b2f32
)
1040 ins
.constants
.f32
[0] = 1.0f
;
1042 ins
.constants
.i32
[0] = 1;
1044 for (unsigned c
= 0; c
< 16; ++c
)
1045 ins
.swizzle
[1][c
] = 0;
1046 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1047 /* Lots of instructions need a 0 plonked in */
1048 ins
.has_inline_constant
= false;
1049 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1050 ins
.has_constants
= true;
1051 ins
.constants
.u32
[0] = 0;
1053 for (unsigned c
= 0; c
< 16; ++c
)
1054 ins
.swizzle
[1][c
] = 0;
1055 } else if (instr
->op
== nir_op_inot
) {
1059 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1060 /* To avoid duplicating the lookup tables (probably), true LUT
1061 * instructions can only operate as if they were scalars. Lower
1062 * them here by changing the component. */
1064 unsigned orig_mask
= ins
.mask
;
1066 for (int i
= 0; i
< nr_components
; ++i
) {
1067 /* Mask the associated component, dropping the
1068 * instruction if needed */
1071 ins
.mask
&= orig_mask
;
1076 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1077 ins
.swizzle
[0][j
] = nirmods
[0]->swizzle
[i
]; /* Pull from the correct component */
1079 emit_mir_instruction(ctx
, ins
);
1082 emit_mir_instruction(ctx
, ins
);
1089 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1091 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1092 unsigned nir_mask
= 0;
1096 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1097 dsize
= nir_dest_bit_size(intr
->dest
);
1099 nir_mask
= nir_intrinsic_write_mask(intr
);
1103 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1104 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1105 mir_set_bytemask(ins
, bytemask
);
1108 ins
->load_64
= true;
1111 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1112 * optimized) versions of UBO #0 */
1114 static midgard_instruction
*
1116 compiler_context
*ctx
,
1120 nir_src
*indirect_offset
,
1121 unsigned indirect_shift
,
1124 /* TODO: half-floats */
1126 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1127 ins
.constants
.u32
[0] = offset
;
1129 if (instr
->type
== nir_instr_type_intrinsic
)
1130 mir_set_intr_mask(instr
, &ins
, true);
1132 if (indirect_offset
) {
1133 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1134 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1136 ins
.load_store
.arg_2
= 0x1E;
1139 ins
.load_store
.arg_1
= index
;
1141 return emit_mir_instruction(ctx
, ins
);
1144 /* Globals are like UBOs if you squint. And shared memory is like globals if
1145 * you squint even harder */
1149 compiler_context
*ctx
,
1158 midgard_instruction ins
;
1161 ins
= m_ld_int4(srcdest
, 0);
1163 ins
= m_st_int4(srcdest
, 0);
1165 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1166 mir_set_intr_mask(instr
, &ins
, is_read
);
1168 emit_mir_instruction(ctx
, ins
);
1173 compiler_context
*ctx
,
1174 unsigned dest
, unsigned offset
,
1175 unsigned nr_comp
, unsigned component
,
1176 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1178 /* XXX: Half-floats? */
1179 /* TODO: swizzle, mask */
1181 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1182 ins
.mask
= mask_of(nr_comp
);
1184 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1185 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1187 midgard_varying_parameter p
= {
1189 .interpolation
= midgard_interp_default
,
1194 memcpy(&u
, &p
, sizeof(p
));
1195 ins
.load_store
.varying_parameters
= u
;
1197 if (indirect_offset
)
1198 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1200 ins
.load_store
.arg_2
= 0x1E;
1202 ins
.load_store
.arg_1
= 0x9E;
1204 /* Use the type appropriate load */
1208 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1211 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1213 case nir_type_float
:
1214 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1217 unreachable("Attempted to load unknown type");
1221 emit_mir_instruction(ctx
, ins
);
1226 compiler_context
*ctx
,
1227 unsigned dest
, unsigned offset
,
1228 unsigned nr_comp
, nir_alu_type t
)
1230 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1231 ins
.load_store
.arg_1
= 0x1E;
1232 ins
.load_store
.arg_2
= 0x1E;
1233 ins
.mask
= mask_of(nr_comp
);
1235 /* Use the type appropriate load */
1239 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1242 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1244 case nir_type_float
:
1245 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1248 unreachable("Attempted to load unknown type");
1252 emit_mir_instruction(ctx
, ins
);
1256 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1257 unsigned nr_components
, unsigned offset
)
1261 /* Figure out which uniform this is */
1262 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1263 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1265 unsigned dest
= nir_dest_index(&nir_dest
);
1267 /* Sysvals are prefix uniforms */
1268 unsigned uniform
= ((uintptr_t) val
) - 1;
1270 /* Emit the read itself -- this is never indirect */
1271 midgard_instruction
*ins
=
1272 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1274 ins
->mask
= mask_of(nr_components
);
1278 compute_builtin_arg(nir_op op
)
1281 case nir_intrinsic_load_work_group_id
:
1283 case nir_intrinsic_load_local_invocation_id
:
1286 unreachable("Invalid compute paramater loaded");
1291 emit_fragment_store(compiler_context
*ctx
, unsigned src
, enum midgard_rt_id rt
)
1293 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1295 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1299 emit_explicit_constant(ctx
, src
, src
);
1301 struct midgard_instruction ins
=
1302 v_branch(false, false);
1304 ins
.writeout
= true;
1306 /* Add dependencies */
1308 ins
.constants
.u32
[0] = rt
== MIDGARD_ZS_RT
?
