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) \
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, \
125 #define M_LOAD(name) M_LOAD_STORE(name, false)
126 #define M_STORE(name) M_LOAD_STORE(name, true)
128 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
129 * the corresponding Midgard source */
131 static midgard_vector_alu_src
132 vector_alu_modifiers(nir_alu_src
*src
, bool is_int
, unsigned broadcast_count
,
133 bool half
, bool sext
)
135 /* Figure out how many components there are so we can adjust.
136 * Specifically we want to broadcast the last channel so things like
140 if (broadcast_count
&& src
) {
141 uint8_t last_component
= src
->swizzle
[broadcast_count
- 1];
143 for (unsigned c
= broadcast_count
; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
144 src
->swizzle
[c
] = last_component
;
148 midgard_vector_alu_src alu_src
= {
155 alu_src
.mod
= midgard_int_normal
;
157 /* Sign/zero-extend if needed */
161 midgard_int_sign_extend
162 : midgard_int_zero_extend
;
165 /* These should have been lowered away */
167 assert(!(src
->abs
|| src
->negate
));
170 alu_src
.mod
= (src
->abs
<< 0) | (src
->negate
<< 1);
181 M_LOAD(ld_color_buffer_32u
);
183 M_LOAD(ld_cubemap_coords
);
184 M_LOAD(ld_compute_id
);
186 static midgard_instruction
187 v_branch(bool conditional
, bool invert
)
189 midgard_instruction ins
= {
191 .unit
= ALU_ENAB_BRANCH
,
192 .compact_branch
= true,
194 .conditional
= conditional
,
195 .invert_conditional
= invert
198 .src
= { ~0, ~0, ~0, ~0 },
204 static midgard_branch_extended
205 midgard_create_branch_extended( midgard_condition cond
,
206 midgard_jmp_writeout_op op
,
208 signed quadword_offset
)
210 /* The condition code is actually a LUT describing a function to
211 * combine multiple condition codes. However, we only support a single
212 * condition code at the moment, so we just duplicate over a bunch of
215 uint16_t duplicated_cond
=
225 midgard_branch_extended branch
= {
227 .dest_tag
= dest_tag
,
228 .offset
= quadword_offset
,
229 .cond
= duplicated_cond
236 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
238 ins
->has_constants
= true;
239 memcpy(&ins
->constants
, constants
, 16);
243 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
245 return glsl_count_attribute_slots(type
, false);
248 /* Lower fdot2 to a vector multiplication followed by channel addition */
250 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
252 if (alu
->op
!= nir_op_fdot2
)
255 b
->cursor
= nir_before_instr(&alu
->instr
);
257 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
258 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
260 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
262 nir_ssa_def
*sum
= nir_fadd(b
,
263 nir_channel(b
, product
, 0),
264 nir_channel(b
, product
, 1));
266 /* Replace the fdot2 with this sum */
267 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
271 midgard_nir_lower_fdot2(nir_shader
*shader
)
273 bool progress
= false;
275 nir_foreach_function(function
, shader
) {
276 if (!function
->impl
) continue;
279 nir_builder
*b
= &_b
;
280 nir_builder_init(b
, function
->impl
);
282 nir_foreach_block(block
, function
->impl
) {
283 nir_foreach_instr_safe(instr
, block
) {
284 if (instr
->type
!= nir_instr_type_alu
) continue;
286 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
287 midgard_nir_lower_fdot2_body(b
, alu
);
293 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
300 /* Midgard can't write depth and stencil separately. It has to happen in a
301 * single store operation containing both. Let's add a panfrost specific
302 * intrinsic and turn all depth/stencil stores into a packed depth+stencil
306 midgard_nir_lower_zs_store(nir_shader
*nir
)
308 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
311 nir_variable
*z_var
= NULL
, *s_var
= NULL
;
313 nir_foreach_variable(var
, &nir
->outputs
) {
314 if (var
->data
.location
== FRAG_RESULT_DEPTH
)
316 else if (var
->data
.location
== FRAG_RESULT_STENCIL
)
320 if (!z_var
&& !s_var
)
323 bool progress
= false;
325 nir_foreach_function(function
, nir
) {
326 if (!function
->impl
) continue;
328 nir_intrinsic_instr
*z_store
= NULL
, *s_store
= NULL
, *last_store
= NULL
;
330 nir_foreach_block(block
, function
->impl
) {
331 nir_foreach_instr_safe(instr
, block
) {
332 if (instr
->type
!= nir_instr_type_intrinsic
)
335 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
336 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
339 if (z_var
&& nir_intrinsic_base(intr
) == z_var
->data
.driver_location
) {
345 if (s_var
&& nir_intrinsic_base(intr
) == s_var
->data
.driver_location
) {
353 if (!z_store
&& !s_store
) continue;
356 nir_builder_init(&b
, function
->impl
);
358 b
.cursor
= nir_before_instr(&last_store
->instr
);
360 nir_ssa_def
*zs_store_src
;
362 if (z_store
&& s_store
) {
363 nir_ssa_def
*srcs
[2] = {
364 nir_ssa_for_src(&b
, z_store
->src
[0], 1),
365 nir_ssa_for_src(&b
, s_store
->src
[0], 1),
368 zs_store_src
= nir_vec(&b
, srcs
, 2);
370 zs_store_src
= nir_ssa_for_src(&b
, last_store
->src
[0], 1);
373 nir_intrinsic_instr
*zs_store
;
375 zs_store
= nir_intrinsic_instr_create(b
.shader
,
376 nir_intrinsic_store_zs_output_pan
);
377 zs_store
->src
[0] = nir_src_for_ssa(zs_store_src
);
378 zs_store
->num_components
= z_store
&& s_store
? 2 : 1;
379 nir_intrinsic_set_component(zs_store
, z_store
? 0 : 1);
381 /* Replace the Z and S store by a ZS store */
382 nir_builder_instr_insert(&b
, &zs_store
->instr
);
385 nir_instr_remove(&z_store
->instr
);
388 nir_instr_remove(&s_store
->instr
);
390 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
397 /* Flushes undefined values to zero */
400 optimise_nir(nir_shader
*nir
, unsigned quirks
)
403 unsigned lower_flrp
=
404 (nir
->options
->lower_flrp16
? 16 : 0) |
405 (nir
->options
->lower_flrp32
? 32 : 0) |
406 (nir
->options
->lower_flrp64
? 64 : 0);
408 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
409 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
411 nir_lower_tex_options lower_tex_options
= {
412 .lower_txs_lod
= true,
414 .lower_tex_without_implicit_lod
=
415 (quirks
& MIDGARD_EXPLICIT_LOD
),
417 /* TODO: we have native gradient.. */
421 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
423 /* Must lower fdot2 after tex is lowered */
424 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
426 /* T720 is broken. */
428 if (quirks
& MIDGARD_BROKEN_LOD
)
429 NIR_PASS_V(nir
, midgard_nir_lod_errata
);
434 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
435 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
437 NIR_PASS(progress
, nir
, nir_copy_prop
);
438 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
439 NIR_PASS(progress
, nir
, nir_opt_dce
);
440 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
441 NIR_PASS(progress
, nir
, nir_opt_cse
);
442 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
443 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
444 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
446 if (lower_flrp
!= 0) {
447 bool lower_flrp_progress
= false;
448 NIR_PASS(lower_flrp_progress
,
452 false /* always_precise */,
453 nir
->options
->lower_ffma
);
454 if (lower_flrp_progress
) {
455 NIR_PASS(progress
, nir
,
456 nir_opt_constant_folding
);
460 /* Nothing should rematerialize any flrps, so we only
461 * need to do this lowering once.
