2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 #include <sys/types.h>
33 #include "main/mtypes.h"
34 #include "compiler/glsl/glsl_to_nir.h"
35 #include "compiler/nir_types.h"
36 #include "compiler/nir/nir_builder.h"
37 #include "util/half_float.h"
38 #include "util/u_math.h"
39 #include "util/u_debug.h"
40 #include "util/u_dynarray.h"
41 #include "util/list.h"
42 #include "main/mtypes.h"
45 #include "midgard_nir.h"
46 #include "midgard_compile.h"
47 #include "midgard_ops.h"
50 #include "midgard_quirks.h"
52 #include "disassemble.h"
54 static const struct debug_named_value debug_options
[] = {
55 {"msgs", MIDGARD_DBG_MSGS
, "Print debug messages"},
56 {"shaders", MIDGARD_DBG_SHADERS
, "Dump shaders in NIR and MIR"},
57 {"shaderdb", MIDGARD_DBG_SHADERDB
, "Prints shader-db statistics"},
61 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug
, "MIDGARD_MESA_DEBUG", debug_options
, 0)
63 unsigned SHADER_DB_COUNT
= 0;
65 int midgard_debug
= 0;
67 #define DBG(fmt, ...) \
68 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
69 fprintf(stderr, "%s:%d: "fmt, \
70 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
71 static midgard_block
*
72 create_empty_block(compiler_context
*ctx
)
74 midgard_block
*blk
= rzalloc(ctx
, midgard_block
);
76 blk
->base
.predecessors
= _mesa_set_create(blk
,
78 _mesa_key_pointer_equal
);
80 blk
->base
.name
= ctx
->block_source_count
++;
86 schedule_barrier(compiler_context
*ctx
)
88 midgard_block
*temp
= ctx
->after_block
;
89 ctx
->after_block
= create_empty_block(ctx
);
91 list_addtail(&ctx
->after_block
->base
.link
, &ctx
->blocks
);
92 list_inithead(&ctx
->after_block
->base
.instructions
);
93 pan_block_add_successor(&ctx
->current_block
->base
, &ctx
->after_block
->base
);
94 ctx
->current_block
= ctx
->after_block
;
95 ctx
->after_block
= temp
;
98 /* Helpers to generate midgard_instruction's using macro magic, since every
99 * driver seems to do it that way */
101 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
103 #define M_LOAD_STORE(name, store, T) \
104 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
105 midgard_instruction i = { \
106 .type = TAG_LOAD_STORE_4, \
109 .src = { ~0, ~0, ~0, ~0 }, \
110 .swizzle = SWIZZLE_IDENTITY_4, \
112 .op = midgard_op_##name, \
119 i.src_types[0] = T; \
127 #define M_LOAD(name, T) M_LOAD_STORE(name, false, T)
128 #define M_STORE(name, T) M_LOAD_STORE(name, true, T)
130 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
131 * the corresponding Midgard source */
133 static midgard_vector_alu_src
134 vector_alu_modifiers(nir_alu_src
*src
, bool is_int
, unsigned broadcast_count
,
135 bool half
, bool sext
)
137 /* Figure out how many components there are so we can adjust.
138 * Specifically we want to broadcast the last channel so things like
142 if (broadcast_count
&& src
) {
143 uint8_t last_component
= src
->swizzle
[broadcast_count
- 1];
145 for (unsigned c
= broadcast_count
; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
146 src
->swizzle
[c
] = last_component
;
150 midgard_vector_alu_src alu_src
= {
157 alu_src
.mod
= midgard_int_normal
;
159 /* Sign/zero-extend if needed */
163 midgard_int_sign_extend
164 : midgard_int_zero_extend
;
167 /* These should have been lowered away */
169 assert(!(src
->abs
|| src
->negate
));
172 alu_src
.mod
= (src
->abs
<< 0) | (src
->negate
<< 1);
178 M_LOAD(ld_attr_32
, nir_type_uint32
);
179 M_LOAD(ld_vary_32
, nir_type_uint32
);
180 M_LOAD(ld_ubo_int4
, nir_type_uint32
);
181 M_LOAD(ld_int4
, nir_type_uint32
);
182 M_STORE(st_int4
, nir_type_uint32
);
183 M_LOAD(ld_color_buffer_32u
, nir_type_uint32
);
184 M_STORE(st_vary_32
, nir_type_uint32
);
185 M_LOAD(ld_cubemap_coords
, nir_type_uint32
);
186 M_LOAD(ld_compute_id
, nir_type_uint32
);
188 static midgard_instruction
189 v_branch(bool conditional
, bool invert
)
191 midgard_instruction ins
= {
193 .unit
= ALU_ENAB_BRANCH
,
194 .compact_branch
= true,
196 .conditional
= conditional
,
197 .invert_conditional
= invert
200 .src
= { ~0, ~0, ~0, ~0 },
206 static midgard_branch_extended
207 midgard_create_branch_extended( midgard_condition cond
,
208 midgard_jmp_writeout_op op
,
210 signed quadword_offset
)
212 /* The condition code is actually a LUT describing a function to
213 * combine multiple condition codes. However, we only support a single
214 * condition code at the moment, so we just duplicate over a bunch of
217 uint16_t duplicated_cond
=
227 midgard_branch_extended branch
= {
229 .dest_tag
= dest_tag
,
230 .offset
= quadword_offset
,
231 .cond
= duplicated_cond
238 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
240 ins
->has_constants
= true;
241 memcpy(&ins
->constants
, constants
, 16);
245 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
247 return glsl_count_attribute_slots(type
, false);
250 /* Lower fdot2 to a vector multiplication followed by channel addition */
252 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
254 if (alu
->op
!= nir_op_fdot2
)
257 b
->cursor
= nir_before_instr(&alu
->instr
);
259 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
260 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
262 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
264 nir_ssa_def
*sum
= nir_fadd(b
,
265 nir_channel(b
, product
, 0),
266 nir_channel(b
, product
, 1));
268 /* Replace the fdot2 with this sum */
269 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
273 midgard_nir_lower_fdot2(nir_shader
*shader
)
275 bool progress
= false;
277 nir_foreach_function(function
, shader
) {
278 if (!function
->impl
) continue;
281 nir_builder
*b
= &_b
;
282 nir_builder_init(b
, function
->impl
);
284 nir_foreach_block(block
, function
->impl
) {
285 nir_foreach_instr_safe(instr
, block
) {
286 if (instr
->type
!= nir_instr_type_alu
) continue;
288 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
289 midgard_nir_lower_fdot2_body(b
, alu
);
295 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
302 /* Midgard can't write depth and stencil separately. It has to happen in a
303 * single store operation containing both. Let's add a panfrost specific
304 * intrinsic and turn all depth/stencil stores into a packed depth+stencil
308 midgard_nir_lower_zs_store(nir_shader
*nir
)
310 if (nir
->info
.stage
!= MESA_SHADER_FRAGMENT
)
313 nir_variable
*z_var
= NULL
, *s_var
= NULL
;
315 nir_foreach_variable(var
, &nir
->outputs
) {
316 if (var
->data
.location
== FRAG_RESULT_DEPTH
)
318 else if (var
->data
.location
== FRAG_RESULT_STENCIL
)
322 if (!z_var
&& !s_var
)
325 bool progress
= false;
327 nir_foreach_function(function
, nir
) {
328 if (!function
->impl
) continue;
330 nir_intrinsic_instr
*z_store
= NULL
, *s_store
= NULL
, *last_store
= NULL
;
332 nir_foreach_block(block
, function
->impl
) {
333 nir_foreach_instr_safe(instr
, block
) {
334 if (instr
->type
!= nir_instr_type_intrinsic
)
337 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
338 if (intr
->intrinsic
!= nir_intrinsic_store_output
)
341 if (z_var
&& nir_intrinsic_base(intr
) == z_var
->data
.driver_location
) {
347 if (s_var
&& nir_intrinsic_base(intr
) == s_var
->data
.driver_location
) {
355 if (!z_store
&& !s_store
) continue;
358 nir_builder_init(&b
, function
->impl
);
360 b
.cursor
= nir_before_instr(&last_store
->instr
);
362 nir_ssa_def
*zs_store_src
;
364 if (z_store
&& s_store
) {
365 nir_ssa_def
*srcs
[2] = {
366 nir_ssa_for_src(&b
, z_store
->src
[0], 1),
367 nir_ssa_for_src(&b
, s_store
->src
[0], 1),
370 zs_store_src
= nir_vec(&b
, srcs
, 2);
372 zs_store_src
= nir_ssa_for_src(&b
, last_store
->src
[0], 1);
375 nir_intrinsic_instr
*zs_store
;
377 zs_store
= nir_intrinsic_instr_create(b
.shader
,
378 nir_intrinsic_store_zs_output_pan
);
379 zs_store
->src
[0] = nir_src_for_ssa(zs_store_src
);
380 zs_store
->num_components
= z_store
&& s_store
? 2 : 1;
381 nir_intrinsic_set_component(zs_store
, z_store
? 0 : 1);
383 /* Replace the Z and S store by a ZS store */
384 nir_builder_instr_insert(&b
, &zs_store
->instr
);
387 nir_instr_remove(&z_store
->instr
);
390 nir_instr_remove(&s_store
->instr
);
392 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
399 /* Flushes undefined values to zero */
402 optimise_nir(nir_shader
*nir
, unsigned quirks
)
405 unsigned lower_flrp
=
406 (nir
->options
->lower_flrp16
? 16 : 0) |
407 (nir
->options
->lower_flrp32
? 32 : 0) |
408 (nir
->options
->lower_flrp64
? 64 : 0);
410 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
411 NIR_PASS(progress
, nir
, nir_lower_idiv
, nir_lower_idiv_fast
);
413 nir_lower_tex_options lower_tex_options
= {
414 .lower_txs_lod
= true,
416 .lower_tex_without_implicit_lod
=
417 (quirks
& MIDGARD_EXPLICIT_LOD
),
419 /* TODO: we have native gradient.. */
423 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
425 /* Must lower fdot2 after tex is lowered */
426 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
428 /* T720 is broken. */
430 if (quirks
& MIDGARD_BROKEN_LOD
)
431 NIR_PASS_V(nir
, midgard_nir_lod_errata
);
436 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
437 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
439 NIR_PASS(progress
, nir
, nir_copy_prop
);
440 NIR_PASS(progress
, nir
, nir_opt_remove_phis
);
441 NIR_PASS(progress
, nir
, nir_opt_dce
);
442 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
443 NIR_PASS(progress
, nir
, nir_opt_cse
);
444 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
445 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
446 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
448 if (lower_flrp
!= 0) {
449 bool lower_flrp_progress
= false;
450 NIR_PASS(lower_flrp_progress
,
454 false /* always_precise */,
455 nir
->options
->lower_ffma
);
456 if (lower_flrp_progress
) {
457 NIR_PASS(progress
, nir
,
458 nir_opt_constant_folding
);
462 /* Nothing should rematerialize any flrps, so we only
463 * need to do this lowering once.