1309 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1311 /* Emit the branch */
1312 br
= emit_mir_instruction(ctx
, ins
);
1313 schedule_barrier(ctx
);
1314 ctx
->writeout_branch
[rt
] = br
;
1316 /* Push our current location = current block count - 1 = where we'll
1317 * jump to. Maybe a bit too clever for my own good */
1319 br
->branch
.target_block
= ctx
->block_count
- 1;
1323 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1325 unsigned reg
= nir_dest_index(&instr
->dest
);
1326 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1327 ins
.mask
= mask_of(3);
1328 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1329 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1330 emit_mir_instruction(ctx
, ins
);
1334 vertex_builtin_arg(nir_op op
)
1337 case nir_intrinsic_load_vertex_id
:
1338 return PAN_VERTEX_ID
;
1339 case nir_intrinsic_load_instance_id
:
1340 return PAN_INSTANCE_ID
;
1342 unreachable("Invalid vertex builtin");
1347 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1349 unsigned reg
= nir_dest_index(&instr
->dest
);
1350 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1354 emit_control_barrier(compiler_context
*ctx
)
1356 midgard_instruction ins
= {
1357 .type
= TAG_TEXTURE_4
,
1358 .src
= { ~0, ~0, ~0, ~0 },
1360 .op
= TEXTURE_OP_BARRIER
,
1362 /* TODO: optimize */
1363 .barrier_buffer
= 1,
1368 emit_mir_instruction(ctx
, ins
);
1371 static const nir_variable
*
1372 search_var(struct exec_list
*vars
, unsigned driver_loc
)
1374 nir_foreach_variable(var
, vars
) {
1375 if (var
->data
.driver_location
== driver_loc
)
1383 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1385 unsigned offset
= 0, reg
;
1387 switch (instr
->intrinsic
) {
1388 case nir_intrinsic_discard_if
:
1389 case nir_intrinsic_discard
: {
1390 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1391 struct midgard_instruction discard
= v_branch(conditional
, false);
1392 discard
.branch
.target_type
= TARGET_DISCARD
;
1395 discard
.src
[0] = nir_src_index(ctx
, &instr
->src
[0]);
1397 emit_mir_instruction(ctx
, discard
);
1398 schedule_barrier(ctx
);
1403 case nir_intrinsic_load_uniform
:
1404 case nir_intrinsic_load_ubo
:
1405 case nir_intrinsic_load_global
:
1406 case nir_intrinsic_load_shared
:
1407 case nir_intrinsic_load_input
:
1408 case nir_intrinsic_load_interpolated_input
: {
1409 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1410 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1411 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1412 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1413 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1414 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1416 /* Get the base type of the intrinsic */
1417 /* TODO: Infer type? Does it matter? */
1419 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1420 (is_interp
) ? nir_type_float
:
1421 nir_intrinsic_type(instr
);
1423 t
= nir_alu_type_get_base_type(t
);
1425 if (!(is_ubo
|| is_global
)) {
1426 offset
= nir_intrinsic_base(instr
);
1429 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1431 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1433 bool direct
= nir_src_is_const(*src_offset
);
1434 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1437 offset
+= nir_src_as_uint(*src_offset
);
1439 /* We may need to apply a fractional offset */
1440 int component
= (is_flat
|| is_interp
) ?
1441 nir_intrinsic_component(instr
) : 0;
1442 reg
= nir_dest_index(&instr
->dest
);
1444 if (is_uniform
&& !ctx
->is_blend
) {
1445 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1446 } else if (is_ubo
) {
1447 nir_src index
= instr
->src
[0];
1449 /* TODO: Is indirect block number possible? */
1450 assert(nir_src_is_const(index
));
1452 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1453 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1454 } else if (is_global
|| is_shared
) {
1455 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1456 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1457 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
, is_flat
);
1458 } else if (ctx
->is_blend
) {
1459 /* For blend shaders, load the input color, which is
1460 * preloaded to r0 */
1462 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), reg
);
1463 emit_mir_instruction(ctx
, move
);
1464 schedule_barrier(ctx
);
1465 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1466 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1468 DBG("Unknown load\n");
1475 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1476 case nir_intrinsic_load_barycentric_pixel
:
1477 case nir_intrinsic_load_barycentric_centroid
:
1480 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1482 case nir_intrinsic_load_raw_output_pan
:
1483 case nir_intrinsic_load_output_u8_as_fp16_pan
:
1484 reg
= nir_dest_index(&instr
->dest
);
1485 assert(ctx
->is_blend
);
1487 /* T720 and below use different blend opcodes with slightly
1488 * different semantics than T760 and up */
1490 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1491 bool old_blend
= ctx
->quirks
& MIDGARD_OLD_BLEND
;
1493 if (instr
->intrinsic
== nir_intrinsic_load_output_u8_as_fp16_pan
) {
1494 ld
.load_store
.op
= old_blend
?
1495 midgard_op_ld_color_buffer_u8_as_fp16_old
:
1496 midgard_op_ld_color_buffer_u8_as_fp16
;
1499 ld
.load_store
.address
= 1;
1500 ld
.load_store
.arg_2
= 0x1E;
1503 for (unsigned c
= 2; c
< 16; ++c
)
1504 ld
.swizzle
[0][c
] = 0;
1507 emit_mir_instruction(ctx
, ld
);
1510 case nir_intrinsic_load_blend_const_color_rgba
: {
1511 assert(ctx
->is_blend
);
1512 reg
= nir_dest_index(&instr
->dest
);
1514 /* Blend constants are embedded directly in the shader and
1515 * patched in, so we use some magic routing */
1517 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1518 ins
.has_constants
= true;
1519 ins
.has_blend_constant
= true;
1520 emit_mir_instruction(ctx
, ins
);
1524 case nir_intrinsic_store_zs_output_pan
: {
1525 assert(ctx
->stage
== MESA_SHADER_FRAGMENT
);
1526 emit_fragment_store(ctx
, nir_src_index(ctx
, &instr
->src
[0]),
1529 midgard_instruction
*br
= ctx
->writeout_branch
[MIDGARD_ZS_RT
];
1531 if (!nir_intrinsic_component(instr
))
1532 br
->writeout_depth
= true;
1533 if (nir_intrinsic_component(instr
) ||
1534 instr
->num_components
)
1535 br
->writeout_stencil
= true;
1536 assert(br
->writeout_depth
| br
->writeout_stencil
);
1540 case nir_intrinsic_store_output
:
1541 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1543 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1545 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1547 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1548 const nir_variable