466 NIR_PASS(progress
, nir
, nir_opt_undef
);
467 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
469 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
472 nir_var_function_temp
);
474 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
477 /* Must be run at the end to prevent creation of fsin/fcos ops */
478 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
483 NIR_PASS(progress
, nir
, nir_opt_dce
);
484 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
485 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
486 NIR_PASS(progress
, nir
, nir_copy_prop
);
489 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
491 /* We implement booleans as 32-bit 0/~0 */
492 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
494 /* Now that booleans are lowered, we can run out late opts */
495 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
497 /* Lower mods for float ops only. Integer ops don't support modifiers
498 * (saturate doesn't make sense on integers, neg/abs require dedicated
501 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
502 NIR_PASS(progress
, nir
, nir_copy_prop
);
503 NIR_PASS(progress
, nir
, nir_opt_dce
);
505 /* Take us out of SSA */
506 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
507 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
509 /* We are a vector architecture; write combine where possible */
510 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
511 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
513 NIR_PASS(progress
, nir
, nir_opt_dce
);
516 /* Do not actually emit a load; instead, cache the constant for inlining */
519 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
521 nir_ssa_def def
= instr
->def
;
523 midgard_constants
*consts
= rzalloc(NULL
, midgard_constants
);
525 assert(instr
->def
.num_components
* instr
->def
.bit_size
<= sizeof(*consts
) * 8);
527 #define RAW_CONST_COPY(bits) \
528 nir_const_value_to_array(consts->u##bits, instr->value, \
529 instr->def.num_components, u##bits)
531 switch (instr
->def
.bit_size
) {
545 unreachable("Invalid bit_size for load_const instruction\n");
548 /* Shifted for SSA, +1 for off-by-one */
549 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, consts
);
552 /* Normally constants are embedded implicitly, but for I/O and such we have to
553 * explicitly emit a move with the constant source */
556 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
558 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
560 if (constant_value
) {
561 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), to
);
562 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
563 emit_mir_instruction(ctx
, ins
);
568 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
570 unsigned comp
= src
->swizzle
[0];
572 for (unsigned c
= 1; c
< nr_components
; ++c
) {
573 if (src
->swizzle
[c
] != comp
)
580 #define ALU_CASE(nir, _op) \
582 op = midgard_alu_op_##_op; \
583 assert(src_bitsize == dst_bitsize); \
586 #define ALU_CASE_BCAST(nir, _op, count) \
588 op = midgard_alu_op_##_op; \
589 broadcast_swizzle = count; \
590 assert(src_bitsize == dst_bitsize); \
593 nir_is_fzero_constant(nir_src src
)
595 if (!nir_src_is_const(src
))
598 for (unsigned c
= 0; c
< nir_src_num_components(src
); ++c
) {
599 if (nir_src_comp_as_float(src
, c
) != 0.0)
606 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
607 * special treatment override this anyway. */
609 static midgard_reg_mode
610 reg_mode_for_nir(nir_alu_instr
*instr
)
612 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
614 switch (src_bitsize
) {
616 return midgard_reg_mode_8
;
618 return midgard_reg_mode_16
;
620 return midgard_reg_mode_32
;
622 return midgard_reg_mode_64
;
624 unreachable("Invalid bit size");
629 mir_copy_src(midgard_instruction
*ins
, nir_alu_instr
*instr
, unsigned i
, unsigned to
)
631 unsigned bits
= nir_src_bit_size(instr
->src
[i
].src
);
633 ins
->src
[to
] = nir_src_index(NULL
, &instr
->src
[i
].src
);
634 ins
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
] | bits
;
638 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
640 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
641 * is handled elsewhere */
643 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
644 midgard_emit_derivatives(ctx
, instr
);
648 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
650 unsigned nr_components
= nir_dest_num_components(instr
->dest
.dest
);
651 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
654 /* Number of components valid to check for the instruction (the rest
655 * will be forced to the last), or 0 to use as-is. Relevant as
656 * ball-type instructions have a channel count in NIR but are all vec4
659 unsigned broadcast_swizzle
= 0;
661 /* What register mode should we operate in? */
662 midgard_reg_mode reg_mode
=
663 reg_mode_for_nir(instr
);
665 /* Do we need a destination override? Used for inline
668 midgard_dest_override dest_override
=
669 midgard_dest_override_none
;
671 /* Should we use a smaller respective source and sign-extend? */
673 bool half_1
= false, sext_1
= false;
674 bool half_2
= false, sext_2
= false;
676 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
677 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
680 ALU_CASE(fadd
, fadd
);
681 ALU_CASE(fmul
, fmul
);
682 ALU_CASE(fmin
, fmin
);
683 ALU_CASE(fmax
, fmax
);
684 ALU_CASE(imin
, imin
);
685 ALU_CASE(imax
, imax
);
686 ALU_CASE(umin
, umin
);
687 ALU_CASE(umax
, umax
);
688 ALU_CASE(ffloor
, ffloor
);
689 ALU_CASE(fround_even
, froundeven
);
690 ALU_CASE(ftrunc
, ftrunc
);
691 ALU_CASE(fceil
, fceil
);
692 ALU_CASE(fdot3
, fdot3
);
693 ALU_CASE(fdot4
, fdot4
);
694 ALU_CASE(iadd
, iadd
);
695 ALU_CASE(isub
, isub
);
696 ALU_CASE(imul
, imul
);
698 /* Zero shoved as second-arg */
699 ALU_CASE(iabs
, iabsdiff
);
703 ALU_CASE(feq32
, feq
);
704 ALU_CASE(fne32
, fne
);
705 ALU_CASE(flt32
, flt
);
706 ALU_CASE(ieq32
, ieq
);
707 ALU_CASE(ine32
, ine
);
708 ALU_CASE(ilt32
, ilt
);
709 ALU_CASE(ult32
, ult
);
711 /* We don't have a native b2f32 instruction. Instead, like many
712 * GPUs, we exploit booleans as 0/~0 for false/true, and
713 * correspondingly AND
714 * by 1.0 to do the type conversion. For the moment, prime us
717 * iand [whatever], #0
719 * At the end of emit_alu (as MIR), we'll fix-up the constant
722 ALU_CASE(b2f32
, iand
);
723 ALU_CASE(b2i32
, iand
);
725 /* Likewise, we don't have a dedicated f2b32 instruction, but
726 * we can do a "not equal to 0.0" test. */
728 ALU_CASE(f2b32
, fne
);
729 ALU_CASE(i2b32
, ine
);
731 ALU_CASE(frcp
, frcp
);
732 ALU_CASE(frsq
, frsqrt
);
733 ALU_CASE(fsqrt
, fsqrt
);
734 ALU_CASE(fexp2
, fexp2
);
735 ALU_CASE(flog2
, flog2
);
737 ALU_CASE(f2i64
, f2i_rtz
);
738 ALU_CASE(f2u64
, f2u_rtz
);
739 ALU_CASE(i2f64
, i2f_rtz
);
740 ALU_CASE(u2f64
, u2f_rtz
);
742 ALU_CASE(f2i32
, f2i_rtz
);
743 ALU_CASE(f2u32
, f2u_rtz
);
744 ALU_CASE(i2f32
, i2f_rtz
);
745 ALU_CASE(u2f32
, u2f_rtz
);
747 ALU_CASE(f2i16
, f2i_rtz
);
748 ALU_CASE(f2u16
, f2u_rtz
);
749 ALU_CASE(i2f16
, i2f_rtz
);
750 ALU_CASE(u2f16
, u2f_rtz
);
752 ALU_CASE(fsin
, fsin
);
753 ALU_CASE(fcos
, fcos
);
755 /* We'll set invert */
756 ALU_CASE(inot
, imov
);
757 ALU_CASE(iand
, iand
);
759 ALU_CASE(ixor
, ixor
);
760 ALU_CASE(ishl
, ishl
);
761 ALU_CASE(ishr
, iasr
);
762 ALU_CASE(ushr
, ilsr
);
764 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
765 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
766 ALU_CASE(b32all_fequal4
, fball_eq
);
768 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
769 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
770 ALU_CASE(b32any_fnequal4
, fbany_neq
);
772 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
773 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
774 ALU_CASE(b32all_iequal4
, iball_eq
);
776 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
777 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
778 ALU_CASE(b32any_inequal4
, ibany_neq
);
780 /* Source mods will be shoved in later */
781 ALU_CASE(fabs
, fmov
);
782 ALU_CASE(fneg
, fmov
);
783 ALU_CASE(fsat
, fmov
);
785 /* For size conversion, we use a move. Ideally though we would squash
786 * these ops together; maybe that has to happen after in NIR as part of
787 * propagation...? An earlier algebraic pass ensured we step down by
788 * only / exactly one size. If stepping down, we use a dest override to
789 * reduce the size; if stepping up, we use a larger-sized move with a
790 * half source and a sign/zero-extension modifier */
796 /* If we end up upscale, we'll need a sign-extend on the
797 * operand (the second argument) */
808 if (instr
->op
== nir_op_f2f16
|| instr
->op
== nir_op_f2f32
||
809 instr
->op
== nir_op_f2f64
)
810 op
= midgard_alu_op_fmov
;
812 op
= midgard_alu_op_imov
;
814 if (dst_bitsize
== (src_bitsize
* 2)) {
818 /* Use a greater register mode */
820 } else if (src_bitsize
== (dst_bitsize
* 2)) {
821 /* Converting down */
822 dest_override
= midgard_dest_override_lower
;
828 /* For greater-or-equal, we lower to less-or-equal and flip the
836 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
837 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
838 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
839 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
842 /* Swap via temporary */
843 nir_alu_src temp
= instr
->src
[1];
844 instr
->src
[1] = instr
->src
[0];
845 instr
->src
[0] = temp
;
850 case nir_op_b32csel
: {
851 /* Midgard features both fcsel and icsel, depending on
852 * the type of the arguments/output. However, as long
853 * as we're careful we can _always_ use icsel and
