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
.src_types
[1] = nir_type_float32
;
1007 ins
.has_constants
= true;
1009 if (instr
->op
== nir_op_b2f32
)
1010 ins
.constants
.f32
[0] = 1.0f
;
1012 ins
.constants
.i32
[0] = 1;
1014 for (unsigned c
= 0; c
< 16; ++c
)
1015 ins
.swizzle
[1][c
] = 0;
1016 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1017 /* Lots of instructions need a 0 plonked in */
1018 ins
.has_inline_constant
= false;
1019 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1020 ins
.src_types
[1] = nir_type_uint32
;
1021 ins
.has_constants
= true;
1022 ins
.constants
.u32
[0] = 0;
1024 for (unsigned c
= 0; c
< 16; ++c
)
1025 ins
.swizzle
[1][c
] = 0;
1026 } else if (instr
->op
== nir_op_inot
) {
1030 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1031 /* To avoid duplicating the lookup tables (probably), true LUT
1032 * instructions can only operate as if they were scalars. Lower
1033 * them here by changing the component. */
1035 unsigned orig_mask
= ins
.mask
;
1037 for (int i
= 0; i
< nr_components
; ++i
) {
1038 /* Mask the associated component, dropping the
1039 * instruction if needed */
1042 ins
.mask
&= orig_mask
;
1047 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1048 ins
.swizzle
[0][j
] = nirmods
[0]->swizzle
[i
]; /* Pull from the correct component */
1050 emit_mir_instruction(ctx
, ins
);
1053 emit_mir_instruction(ctx
, ins
);
1060 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1062 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1063 unsigned nir_mask
= 0;
1067 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1068 dsize
= nir_dest_bit_size(intr
->dest
);
1070 nir_mask
= nir_intrinsic_write_mask(intr
);
1074 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1075 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1076 mir_set_bytemask(ins
, bytemask
);
1079 ins
->load_64
= true;
1082 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1083 * optimized) versions of UBO #0 */
1085 static midgard_instruction
*
1087 compiler_context
*ctx
,
1091 nir_src
*indirect_offset
,
1092 unsigned indirect_shift
,
1095 /* TODO: half-floats */
1097 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1098 ins
.constants
.u32
[0] = offset
;
1100 if (instr
->type
== nir_instr_type_intrinsic
)
1101 mir_set_intr_mask(instr
, &ins
, true);
1103 if (indirect_offset
) {
1104 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1105 ins
.src_types
[2] = nir_type_uint32
;
1106 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1108 ins
.load_store
.arg_2
= 0x1E;
1111 ins
.load_store
.arg_1
= index
;
1113 return emit_mir_instruction(ctx
, ins
);
1116 /* Globals are like UBOs if you squint. And shared memory is like globals if
1117 * you squint even harder */
1121 compiler_context
*ctx
,
1130 midgard_instruction ins
;
1133 ins
= m_ld_int4(srcdest
, 0);
1135 ins
= m_st_int4(srcdest
, 0);
1137 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1138 mir_set_intr_mask(instr
, &ins
, is_read
);
1140 emit_mir_instruction(ctx
, ins
);
1145 compiler_context
*ctx
,
1146 unsigned dest
, unsigned offset
,
1147 unsigned nr_comp
, unsigned component
,
1148 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1150 /* XXX: Half-floats? */
1151 /* TODO: swizzle, mask */
1153 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1154 ins
.mask
= mask_of(nr_comp
);
1156 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1157 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1159 midgard_varying_parameter p
= {
1161 .interpolation
= midgard_interp_default
,
1166 memcpy(&u
, &p
, sizeof(p
));
1167 ins
.load_store
.varying_parameters
= u
;
1169 if (indirect_offset
) {
1170 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1171 ins
.src_types
[2] = nir_type_uint32
;
1173 ins
.load_store
.arg_2
= 0x1E;
1175 ins
.load_store
.arg_1
= 0x9E;
1177 /* Use the type appropriate load */
1181 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1184 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1186 case nir_type_float
:
1187 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1190 unreachable("Attempted to load unknown type");
1194 emit_mir_instruction(ctx
, ins
);
1199 compiler_context
*ctx
,
1200 unsigned dest
, unsigned offset
,
1201 unsigned nr_comp
, nir_alu_type t
)
1203 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1204 ins
.load_store
.arg_1
= 0x1E;
1205 ins
.load_store
.arg_2
= 0x1E;
1206 ins
.mask
= mask_of(nr_comp
);
1208 /* Use the type appropriate load */
1212 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1215 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1217 case nir_type_float