*var
;
1549 enum midgard_rt_id rt
;
1551 var
= search_var(&ctx
->nir
->outputs
,
1552 nir_intrinsic_base(instr
));
1554 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1555 rt
= MIDGARD_COLOR_RT0
;
1556 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1557 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1562 emit_fragment_store(ctx
, reg
, rt
);
1563 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1564 /* We should have been vectorized, though we don't
1565 * currently check that st_vary is emitted only once
1566 * per slot (this is relevant, since there's not a mask
1567 * parameter available on the store [set to 0 by the
1568 * blob]). We do respect the component by adjusting the
1569 * swizzle. If this is a constant source, we'll need to
1570 * emit that explicitly. */
1572 emit_explicit_constant(ctx
, reg
, reg
);
1574 unsigned dst_component
= nir_intrinsic_component(instr
);
1575 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1577 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1578 st
.load_store
.arg_1
= 0x9E;
1579 st
.load_store
.arg_2
= 0x1E;
1581 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1584 st
.load_store
.op
= midgard_op_st_vary_32u
;
1587 st
.load_store
.op
= midgard_op_st_vary_32i
;
1589 case nir_type_float
:
1590 st
.load_store
.op
= midgard_op_st_vary_32
;
1593 unreachable("Attempted to store unknown type");
1597 /* nir_intrinsic_component(store_intr) encodes the
1598 * destination component start. Source component offset
1599 * adjustment is taken care of in
1600 * install_registers_instr(), when offset_swizzle() is
1603 unsigned src_component
= COMPONENT_X
;
1605 assert(nr_comp
> 0);
1606 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1607 st
.swizzle
[0][i
] = src_component
;
1608 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1612 emit_mir_instruction(ctx
, st
);
1614 DBG("Unknown store\n");
1620 /* Special case of store_output for lowered blend shaders */
1621 case nir_intrinsic_store_raw_output_pan
:
1622 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1623 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1625 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1626 /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
1627 * reg = r0.x. We want to splatter. So we can do a 32-bit move
1630 * imov r0.xyzw, r0.xxxx
1633 unsigned expanded
= make_compiler_temp(ctx
);
1635 midgard_instruction splatter
= v_mov(reg
, expanded
);
1637 for (unsigned c
= 0; c
< 16; ++c
)
1638 splatter
.swizzle
[1][c
] = 0;
1640 emit_mir_instruction(ctx
, splatter
);
1641 emit_fragment_store(ctx
, expanded
, ctx
->blend_rt
);
1643 emit_fragment_store(ctx
, reg
, ctx
->blend_rt
);
1647 case nir_intrinsic_store_global
:
1648 case nir_intrinsic_store_shared
:
1649 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1650 emit_explicit_constant(ctx
, reg
, reg
);
1652 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1655 case nir_intrinsic_load_ssbo_address
:
1656 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1659 case nir_intrinsic_get_buffer_size
:
1660 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1663 case nir_intrinsic_load_viewport_scale
:
1664 case nir_intrinsic_load_viewport_offset
:
1665 case nir_intrinsic_load_num_work_groups
:
1666 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1667 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1670 case nir_intrinsic_load_work_group_id
:
1671 case nir_intrinsic_load_local_invocation_id
:
1672 emit_compute_builtin(ctx
, instr
);
1675 case nir_intrinsic_load_vertex_id
:
1676 case nir_intrinsic_load_instance_id
:
1677 emit_vertex_builtin(ctx
, instr
);
1680 case nir_intrinsic_memory_barrier_buffer
:
1681 case nir_intrinsic_memory_barrier_shared
:
1684 case nir_intrinsic_control_barrier
:
1685 schedule_barrier(ctx
);
1686 emit_control_barrier(ctx
);
1687 schedule_barrier(ctx
);
1691 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1698 midgard_tex_format(enum glsl_sampler_dim dim
)
1701 case GLSL_SAMPLER_DIM_1D
:
1702 case GLSL_SAMPLER_DIM_BUF
:
1705 case GLSL_SAMPLER_DIM_2D
:
1706 case GLSL_SAMPLER_DIM_EXTERNAL
:
1707 case GLSL_SAMPLER_DIM_RECT
:
1710 case GLSL_SAMPLER_DIM_3D
:
1713 case GLSL_SAMPLER_DIM_CUBE
:
1714 return MALI_TEX_CUBE
;
1717 DBG("Unknown sampler dim type\n");
1723 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1727 pan_attach_constant_bias(
1728 compiler_context
*ctx
,
1730 midgard_texture_word
*word
)
1732 /* To attach as constant, it has to *be* constant */
1734 if (!nir_src_is_const(lod
))
1737 float f
= nir_src_as_float(lod
);
1739 /* Break into fixed-point */
1741 float lod_frac
= f
- lod_int
;
1743 /* Carry over negative fractions */
1744 if (lod_frac
< 0.0) {
1750 word
->bias
= float_to_ubyte(lod_frac
);
1751 word
->bias_int
= lod_int
;
1756 static enum mali_sampler_type
1757 midgard_sampler_type(nir_alu_type t
) {
1758 switch (nir_alu_type_get_base_type(t
))
1760 case nir_type_float
:
1761 return MALI_SAMPLER_FLOAT
;
1763 return MALI_SAMPLER_SIGNED
;
1765 return MALI_SAMPLER_UNSIGNED
;
1767 unreachable("Unknown sampler type");
1772 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1773 unsigned midgard_texop
)
1776 //assert (!instr->sampler);
1778 int texture_index
= instr
->texture_index
;
1779 int sampler_index
= texture_index
;
1781 /* No helper to build texture words -- we do it all here */
1782 midgard_instruction ins
= {
1783 .type
= TAG_TEXTURE_4
,
1785 .dest
= nir_dest_index(&instr
->dest
),
1786 .src
= { ~0, ~0, ~0, ~0 },
1787 .swizzle
= SWIZZLE_IDENTITY_4
,
1789 .op
= midgard_texop
,
1790 .format
= midgard_tex_format(instr
->sampler_dim
),
1791 .texture_handle
= texture_index
,
1792 .sampler_handle
= sampler_index
,
1798 .sampler_type
= midgard_sampler_type(instr
->dest_type
),
1799 .shadow
= instr
->is_shadow
,
1803 /* We may need a temporary for the coordinate */
1805 bool needs_temp_coord
=
1806 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
1807 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
1810 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
1812 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1813 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1814 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1816 switch (instr
->src
[i
].src_type
) {
1817 case nir_tex_src_coord
: {
1818 emit_explicit_constant(ctx
, index
, index
);
1820 unsigned coord_mask
= mask_of(instr
->coord_components
);
1822 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
1825 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