854 * _never_ need fcsel, since the latter does additional
855 * floating-point-specific processing whereas the
856 * former just moves bits on the wire. It's not obvious
857 * why these are separate opcodes, save for the ability
858 * to do things like sat/pos/abs/neg for free */
860 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
861 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
863 /* The condition is the first argument; move the other
864 * arguments up one to be a binary instruction for
865 * Midgard with the condition last */
867 nir_alu_src temp
= instr
->src
[2];
869 instr
->src
[2] = instr
->src
[0];
870 instr
->src
[0] = instr
->src
[1];
871 instr
->src
[1] = temp
;
877 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
882 /* Midgard can perform certain modifiers on output of an ALU op */
885 if (midgard_is_integer_out_op(op
)) {
886 outmod
= midgard_outmod_int_wrap
;
888 bool sat
= instr
->dest
.saturate
|| instr
->op
== nir_op_fsat
;
889 outmod
= sat
? midgard_outmod_sat
: midgard_outmod_none
;
892 /* fmax(a, 0.0) can turn into a .pos modifier as an optimization */
894 if (instr
->op
== nir_op_fmax
) {
895 if (nir_is_fzero_constant(instr
->src
[0].src
)) {
896 op
= midgard_alu_op_fmov
;
898 outmod
= midgard_outmod_pos
;
899 instr
->src
[0] = instr
->src
[1];
900 } else if (nir_is_fzero_constant(instr
->src
[1].src
)) {
901 op
= midgard_alu_op_fmov
;
903 outmod
= midgard_outmod_pos
;
907 /* Fetch unit, quirks, etc information */
908 unsigned opcode_props
= alu_opcode_props
[op
].props
;
909 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
911 midgard_instruction ins
= {
913 .dest
= nir_dest_index(&instr
->dest
.dest
),
914 .dest_type
= nir_op_infos
[instr
->op
].output_type
915 | nir_dest_bit_size(instr
->dest
.dest
),
918 for (unsigned i
= nr_inputs
; i
< ARRAY_SIZE(ins
.src
); ++i
)
921 if (quirk_flipped_r24
) {
923 mir_copy_src(&ins
, instr
, 0, 1);
925 for (unsigned i
= 0; i
< nr_inputs
; ++i
)
926 mir_copy_src(&ins
, instr
, i
, quirk_flipped_r24
? 1 : i
);
929 nir_alu_src
*nirmods
[3] = { NULL
};
931 if (nr_inputs
>= 2) {
932 nirmods
[0] = &instr
->src
[0];
933 nirmods
[1] = &instr
->src
[1];
934 } else if (nr_inputs
== 1) {
935 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
941 nirmods
[2] = &instr
->src
[2];
943 /* These were lowered to a move, so apply the corresponding mod */
945 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
946 nir_alu_src
*s
= nirmods
[quirk_flipped_r24
];
948 if (instr
->op
== nir_op_fneg
)
949 s
->negate
= !s
->negate
;
951 if (instr
->op
== nir_op_fabs
)
955 bool is_int
= midgard_is_integer_op(op
);
957 ins
.mask
= mask_of(nr_components
);
959 midgard_vector_alu alu
= {
961 .reg_mode
= reg_mode
,
962 .dest_override
= dest_override
,
965 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, sext_1
)),
966 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
, broadcast_swizzle
, half_2
, sext_2
)),
969 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
972 ins
.mask
&= instr
->dest
.write_mask
;
974 for (unsigned m
= 0; m
< 3; ++m
) {
978 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; ++c
)
979 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[c
];
981 /* Replicate. TODO: remove when vec16 lands */
982 for (unsigned c
= NIR_MAX_VEC_COMPONENTS
; c
< MIR_VEC_COMPONENTS
; ++c
)
983 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[NIR_MAX_VEC_COMPONENTS
- 1];
986 if (nr_inputs
== 3) {
987 /* Conditions can't have mods */
988 assert(!nirmods
[2]->abs
);
989 assert(!nirmods
[2]->negate
);
994 /* Late fixup for emulated instructions */
996 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
997 /* Presently, our second argument is an inline #0 constant.
998 * Switch over to an embedded 1.0 constant (that can't fit
999 * inline, since we're 32-bit, not 16-bit like the inline
1002 ins
.has_inline_constant
= false;
1003 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1004 ins
.has_constants
= true;
1006 if (instr
->op
== nir_op_b2f32
)
1007 ins
.constants
.f32
[0] = 1.0f
;
1009 ins
.constants
.i32
[0] = 1;
1011 for (unsigned c
= 0; c
< 16; ++c
)
1012 ins
.swizzle
[1][c
] = 0;
1013 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1014 /* Lots of instructions need a 0 plonked in */
1015 ins
.has_inline_constant
= false;
1016 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
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;
1022 } else if (instr
->op
== nir_op_inot
) {
1026 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1027 /* To avoid duplicating the lookup tables (probably), true LUT
1028 * instructions can only operate as if they were scalars. Lower
1029 * them here by changing the component. */
1031 unsigned orig_mask
= ins
.mask
;
1033 for (int i
= 0; i
< nr_components
; ++i
) {
1034 /* Mask the associated component, dropping the
1035 * instruction if needed */
1038 ins
.mask
&= orig_mask
;
1043 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1044 ins
.swizzle
[0][j
] = nirmods
[0]->swizzle
[i
]; /* Pull from the correct component */
1046 emit_mir_instruction(ctx
, ins
);
1049 emit_mir_instruction(ctx
, ins
);
1056 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1058 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1059 unsigned nir_mask
= 0;
1063 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1064 dsize
= nir_dest_bit_size(intr
->dest
);
1066 nir_mask
= nir_intrinsic_write_mask(intr
);
1070 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1071 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1072 mir_set_bytemask(ins
, bytemask
);
1075 ins
->load_64
= true;
1078 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1079 * optimized) versions of UBO #0 */
1081 static midgard_instruction
*
1083 compiler_context
*ctx
,
1087 nir_src
*indirect_offset
,
1088 unsigned indirect_shift
,
1091 /* TODO: half-floats */
1093 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1094 ins
.constants
.u32
[0] = offset
;
1096 if (instr
->type
== nir_instr_type_intrinsic
)
1097 mir_set_intr_mask(instr
, &ins
, true);
1099 if (indirect_offset
) {
1100 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1101 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1103 ins
.load_store
.arg_2
= 0x1E;
1106 ins
.load_store
.arg_1
= index
;
1108 return emit_mir_instruction(ctx
, ins
);
1111 /* Globals are like UBOs if you squint. And shared memory is like globals if
1112 * you squint even harder */
1116 compiler_context
*ctx
,
1125 midgard_instruction ins
;
1128 ins
= m_ld_int4(srcdest
, 0);
1130 ins
= m_st_int4(srcdest
, 0);
1132 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1133 mir_set_intr_mask(instr
, &ins
, is_read
);
1135 emit_mir_instruction(ctx
, ins
);
1140 compiler_context
*ctx
,
1141 unsigned dest
, unsigned offset
,
1142 unsigned nr_comp
, unsigned component
,
1143 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1145 /* XXX: Half-floats? */
1146 /* TODO: swizzle, mask */
1148 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1149 ins
.mask
= mask_of(nr_comp
);
1151 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1152 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1154 midgard_varying_parameter p
= {
1156 .interpolation
= midgard_interp_default
,
1161 memcpy(&u
, &p
, sizeof(p
));
1162 ins
.load_store
.varying_parameters
= u
;
1164 if (indirect_offset
)
1165 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1167 ins
.load_store
.arg_2
= 0x1E;
1169 ins
.load_store
.arg_1
= 0x9E;
1171 /* Use the type appropriate load */
1175 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1178 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1180 case nir_type_float
:
1181 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1184 unreachable("Attempted to load unknown type");
1188 emit_mir_instruction(ctx
, ins
);
1193 compiler_context
*ctx
,
1194 unsigned dest
, unsigned offset
,
1195 unsigned nr_comp
, nir_alu_type t
)
1197 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1198 ins
.load_store
.arg_1
= 0x1E;
1199 ins
.load_store
.arg_2
= 0x1E;
1200 ins
.mask
= mask_of(nr_comp
);
1202 /* Use the type appropriate load */
1206 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1209 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1211 case nir_type_float
:
1212 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1215 unreachable("Attempted to load unknown type");
1219 emit_mir_instruction(ctx
, ins
);
1223 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1224 unsigned nr_components
, unsigned offset
)
1228 /* Figure out which uniform this is */
1229 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1230 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1232 unsigned dest
= nir_dest_index(&nir_dest
);
1234 /* Sysvals are prefix uniforms */
1235 unsigned uniform
= ((uintptr_t) val
) - 1;
1237 /* Emit the read itself -- this is never indirect */
1238 midgard_instruction
*ins
=
1239 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1241 ins
->mask
= mask_of(nr_components
);
1245 compute_builtin_arg(nir_op op
)
1248 case nir_intrinsic_load_work_group_id
:
1250 case nir_intrinsic_load_local_invocation_id
:
1253 unreachable("Invalid compute paramater loaded");
1258 emit_fragment_store(compiler_context
*ctx
, unsigned src
, enum midgard_rt_id rt
)
1260 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1262 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1266 emit_explicit_constant(ctx
, src
, src
);
1268 struct midgard_instruction ins
=
1269 v_branch(false, false);
1271 ins
.writeout
= true;
1273 /* Add dependencies */
1275 ins
.constants
.u32
[0] = rt
== MIDGARD_ZS_RT
?