468 NIR_PASS(progress
, nir
, nir_opt_undef
);
469 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
471 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
474 nir_var_function_temp
);
476 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
479 /* Must be run at the end to prevent creation of fsin/fcos ops */
480 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
485 NIR_PASS(progress
, nir
, nir_opt_dce
);
486 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
487 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
488 NIR_PASS(progress
, nir
, nir_copy_prop
);
491 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
493 /* We implement booleans as 32-bit 0/~0 */
494 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
496 /* Now that booleans are lowered, we can run out late opts */
497 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
499 /* Lower mods for float ops only. Integer ops don't support modifiers
500 * (saturate doesn't make sense on integers, neg/abs require dedicated
503 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
504 NIR_PASS(progress
, nir
, nir_copy_prop
);
505 NIR_PASS(progress
, nir
, nir_opt_dce
);
507 /* Take us out of SSA */
508 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
509 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
511 /* We are a vector architecture; write combine where possible */
512 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
513 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
515 NIR_PASS(progress
, nir
, nir_opt_dce
);
518 /* Do not actually emit a load; instead, cache the constant for inlining */
521 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
523 nir_ssa_def def
= instr
->def
;
525 midgard_constants
*consts
= rzalloc(NULL
, midgard_constants
);
527 assert(instr
->def
.num_components
* instr
->def
.bit_size
<= sizeof(*consts
) * 8);
529 #define RAW_CONST_COPY(bits) \
530 nir_const_value_to_array(consts->u##bits, instr->value, \
531 instr->def.num_components, u##bits)
533 switch (instr
->def
.bit_size
) {
547 unreachable("Invalid bit_size for load_const instruction\n");
550 /* Shifted for SSA, +1 for off-by-one */
551 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, consts
);
554 /* Normally constants are embedded implicitly, but for I/O and such we have to
555 * explicitly emit a move with the constant source */
558 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
560 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
562 if (constant_value
) {
563 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), to
);
564 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
565 emit_mir_instruction(ctx
, ins
);
570 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
572 unsigned comp
= src
->swizzle
[0];
574 for (unsigned c
= 1; c
< nr_components
; ++c
) {
575 if (src
->swizzle
[c
] != comp
)
582 #define ALU_CASE(nir, _op) \
584 op = midgard_alu_op_##_op; \
585 assert(src_bitsize == dst_bitsize); \
588 #define ALU_CASE_BCAST(nir, _op, count) \
590 op = midgard_alu_op_##_op; \
591 broadcast_swizzle = count; \
592 assert(src_bitsize == dst_bitsize); \
595 nir_is_fzero_constant(nir_src src
)
597 if (!nir_src_is_const(src
))
600 for (unsigned c
= 0; c
< nir_src_num_components(src
); ++c
) {
601 if (nir_src_comp_as_float(src
, c
) != 0.0)
608 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
609 * special treatment override this anyway. */
611 static midgard_reg_mode
612 reg_mode_for_nir(nir_alu_instr
*instr
)
614 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
616 switch (src_bitsize
) {
618 return midgard_reg_mode_8
;
620 return midgard_reg_mode_16
;
622 return midgard_reg_mode_32
;
624 return midgard_reg_mode_64
;
626 unreachable("Invalid bit size");
631 mir_copy_src(midgard_instruction
*ins
, nir_alu_instr
*instr
, unsigned i
, unsigned to
)
633 unsigned bits
= nir_src_bit_size(instr
->src
[i
].src
);
635 ins
->src
[to
] = nir_src_index(NULL
, &instr
->src
[i
].src
);
636 ins
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
] | bits
;
640 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
642 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
643 * is handled elsewhere */
645 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
646 midgard_emit_derivatives(ctx
, instr
);
650 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
652 unsigned nr_components
= nir_dest_num_components(instr
->dest
.dest
);
653 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
656 /* Number of components valid to check for the instruction (the rest
657 * will be forced to the last), or 0 to use as-is. Relevant as
658 * ball-type instructions have a channel count in NIR but are all vec4
661 unsigned broadcast_swizzle
= 0;
663 /* What register mode should we operate in? */
664 midgard_reg_mode reg_mode
=
665 reg_mode_for_nir(instr
);
667 /* Do we need a destination override? Used for inline
670 midgard_dest_override dest_override
=
671 midgard_dest_override_none
;
673 /* Should we use a smaller respective source and sign-extend? */
675 bool half_1
= false, sext_1
= false;
676 bool half_2
= false, sext_2
= false;
678 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
679 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
682 ALU_CASE(fadd
, fadd
);
683 ALU_CASE(fmul
, fmul
);
684 ALU_CASE(fmin
, fmin
);
685 ALU_CASE(fmax
, fmax
);
686 ALU_CASE(imin
, imin
);
687 ALU_CASE(imax
, imax
);
688 ALU_CASE(umin
, umin
);
689 ALU_CASE(umax
, umax
);
690 ALU_CASE(ffloor
, ffloor
);
691 ALU_CASE(fround_even
, froundeven
);
692 ALU_CASE(ftrunc
, ftrunc
);
693 ALU_CASE(fceil
, fceil
);
694 ALU_CASE(fdot3
, fdot3
);
695 ALU_CASE(fdot4
, fdot4
);
696 ALU_CASE(iadd
, iadd
);
697 ALU_CASE(isub
, isub
);
698 ALU_CASE(imul
, imul
);
700 /* Zero shoved as second-arg */
701 ALU_CASE(iabs
, iabsdiff
);
705 ALU_CASE(feq32
, feq
);
706 ALU_CASE(fne32
, fne
);
707 ALU_CASE(flt32
, flt
);
708 ALU_CASE(ieq32
, ieq
);
709 ALU_CASE(ine32
, ine
);
710 ALU_CASE(ilt32
, ilt
);
711 ALU_CASE(ult32
, ult
);
713 /* We don't have a native b2f32 instruction. Instead, like many
714 * GPUs, we exploit booleans as 0/~0 for false/true, and
715 * correspondingly AND
716 * by 1.0 to do the type conversion. For the moment, prime us
719 * iand [whatever], #0
721 * At the end of emit_alu (as MIR), we'll fix-up the constant
724 ALU_CASE(b2f32
, iand
);
725 ALU_CASE(b2i32
, iand
);
727 /* Likewise, we don't have a dedicated f2b32 instruction, but
728 * we can do a "not equal to 0.0" test. */
730 ALU_CASE(f2b32
, fne
);
731 ALU_CASE(i2b32
, ine
);
733 ALU_CASE(frcp
, frcp
);
734 ALU_CASE(frsq
, frsqrt
);
735 ALU_CASE(fsqrt
, fsqrt
);
736 ALU_CASE(fexp2
, fexp2
);