1827 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1828 /* texelFetch is undefined on samplerCube */
1829 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1831 /* For cubemaps, we use a special ld/st op to
1832 * select the face and copy the xy into the
1833 * texture register */
1835 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
1837 ld
.mask
= 0x3; /* xy */
1838 ld
.load_store
.arg_1
= 0x20;
1839 ld
.swizzle
[1][3] = COMPONENT_X
;
1840 emit_mir_instruction(ctx
, ld
);
1843 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1844 ins
.swizzle
[1][3] = COMPONENT_X
;
1845 } else if (needs_temp_coord
) {
1846 /* mov coord_temp, coords */
1847 midgard_instruction mov
= v_mov(index
, coords
);
1848 mov
.mask
= coord_mask
;
1851 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
1853 emit_mir_instruction(ctx
, mov
);
1858 ins
.src
[1] = coords
;
1860 /* Texelfetch coordinates uses all four elements
1861 * (xyz/index) regardless of texture dimensionality,
1862 * which means it's necessary to zero the unused
1863 * components to keep everything happy */
1865 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1866 /* mov index.zw, #0, or generalized */
1867 midgard_instruction mov
=
1868 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
1869 mov
.has_constants
= true;
1870 mov
.mask
= coord_mask
^ 0xF;
1871 emit_mir_instruction(ctx
, mov
);
1874 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1875 /* Array component in w but NIR wants it in z,
1876 * but if we have a temp coord we already fixed
1879 if (nr_components
== 3) {
1880 ins
.swizzle
[1][2] = COMPONENT_Z
;
1881 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
1882 } else if (nr_components
== 2) {
1884 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1885 ins
.swizzle
[1][3] = COMPONENT_X
;
1887 unreachable("Invalid texture 2D components");
1890 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1892 ins
.swizzle
[1][2] = COMPONENT_Z
;
1893 ins
.swizzle
[1][3] = COMPONENT_W
;
1899 case nir_tex_src_bias
:
1900 case nir_tex_src_lod
: {
1901 /* Try as a constant if we can */
1903 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1904 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1907 ins
.texture
.lod_register
= true;
1910 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1911 ins
.swizzle
[2][c
] = COMPONENT_X
;
1913 emit_explicit_constant(ctx
, index
, index
);
1918 case nir_tex_src_offset
: {
1919 ins
.texture
.offset_register
= true;
1922 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1923 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
1925 emit_explicit_constant(ctx
, index
, index
);
1929 case nir_tex_src_comparator
: {
1930 unsigned comp
= COMPONENT_Z
;
1932 /* mov coord_temp.foo, coords */
1933 midgard_instruction mov
= v_mov(index
, coords
);
1934 mov
.mask
= 1 << comp
;
1936 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
1937 mov
.swizzle
[1][i
] = COMPONENT_X
;
1939 emit_mir_instruction(ctx
, mov
);
1944 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
1950 emit_mir_instruction(ctx
, ins
);
1952 /* Used for .cont and .last hinting */
1953 ctx
->texture_op_count
++;
1957 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1959 switch (instr
->op
) {
1962 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1965 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1968 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1971 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
1974 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
1981 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1983 switch (instr
->type
) {
1984 case nir_jump_break
: {
1985 /* Emit a branch out of the loop */
1986 struct midgard_instruction br
= v_branch(false, false);
1987 br
.branch
.target_type
= TARGET_BREAK
;
1988 br
.branch
.target_break
= ctx
->current_loop_depth
;
1989 emit_mir_instruction(ctx
, br
);
1994 DBG("Unknown jump type %d\n", instr
->type
);
2000 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
2002 switch (instr
->type
) {
2003 case nir_instr_type_load_const
:
2004 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
2007 case nir_instr_type_intrinsic
:
2008 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
2011 case nir_instr_type_alu
:
2012 emit_alu(ctx
, nir_instr_as_alu(instr
));
2015 case nir_instr_type_tex
:
2016 emit_tex(ctx
, nir_instr_as_tex(instr
));
2019 case nir_instr_type_jump
:
2020 emit_jump(ctx
, nir_instr_as_jump(instr
));
2023 case nir_instr_type_ssa_undef
:
2028 DBG("Unhandled instruction type\n");
2034 /* ALU instructions can inline or embed constants, which decreases register
2035 * pressure and saves space. */
2037 #define CONDITIONAL_ATTACH(idx) { \
2038 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
2041 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
2042 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2047 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
2049 mir_foreach_instr_in_block(block
, alu
) {
2050 /* Other instructions cannot inline constants */
2051 if (alu
->type
!= TAG_ALU_4
) continue;
2052 if (alu
->compact_branch
) continue;
2054 /* If there is already a constant here, we can do nothing */
2055 if (alu
->has_constants
) continue;
2057 CONDITIONAL_ATTACH(0);
2059 if (!alu
->has_constants
) {
2060 CONDITIONAL_ATTACH(1)
2061 } else if (!alu
->inline_constant
) {
2062 /* Corner case: _two_ vec4 constants, for instance with a
2063 * csel. For this case, we can only use a constant
2064 * register for one, we'll have to emit a move for the
2065 * other. Note, if both arguments are constants, then
2066 * necessarily neither argument depends on the value of
2067 * any particular register. As the destination register
2068 * will be wiped, that means we can spill the constant
2069 * to the destination register.
2072 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2073 unsigned scratch
= alu
->dest
;
2076 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2077 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2079 /* Set the source */
2080 alu
->src
[1] = scratch
;
2082 /* Inject us -before- the last instruction which set r31 */
2083 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2089 /* Being a little silly with the names, but returns the op that is the bitwise
2090 * inverse of the op with the argument switched. I.e. (f and g are
2093 * f(a, b) = ~g(b, a)
2095 * Corollary: if g is the contrapositve of f, f is the contrapositive of g:
2097 * f(a, b) = ~g(b, a)
2098 * ~f(a, b) = g(b, a)
2099 * ~f(a, b) = ~h(a, b) where h is the contrapositive of g
2102 * Thus we define this function in pairs.