1276 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1278 /* Emit the branch */
1279 br
= emit_mir_instruction(ctx
, ins
);
1280 schedule_barrier(ctx
);
1281 ctx
->writeout_branch
[rt
] = br
;
1283 /* Push our current location = current block count - 1 = where we'll
1284 * jump to. Maybe a bit too clever for my own good */
1286 br
->branch
.target_block
= ctx
->block_count
- 1;
1290 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1292 unsigned reg
= nir_dest_index(&instr
->dest
);
1293 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1294 ins
.mask
= mask_of(3);
1295 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1296 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1297 emit_mir_instruction(ctx
, ins
);
1301 vertex_builtin_arg(nir_op op
)
1304 case nir_intrinsic_load_vertex_id
:
1305 return PAN_VERTEX_ID
;
1306 case nir_intrinsic_load_instance_id
:
1307 return PAN_INSTANCE_ID
;
1309 unreachable("Invalid vertex builtin");
1314 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1316 unsigned reg
= nir_dest_index(&instr
->dest
);
1317 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1321 emit_control_barrier(compiler_context
*ctx
)
1323 midgard_instruction ins
= {
1324 .type
= TAG_TEXTURE_4
,
1325 .src
= { ~0, ~0, ~0, ~0 },
1327 .op
= TEXTURE_OP_BARRIER
,
1329 /* TODO: optimize */
1330 .barrier_buffer
= 1,
1335 emit_mir_instruction(ctx
, ins
);
1338 static const nir_variable
*
1339 search_var(struct exec_list
*vars
, unsigned driver_loc
)
1341 nir_foreach_variable(var
, vars
) {
1342 if (var
->data
.driver_location
== driver_loc
)
1350 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1352 unsigned offset
= 0, reg
;
1354 switch (instr
->intrinsic
) {
1355 case nir_intrinsic_discard_if
:
1356 case nir_intrinsic_discard
: {
1357 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1358 struct midgard_instruction discard
= v_branch(conditional
, false);
1359 discard
.branch
.target_type
= TARGET_DISCARD
;
1362 discard
.src
[0] = nir_src_index(ctx
, &instr
->src
[0]);
1364 emit_mir_instruction(ctx
, discard
);
1365 schedule_barrier(ctx
);
1370 case nir_intrinsic_load_uniform
:
1371 case nir_intrinsic_load_ubo
:
1372 case nir_intrinsic_load_global
:
1373 case nir_intrinsic_load_shared
:
1374 case nir_intrinsic_load_input
:
1375 case nir_intrinsic_load_interpolated_input
: {
1376 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1377 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1378 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1379 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1380 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1381 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1383 /* Get the base type of the intrinsic */
1384 /* TODO: Infer type? Does it matter? */
1386 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1387 (is_interp
) ? nir_type_float
:
1388 nir_intrinsic_type(instr
);
1390 t
= nir_alu_type_get_base_type(t
);
1392 if (!(is_ubo
|| is_global
)) {
1393 offset
= nir_intrinsic_base(instr
);
1396 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1398 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1400 bool direct
= nir_src_is_const(*src_offset
);
1401 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1404 offset
+= nir_src_as_uint(*src_offset
);
1406 /* We may need to apply a fractional offset */
1407 int component
= (is_flat
|| is_interp
) ?
1408 nir_intrinsic_component(instr
) : 0;
1409 reg
= nir_dest_index(&instr
->dest
);
1411 if (is_uniform
&& !ctx
->is_blend
) {
1412 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1413 } else if (is_ubo
) {
1414 nir_src index
= instr
->src
[0];
1416 /* TODO: Is indirect block number possible? */
1417 assert(nir_src_is_const(index
));
1419 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1420 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1421 } else if (is_global
|| is_shared
) {
1422 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1423 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1424 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
, is_flat
);
1425 } else if (ctx
->is_blend
) {
1426 /* For blend shaders, load the input color, which is
1427 * preloaded to r0 */
1429 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), reg
);
1430 emit_mir_instruction(ctx
, move
);
1431 schedule_barrier(ctx
);
1432 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1433 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1435 DBG("Unknown load\n");
1442 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1443 case nir_intrinsic_load_barycentric_pixel
:
1444 case nir_intrinsic_load_barycentric_centroid
:
1447 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1449 case nir_intrinsic_load_raw_output_pan
:
1450 case nir_intrinsic_load_output_u8_as_fp16_pan
:
1451 reg
= nir_dest_index(&instr
->dest
);
1452 assert(ctx
->is_blend
);
1454 /* T720 and below use different blend opcodes with slightly
1455 * different semantics than T760 and up */
1457 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1458 bool old_blend
= ctx
->quirks
& MIDGARD_OLD_BLEND
;
1460 if (instr
->intrinsic
== nir_intrinsic_load_output_u8_as_fp16_pan
) {
1461 ld
.load_store
.op
= old_blend
?