737 ALU_CASE(flog2
, flog2
);
739 ALU_CASE(f2i64
, f2i_rtz
);
740 ALU_CASE(f2u64
, f2u_rtz
);
741 ALU_CASE(i2f64
, i2f_rtz
);
742 ALU_CASE(u2f64
, u2f_rtz
);
744 ALU_CASE(f2i32
, f2i_rtz
);
745 ALU_CASE(f2u32
, f2u_rtz
);
746 ALU_CASE(i2f32
, i2f_rtz
);
747 ALU_CASE(u2f32
, u2f_rtz
);
749 ALU_CASE(f2i16
, f2i_rtz
);
750 ALU_CASE(f2u16
, f2u_rtz
);
751 ALU_CASE(i2f16
, i2f_rtz
);
752 ALU_CASE(u2f16
, u2f_rtz
);
754 ALU_CASE(fsin
, fsin
);
755 ALU_CASE(fcos
, fcos
);
757 /* We'll set invert */
758 ALU_CASE(inot
, imov
);
759 ALU_CASE(iand
, iand
);
761 ALU_CASE(ixor
, ixor
);
762 ALU_CASE(ishl
, ishl
);
763 ALU_CASE(ishr
, iasr
);
764 ALU_CASE(ushr
, ilsr
);
766 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
767 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
768 ALU_CASE(b32all_fequal4
, fball_eq
);
770 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
771 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
772 ALU_CASE(b32any_fnequal4
, fbany_neq
);
774 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
775 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
776 ALU_CASE(b32all_iequal4
, iball_eq
);
778 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
779 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
780 ALU_CASE(b32any_inequal4
, ibany_neq
);
782 /* Source mods will be shoved in later */
783 ALU_CASE(fabs
, fmov
);
784 ALU_CASE(fneg
, fmov
);
785 ALU_CASE(fsat
, fmov
);
787 /* For size conversion, we use a move. Ideally though we would squash
788 * these ops together; maybe that has to happen after in NIR as part of
789 * propagation...? An earlier algebraic pass ensured we step down by
790 * only / exactly one size. If stepping down, we use a dest override to
791 * reduce the size; if stepping up, we use a larger-sized move with a
792 * half source and a sign/zero-extension modifier */
798 /* If we end up upscale, we'll need a sign-extend on the
799 * operand (the second argument) */
810 if (instr
->op
== nir_op_f2f16
|| instr
->op
== nir_op_f2f32
||
811 instr
->op
== nir_op_f2f64
)
812 op
= midgard_alu_op_fmov
;
814 op
= midgard_alu_op_imov
;
816 if (dst_bitsize
== (src_bitsize
* 2)) {
820 /* Use a greater register mode */
822 } else if (src_bitsize
== (dst_bitsize
* 2)) {
823 /* Converting down */
824 dest_override
= midgard_dest_override_lower
;
830 /* For greater-or-equal, we lower to less-or-equal and flip the
838 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
839 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
840 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
841 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
844 /* Swap via temporary */
845 nir_alu_src temp
= instr
->src
[1];
846 instr
->src
[1] = instr
->src
[0];
847 instr
->src
[0] = temp
;
852 case nir_op_b32csel
: {
853 /* Midgard features both fcsel and icsel, depending on
854 * the type of the arguments/output. However, as long
855 * as we're careful we can _always_ use icsel and
856 * _never_ need fcsel, since the latter does additional
857 * floating-point-specific processing whereas the
858 * former just moves bits on the wire. It's not obvious
859 * why these are separate opcodes, save for the ability
860 * to do things like sat/pos/abs/neg for free */
862 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
863 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
865 /* The condition is the first argument; move the other
866 * arguments up one to be a binary instruction for
867 * Midgard with the condition last */
869 nir_alu_src temp
= instr
->src
[2];
871 instr
->src
[2] = instr
->src
[0];
872 instr
->src
[0] = instr
->src
[1];
873 instr
->src
[1] = temp
;
879 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
884 /* Midgard can perform certain modifiers on output of an ALU op */
887 if (midgard_is_integer_out_op(op
)) {
888 outmod
= midgard_outmod_int_wrap
;
890 bool sat
= instr
->dest
.saturate
|| instr
->op
== nir_op_fsat
;
891 outmod
= sat
? midgard_outmod_sat
: midgard_outmod_none
;
894 /* fmax(a, 0.0) can turn into a .pos modifier as an optimization */
896 if (instr
->op
== nir_op_fmax
) {
897 if (nir_is_fzero_constant(instr
->src
[0].src
)) {
898 op
= midgard_alu_op_fmov
;
900 outmod
= midgard_outmod_pos
;
901 instr
->src
[0] = instr
->src
[1];
902 } else if (nir_is_fzero_constant(instr
->src
[1].src
)) {
903 op
= midgard_alu_op_fmov
;
905 outmod
= midgard_outmod_pos
;
909 /* Fetch unit, quirks, etc information */
910 unsigned opcode_props
= alu_opcode_props
[op
].props
;
911 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
913 midgard_instruction ins
= {
915 .dest
= nir_dest_index(&instr
->dest
.dest
),
916 .dest_type
= nir_op_infos
[instr
->op
].output_type
917 | nir_dest_bit_size(instr
->dest
.dest
),
920 for (unsigned i
= nr_inputs
; i
< ARRAY_SIZE(ins
.src
); ++i
)
923 if (quirk_flipped_r24
) {
925 mir_copy_src(&ins
, instr
, 0, 1);
927 for (unsigned i
= 0; i
< nr_inputs
; ++i
)
928 mir_copy_src(&ins
, instr
, i
, quirk_flipped_r24
? 1 : i
);
931 nir_alu_src
*nirmods
[3] = { NULL
};
933 if (nr_inputs
>= 2) {
934 nirmods
[0] = &instr
->src
[0];
935 nirmods
[1] = &instr
->src
[1];
936 } else if (nr_inputs
== 1) {
937 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
943 nirmods
[2] = &instr
->src
[2];
945 /* These were lowered to a move, so apply the corresponding mod */
947 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
948 nir_alu_src
*s
= nirmods
[quirk_flipped_r24
];
950 if (instr
->op
== nir_op_fneg
)
951 s
->negate
= !s
->negate
;
953 if (instr
->op
== nir_op_fabs
)
957 bool is_int
= midgard_is_integer_op(op
);
959 ins
.mask
= mask_of(nr_components
);
961 midgard_vector_alu alu
= {
963 .reg_mode
= reg_mode
,
964 .dest_override
= dest_override
,
967 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, sext_1
)),
968 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
, broadcast_swizzle
, half_2
, sext_2
)),
971 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
974 ins
.mask
&= instr
->dest
.write_mask
;
976 for (unsigned m
= 0; m
< 3; ++m
) {
980 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; ++c
)
981 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[c
];
983 /* Replicate. TODO: remove when vec16 lands */
984 for (unsigned c
= NIR_MAX_VEC_COMPONENTS
; c
< MIR_VEC_COMPONENTS
; ++c
)
985 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[NIR_MAX_VEC_COMPONENTS
- 1];
988 if (nr_inputs
== 3) {
989 /* Conditions can't have mods */
990 assert(!nirmods
[2]->abs
);
991 assert(!nirmods
[2]->negate
);
996 /* Late fixup for emulated instructions */
998 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
999 /* Presently, our second argument is an inline #0 constant.