2105 static inline midgard_alu_op
2106 mir_contrapositive(midgard_alu_op op
)
2109 case midgard_alu_op_flt
:
2110 return midgard_alu_op_fle
;
2111 case midgard_alu_op_fle
:
2112 return midgard_alu_op_flt
;
2114 case midgard_alu_op_ilt
:
2115 return midgard_alu_op_ile
;
2116 case midgard_alu_op_ile
:
2117 return midgard_alu_op_ilt
;
2120 unreachable("No known contrapositive");
2124 /* Midgard supports two types of constants, embedded constants (128-bit) and
2125 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2126 * constants can be demoted to inline constants, for space savings and
2127 * sometimes a performance boost */
2130 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2132 mir_foreach_instr_in_block(block
, ins
) {
2133 if (!ins
->has_constants
) continue;
2134 if (ins
->has_inline_constant
) continue;
2136 /* Blend constants must not be inlined by definition */
2137 if (ins
->has_blend_constant
) continue;
2139 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2140 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2141 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2143 if (!(is_16
|| is_32
))
2146 /* src1 cannot be an inline constant due to encoding
2147 * restrictions. So, if possible we try to flip the arguments
2150 int op
= ins
->alu
.op
;
2152 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2153 bool flip
= alu_opcode_props
[op
].props
& OP_COMMUTES
;
2156 /* Conditionals can be inverted */
2157 case midgard_alu_op_flt
:
2158 case midgard_alu_op_ilt
:
2159 case midgard_alu_op_fle
:
2160 case midgard_alu_op_ile
:
2161 ins
->alu
.op
= mir_contrapositive(ins
->alu
.op
);
2166 case midgard_alu_op_fcsel
:
2167 case midgard_alu_op_icsel
:
2168 DBG("Missed non-commutative flip (%s)\n", alu_opcode_props
[op
].name
);
2177 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2178 /* Extract the source information */
2180 midgard_vector_alu_src
*src
;
2181 int q
= ins
->alu
.src2
;
2182 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2185 /* Component is from the swizzle. Take a nonzero component */
2187 unsigned first_comp
= ffs(ins
->mask
) - 1;
2188 unsigned component
= ins
->swizzle
[1][first_comp
];
2190 /* Scale constant appropriately, if we can legally */
2191 uint16_t scaled_constant
= 0;
2194 scaled_constant
= ins
->constants
.u16
[component
];
2195 } else if (midgard_is_integer_op(op
)) {
2196 scaled_constant
= ins
->constants
.u32
[component
];
2198 /* Constant overflow after resize */
2199 if (scaled_constant
!= ins
->constants
.u32
[component
])
2202 float original
= ins
->constants
.f32
[component
];
2203 scaled_constant
= _mesa_float_to_half(original
);
2205 /* Check for loss of precision. If this is
2206 * mediump, we don't care, but for a highp
2207 * shader, we need to pay attention. NIR
2208 * doesn't yet tell us which mode we're in!
2209 * Practically this prevents most constants
2210 * from being inlined, sadly. */
2212 float fp32
= _mesa_half_to_float(scaled_constant
);
2214 if (fp32
!= original
)
2218 /* We don't know how to handle these with a constant */
2220 if (mir_nontrivial_source2_mod_simple(ins
) || src
->rep_low
|| src
->rep_high
) {
2221 DBG("Bailing inline constant...\n");
2225 /* Make sure that the constant is not itself a vector
2226 * by checking if all accessed values are the same. */
2228 const midgard_constants
*cons
= &ins
->constants
;
2229 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2231 bool is_vector
= false;
2232 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2234 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2235 /* We only care if this component is actually used */
2236 if (!(mask
& (1 << c
)))
2239 uint32_t test
= is_16
?
2240 cons
->u16
[ins
->swizzle
[1][c
]] :
2241 cons
->u32
[ins
->swizzle
[1][c
]];
2243 if (test
!= value
) {
2252 /* Get rid of the embedded constant */
2253 ins
->has_constants
= false;
2255 ins
->has_inline_constant
= true;
2256 ins
->inline_constant
= scaled_constant
;
2261 /* Dead code elimination for branches at the end of a block - only one branch
2262 * per block is legal semantically */
2265 midgard_opt_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2267 bool branched
= false;
2269 mir_foreach_instr_in_block_safe(block
, ins
) {
2270 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2273 mir_remove_instruction(ins
);
2279 /* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
2280 * the move can be propagated away entirely */
2283 mir_compose_float_outmod(midgard_outmod_float
*outmod
, midgard_outmod_float comp
)
2286 if (comp
== midgard_outmod_none
)
2289 if (*outmod
== midgard_outmod_none
) {
2294 /* TODO: Compose rules */
2299 midgard_opt_pos_propagate(compiler_context
*ctx
, midgard_block
*block
)
2301 bool progress
= false;
2303 mir_foreach_instr_in_block_safe(block
, ins
) {
2304 if (ins
->type
!= TAG_ALU_4
) continue;
2305 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2306 if (ins
->alu
.outmod
!= midgard_outmod_pos
) continue;
2308 /* TODO: Registers? */
2309 unsigned src
= ins
->src
[1];
2310 if (src
& IS_REG
) continue;
2312 /* There might be a source modifier, too */
2313 if (mir_nontrivial_source2_mod(ins
)) continue;
2315 /* Backpropagate the modifier */