1462 midgard_op_ld_color_buffer_u8_as_fp16_old
:
1463 midgard_op_ld_color_buffer_u8_as_fp16
;
1466 ld
.load_store
.address
= 1;
1467 ld
.load_store
.arg_2
= 0x1E;
1470 for (unsigned c
= 2; c
< 16; ++c
)
1471 ld
.swizzle
[0][c
] = 0;
1474 emit_mir_instruction(ctx
, ld
);
1477 case nir_intrinsic_load_blend_const_color_rgba
: {
1478 assert(ctx
->is_blend
);
1479 reg
= nir_dest_index(&instr
->dest
);
1481 /* Blend constants are embedded directly in the shader and
1482 * patched in, so we use some magic routing */
1484 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1485 ins
.has_constants
= true;
1486 ins
.has_blend_constant
= true;
1487 emit_mir_instruction(ctx
, ins
);
1491 case nir_intrinsic_store_zs_output_pan
: {
1492 assert(ctx
->stage
== MESA_SHADER_FRAGMENT
);
1493 emit_fragment_store(ctx
, nir_src_index(ctx
, &instr
->src
[0]),
1496 midgard_instruction
*br
= ctx
->writeout_branch
[MIDGARD_ZS_RT
];
1498 if (!nir_intrinsic_component(instr
))
1499 br
->writeout_depth
= true;
1500 if (nir_intrinsic_component(instr
) ||
1501 instr
->num_components
)
1502 br
->writeout_stencil
= true;
1503 assert(br
->writeout_depth
| br
->writeout_stencil
);
1507 case nir_intrinsic_store_output
:
1508 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1510 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1512 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1514 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1515 const nir_variable
*var
;
1516 enum midgard_rt_id rt
;
1518 var
= search_var(&ctx
->nir
->outputs
,
1519 nir_intrinsic_base(instr
));
1521 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1522 rt
= MIDGARD_COLOR_RT0
;
1523 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1524 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1529 emit_fragment_store(ctx
, reg
, rt
);
1530 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1531 /* We should have been vectorized, though we don't
1532 * currently check that st_vary is emitted only once
1533 * per slot (this is relevant, since there's not a mask
1534 * parameter available on the store [set to 0 by the
1535 * blob]). We do respect the component by adjusting the
1536 * swizzle. If this is a constant source, we'll need to
1537 * emit that explicitly. */
1539 emit_explicit_constant(ctx
, reg
, reg
);
1541 unsigned dst_component
= nir_intrinsic_component(instr
);
1542 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1544 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1545 st
.load_store
.arg_1
= 0x9E;
1546 st
.load_store
.arg_2
= 0x1E;
1548 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1551 st
.load_store
.op
= midgard_op_st_vary_32u
;
1554 st
.load_store
.op
= midgard_op_st_vary_32i
;
1556 case nir_type_float
:
1557 st
.load_store
.op
= midgard_op_st_vary_32
;
1560 unreachable("Attempted to store unknown type");
1564 /* nir_intrinsic_component(store_intr) encodes the
1565 * destination component start. Source component offset
1566 * adjustment is taken care of in
1567 * install_registers_instr(), when offset_swizzle() is
1570 unsigned src_component
= COMPONENT_X
;
1572 assert(nr_comp
> 0);
1573 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1574 st
.swizzle
[0][i
] = src_component
;
1575 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1579 emit_mir_instruction(ctx
, st
);
1581 DBG("Unknown store\n");
1587 /* Special case of store_output for lowered blend shaders */
1588 case nir_intrinsic_store_raw_output_pan
:
1589 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1590 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1592 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1593 /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
1594 * reg = r0.x. We want to splatter. So we can do a 32-bit move
1597 * imov r0.xyzw, r0.xxxx
1600 unsigned expanded
= make_compiler_temp(ctx
);
1602 midgard_instruction splatter
= v_mov(reg
, expanded
);
1604 for (unsigned c
= 0; c
< 16; ++c
)
1605 splatter
.swizzle
[1][c
] = 0;
1607 emit_mir_instruction(ctx
, splatter
);
1608 emit_fragment_store(ctx
, expanded
, ctx
->blend_rt
);
1610 emit_fragment_store(ctx
, reg
, ctx
->blend_rt
);
1614 case nir_intrinsic_store_global
:
1615 case nir_intrinsic_store_shared
:
1616 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1617 emit_explicit_constant(ctx
, reg
, reg
);
1619 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1622 case nir_intrinsic_load_ssbo_address
:
1623 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1626 case nir_intrinsic_get_buffer_size
:
1627 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1630 case nir_intrinsic_load_viewport_scale
:
1631 case nir_intrinsic_load_viewport_offset
:
1632 case nir_intrinsic_load_num_work_groups
:
1633 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1634 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1637 case nir_intrinsic_load_work_group_id
:
1638 case nir_intrinsic_load_local_invocation_id
:
1639 emit_compute_builtin(ctx
, instr
);
1642 case nir_intrinsic_load_vertex_id
:
1643 case nir_intrinsic_load_instance_id
:
1644 emit_vertex_builtin(ctx
, instr
);
1647 case nir_intrinsic_memory_barrier_buffer
:
1648 case nir_intrinsic_memory_barrier_shared
:
1651 case nir_intrinsic_control_barrier
:
1652 schedule_barrier(ctx
);
1653 emit_control_barrier(ctx
);
1654 schedule_barrier(ctx
);
1658 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1665 midgard_tex_format(enum glsl_sampler_dim dim
)
1668 case GLSL_SAMPLER_DIM_1D
:
1669 case GLSL_SAMPLER_DIM_BUF
:
1672 case GLSL_SAMPLER_DIM_2D
:
1673 case GLSL_SAMPLER_DIM_EXTERNAL
:
1674 case GLSL_SAMPLER_DIM_RECT
:
1677 case GLSL_SAMPLER_DIM_3D
:
1680 case GLSL_SAMPLER_DIM_CUBE
:
1681 return MALI_TEX_CUBE
;
1684 DBG("Unknown sampler dim type\n");
1690 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1694 pan_attach_constant_bias(
1695 compiler_context
*ctx
,
1697 midgard_texture_word
*word
)
1699 /* To attach as constant, it has to *be* constant */
1701 if (!nir_src_is_const(lod
))
1704 float f
= nir_src_as_float(lod
);
1706 /* Break into fixed-point */
1708 float lod_frac
= f
- lod_int
;
1710 /* Carry over negative fractions */
1711 if (lod_frac
< 0.0) {
1717 word
->bias
= float_to_ubyte(lod_frac
);
1718 word
->bias_int
= lod_int
;
1723 static enum mali_sampler_type
1724 midgard_sampler_type(nir_alu_type t
) {
1725 switch (nir_alu_type_get_base_type(t
))
1727 case nir_type_float
:
1728 return MALI_SAMPLER_FLOAT
;
1730 return MALI_SAMPLER_SIGNED
;
1732 return MALI_SAMPLER_UNSIGNED
;
1734 unreachable("Unknown sampler type");
1739 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1740 unsigned midgard_texop
)
1743 //assert (!instr->sampler);
1745 int texture_index
= instr
->texture_index
;
1746 int sampler_index
= texture_index
;
1748 /* No helper to build texture words -- we do it all here */
1749 midgard_instruction ins
= {
1750 .type
= TAG_TEXTURE_4
,
1752 .dest
= nir_dest_index(&instr
->dest
),
1753 .src
= { ~0, ~0, ~0, ~0 },
1754 .swizzle
= SWIZZLE_IDENTITY_4
,
1756 .op
= midgard_texop
,
1757 .format
= midgard_tex_format(instr
->sampler_dim
),
1758 .texture_handle
= texture_index
,
1759 .sampler_handle
= sampler_index
,
1765 .sampler_type
= midgard_sampler_type(instr
->dest_type
),
1766 .shadow
= instr
->is_shadow
,
1770 /* We may need a temporary for the coordinate */
1772 bool needs_temp_coord
=
1773 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
1774 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
1777 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
1779 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1780 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1781 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1783 switch (instr
->src
[i
].src_type
) {
1784 case nir_tex_src_coord
: {
1785 emit_explicit_constant(ctx
, index
, index
);
1787 unsigned coord_mask
= mask_of(instr
->coord_components
);
1789 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
1792 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
1794 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1795 /* texelFetch is undefined on samplerCube */
1796 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1798 /* For cubemaps, we use a special ld/st op to
1799 * select the face and copy the xy into the
1800 * texture register */
1802 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
1804 ld
.mask
= 0x3; /* xy */
1805 ld
.load_store
.arg_1
= 0x20;
1806 ld
.swizzle
[1][3] = COMPONENT_X
;
1807 emit_mir_instruction(ctx
, ld
);
1810 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1811 ins
.swizzle
[1][3] = COMPONENT_X
;
1812 } else if (needs_temp_coord
) {
1813 /* mov coord_temp, coords */
1814 midgard_instruction mov
= v_mov(index
, coords
);
1815 mov
.mask
= coord_mask
;
1818 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
1820 emit_mir_instruction(ctx
, mov
);
1825 ins
.src
[1] = coords
;
1827 /* Texelfetch coordinates uses all four elements
1828 * (xyz/index) regardless of texture dimensionality,
1829 * which means it's necessary to zero the unused