1000 * Switch over to an embedded 1.0 constant (that can't fit
1001 * inline, since we're 32-bit, not 16-bit like the inline
1004 ins
.has_inline_constant
= false;
1005 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1006 ins
.has_constants
= true;
1008 if (instr
->op
== nir_op_b2f32
)
1009 ins
.constants
.f32
[0] = 1.0f
;
1011 ins
.constants
.i32
[0] = 1;
1013 for (unsigned c
= 0; c
< 16; ++c
)
1014 ins
.swizzle
[1][c
] = 0;
1015 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1016 /* Lots of instructions need a 0 plonked in */
1017 ins
.has_inline_constant
= false;
1018 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1019 ins
.has_constants
= true;
1020 ins
.constants
.u32
[0] = 0;
1022 for (unsigned c
= 0; c
< 16; ++c
)
1023 ins
.swizzle
[1][c
] = 0;
1024 } else if (instr
->op
== nir_op_inot
) {
1028 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1029 /* To avoid duplicating the lookup tables (probably), true LUT
1030 * instructions can only operate as if they were scalars. Lower
1031 * them here by changing the component. */
1033 unsigned orig_mask
= ins
.mask
;
1035 for (int i
= 0; i
< nr_components
; ++i
) {
1036 /* Mask the associated component, dropping the
1037 * instruction if needed */
1040 ins
.mask
&= orig_mask
;
1045 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1046 ins
.swizzle
[0][j
] = nirmods
[0]->swizzle
[i
]; /* Pull from the correct component */
1048 emit_mir_instruction(ctx
, ins
);
1051 emit_mir_instruction(ctx
, ins
);
1058 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1060 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1061 unsigned nir_mask
= 0;
1065 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1066 dsize
= nir_dest_bit_size(intr
->dest
);
1068 nir_mask
= nir_intrinsic_write_mask(intr
);
1072 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1073 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1074 mir_set_bytemask(ins
, bytemask
);
1077 ins
->load_64
= true;
1080 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1081 * optimized) versions of UBO #0 */
1083 static midgard_instruction
*
1085 compiler_context
*ctx
,
1089 nir_src
*indirect_offset
,
1090 unsigned indirect_shift
,
1093 /* TODO: half-floats */
1095 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1096 ins
.constants
.u32
[0] = offset
;
1098 if (instr
->type
== nir_instr_type_intrinsic
)
1099 mir_set_intr_mask(instr
, &ins
, true);
1101 if (indirect_offset
) {
1102 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1103 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1105 ins
.load_store
.arg_2
= 0x1E;
1108 ins
.load_store
.arg_1
= index
;
1110 return emit_mir_instruction(ctx
, ins
);
1113 /* Globals are like UBOs if you squint. And shared memory is like globals if
1114 * you squint even harder */
1118 compiler_context
*ctx
,
1127 midgard_instruction ins
;
1130 ins
= m_ld_int4(srcdest
, 0);
1132 ins
= m_st_int4(srcdest
, 0);
1134 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1135 mir_set_intr_mask(instr
, &ins
, is_read
);
1137 emit_mir_instruction(ctx
, ins
);
1142 compiler_context
*ctx
,
1143 unsigned dest
, unsigned offset
,
1144 unsigned nr_comp
, unsigned component
,
1145 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1147 /* XXX: Half-floats? */
1148 /* TODO: swizzle, mask */
1150 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1151 ins
.mask
= mask_of(nr_comp
);
1153 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1154 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1156 midgard_varying_parameter p
= {
1158 .interpolation
= midgard_interp_default
,
1163 memcpy(&u
, &p
, sizeof(p
));
1164 ins
.load_store
.varying_parameters
= u
;
1166 if (indirect_offset
)
1167 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1169 ins
.load_store
.arg_2
= 0x1E;
1171 ins
.load_store
.arg_1
= 0x9E;
1173 /* Use the type appropriate load */
1177 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1180 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1182 case nir_type_float
:
1183 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1186 unreachable("Attempted to load unknown type");
1190 emit_mir_instruction(ctx
, ins
);
1195 compiler_context
*ctx
,
1196 unsigned dest
, unsigned offset
,
1197 unsigned nr_comp
, nir_alu_type t
)
1199 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1200 ins
.load_store
.arg_1
= 0x1E;
1201 ins
.load_store
.arg_2
= 0x1E;
1202 ins
.mask
= mask_of(nr_comp
);
1204 /* Use the type appropriate load */
1208 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1211 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1213 case nir_type_float
:
1214 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1217 unreachable("Attempted to load unknown type");
1221 emit_mir_instruction(ctx
, ins
);
1225 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1226 unsigned nr_components
, unsigned offset
)
1230 /* Figure out which uniform this is */
1231 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1232 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1234 unsigned dest
= nir_dest_index(&nir_dest
);
1236 /* Sysvals are prefix uniforms */
1237 unsigned uniform
= ((uintptr_t) val
) - 1;
1239 /* Emit the read itself -- this is never indirect */
1240 midgard_instruction
*ins
=
1241 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1243 ins
->mask
= mask_of(nr_components
);
1247 compute_builtin_arg(nir_op op
)
1250 case nir_intrinsic_load_work_group_id
:
1252 case nir_intrinsic_load_local_invocation_id
:
1255 unreachable("Invalid compute paramater loaded");
1260 emit_fragment_store(compiler_context
*ctx
, unsigned src
, enum midgard_rt_id rt
)
1262 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1264 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1268 emit_explicit_constant(ctx
, src
, src
);
1270 struct midgard_instruction ins
=
1271 v_branch(false, false);
1273 ins
.writeout
= true;
1275 /* Add dependencies */
1277 ins
.constants
.u32
[0] = rt
== MIDGARD_ZS_RT
?
1278 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1280 /* Emit the branch */
1281 br
= emit_mir_instruction(ctx
, ins
);
1282 schedule_barrier(ctx
);
1283 ctx
->writeout_branch
[rt
] = br
;
1285 /* Push our current location = current block count - 1 = where we'll
1286 * jump to. Maybe a bit too clever for my own good */
1288 br
->branch
.target_block
= ctx
->block_count
- 1;
1292 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1294 unsigned reg
= nir_dest_index(&instr
->dest
);
1295 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1296 ins
.mask
= mask_of(3);
1297 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1298 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1299 emit_mir_instruction(ctx
, ins
);
1303 vertex_builtin_arg(nir_op op
)
1306 case nir_intrinsic_load_vertex_id
:
1307 return PAN_VERTEX_ID
;
1308 case nir_intrinsic_load_instance_id
:
1309 return PAN_INSTANCE_ID
;
1311 unreachable("Invalid vertex builtin");
1316 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1318 unsigned reg
= nir_dest_index(&instr
->dest
);
1319 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1323 emit_control_barrier(compiler_context
*ctx
)
1325 midgard_instruction ins
= {
1326 .type
= TAG_TEXTURE_4
,
1327 .src
= { ~0, ~0, ~0, ~0 },
1329 .op
= TEXTURE_OP_BARRIER
,
1331 /* TODO: optimize */
1332 .barrier_buffer
= 1,
1337 emit_mir_instruction(ctx
, ins
);
1340 static const nir_variable
*
1341 search_var(struct exec_list
*vars
, unsigned driver_loc
)
1343 nir_foreach_variable(var
, vars
) {
1344 if (var
->data
.driver_location
== driver_loc
)
1352 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1354 unsigned offset
= 0, reg
;
1356 switch (instr
->intrinsic
) {
1357 case nir_intrinsic_discard_if
:
1358 case nir_intrinsic_discard
: {
1359 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1360 struct midgard_instruction discard
= v_branch(conditional
, false);
1361 discard
.branch
.target_type
= TARGET_DISCARD
;
1364 discard
.src
[0] = nir_src_index(ctx
, &instr
->src
[0]);
1366 emit_mir_instruction(ctx
, discard
);
1367 schedule_barrier(ctx
);
1372 case nir_intrinsic_load_uniform
:
1373 case nir_intrinsic_load_ubo
:
1374 case nir_intrinsic_load_global
:
1375 case nir_intrinsic_load_shared
:
1376 case nir_intrinsic_load_input
:
1377 case nir_intrinsic_load_interpolated_input
: {
1378 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1379 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1380 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1381 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1382 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1383 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1385 /* Get the base type of the intrinsic */
1386 /* TODO: Infer type? Does it matter? */
1388 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1389 (is_interp
) ? nir_type_float
:
1390 nir_intrinsic_type(instr
);
1392 t
= nir_alu_type_get_base_type(t
);
1394 if (!(is_ubo
|| is_global
)) {
1395 offset
= nir_intrinsic_base(instr
);
1398 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1400 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1402 bool direct
= nir_src_is_const(*src_offset
);
1403 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1406 offset
+= nir_src_as_uint(*src_offset
);
1408 /* We may need to apply a fractional offset */
1409 int component
= (is_flat
|| is_interp
) ?
1410 nir_intrinsic_component(instr
) : 0;
1411 reg
= nir_dest_index(&instr
->dest
);
1413 if (is_uniform
&& !ctx
->is_blend
) {
1414 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1415 } else if (is_ubo
) {
1416 nir_src index
= instr
->src
[0];
1418 /* TODO: Is indirect block number possible? */
1419 assert(nir_src_is_const(index
));
1421 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1422 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1423 } else if (is_global
|| is_shared
) {
1424 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1425 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1426 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
, is_flat
);
1427 } else if (ctx
->is_blend
) {
1428 /* For blend shaders, load the input color, which is
1429 * preloaded to r0 */
1431 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), reg
);
1432 emit_mir_instruction(ctx
, move
);
1433 schedule_barrier(ctx
);
1434 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1435 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1437 DBG("Unknown load\n");
1444 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1445 case nir_intrinsic_load_barycentric_pixel
:
1446 case nir_intrinsic_load_barycentric_centroid
:
1449 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1451 case nir_intrinsic_load_raw_output_pan
:
1452 case nir_intrinsic_load_output_u8_as_fp16_pan
:
1453 reg
= nir_dest_index(&instr
->dest
);
1454 assert(ctx
->is_blend
);
1456 /* T720 and below use different blend opcodes with slightly
1457 * different semantics than T760 and up */
1459 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1460 bool old_blend
= ctx
->quirks
& MIDGARD_OLD_BLEND
;
1462 if (instr
->intrinsic
== nir_intrinsic_load_output_u8_as_fp16_pan
) {
1463 ld
.load_store
.op
= old_blend
?