2316 mir_foreach_instr_in_block_from_rev(block
, v
, mir_prev_op(ins
)) {
2317 if (v
->type
!= TAG_ALU_4
) continue;
2318 if (v
->dest
!= src
) continue;
2320 /* Can we even take a float outmod? */
2321 if (midgard_is_integer_out_op(v
->alu
.op
)) continue;
2323 midgard_outmod_float temp
= v
->alu
.outmod
;
2324 progress
|= mir_compose_float_outmod(&temp
, ins
->alu
.outmod
);
2326 /* Throw in the towel.. */
2327 if (!progress
) break;
2329 /* Otherwise, transfer the modifier */
2330 v
->alu
.outmod
= temp
;
2331 ins
->alu
.outmod
= midgard_outmod_none
;
2341 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2343 /* Loop to ourselves */
2344 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2345 struct midgard_instruction ins
= v_branch(false, false);
2346 ins
.writeout
= true;
2347 ins
.writeout_depth
= br
->writeout_depth
;
2348 ins
.writeout_stencil
= br
->writeout_stencil
;
2349 ins
.branch
.target_block
= ctx
->block_count
- 1;
2350 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2351 emit_mir_instruction(ctx
, ins
);
2353 ctx
->current_block
->epilogue
= true;
2354 schedule_barrier(ctx
);
2355 return ins
.branch
.target_block
;
2358 static midgard_block
*
2359 emit_block(compiler_context
*ctx
, nir_block
*block
)
2361 midgard_block
*this_block
= ctx
->after_block
;
2362 ctx
->after_block
= NULL
;
2365 this_block
= create_empty_block(ctx
);
2367 list_addtail(&this_block
->link
, &ctx
->blocks
);
2369 this_block
->scheduled
= false;
2372 /* Set up current block */
2373 list_inithead(&this_block
->instructions
);
2374 ctx
->current_block
= this_block
;
2376 nir_foreach_instr(instr
, block
) {
2377 emit_instr(ctx
, instr
);
2378 ++ctx
->instruction_count
;
2384 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2387 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2389 midgard_block
*before_block
= ctx
->current_block
;
2391 /* Speculatively emit the branch, but we can't fill it in until later */
2392 EMIT(branch
, true, true);
2393 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2394 then_branch
->src
[0] = nir_src_index(ctx
, &nif
->condition
);
2396 /* Emit the two subblocks. */
2397 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2398 midgard_block
*end_then_block
= ctx
->current_block
;
2400 /* Emit a jump from the end of the then block to the end of the else */
2401 EMIT(branch
, false, false);
2402 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2404 /* Emit second block, and check if it's empty */
2406 int else_idx
= ctx
->block_count
;
2407 int count_in
= ctx
->instruction_count
;
2408 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2409 midgard_block
*end_else_block
= ctx
->current_block
;
2410 int after_else_idx
= ctx
->block_count
;
2412 /* Now that we have the subblocks emitted, fix up the branches */
2417 if (ctx
->instruction_count
== count_in
) {
2418 /* The else block is empty, so don't emit an exit jump */
2419 mir_remove_instruction(then_exit
);
2420 then_branch
->branch
.target_block
= after_else_idx
;
2422 then_branch
->branch
.target_block
= else_idx
;
2423 then_exit
->branch
.target_block
= after_else_idx
;
2426 /* Wire up the successors */
2428 ctx
->after_block
= create_empty_block(ctx
);
2430 midgard_block_add_successor(before_block
, then_block
);
2431 midgard_block_add_successor(before_block
, else_block
);
2433 midgard_block_add_successor(end_then_block
, ctx
->after_block
);
2434 midgard_block_add_successor(end_else_block
, ctx
->after_block
);
2438 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2440 /* Remember where we are */
2441 midgard_block
*start_block
= ctx
->current_block
;
2443 /* Allocate a loop number, growing the current inner loop depth */
2444 int loop_idx
= ++ctx
->current_loop_depth
;
2446 /* Get index from before the body so we can loop back later */
2447 int start_idx
= ctx
->block_count
;
2449 /* Emit the body itself */
2450 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2452 /* Branch back to loop back */
2453 struct midgard_instruction br_back
= v_branch(false, false);
2454 br_back
.branch
.target_block
= start_idx
;
2455 emit_mir_instruction(ctx
, br_back
);
2457 /* Mark down that branch in the graph. */
2458 midgard_block_add_successor(start_block
, loop_block
);
2459 midgard_block_add_successor(ctx
->current_block
, loop_block
);
2461 /* Find the index of the block about to follow us (note: we don't add
2462 * one; blocks are 0-indexed so we get a fencepost problem) */
2463 int break_block_idx
= ctx
->block_count
;
2465 /* Fix up the break statements we emitted to point to the right place,
2466 * now that we can allocate a block number for them */
2467 ctx
->after_block
= create_empty_block(ctx
);
2469 list_for_each_entry_from(struct midgard_block
, block
, start_block
, &ctx
->blocks
, link
) {
2470 mir_foreach_instr_in_block(block
, ins
) {
2471 if (ins
->type
!= TAG_ALU_4
) continue;
2472 if (!ins
->compact_branch
) continue;
2474 /* We found a branch -- check the type to see if we need to do anything */
2475 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2477 /* It's a break! Check if it's our break */
2478 if (ins
->branch
.target_break
!= loop_idx
) continue;
2480 /* Okay, cool, we're breaking out of this loop.