1830 * components to keep everything happy */
1832 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1833 /* mov index.zw, #0, or generalized */
1834 midgard_instruction mov
=
1835 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
1836 mov
.has_constants
= true;
1837 mov
.mask
= coord_mask
^ 0xF;
1838 emit_mir_instruction(ctx
, mov
);
1841 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1842 /* Array component in w but NIR wants it in z,
1843 * but if we have a temp coord we already fixed
1846 if (nr_components
== 3) {
1847 ins
.swizzle
[1][2] = COMPONENT_Z
;
1848 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
1849 } else if (nr_components
== 2) {
1851 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1852 ins
.swizzle
[1][3] = COMPONENT_X
;
1854 unreachable("Invalid texture 2D components");
1857 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1859 ins
.swizzle
[1][2] = COMPONENT_Z
;
1860 ins
.swizzle
[1][3] = COMPONENT_W
;
1866 case nir_tex_src_bias
:
1867 case nir_tex_src_lod
: {
1868 /* Try as a constant if we can */
1870 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1871 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1874 ins
.texture
.lod_register
= true;
1877 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1878 ins
.swizzle
[2][c
] = COMPONENT_X
;
1880 emit_explicit_constant(ctx
, index
, index
);
1885 case nir_tex_src_offset
: {
1886 ins
.texture
.offset_register
= true;
1889 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1890 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
1892 emit_explicit_constant(ctx
, index
, index
);
1896 case nir_tex_src_comparator
: {
1897 unsigned comp
= COMPONENT_Z
;
1899 /* mov coord_temp.foo, coords */
1900 midgard_instruction mov
= v_mov(index
, coords
);
1901 mov
.mask
= 1 << comp
;
1903 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
1904 mov
.swizzle
[1][i
] = COMPONENT_X
;
1906 emit_mir_instruction(ctx
, mov
);
1911 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
1917 emit_mir_instruction(ctx
, ins
);
1919 /* Used for .cont and .last hinting */
1920 ctx
->texture_op_count
++;
1924 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1926 switch (instr
->op
) {
1929 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1932 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1935 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1938 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
1941 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
1948 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1950 switch (instr
->type
) {
1951 case nir_jump_break
: {
1952 /* Emit a branch out of the loop */
1953 struct midgard_instruction br
= v_branch(false, false);
1954 br
.branch
.target_type
= TARGET_BREAK
;
1955 br
.branch
.target_break
= ctx
->current_loop_depth
;
1956 emit_mir_instruction(ctx
, br
);
1961 DBG("Unknown jump type %d\n", instr
->type
);
1967 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1969 switch (instr
->type
) {
1970 case nir_instr_type_load_const
:
1971 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1974 case nir_instr_type_intrinsic
:
1975 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1978 case nir_instr_type_alu
:
1979 emit_alu(ctx
, nir_instr_as_alu(instr
));
1982 case nir_instr_type_tex
:
1983 emit_tex(ctx
, nir_instr_as_tex(instr
));
1986 case nir_instr_type_jump
:
1987 emit_jump(ctx
, nir_instr_as_jump(instr
));
1990 case nir_instr_type_ssa_undef
:
1995 DBG("Unhandled instruction type\n");
2001 /* ALU instructions can inline or embed constants, which decreases register
2002 * pressure and saves space. */
2004 #define CONDITIONAL_ATTACH(idx) { \
2005 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
2008 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
2009 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2014 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
2016 mir_foreach_instr_in_block(block
, alu
) {
2017 /* Other instructions cannot inline constants */
2018 if (alu
->type
!= TAG_ALU_4
) continue;
2019 if (alu
->compact_branch
) continue;
2021 /* If there is already a constant here, we can do nothing */
2022 if (alu
->has_constants
) continue;
2024 CONDITIONAL_ATTACH(0);
2026 if (!alu
->has_constants
) {
2027 CONDITIONAL_ATTACH(1)
2028 } else if (!alu
->inline_constant
) {
2029 /* Corner case: _two_ vec4 constants, for instance with a
2030 * csel. For this case, we can only use a constant
2031 * register for one, we'll have to emit a move for the
2034 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2035 unsigned scratch
= make_compiler_temp(ctx
);
2038 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2039 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2041 /* Set the source */
2042 alu
->src
[1] = scratch
;
2044 /* Inject us -before- the last instruction which set r31 */
2045 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2051 /* Being a little silly with the names, but returns the op that is the bitwise
2052 * inverse of the op with the argument switched. I.e. (f and g are
2055 * f(a, b) = ~g(b, a)
2057 * Corollary: if g is the contrapositve of f, f is the contrapositive of g:
2059 * f(a, b) = ~g(b, a)
2060 * ~f(a, b) = g(b, a)
2061 * ~f(a, b) = ~h(a, b) where h is the contrapositive of g
2064 * Thus we define this function in pairs.
2067 static inline midgard_alu_op
2068 mir_contrapositive(midgard_alu_op op
)
2071 case midgard_alu_op_flt
:
2072 return midgard_alu_op_fle
;
2073 case midgard_alu_op_fle
:
2074 return midgard_alu_op_flt
;
2076 case midgard_alu_op_ilt
:
2077 return midgard_alu_op_ile
;
2078 case midgard_alu_op_ile
:
2079 return midgard_alu_op_ilt
;
2082 unreachable("No known contrapositive");
2086 /* Midgard supports two types of constants, embedded constants (128-bit) and
2087 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2088 * constants can be demoted to inline constants, for space savings and
2089 * sometimes a performance boost */
2092 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2094 mir_foreach_instr_in_block(block
, ins
) {
2095 if (!ins
->has_constants
) continue;
2096 if (ins
->has_inline_constant
) continue;
2098 /* Blend constants must not be inlined by definition */
2099 if (ins
->has_blend_constant
) continue;
2101 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2102 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2103 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2105 if (!(is_16
|| is_32
))
2108 /* src1 cannot be an inline constant due to encoding
2109 * restrictions. So, if possible we try to flip the arguments
2112 int op
= ins
->alu
.op
;
2114 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2115 bool flip
= alu_opcode_props
[op
].props
& OP_COMMUTES
;
2118 /* Conditionals can be inverted */
2119 case midgard_alu_op_flt
:
2120 case midgard_alu_op_ilt
:
2121 case midgard_alu_op_fle
:
2122 case midgard_alu_op_ile
:
2123 ins
->alu
.op
= mir_contrapositive(ins
->alu
.op
);
2128 case midgard_alu_op_fcsel
:
2129 case midgard_alu_op_icsel
:
2130 DBG("Missed non-commutative flip (%s)\n", alu_opcode_props
[op
].name
);
2139 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2140 /* Extract the source information */
2142 midgard_vector_alu_src
*src
;
2143 int q
= ins
->alu
.src2
;
2144 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2147 /* Component is from the swizzle. Take a nonzero component */
2149 unsigned first_comp
= ffs(ins
->mask
) - 1;
2150 unsigned component
= ins
->swizzle
[1][first_comp
];
2152 /* Scale constant appropriately, if we can legally */
2153 uint16_t scaled_constant
= 0;
2156 scaled_constant
= ins
->constants
.u16
[component
];
2157 } else if (midgard_is_integer_op(op
)) {
2158 scaled_constant
= ins
->constants
.u32
[component
];
2160 /* Constant overflow after resize */
2161 if (scaled_constant
!= ins
->constants
.u32
[component
])
2164 float original
= ins
->constants
.f32
[component
];
2165 scaled_constant
= _mesa_float_to_half(original
);
2167 /* Check for loss of precision. If this is
2168 * mediump, we don't care, but for a highp
2169 * shader, we need to pay attention. NIR
2170 * doesn't yet tell us which mode we're in!
2171 * Practically this prevents most constants
2172 * from being inlined, sadly. */
2174 float fp32
= _mesa_half_to_float(scaled_constant
);
2176 if (fp32
!= original
)
2180 /* We don't know how to handle these with a constant */
2182 if (mir_nontrivial_source2_mod_simple(ins
) || src
->rep_low
|| src
->rep_high
) {
2183 DBG("Bailing inline constant...\n");
2187 /* Make sure that the constant is not itself a vector
2188 * by checking if all accessed values are the same. */
2190 const midgard_constants
*cons
= &ins
->constants
;
2191 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2193 bool is_vector
= false;
2194 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2196 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2197 /* We only care if this component is actually used */
2198 if (!(mask
& (1 << c
)))
2201 uint32_t test
= is_16
?