1464 midgard_op_ld_color_buffer_u8_as_fp16_old
:
1465 midgard_op_ld_color_buffer_u8_as_fp16
;
1468 ld
.load_store
.address
= 1;
1469 ld
.load_store
.arg_2
= 0x1E;
1472 for (unsigned c
= 2; c
< 16; ++c
)
1473 ld
.swizzle
[0][c
] = 0;
1476 emit_mir_instruction(ctx
, ld
);
1479 case nir_intrinsic_load_blend_const_color_rgba
: {
1480 assert(ctx
->is_blend
);
1481 reg
= nir_dest_index(&instr
->dest
);
1483 /* Blend constants are embedded directly in the shader and
1484 * patched in, so we use some magic routing */
1486 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1487 ins
.has_constants
= true;
1488 ins
.has_blend_constant
= true;
1489 emit_mir_instruction(ctx
, ins
);
1493 case nir_intrinsic_store_zs_output_pan
: {
1494 assert(ctx
->stage
== MESA_SHADER_FRAGMENT
);
1495 emit_fragment_store(ctx
, nir_src_index(ctx
, &instr
->src
[0]),
1498 midgard_instruction
*br
= ctx
->writeout_branch
[MIDGARD_ZS_RT
];
1500 if (!nir_intrinsic_component(instr
))
1501 br
->writeout_depth
= true;
1502 if (nir_intrinsic_component(instr
) ||
1503 instr
->num_components
)
1504 br
->writeout_stencil
= true;
1505 assert(br
->writeout_depth
| br
->writeout_stencil
);
1509 case nir_intrinsic_store_output
:
1510 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1512 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1514 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1516 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1517 const nir_variable
*var
;
1518 enum midgard_rt_id rt
;
1520 var
= search_var(&ctx
->nir
->outputs
,
1521 nir_intrinsic_base(instr
));
1523 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1524 rt
= MIDGARD_COLOR_RT0
;
1525 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1526 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1531 emit_fragment_store(ctx
, reg
, rt
);
1532 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1533 /* We should have been vectorized, though we don't
1534 * currently check that st_vary is emitted only once
1535 * per slot (this is relevant, since there's not a mask
1536 * parameter available on the store [set to 0 by the
1537 * blob]). We do respect the component by adjusting the
1538 * swizzle. If this is a constant source, we'll need to
1539 * emit that explicitly. */
1541 emit_explicit_constant(ctx
, reg
, reg
);
1543 unsigned dst_component
= nir_intrinsic_component(instr
);
1544 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1546 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1547 st
.load_store
.arg_1
= 0x9E;
1548 st
.load_store
.arg_2
= 0x1E;
1550 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1553 st
.load_store
.op
= midgard_op_st_vary_32u
;
1556 st
.load_store
.op
= midgard_op_st_vary_32i
;
1558 case nir_type_float
:
1559 st
.load_store
.op
= midgard_op_st_vary_32
;
1562 unreachable("Attempted to store unknown type");
1566 /* nir_intrinsic_component(store_intr) encodes the
1567 * destination component start. Source component offset
1568 * adjustment is taken care of in
1569 * install_registers_instr(), when offset_swizzle() is
1572 unsigned src_component
= COMPONENT_X
;
1574 assert(nr_comp
> 0);
1575 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1576 st
.swizzle
[0][i
] = src_component
;
1577 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1581 emit_mir_instruction(ctx
, st
);
1583 DBG("Unknown store\n");
1589 /* Special case of store_output for lowered blend shaders */
1590 case nir_intrinsic_store_raw_output_pan
:
1591 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1592 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1594 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1595 /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
1596 * reg = r0.x. We want to splatter. So we can do a 32-bit move
1599 * imov r0.xyzw, r0.xxxx
1602 unsigned expanded
= make_compiler_temp(ctx
);
1604 midgard_instruction splatter
= v_mov(reg
, expanded
);
1606 for (unsigned c
= 0; c
< 16; ++c
)
1607 splatter
.swizzle
[1][c
] = 0;
1609 emit_mir_instruction(ctx
, splatter
);
1610 emit_fragment_store(ctx
, expanded
, ctx
->blend_rt
);
1612 emit_fragment_store(ctx
, reg
, ctx
->blend_rt
);
1616 case nir_intrinsic_store_global
:
1617 case nir_intrinsic_store_shared
:
1618 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1619 emit_explicit_constant(ctx
, reg
, reg
);
1621 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1624 case nir_intrinsic_load_ssbo_address
:
1625 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1628 case nir_intrinsic_get_buffer_size
:
1629 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1632 case nir_intrinsic_load_viewport_scale
:
1633 case nir_intrinsic_load_viewport_offset
:
1634 case nir_intrinsic_load_num_work_groups
:
1635 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1636 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1639 case nir_intrinsic_load_work_group_id
:
1640 case nir_intrinsic_load_local_invocation_id
:
1641 emit_compute_builtin(ctx
, instr
);
1644 case nir_intrinsic_load_vertex_id
:
1645 case nir_intrinsic_load_instance_id
:
1646 emit_vertex_builtin(ctx
, instr
);
1649 case nir_intrinsic_memory_barrier_buffer
:
1650 case nir_intrinsic_memory_barrier_shared
:
1653 case nir_intrinsic_control_barrier
:
1654 schedule_barrier(ctx
);
1655 emit_control_barrier(ctx
);
1656 schedule_barrier(ctx
);
1660 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1667 midgard_tex_format(enum glsl_sampler_dim dim
)
1670 case GLSL_SAMPLER_DIM_1D
:
1671 case GLSL_SAMPLER_DIM_BUF
:
1674 case GLSL_SAMPLER_DIM_2D
:
1675 case GLSL_SAMPLER_DIM_EXTERNAL
:
1676 case GLSL_SAMPLER_DIM_RECT
:
1679 case GLSL_SAMPLER_DIM_3D
:
1682 case GLSL_SAMPLER_DIM_CUBE
:
1683 return MALI_TEX_CUBE
;
1686 DBG("Unknown sampler dim type\n");
1692 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1696 pan_attach_constant_bias(
1697 compiler_context
*ctx
,
1699 midgard_texture_word
*word
)
1701 /* To attach as constant, it has to *be* constant */
1703 if (!nir_src_is_const(lod
))
1706 float f
= nir_src_as_float(lod
);
1708 /* Break into fixed-point */
1710 float lod_frac
= f
- lod_int
;
1712 /* Carry over negative fractions */
1713 if (lod_frac
< 0.0) {
1719 word
->bias
= float_to_ubyte(lod_frac
);
1720 word
->bias_int
= lod_int
;
1726 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1727 unsigned midgard_texop
)
1730 //assert (!instr->sampler);
1732 int texture_index
= instr
->texture_index
;
1733 int sampler_index
= texture_index
;
1735 nir_alu_type dest_base
= nir_alu_type_get_base_type(instr
->dest_type
);
1736 nir_alu_type dest_type
= dest_base
| nir_dest_bit_size(instr
->dest
);
1738 midgard_instruction ins
= {
1739 .type
= TAG_TEXTURE_4
,
1741 .dest
= nir_dest_index(&instr
->dest
),
1742 .src
= { ~0, ~0, ~0, ~0 },
1743 .dest_type
= dest_type
,
1744 .swizzle
= SWIZZLE_IDENTITY_4
,
1746 .op
= midgard_texop
,
1747 .format
= midgard_tex_format(instr
->sampler_dim
),
1748 .texture_handle
= texture_index
,
1749 .sampler_handle
= sampler_index
,
1750 .shadow
= instr
->is_shadow
,
1754 /* We may need a temporary for the coordinate */
1756 bool needs_temp_coord
=
1757 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
1758 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
1761 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
1763 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1764 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1765 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1766 unsigned sz
= nir_src_bit_size(instr
->src
[i
].src
);
1767 nir_alu_type T
= nir_tex_instr_src_type(instr
, i
) | sz
;
1769 switch (instr
->src
[i
].src_type
) {
1770 case nir_tex_src_coord
: {
1771 emit_explicit_constant(ctx
, index
, index
);
1773 unsigned coord_mask
= mask_of(instr
->coord_components
);
1775 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
1778 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
1780 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1781 /* texelFetch is undefined on samplerCube */
1782 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1784 /* For cubemaps, we use a special ld/st op to
1785 * select the face and copy the xy into the
1786 * texture register */
1788 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
1790 ld
.src_types
[1] = T
;
1791 ld
.mask
= 0x3; /* xy */
1792 ld
.load_store
.arg_1
= 0x20;
1793 ld
.swizzle
[1][3] = COMPONENT_X
;
1794 emit_mir_instruction(ctx
, ld
);
1797 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1798 ins
.swizzle
[1][3] = COMPONENT_X
;
1799 } else if (needs_temp_coord
) {
1800 /* mov coord_temp, coords */
1801 midgard_instruction mov
= v_mov(index
, coords
);
1802 mov
.mask
= coord_mask
;
1805 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
1807 emit_mir_instruction(ctx
, mov
);
1812 ins
.src
[1] = coords
;
1813 ins
.src_types
[1] = T
;
1815 /* Texelfetch coordinates uses all four elements
1816 * (xyz/index) regardless of texture dimensionality,
1817 * which means it's necessary to zero the unused
1818 * components to keep everything happy */