2481 * Rewrite from a break to a goto */
2483 ins
->branch
.target_type
= TARGET_GOTO
;
2484 ins
->branch
.target_block
= break_block_idx
;
2486 midgard_block_add_successor(block
, ctx
->after_block
);
2490 /* Now that we've finished emitting the loop, free up the depth again
2491 * so we play nice with recursion amid nested loops */
2492 --ctx
->current_loop_depth
;
2494 /* Dump loop stats */
2498 static midgard_block
*
2499 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2501 midgard_block
*start_block
= NULL
;
2503 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2504 switch (node
->type
) {
2505 case nir_cf_node_block
: {
2506 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2509 start_block
= block
;
2514 case nir_cf_node_if
:
2515 emit_if(ctx
, nir_cf_node_as_if(node
));
2518 case nir_cf_node_loop
:
2519 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2522 case nir_cf_node_function
:
2531 /* Due to lookahead, we need to report the first tag executed in the command
2532 * stream and in branch targets. An initial block might be empty, so iterate
2533 * until we find one that 'works' */
2536 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2538 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2540 mir_foreach_block_from(ctx
, initial_block
, v
) {
2541 if (v
->quadword_count
) {
2542 midgard_bundle
*initial_bundle
=
2543 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2545 return initial_bundle
->tag
;
2549 /* Default to a tag 1 which will break from the shader, in case we jump
2550 * to the exit block (i.e. `return` in a compute shader) */
2556 pan_format_from_nir_base(nir_alu_type base
)
2560 return MALI_FORMAT_SINT
;
2563 return MALI_FORMAT_UINT
;
2564 case nir_type_float
:
2565 return MALI_CHANNEL_FLOAT
;
2567 unreachable("Invalid base");
2572 pan_format_from_nir_size(nir_alu_type base
, unsigned size
)
2574 if (base
== nir_type_float
) {
2576 case 16: return MALI_FORMAT_SINT
;
2577 case 32: return MALI_FORMAT_UNORM
;
2579 unreachable("Invalid float size for format");
2584 case 8: return MALI_CHANNEL_8
;
2585 case 16: return MALI_CHANNEL_16
;
2586 case 32: return MALI_CHANNEL_32
;
2588 unreachable("Invalid int size for format");
2593 static enum mali_format
2594 pan_format_from_glsl(const struct glsl_type
*type
)
2596 enum glsl_base_type glsl_base
= glsl_get_base_type(glsl_without_array(type
));
2597 nir_alu_type t
= nir_get_nir_type_for_glsl_base_type(glsl_base
);
2599 unsigned base
= nir_alu_type_get_base_type(t
);
2600 unsigned size
= nir_alu_type_get_type_size(t
);
2602 return pan_format_from_nir_base(base
) |
2603 pan_format_from_nir_size(base
, size
) |
2604 MALI_NR_CHANNELS(4);
2607 /* For each fragment writeout instruction, generate a writeout loop to
2608 * associate with it */
2611 mir_add_writeout_loops(compiler_context
*ctx
)
2613 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2614 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2617 unsigned popped
= br
->branch
.target_block
;
2618 midgard_block_add_successor(mir_get_block(ctx
, popped
- 1), ctx
->current_block
);
2619 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2621 /* If we have more RTs, we'll need to restore back after our
2622 * loop terminates */
2624 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2625 midgard_instruction uncond
= v_branch(false, false);
2626 uncond
.branch
.target_block
= popped
;
2627 emit_mir_instruction(ctx
, uncond
);
2628 midgard_block_add_successor(ctx
->current_block
, mir_get_block(ctx
, popped
));
2629 schedule_barrier(ctx
);
2631 /* We're last, so we can terminate here */
2632 br
->last_writeout
= true;
2638 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2640 struct util_dynarray
*compiled
= &program
->compiled
;
2642 midgard_debug
= debug_get_option_midgard_debug();
2644 /* TODO: Bound against what? */
2645 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2648 ctx
->stage
= nir
->info
.stage
;
2649 ctx
->is_blend
= is_blend
;
2650 ctx
->alpha_ref
= program
->alpha_ref
;
2651 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2652 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2654 /* Start off with a safe cutoff, allowing usage of all 16 work
2655 * registers. Later, we'll promote uniform reads to uniform registers
2656 * if we determine it is beneficial to do so */
2657 ctx
->uniform_cutoff
= 8;
2659 /* Initialize at a global (not block) level hash tables */
2661 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2662 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2664 /* Record the varying mapping for the command stream's bookkeeping */
2666 struct exec_list
*varyings
=
2667 ctx
->stage
== MESA_SHADER_VERTEX
? &nir
->outputs
: &nir
->inputs
;
2669 unsigned max_varying
= 0;
2670 nir_foreach_variable(var
, varyings
) {
2671 unsigned loc
= var
->data
.driver_location
;
2672 unsigned sz
= glsl_type_size(var
->type
, FALSE
);
2674 for (int c
= 0; c
< sz
; ++c
) {
2675 program
->varyings
[loc
+ c
] = var
->data
.location
+ c
;
2676 program
->varying_type
[loc
+ c
] = pan_format_from_glsl(var
->type
);
2677 max_varying
= MAX2(max_varying
, loc
+ c
);
2681 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2682 * (so we don't accidentally duplicate the epilogue since mesa/st has
2683 * messed with our I/O quite a bit already) */
2685 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2687 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2688 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2689 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2692 NIR_PASS_V(nir
, nir_lower_var_copies
);
2693 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2694 NIR_PASS_V(nir
, nir_split_var_copies
);
2695 NIR_PASS_V(nir
, nir_lower_var_copies
);
2696 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2697 NIR_PASS_V(nir
, nir_lower_var_copies
);
2698 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2700 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2701 NIR_PASS_V(nir
, nir_lower_ssbo
);
2702 NIR_PASS_V(nir
, midgard_nir_lower_zs_store
);
2704 /* Optimisation passes */
2706 optimise_nir(nir
, ctx
->quirks
);
2708 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2709 nir_print_shader(nir
, stdout
);
2712 /* Assign sysvals and counts, now that we're sure
2713 * (post-optimisation) */
2715 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2717 program
->uniform_count
= nir
->num_uniforms
;
2718 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2719 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2721 nir_foreach_function(func
, nir
) {
2725 list_inithead(&ctx
->blocks
);
2726 ctx
->block_count
= 0;
2729 emit_cf_list(ctx
, &func
->impl
->body
);
2730 break; /* TODO: Multi-function shaders */
2733 util_dynarray_init(compiled
, NULL
);
2735 /* Per-block lowering before opts */
2737 mir_foreach_block(ctx
, block
) {
2738 inline_alu_constants(ctx
, block
);