2202 cons
->u16
[ins
->swizzle
[1][c
]] :
2203 cons
->u32
[ins
->swizzle
[1][c
]];
2205 if (test
!= value
) {
2214 /* Get rid of the embedded constant */
2215 ins
->has_constants
= false;
2217 ins
->has_inline_constant
= true;
2218 ins
->inline_constant
= scaled_constant
;
2223 /* Dead code elimination for branches at the end of a block - only one branch
2224 * per block is legal semantically */
2227 midgard_opt_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2229 bool branched
= false;
2231 mir_foreach_instr_in_block_safe(block
, ins
) {
2232 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2235 mir_remove_instruction(ins
);
2241 /* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
2242 * the move can be propagated away entirely */
2245 mir_compose_float_outmod(midgard_outmod_float
*outmod
, midgard_outmod_float comp
)
2248 if (comp
== midgard_outmod_none
)
2251 if (*outmod
== midgard_outmod_none
) {
2256 /* TODO: Compose rules */
2261 midgard_opt_pos_propagate(compiler_context
*ctx
, midgard_block
*block
)
2263 bool progress
= false;
2265 mir_foreach_instr_in_block_safe(block
, ins
) {
2266 if (ins
->type
!= TAG_ALU_4
) continue;
2267 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2268 if (ins
->alu
.outmod
!= midgard_outmod_pos
) continue;
2270 /* TODO: Registers? */
2271 unsigned src
= ins
->src
[1];
2272 if (src
& PAN_IS_REG
) continue;
2274 /* There might be a source modifier, too */
2275 if (mir_nontrivial_source2_mod(ins
)) continue;
2277 /* Backpropagate the modifier */
2278 mir_foreach_instr_in_block_from_rev(block
, v
, mir_prev_op(ins
)) {
2279 if (v
->type
!= TAG_ALU_4
) continue;
2280 if (v
->dest
!= src
) continue;
2282 /* Can we even take a float outmod? */
2283 if (midgard_is_integer_out_op(v
->alu
.op
)) continue;
2285 midgard_outmod_float temp
= v
->alu
.outmod
;
2286 progress
|= mir_compose_float_outmod(&temp
, ins
->alu
.outmod
);
2288 /* Throw in the towel.. */
2289 if (!progress
) break;
2291 /* Otherwise, transfer the modifier */
2292 v
->alu
.outmod
= temp
;
2293 ins
->alu
.outmod
= midgard_outmod_none
;
2303 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2305 /* Loop to ourselves */
2306 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2307 struct midgard_instruction ins
= v_branch(false, false);
2308 ins
.writeout
= true;
2309 ins
.writeout_depth
= br
->writeout_depth
;
2310 ins
.writeout_stencil
= br
->writeout_stencil
;
2311 ins
.branch
.target_block
= ctx
->block_count
- 1;
2312 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2313 emit_mir_instruction(ctx
, ins
);
2315 ctx
->current_block
->epilogue
= true;
2316 schedule_barrier(ctx
);
2317 return ins
.branch
.target_block
;
2320 static midgard_block
*
2321 emit_block(compiler_context
*ctx
, nir_block
*block
)
2323 midgard_block
*this_block
= ctx
->after_block
;
2324 ctx
->after_block
= NULL
;
2327 this_block
= create_empty_block(ctx
);
2329 list_addtail(&this_block
->base
.link
, &ctx
->blocks
);
2331 this_block
->scheduled
= false;
2334 /* Set up current block */
2335 list_inithead(&this_block
->base
.instructions
);
2336 ctx
->current_block
= this_block
;
2338 nir_foreach_instr(instr
, block
) {
2339 emit_instr(ctx
, instr
);
2340 ++ctx
->instruction_count
;
2346 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2349 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2351 midgard_block
*before_block
= ctx
->current_block
;
2353 /* Speculatively emit the branch, but we can't fill it in until later */
2354 EMIT(branch
, true, true);
2355 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2356 then_branch
->src
[0] = nir_src_index(ctx
, &nif
->condition
);
2358 /* Emit the two subblocks. */
2359 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2360 midgard_block
*end_then_block
= ctx
->current_block
;
2362 /* Emit a jump from the end of the then block to the end of the else */
2363 EMIT(branch
, false, false);
2364 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2366 /* Emit second block, and check if it's empty */
2368 int else_idx
= ctx
->block_count
;
2369 int count_in
= ctx
->instruction_count
;
2370 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2371 midgard_block
*end_else_block
= ctx
->current_block
;
2372 int after_else_idx
= ctx
->block_count
;
2374 /* Now that we have the subblocks emitted, fix up the branches */
2379 if (ctx
->instruction_count
== count_in
) {
2380 /* The else block is empty, so don't emit an exit jump */
2381 mir_remove_instruction(then_exit
);
2382 then_branch
->branch
.target_block
= after_else_idx
;
2384 then_branch
->branch
.target_block
= else_idx
;
2385 then_exit
->branch
.target_block
= after_else_idx
;
2388 /* Wire up the successors */
2390 ctx
->after_block
= create_empty_block(ctx
);
2392 pan_block_add_successor(&before_block
->base
, &then_block
->base
);
2393 pan_block_add_successor(&before_block
->base
, &else_block
->base
);
2395 pan_block_add_successor(&end_then_block
->base
, &ctx
->after_block
->base
);
2396 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
);
2400 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2402 /* Remember where we are */
2403 midgard_block
*start_block
= ctx
->current_block
;
2405 /* Allocate a loop number, growing the current inner loop depth */
2406 int loop_idx
= ++ctx
->current_loop_depth
;
2408 /* Get index from before the body so we can loop back later */
2409 int start_idx
= ctx
->block_count
;
2411 /* Emit the body itself */
2412 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2414 /* Branch back to loop back */
2415 struct midgard_instruction br_back
= v_branch(false, false);
2416 br_back
.branch
.target_block
= start_idx
;
2417 emit_mir_instruction(ctx
, br_back
);
2419 /* Mark down that branch in the graph. */
2420 pan_block_add_successor(&start_block
->base
, &loop_block
->base
);
2421 pan_block_add_successor(&ctx
->current_block
->base
, &loop_block
->base
);
2423 /* Find the index of the block about to follow us (note: we don't add
2424 * one; blocks are 0-indexed so we get a fencepost problem) */
2425 int break_block_idx
= ctx
->block_count
;
2427 /* Fix up the break statements we emitted to point to the right place,
2428 * now that we can allocate a block number for them */
2429 ctx
->after_block
= create_empty_block(ctx
);
2431 mir_foreach_block_from(ctx
, start_block
, _block
) {
2432 mir_foreach_instr_in_block(((midgard_block
*) _block
), ins
) {
2433 if (ins
->type
!= TAG_ALU_4
) continue;
2434 if (!ins
->compact_branch
) continue;
2436 /* We found a branch -- check the type to see if we need to do anything */
2437 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2439 /* It's a break! Check if it's our break */
2440 if (ins
->branch
.target_break
!= loop_idx
) continue;
2442 /* Okay, cool, we're breaking out of this loop.
2443 * Rewrite from a break to a goto */
2445 ins
->branch
.target_type
= TARGET_GOTO
;
2446 ins
->branch
.target_block
= break_block_idx
;
2448 pan_block_add_successor(_block
, &ctx
->after_block
->base
);
2452 /* Now that we've finished emitting the loop, free up the depth again
2453 * so we play nice with recursion amid nested loops */
2454 --ctx
->current_loop_depth
;
2456 /* Dump loop stats */
2460 static midgard_block
*
2461 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2463 midgard_block
*start_block
= NULL
;
2465 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2466 switch (node
->type
) {
2467 case nir_cf_node_block
: {
2468 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2471 start_block
= block
;
2476 case nir_cf_node_if
:
2477 emit_if(ctx
, nir_cf_node_as_if(node
));
2480 case nir_cf_node_loop
:
2481 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2484 case nir_cf_node_function
:
2493 /* Due to lookahead, we need to report the first tag executed in the command
2494 * stream and in branch targets. An initial block might be empty, so iterate
2495 * until we find one that 'works' */
2498 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2500 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2502 mir_foreach_block_from(ctx
, initial_block
, _v
) {
2503 midgard_block
*v
= (midgard_block
*) _v
;
2504 if (v
->quadword_count
) {
2505 midgard_bundle
*initial_bundle
=
2506 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2508 return initial_bundle
->tag
;
2512 /* Default to a tag 1 which will break from the shader, in case we jump
2513 * to the exit block (i.e. `return` in a compute shader) */
2518 /* For each fragment writeout instruction, generate a writeout loop to
2519 * associate with it */
2522 mir_add_writeout_loops(compiler_context
*ctx
)
2524 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2525 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2528 unsigned popped
= br
->branch
.target_block
;
2529 pan_block_add_successor(&(mir_get_block(ctx
, popped
- 1)->base
), &ctx
->current_block
->base
);
2530 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2532 /* If we have more RTs, we'll need to restore back after our
2533 * loop terminates */
2535 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2536 midgard_instruction uncond
= v_branch(false, false);
2537 uncond
.branch
.target_block
= popped
;
2538 emit_mir_instruction(ctx
, uncond
);
2539 pan_block_add_successor(&ctx
->current_block
->base
, &(mir_get_block(ctx
, popped
)->base
));
2540 schedule_barrier(ctx
);
2542 /* We're last, so we can terminate here */
2543 br
->last_writeout
= true;
2549 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2551 struct util_dynarray
*compiled
= &program
->compiled
;
2553 midgard_debug
= debug_get_option_midgard_debug();
2555 /* TODO: Bound against what? */
2556 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2559 ctx
->stage
= nir