1820 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1821 /* mov index.zw, #0, or generalized */
1822 midgard_instruction mov
=
1823 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
1824 mov
.has_constants
= true;
1825 mov
.mask
= coord_mask
^ 0xF;
1826 emit_mir_instruction(ctx
, mov
);
1829 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1830 /* Array component in w but NIR wants it in z,
1831 * but if we have a temp coord we already fixed
1834 if (nr_components
== 3) {
1835 ins
.swizzle
[1][2] = COMPONENT_Z
;
1836 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
1837 } else if (nr_components
== 2) {
1839 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1840 ins
.swizzle
[1][3] = COMPONENT_X
;
1842 unreachable("Invalid texture 2D components");
1845 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1847 ins
.swizzle
[1][2] = COMPONENT_Z
;
1848 ins
.swizzle
[1][3] = COMPONENT_W
;
1854 case nir_tex_src_bias
:
1855 case nir_tex_src_lod
: {
1856 /* Try as a constant if we can */
1858 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1859 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1862 ins
.texture
.lod_register
= true;
1864 ins
.src_types
[2] = T
;
1866 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1867 ins
.swizzle
[2][c
] = COMPONENT_X
;
1869 emit_explicit_constant(ctx
, index
, index
);
1874 case nir_tex_src_offset
: {
1875 ins
.texture
.offset_register
= true;
1877 ins
.src_types
[3] = T
;
1879 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1880 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
1882 emit_explicit_constant(ctx
, index
, index
);
1886 case nir_tex_src_comparator
: {
1887 unsigned comp
= COMPONENT_Z
;
1889 /* mov coord_temp.foo, coords */
1890 midgard_instruction mov
= v_mov(index
, coords
);
1891 mov
.mask
= 1 << comp
;
1893 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
1894 mov
.swizzle
[1][i
] = COMPONENT_X
;
1896 emit_mir_instruction(ctx
, mov
);
1901 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
1907 emit_mir_instruction(ctx
, ins
);
1909 /* Used for .cont and .last hinting */
1910 ctx
->texture_op_count
++;
1914 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1916 switch (instr
->op
) {
1919 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1922 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1925 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1928 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
1931 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
1938 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1940 switch (instr
->type
) {
1941 case nir_jump_break
: {
1942 /* Emit a branch out of the loop */
1943 struct midgard_instruction br
= v_branch(false, false);
1944 br
.branch
.target_type
= TARGET_BREAK
;
1945 br
.branch
.target_break
= ctx
->current_loop_depth
;
1946 emit_mir_instruction(ctx
, br
);
1951 DBG("Unknown jump type %d\n", instr
->type
);
1957 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1959 switch (instr
->type
) {
1960 case nir_instr_type_load_const
:
1961 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1964 case nir_instr_type_intrinsic
:
1965 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1968 case nir_instr_type_alu
:
1969 emit_alu(ctx
, nir_instr_as_alu(instr
));
1972 case nir_instr_type_tex
:
1973 emit_tex(ctx
, nir_instr_as_tex(instr
));
1976 case nir_instr_type_jump
:
1977 emit_jump(ctx
, nir_instr_as_jump(instr
));
1980 case nir_instr_type_ssa_undef
:
1985 DBG("Unhandled instruction type\n");
1991 /* ALU instructions can inline or embed constants, which decreases register
1992 * pressure and saves space. */
1994 #define CONDITIONAL_ATTACH(idx) { \
1995 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
1998 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
1999 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2004 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
2006 mir_foreach_instr_in_block(block
, alu
) {
2007 /* Other instructions cannot inline constants */
2008 if (alu
->type
!= TAG_ALU_4
) continue;
2009 if (alu
->compact_branch
) continue;
2011 /* If there is already a constant here, we can do nothing */
2012 if (alu
->has_constants
) continue;
2014 CONDITIONAL_ATTACH(0);
2016 if (!alu
->has_constants
) {
2017 CONDITIONAL_ATTACH(1)
2018 } else if (!alu
->inline_constant
) {
2019 /* Corner case: _two_ vec4 constants, for instance with a
2020 * csel. For this case, we can only use a constant
2021 * register for one, we'll have to emit a move for the
2024 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2025 unsigned scratch
= make_compiler_temp(ctx
);
2028 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2029 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2031 /* Set the source */
2032 alu
->src
[1] = scratch
;
2034 /* Inject us -before- the last instruction which set r31 */
2035 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2041 /* Being a little silly with the names, but returns the op that is the bitwise
2042 * inverse of the op with the argument switched. I.e. (f and g are
2045 * f(a, b) = ~g(b, a)
2047 * Corollary: if g is the contrapositve of f, f is the contrapositive of g:
2049 * f(a, b) = ~g(b, a)
2050 * ~f(a, b) = g(b, a)
2051 * ~f(a, b) = ~h(a, b) where h is the contrapositive of g
2054 * Thus we define this function in pairs.
2057 static inline midgard_alu_op
2058 mir_contrapositive(midgard_alu_op op
)
2061 case midgard_alu_op_flt
:
2062 return midgard_alu_op_fle
;
2063 case midgard_alu_op_fle
:
2064 return midgard_alu_op_flt
;
2066 case midgard_alu_op_ilt
:
2067 return midgard_alu_op_ile
;
2068 case midgard_alu_op_ile
:
2069 return midgard_alu_op_ilt
;
2072 unreachable("No known contrapositive");
2076 /* Midgard supports two types of constants, embedded constants (128-bit) and
2077 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2078 * constants can be demoted to inline constants, for space savings and
2079 * sometimes a performance boost */
2082 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2084 mir_foreach_instr_in_block(block
, ins
) {
2085 if (!ins
->has_constants
) continue;
2086 if (ins
->has_inline_constant
) continue;
2088 /* Blend constants must not be inlined by definition */
2089 if (ins
->has_blend_constant
) continue;
2091 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2092 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2093 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2095 if (!(is_16
|| is_32
))
2098 /* src1 cannot be an inline constant due to encoding
2099 * restrictions. So, if possible we try to flip the arguments
2102 int op
= ins
->alu
.op
;
2104 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2105 bool flip
= alu_opcode_props
[op
].props
& OP_COMMUTES
;
2108 /* Conditionals can be inverted */
2109 case midgard_alu_op_flt
:
2110 case midgard_alu_op_ilt
:
2111 case midgard_alu_op_fle
:
2112 case midgard_alu_op_ile
:
2113 ins
->alu
.op
= mir_contrapositive(ins
->alu
.op
);
2118 case midgard_alu_op_fcsel
:
2119 case midgard_alu_op_icsel
:
2120 DBG("Missed non-commutative flip (%s)\n", alu_opcode_props
[op
].name
);
2129 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2130 /* Extract the source information */
2132 midgard_vector_alu_src
*src
;
2133 int q
= ins
->alu
.src2
;
2134 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2137 /* Component is from the swizzle. Take a nonzero component */
2139 unsigned first_comp
= ffs(ins
->mask
) - 1;
2140 unsigned component
= ins
->swizzle
[1][first_comp
];
2142 /* Scale constant appropriately, if we can legally */
2143 uint16_t scaled_constant
= 0;
2146 scaled_constant
= ins
->constants
.u16
[component
];
2147 } else if (midgard_is_integer_op(op
)) {
2148 scaled_constant
= ins
->constants
.u32
[component
];
2150 /* Constant overflow after resize */
2151 if (scaled_constant
!= ins
->constants
.u32
[component
])
2154 float original
= ins
->constants
.f32
[component
];
2155 scaled_constant
= _mesa_float_to_half(original
);
2157 /* Check for loss of precision. If this is
2158 * mediump, we don't care, but for a highp
2159 * shader, we need to pay attention. NIR
2160 * doesn't yet tell us which mode we're in!
2161 * Practically this prevents most constants
2162 * from being inlined, sadly. */
2164 float fp32
= _mesa_half_to_float(scaled_constant
);
2166 if (fp32
!= original
)
2170 /* We don't know how to handle these with a constant */
2172 if (mir_nontrivial_source2_mod_simple(ins
) || src
->rep_low
|| src
->rep_high
) {
2173 DBG("Bailing inline constant...\n");
2177 /* Make sure that the constant is not itself a vector
2178 * by checking if all accessed values are the same. */
2180 const midgard_constants
*cons
= &ins
->constants
;
2181 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2183 bool is_vector
= false;
2184 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2186 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2187 /* We only care if this component is actually used */
2188 if (!(mask
& (1 << c
)))
2191 uint32_t test
= is_16
?