2739 midgard_opt_promote_fmov(ctx
, block
);
2740 embedded_to_inline_constant(ctx
, block
);
2742 /* MIR-level optimizations */
2744 bool progress
= false;
2749 mir_foreach_block(ctx
, block
) {
2750 progress
|= midgard_opt_pos_propagate(ctx
, block
);
2751 progress
|= midgard_opt_copy_prop(ctx
, block
);
2752 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2753 progress
|= midgard_opt_combine_projection(ctx
, block
);
2754 progress
|= midgard_opt_varying_projection(ctx
, block
);
2755 progress
|= midgard_opt_not_propagate(ctx
, block
);
2756 progress
|= midgard_opt_fuse_src_invert(ctx
, block
);
2757 progress
|= midgard_opt_fuse_dest_invert(ctx
, block
);
2758 progress
|= midgard_opt_csel_invert(ctx
, block
);
2759 progress
|= midgard_opt_drop_cmp_invert(ctx
, block
);
2760 progress
|= midgard_opt_invert_branch(ctx
, block
);
2764 mir_foreach_block(ctx
, block
) {
2765 midgard_lower_invert(ctx
, block
);
2766 midgard_lower_derivatives(ctx
, block
);
2769 /* Nested control-flow can result in dead branches at the end of the
2770 * block. This messes with our analysis and is just dead code, so cull
2772 mir_foreach_block(ctx
, block
) {
2773 midgard_opt_cull_dead_branch(ctx
, block
);
2776 /* Ensure we were lowered */
2777 mir_foreach_instr_global(ctx
, ins
) {
2778 assert(!ins
->invert
);
2781 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2782 mir_add_writeout_loops(ctx
);
2785 midgard_schedule_program(ctx
);
2788 /* Now that all the bundles are scheduled and we can calculate block
2789 * sizes, emit actual branch instructions rather than placeholders */
2791 int br_block_idx
= 0;
2793 mir_foreach_block(ctx
, block
) {
2794 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2795 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2796 midgard_instruction
*ins
= bundle
->instructions
[c
];
2798 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2800 /* Parse some basic branch info */
2801 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2802 bool is_conditional
= ins
->branch
.conditional
;
2803 bool is_inverted
= ins
->branch
.invert_conditional
;
2804 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2805 bool is_writeout
= ins
->writeout
;
2807 /* Determine the block we're jumping to */
2808 int target_number
= ins
->branch
.target_block
;
2810 /* Report the destination tag */
2811 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2813 /* Count up the number of quadwords we're
2814 * jumping over = number of quadwords until
2815 * (br_block_idx, target_number) */
2817 int quadword_offset
= 0;
2821 } else if (target_number
> br_block_idx
) {
2824 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2825 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2828 quadword_offset
+= blk
->quadword_count
;
2831 /* Jump backwards */
2833 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2834 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2837 quadword_offset
-= blk
->quadword_count
;
2841 /* Unconditional extended branches (far jumps)
2842 * have issues, so we always use a conditional
2843 * branch, setting the condition to always for
2844 * unconditional. For compact unconditional
2845 * branches, cond isn't used so it doesn't
2846 * matter what we pick. */
2848 midgard_condition cond
=
2849 !is_conditional
? midgard_condition_always
:
2850 is_inverted
? midgard_condition_false
:
2851 midgard_condition_true
;
2853 midgard_jmp_writeout_op op
=
2854 is_discard
? midgard_jmp_writeout_op_discard
:
2855 is_writeout
? midgard_jmp_writeout_op_writeout
:
2856 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2857 midgard_jmp_writeout_op_branch_cond
;
2860 midgard_branch_extended branch
=
2861 midgard_create_branch_extended(
2866 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2867 } else if (is_conditional
|| is_discard
) {
2868 midgard_branch_cond branch
= {
2870 .dest_tag
= dest_tag
,
2871 .offset
= quadword_offset
,
2875 assert(branch
.offset
== quadword_offset
);
2877 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2879 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2881 midgard_branch_uncond branch
= {
2883 .dest_tag
= dest_tag
,
2884 .offset
= quadword_offset
,
2888 assert(branch
.offset
== quadword_offset
);
2890 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2898 /* Emit flat binary from the instruction arrays. Iterate each block in
2899 * sequence. Save instruction boundaries such that lookahead tags can
2900 * be assigned easily */
2902 /* Cache _all_ bundles in source order for lookahead across failed branches */
2904 int bundle_count
= 0;
2905 mir_foreach_block(ctx
, block
) {
2906 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2908 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2910 mir_foreach_block(ctx
, block
) {
2911 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2912 source_order_bundles
[bundle_idx
++] = bundle
;
2916 int current_bundle
= 0;
2918 /* Midgard prefetches instruction types, so during emission we
2919 * need to lookahead. Unless this is the last instruction, in
2920 * which we return 1. */
2922 mir_foreach_block(ctx
, block
) {
2923 mir_foreach_bundle_in_block(block
, bundle
) {
2926 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2927 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2929 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2933 /* TODO: Free deeper */
2934 //util_dynarray_fini(&block->instructions);
2937 free(source_order_bundles
);
2939 /* Report the very first tag executed */
2940 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2942 /* Deal with off-by-one related to the fencepost problem */
2943 program
->work_register_count
= ctx
->work_registers
+ 1;
2944 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2946 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2947 program
->tls_size
= ctx
->tls_size
;
2949 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2950 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
2952 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
2953 unsigned nr_bundles
= 0, nr_ins
= 0;
2955 /* Count instructions and bundles */
2957 mir_foreach_block(ctx
, block
) {
2958 nr_bundles
+= util_dynarray_num_elements(
2959 &block
->bundles
, midgard_bundle
);
2961 mir_foreach_bundle_in_block(block
, bun
)
2962 nr_ins
+= bun
->instruction_count
;
2965 /* Calculate thread count. There are certain cutoffs by
2966 * register count for thread count */
2968 unsigned nr_registers
= program
->work_register_count
;
2970 unsigned nr_threads
=
2971 (nr_registers
<= 4) ? 4 :
2972 (nr_registers
<= 8) ? 2 :
2977 fprintf(stderr
, "shader%d - %s shader: "
2978 "%u inst, %u bundles, %u quadwords, "
2979 "%u registers, %u threads, %u loops, "
2980 "%u:%u spills:fills\n",
2982 gl_shader_stage_name(ctx
->stage
),
2983 nr_ins
, nr_bundles
, ctx
->quadword_count
,
2984 nr_registers
, nr_threads
,
2986 ctx
->spills
, ctx
->fills
);