->info
.stage
;
2560 ctx
->is_blend
= is_blend
;
2561 ctx
->alpha_ref
= program
->alpha_ref
;
2562 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2563 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2565 /* Start off with a safe cutoff, allowing usage of all 16 work
2566 * registers. Later, we'll promote uniform reads to uniform registers
2567 * if we determine it is beneficial to do so */
2568 ctx
->uniform_cutoff
= 8;
2570 /* Initialize at a global (not block) level hash tables */
2572 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2573 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2575 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2576 * (so we don't accidentally duplicate the epilogue since mesa/st has
2577 * messed with our I/O quite a bit already) */
2579 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2581 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2582 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2583 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2586 NIR_PASS_V(nir
, nir_lower_var_copies
);
2587 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2588 NIR_PASS_V(nir
, nir_split_var_copies
);
2589 NIR_PASS_V(nir
, nir_lower_var_copies
);
2590 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2591 NIR_PASS_V(nir
, nir_lower_var_copies
);
2592 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2594 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2595 NIR_PASS_V(nir
, nir_lower_ssbo
);
2596 NIR_PASS_V(nir
, midgard_nir_lower_zs_store
);
2598 /* Optimisation passes */
2600 optimise_nir(nir
, ctx
->quirks
);
2602 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2603 nir_print_shader(nir
, stdout
);
2606 /* Assign sysvals and counts, now that we're sure
2607 * (post-optimisation) */
2609 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2610 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2611 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2613 nir_foreach_function(func
, nir
) {
2617 list_inithead(&ctx
->blocks
);
2618 ctx
->block_count
= 0;
2621 emit_cf_list(ctx
, &func
->impl
->body
);
2622 break; /* TODO: Multi-function shaders */
2625 util_dynarray_init(compiled
, NULL
);
2627 /* Per-block lowering before opts */
2629 mir_foreach_block(ctx
, _block
) {
2630 midgard_block
*block
= (midgard_block
*) _block
;
2631 inline_alu_constants(ctx
, block
);
2632 midgard_opt_promote_fmov(ctx
, block
);
2633 embedded_to_inline_constant(ctx
, block
);
2635 /* MIR-level optimizations */
2637 bool progress
= false;
2642 mir_foreach_block(ctx
, _block
) {
2643 midgard_block
*block
= (midgard_block
*) _block
;
2644 progress
|= midgard_opt_pos_propagate(ctx
, block
);
2645 progress
|= midgard_opt_copy_prop(ctx
, block
);
2646 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2647 progress
|= midgard_opt_combine_projection(ctx
, block
);
2648 progress
|= midgard_opt_varying_projection(ctx
, block
);
2649 progress
|= midgard_opt_not_propagate(ctx
, block
);
2650 progress
|= midgard_opt_fuse_src_invert(ctx
, block
);
2651 progress
|= midgard_opt_fuse_dest_invert(ctx
, block
);
2652 progress
|= midgard_opt_csel_invert(ctx
, block
);
2653 progress
|= midgard_opt_drop_cmp_invert(ctx
, block
);
2654 progress
|= midgard_opt_invert_branch(ctx
, block
);
2658 mir_foreach_block(ctx
, _block
) {
2659 midgard_block
*block
= (midgard_block
*) _block
;
2660 midgard_lower_invert(ctx
, block
);
2661 midgard_lower_derivatives(ctx
, block
);
2664 /* Nested control-flow can result in dead branches at the end of the
2665 * block. This messes with our analysis and is just dead code, so cull
2667 mir_foreach_block(ctx
, _block
) {
2668 midgard_block
*block
= (midgard_block
*) _block
;
2669 midgard_opt_cull_dead_branch(ctx
, block
);
2672 /* Ensure we were lowered */
2673 mir_foreach_instr_global(ctx
, ins
) {
2674 assert(!ins
->invert
);
2677 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2678 mir_add_writeout_loops(ctx
);
2681 midgard_schedule_program(ctx
);
2684 /* Now that all the bundles are scheduled and we can calculate block
2685 * sizes, emit actual branch instructions rather than placeholders */
2687 int br_block_idx
= 0;
2689 mir_foreach_block(ctx
, _block
) {
2690 midgard_block
*block
= (midgard_block
*) _block
;
2691 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2692 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2693 midgard_instruction
*ins
= bundle
->instructions
[c
];
2695 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2697 /* Parse some basic branch info */
2698 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2699 bool is_conditional
= ins
->branch
.conditional
;
2700 bool is_inverted
= ins
->branch
.invert_conditional
;
2701 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2702 bool is_writeout
= ins
->writeout
;
2704 /* Determine the block we're jumping to */
2705 int target_number
= ins
->branch
.target_block
;
2707 /* Report the destination tag */
2708 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2710 /* Count up the number of quadwords we're
2711 * jumping over = number of quadwords until
2712 * (br_block_idx, target_number) */
2714 int quadword_offset
= 0;
2718 } else if (target_number
> br_block_idx
) {
2721 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2722 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2725 quadword_offset
+= blk
->quadword_count
;
2728 /* Jump backwards */
2730 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2731 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2734 quadword_offset
-= blk
->quadword_count
;
2738 /* Unconditional extended branches (far jumps)
2739 * have issues, so we always use a conditional
2740 * branch, setting the condition to always for
2741 * unconditional. For compact unconditional
2742 * branches, cond isn't used so it doesn't
2743 * matter what we pick. */
2745 midgard_condition cond
=
2746 !is_conditional
? midgard_condition_always
:
2747 is_inverted
? midgard_condition_false
:
2748 midgard_condition_true
;
2750 midgard_jmp_writeout_op op
=
2751 is_discard
? midgard_jmp_writeout_op_discard
:
2752 is_writeout
? midgard_jmp_writeout_op_writeout
:
2753 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2754 midgard_jmp_writeout_op_branch_cond
;
2757 midgard_branch_extended branch
=
2758 midgard_create_branch_extended(
2763 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2764 } else if (is_conditional
|| is_discard
) {
2765 midgard_branch_cond branch
= {
2767 .dest_tag
= dest_tag
,
2768 .offset
= quadword_offset
,
2772 assert(branch
.offset
== quadword_offset
);
2774 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2776 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2778 midgard_branch_uncond branch
= {
2780 .dest_tag
= dest_tag
,
2781 .offset
= quadword_offset
,
2785 assert(branch
.offset
== quadword_offset
);
2787 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2795 /* Emit flat binary from the instruction arrays. Iterate each block in
2796 * sequence. Save instruction boundaries such that lookahead tags can
2797 * be assigned easily */
2799 /* Cache _all_ bundles in source order for lookahead across failed branches */
2801 int bundle_count
= 0;
2802 mir_foreach_block(ctx
, _block
) {
2803 midgard_block
*block
= (midgard_block
*) _block
;
2804 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2806 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2808 mir_foreach_block(ctx
, _block
) {
2809 midgard_block
*block
= (midgard_block
*) _block
;
2810 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2811 source_order_bundles
[bundle_idx
++] = bundle
;
2815 int current_bundle
= 0;
2817 /* Midgard prefetches instruction types, so during emission we
2818 * need to lookahead. Unless this is the last instruction, in
2819 * which we return 1. */
2821 mir_foreach_block(ctx
, _block
) {
2822 midgard_block
*block
= (midgard_block
*) _block
;
2823 mir_foreach_bundle_in_block(block
, bundle
) {
2826 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2827 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2829 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2833 /* TODO: Free deeper */
2834 //util_dynarray_fini(&block->instructions);
2837 free(source_order_bundles
);
2839 /* Report the very first tag executed */
2840 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2842 /* Deal with off-by-one related to the fencepost problem */
2843 program
->work_register_count
= ctx
->work_registers
+ 1;
2844 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2846 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2847 program
->tls_size
= ctx
->tls_size
;
2849 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2850 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
2852 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
2853 unsigned nr_bundles
= 0, nr_ins
= 0;
2855 /* Count instructions and bundles */
2857 mir_foreach_block(ctx
, _block
) {
2858 midgard_block
*block
= (midgard_block
*) _block
;
2859 nr_bundles
+= util_dynarray_num_elements(
2860 &block
->bundles
, midgard_bundle
);
2862 mir_foreach_bundle_in_block(block
, bun
)
2863 nr_ins
+= bun
->instruction_count
;
2866 /* Calculate thread count. There are certain cutoffs by
2867 * register count for thread count */
2869 unsigned nr_registers
= program
->work_register_count
;
2871 unsigned nr_threads
=
2872 (nr_registers
<= 4) ? 4 :
2873 (nr_registers
<= 8) ? 2 :
2878 fprintf(stderr
, "shader%d - %s shader: "
2879 "%u inst, %u bundles, %u quadwords, "
2880 "%u registers, %u threads, %u loops, "
2881 "%u:%u spills:fills\n",
2883 gl_shader_stage_name(ctx
->stage
),
2884 nr_ins
, nr_bundles
, ctx
->quadword_count
,
2885 nr_registers
, nr_threads
,
2887 ctx
->spills
, ctx
->fills
);