2192 cons
->u16
[ins
->swizzle
[1][c
]] :
2193 cons
->u32
[ins
->swizzle
[1][c
]];
2195 if (test
!= value
) {
2204 /* Get rid of the embedded constant */
2205 ins
->has_constants
= false;
2207 ins
->has_inline_constant
= true;
2208 ins
->inline_constant
= scaled_constant
;
2213 /* Dead code elimination for branches at the end of a block - only one branch
2214 * per block is legal semantically */
2217 midgard_opt_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2219 bool branched
= false;
2221 mir_foreach_instr_in_block_safe(block
, ins
) {
2222 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2225 mir_remove_instruction(ins
);
2231 /* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
2232 * the move can be propagated away entirely */
2235 mir_compose_float_outmod(midgard_outmod_float
*outmod
, midgard_outmod_float comp
)
2238 if (comp
== midgard_outmod_none
)
2241 if (*outmod
== midgard_outmod_none
) {
2246 /* TODO: Compose rules */
2251 midgard_opt_pos_propagate(compiler_context
*ctx
, midgard_block
*block
)
2253 bool progress
= false;
2255 mir_foreach_instr_in_block_safe(block
, ins
) {
2256 if (ins
->type
!= TAG_ALU_4
) continue;
2257 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2258 if (ins
->alu
.outmod
!= midgard_outmod_pos
) continue;
2260 /* TODO: Registers? */
2261 unsigned src
= ins
->src
[1];
2262 if (src
& PAN_IS_REG
) continue;
2264 /* There might be a source modifier, too */
2265 if (mir_nontrivial_source2_mod(ins
)) continue;
2267 /* Backpropagate the modifier */
2268 mir_foreach_instr_in_block_from_rev(block
, v
, mir_prev_op(ins
)) {
2269 if (v
->type
!= TAG_ALU_4
) continue;
2270 if (v
->dest
!= src
) continue;
2272 /* Can we even take a float outmod? */
2273 if (midgard_is_integer_out_op(v
->alu
.op
)) continue;
2275 midgard_outmod_float temp
= v
->alu
.outmod
;
2276 progress
|= mir_compose_float_outmod(&temp
, ins
->alu
.outmod
);
2278 /* Throw in the towel.. */
2279 if (!progress
) break;
2281 /* Otherwise, transfer the modifier */
2282 v
->alu
.outmod
= temp
;
2283 ins
->alu
.outmod
= midgard_outmod_none
;
2293 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2295 /* Loop to ourselves */
2296 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2297 struct midgard_instruction ins
= v_branch(false, false);
2298 ins
.writeout
= true;
2299 ins
.writeout_depth
= br
->writeout_depth
;
2300 ins
.writeout_stencil
= br
->writeout_stencil
;
2301 ins
.branch
.target_block
= ctx
->block_count
- 1;
2302 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2303 emit_mir_instruction(ctx
, ins
);
2305 ctx
->current_block
->epilogue
= true;
2306 schedule_barrier(ctx
);
2307 return ins
.branch
.target_block
;
2310 static midgard_block
*
2311 emit_block(compiler_context
*ctx
, nir_block
*block
)
2313 midgard_block
*this_block
= ctx
->after_block
;
2314 ctx
->after_block
= NULL
;
2317 this_block
= create_empty_block(ctx
);
2319 list_addtail(&this_block
->base
.link
, &ctx
->blocks
);
2321 this_block
->scheduled
= false;
2324 /* Set up current block */
2325 list_inithead(&this_block
->base
.instructions
);
2326 ctx
->current_block
= this_block
;
2328 nir_foreach_instr(instr
, block
) {
2329 emit_instr(ctx
, instr
);
2330 ++ctx
->instruction_count
;
2336 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2339 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2341 midgard_block
*before_block
= ctx
->current_block
;
2343 /* Speculatively emit the branch, but we can't fill it in until later */
2344 EMIT(branch
, true, true);
2345 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2346 then_branch
->src
[0] = nir_src_index(ctx
, &nif
->condition
);
2348 /* Emit the two subblocks. */
2349 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2350 midgard_block
*end_then_block
= ctx
->current_block
;
2352 /* Emit a jump from the end of the then block to the end of the else */
2353 EMIT(branch
, false, false);
2354 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2356 /* Emit second block, and check if it's empty */
2358 int else_idx
= ctx
->block_count
;
2359 int count_in
= ctx
->instruction_count
;
2360 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2361 midgard_block
*end_else_block
= ctx
->current_block
;
2362 int after_else_idx
= ctx
->block_count
;
2364 /* Now that we have the subblocks emitted, fix up the branches */
2369 if (ctx
->instruction_count
== count_in
) {
2370 /* The else block is empty, so don't emit an exit jump */
2371 mir_remove_instruction(then_exit
);
2372 then_branch
->branch
.target_block
= after_else_idx
;
2374 then_branch
->branch
.target_block
= else_idx
;
2375 then_exit
->branch
.target_block
= after_else_idx
;
2378 /* Wire up the successors */
2380 ctx
->after_block
= create_empty_block(ctx
);
2382 pan_block_add_successor(&before_block
->base
, &then_block
->base
);
2383 pan_block_add_successor(&before_block
->base
, &else_block
->base
);
2385 pan_block_add_successor(&end_then_block
->base
, &ctx
->after_block
->base
);
2386 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
);
2390 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2392 /* Remember where we are */
2393 midgard_block
*start_block
= ctx
->current_block
;
2395 /* Allocate a loop number, growing the current inner loop depth */
2396 int loop_idx
= ++ctx
->current_loop_depth
;
2398 /* Get index from before the body so we can loop back later */
2399 int start_idx
= ctx
->block_count
;
2401 /* Emit the body itself */
2402 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2404 /* Branch back to loop back */
2405 struct midgard_instruction br_back
= v_branch(false, false);
2406 br_back
.branch
.target_block
= start_idx
;
2407 emit_mir_instruction(ctx
, br_back
);
2409 /* Mark down that branch in the graph. */
2410 pan_block_add_successor(&start_block
->base
, &loop_block
->base
);
2411 pan_block_add_successor(&ctx
->current_block
->base
, &loop_block
->base
);
2413 /* Find the index of the block about to follow us (note: we don't add
2414 * one; blocks are 0-indexed so we get a fencepost problem) */
2415 int break_block_idx
= ctx
->block_count
;
2417 /* Fix up the break statements we emitted to point to the right place,
2418 * now that we can allocate a block number for them */
2419 ctx
->after_block
= create_empty_block(ctx
);
2421 mir_foreach_block_from(ctx
, start_block
, _block
) {
2422 mir_foreach_instr_in_block(((midgard_block
*) _block
), ins
) {
2423 if (ins
->type
!= TAG_ALU_4
) continue;
2424 if (!ins
->compact_branch
) continue;
2426 /* We found a branch -- check the type to see if we need to do anything */
2427 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2429 /* It's a break! Check if it's our break */
2430 if (ins
->branch
.target_break
!= loop_idx
) continue;
2432 /* Okay, cool, we're breaking out of this loop.
2433 * Rewrite from a break to a goto */
2435 ins
->branch
.target_type
= TARGET_GOTO
;
2436 ins
->branch
.target_block
= break_block_idx
;
2438 pan_block_add_successor(_block
, &ctx
->after_block
->base
);
2442 /* Now that we've finished emitting the loop, free up the depth again
2443 * so we play nice with recursion amid nested loops */
2444 --ctx
->current_loop_depth
;
2446 /* Dump loop stats */
2450 static midgard_block
*
2451 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2453 midgard_block
*start_block
= NULL
;
2455 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2456 switch (node
->type
) {
2457 case nir_cf_node_block
: {
2458 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2461 start_block
= block
;
2466 case nir_cf_node_if
:
2467 emit_if(ctx
, nir_cf_node_as_if(node
));
2470 case nir_cf_node_loop
:
2471 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2474 case nir_cf_node_function
:
2483 /* Due to lookahead, we need to report the first tag executed in the command
2484 * stream and in branch targets. An initial block might be empty, so iterate
2485 * until we find one that 'works' */
2488 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2490 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2492 mir_foreach_block_from(ctx
, initial_block
, _v
) {
2493 midgard_block
*v
= (midgard_block
*) _v
;
2494 if (v
->quadword_count
) {
2495 midgard_bundle
*initial_bundle
=
2496 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2498 return initial_bundle
->tag
;
2502 /* Default to a tag 1 which will break from the shader, in case we jump
2503 * to the exit block (i.e. `return` in a compute shader) */
2508 /* For each fragment writeout instruction, generate a writeout loop to
2509 * associate with it */
2512 mir_add_writeout_loops(compiler_context
*ctx
)
2514 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2515 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2518 unsigned popped
= br
->branch
.target_block
;
2519 pan_block_add_successor(&(mir_get_block(ctx
, popped
- 1)->base
), &ctx
->current_block
->base
);
2520 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2522 /* If we have more RTs, we'll need to restore back after our
2523 * loop terminates */
2525 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2526 midgard_instruction uncond
= v_branch(false, false);
2527 uncond
.branch
.target_block
= popped
;
2528 emit_mir_instruction(ctx
, uncond
);
2529 pan_block_add_successor(&ctx
->current_block
->base
, &(mir_get_block(ctx
, popped
)->base
));
2530 schedule_barrier(ctx
);
2532 /* We're last, so we can terminate here */
2533 br
->last_writeout
= true;
2539 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2541 struct util_dynarray
*compiled
= &program
->compiled
;
2543 midgard_debug
= debug_get_option_midgard_debug();
2545 /* TODO: Bound against what? */
2546 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2549 ctx
->stage
= nir
->info
.stage
;
2550 ctx
->is_blend
= is_blend
;
2551 ctx
->alpha_ref
= program
->alpha_ref
;
2552 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2553 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2555 /* Start off with a safe cutoff, allowing usage of all 16 work
2556 * registers. Later, we'll promote uniform reads to uniform registers
2557 * if we determine it is beneficial to do so */
2558 ctx
->uniform_cutoff
= 8;
2560 /* Initialize at a global (not block) level hash tables */
2562 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2563 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2565 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2566 * (so we don't accidentally duplicate the epilogue since mesa/st has
2567 * messed with our I/O quite a bit already) */
2569 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2571 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2572 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2573 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2576 NIR_PASS_V(nir
, nir_lower_var_copies
);
2577 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2578 NIR_PASS_V(nir
, nir_split_var_copies
);
2579 NIR_PASS_V(nir
, nir_lower_var_copies
);
2580 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2581 NIR_PASS_V(nir
, nir_lower_var_copies
);
2582 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2584 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2585 NIR_PASS_V(nir
, nir_lower_ssbo
);
2586 NIR_PASS_V(nir
, midgard_nir_lower_zs_store
);
2588 /* Optimisation passes */
2590 optimise_nir(nir
, ctx
->quirks
);
2592 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2593 nir_print_shader(nir
, stdout
);
2596 /* Assign sysvals and counts, now that we're sure
2597 * (post-optimisation) */
2599 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2600 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2601 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2603 nir_foreach_function(func
, nir
) {
2607 list_inithead(&ctx
->blocks
);
2608 ctx
->block_count
= 0;
2611 emit_cf_list(ctx
, &func
->impl
->body
);
2612 break; /* TODO: Multi-function shaders */
2615 util_dynarray_init(compiled
, NULL
);
2617 /* Per-block lowering before opts */
2619 mir_foreach_block(ctx
, _block
) {
2620 midgard_block
*block
= (midgard_block
*) _block
;
2621 inline_alu_constants(ctx
, block
);
2622 midgard_opt_promote_fmov(ctx
, block
);
2623 embedded_to_inline_constant(ctx
, block
);
2625 /* MIR-level optimizations */
2627 bool progress
= false;
2632 mir_foreach_block(ctx
, _block
) {
2633 midgard_block
*block
= (midgard_block
*) _block
;
2634 progress
|= midgard_opt_pos_propagate(ctx
, block
);
2635 progress
|= midgard_opt_copy_prop(ctx
, block
);
2636 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2637 progress
|= midgard_opt_combine_projection(ctx
, block
);
2638 progress
|= midgard_opt_varying_projection(ctx
, block
);
2639 progress
|= midgard_opt_not_propagate(ctx
, block
);
2640 progress
|= midgard_opt_fuse_src_invert(ctx
, block
);
2641 progress
|= midgard_opt_fuse_dest_invert(ctx
, block
);
2642 progress
|= midgard_opt_csel_invert(ctx
, block
);
2643 progress
|= midgard_opt_drop_cmp_invert(ctx
, block
);
2644 progress
|= midgard_opt_invert_branch(ctx
, block
);
2648 mir_foreach_block(ctx
, _block
) {
2649 midgard_block
*block
= (midgard_block
*) _block
;
2650 midgard_lower_invert(ctx
, block
);
2651 midgard_lower_derivatives(ctx
, block
);
2654 /* Nested control-flow can result in dead branches at the end of the
2655 * block. This messes with our analysis and is just dead code, so cull
2657 mir_foreach_block(ctx
, _block
) {
2658 midgard_block
*block
= (midgard_block
*) _block
;
2659 midgard_opt_cull_dead_branch(ctx
, block
);
2662 /* Ensure we were lowered */
2663 mir_foreach_instr_global(ctx
, ins
) {
2664 assert(!ins
->invert
);
2667 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2668 mir_add_writeout_loops(ctx
);
2671 midgard_schedule_program(ctx
);
2674 /* Now that all the bundles are scheduled and we can calculate block
2675 * sizes, emit actual branch instructions rather than placeholders */
2677 int br_block_idx
= 0;
2679 mir_foreach_block(ctx
, _block
) {
2680 midgard_block
*block
= (midgard_block
*) _block
;
2681 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2682 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2683 midgard_instruction
*ins
= bundle
->instructions
[c
];
2685 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2687 /* Parse some basic branch info */
2688 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2689 bool is_conditional
= ins
->branch
.conditional
;
2690 bool is_inverted
= ins
->branch
.invert_conditional
;
2691 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2692 bool is_writeout
= ins
->writeout
;
2694 /* Determine the block we're jumping to */
2695 int target_number
= ins
->branch
.target_block
;
2697 /* Report the destination tag */
2698 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2700 /* Count up the number of quadwords we're
2701 * jumping over = number of quadwords until
2702 * (br_block_idx, target_number) */
2704 int quadword_offset
= 0;
2708 } else if (target_number
> br_block_idx
) {
2711 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2712 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2715 quadword_offset
+= blk
->quadword_count
;
2718 /* Jump backwards */
2720 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2721 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2724 quadword_offset
-= blk
->quadword_count
;
2728 /* Unconditional extended branches (far jumps)
2729 * have issues, so we always use a conditional
2730 * branch, setting the condition to always for
2731 * unconditional. For compact unconditional
2732 * branches, cond isn't used so it doesn't
2733 * matter what we pick. */
2735 midgard_condition cond
=
2736 !is_conditional
? midgard_condition_always
:
2737 is_inverted
? midgard_condition_false
:
2738 midgard_condition_true
;
2740 midgard_jmp_writeout_op op
=
2741 is_discard
? midgard_jmp_writeout_op_discard
:
2742 is_writeout
? midgard_jmp_writeout_op_writeout
:
2743 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2744 midgard_jmp_writeout_op_branch_cond
;
2747 midgard_branch_extended branch
=
2748 midgard_create_branch_extended(
2753 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2754 } else if (is_conditional
|| is_discard
) {
2755 midgard_branch_cond branch
= {
2757 .dest_tag
= dest_tag
,
2758 .offset
= quadword_offset
,
2762 assert(branch
.offset
== quadword_offset
);
2764 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2766 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2768 midgard_branch_uncond branch
= {
2770 .dest_tag
= dest_tag
,
2771 .offset
= quadword_offset
,
2775 assert(branch
.offset
== quadword_offset
);
2777 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2785 /* Emit flat binary from the instruction arrays. Iterate each block in
2786 * sequence. Save instruction boundaries such that lookahead tags can
2787 * be assigned easily */
2789 /* Cache _all_ bundles in source order for lookahead across failed branches */
2791 int bundle_count
= 0;
2792 mir_foreach_block(ctx
, _block
) {
2793 midgard_block
*block
= (midgard_block
*) _block
;
2794 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2796 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2798 mir_foreach_block(ctx
, _block
) {
2799 midgard_block
*block
= (midgard_block
*) _block
;
2800 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2801 source_order_bundles
[bundle_idx
++] = bundle
;
2805 int current_bundle
= 0;
2807 /* Midgard prefetches instruction types, so during emission we
2808 * need to lookahead. Unless this is the last instruction, in
2809 * which we return 1. */
2811 mir_foreach_block(ctx
, _block
) {
2812 midgard_block
*block
= (midgard_block
*) _block
;
2813 mir_foreach_bundle_in_block(block
, bundle
) {
2816 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2817 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2819 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2823 /* TODO: Free deeper */
2824 //util_dynarray_fini(&block->instructions);
2827 free(source_order_bundles
);
2829 /* Report the very first tag executed */
2830 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2832 /* Deal with off-by-one related to the fencepost problem */
2833 program
->work_register_count
= ctx
->work_registers
+ 1;
2834 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2836 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2837 program
->tls_size
= ctx
->tls_size
;
2839 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2840 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
2842 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
2843 unsigned nr_bundles
= 0, nr_ins
= 0;
2845 /* Count instructions and bundles */
2847 mir_foreach_block(ctx
, _block
) {
2848 midgard_block
*block
= (midgard_block
*) _block
;
2849 nr_bundles
+= util_dynarray_num_elements(
2850 &block
->bundles
, midgard_bundle
);
2852 mir_foreach_bundle_in_block(block
, bun
)
2853 nr_ins
+= bun
->instruction_count
;
2856 /* Calculate thread count. There are certain cutoffs by
2857 * register count for thread count */
2859 unsigned nr_registers
= program
->work_register_count
;
2861 unsigned nr_threads
=
2862 (nr_registers
<= 4) ? 4 :
2863 (nr_registers
<= 8) ? 2 :
2868 fprintf(stderr
, "shader%d - %s shader: "
2869 "%u inst, %u bundles, %u quadwords, "
2870 "%u registers, %u threads, %u loops, "
2871 "%u:%u spills:fills\n",
2873 gl_shader_stage_name(ctx
->stage
),
2874 nr_ins
, nr_bundles
, ctx
->quadword_count
,
2875 nr_registers
, nr_threads
,
2877 ctx
->spills
, ctx
->fills
);