:
1218 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1221 unreachable("Attempted to load unknown type");
1225 emit_mir_instruction(ctx
, ins
);
1229 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1230 unsigned nr_components
, unsigned offset
)
1234 /* Figure out which uniform this is */
1235 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1236 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1238 unsigned dest
= nir_dest_index(&nir_dest
);
1240 /* Sysvals are prefix uniforms */
1241 unsigned uniform
= ((uintptr_t) val
) - 1;
1243 /* Emit the read itself -- this is never indirect */
1244 midgard_instruction
*ins
=
1245 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1247 ins
->mask
= mask_of(nr_components
);
1251 compute_builtin_arg(nir_op op
)
1254 case nir_intrinsic_load_work_group_id
:
1256 case nir_intrinsic_load_local_invocation_id
:
1259 unreachable("Invalid compute paramater loaded");
1264 emit_fragment_store(compiler_context
*ctx
, unsigned src
, enum midgard_rt_id rt
)
1266 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1268 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1272 emit_explicit_constant(ctx
, src
, src
);
1274 struct midgard_instruction ins
=
1275 v_branch(false, false);
1277 ins
.writeout
= true;
1279 /* Add dependencies */
1281 ins
.src_types
[0] = nir_type_uint32
;
1282 ins
.constants
.u32
[0] = rt
== MIDGARD_ZS_RT
?
1283 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1285 /* Emit the branch */
1286 br
= emit_mir_instruction(ctx
, ins
);
1287 schedule_barrier(ctx
);
1288 ctx
->writeout_branch
[rt
] = br
;
1290 /* Push our current location = current block count - 1 = where we'll
1291 * jump to. Maybe a bit too clever for my own good */
1293 br
->branch
.target_block
= ctx
->block_count
- 1;
1297 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1299 unsigned reg
= nir_dest_index(&instr
->dest
);
1300 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1301 ins
.mask
= mask_of(3);
1302 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1303 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1304 emit_mir_instruction(ctx
, ins
);
1308 vertex_builtin_arg(nir_op op
)
1311 case nir_intrinsic_load_vertex_id
:
1312 return PAN_VERTEX_ID
;
1313 case nir_intrinsic_load_instance_id
:
1314 return PAN_INSTANCE_ID
;
1316 unreachable("Invalid vertex builtin");
1321 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1323 unsigned reg
= nir_dest_index(&instr
->dest
);
1324 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1328 emit_control_barrier(compiler_context
*ctx
)
1330 midgard_instruction ins
= {
1331 .type
= TAG_TEXTURE_4
,
1332 .src
= { ~0, ~0, ~0, ~0 },
1334 .op
= TEXTURE_OP_BARRIER
,
1336 /* TODO: optimize */
1337 .barrier_buffer
= 1,
1342 emit_mir_instruction(ctx
, ins
);
1345 static const nir_variable
*
1346 search_var(struct exec_list
*vars
, unsigned driver_loc
)
1348 nir_foreach_variable(var
, vars
) {
1349 if (var
->data
.driver_location
== driver_loc
)
1357 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1359 unsigned offset
= 0, reg
;
1361 switch (instr
->intrinsic
) {
1362 case nir_intrinsic_discard_if
:
1363 case nir_intrinsic_discard
: {
1364 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1365 struct midgard_instruction discard
= v_branch(conditional
, false);
1366 discard
.branch
.target_type
= TARGET_DISCARD
;
1369 discard
.src
[0] = nir_src_index(ctx
, &instr
->src
[0]);
1370 discard
.src_types
[0] = nir_type_uint32
;
1373 emit_mir_instruction(ctx
, discard
);
1374 schedule_barrier(ctx
);
1379 case nir_intrinsic_load_uniform
:
1380 case nir_intrinsic_load_ubo
:
1381 case nir_intrinsic_load_global
:
1382 case nir_intrinsic_load_shared
:
1383 case nir_intrinsic_load_input
:
1384 case nir_intrinsic_load_interpolated_input
: {
1385 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1386 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1387 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1388 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1389 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1390 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1392 /* Get the base type of the intrinsic */
1393 /* TODO: Infer type? Does it matter? */
1395 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1396 (is_interp
) ? nir_type_float
:
1397 nir_intrinsic_type(instr
);
1399 t
= nir_alu_type_get_base_type(t
);
1401 if (!(is_ubo
|| is_global
)) {
1402 offset
= nir_intrinsic_base(instr
);
1405 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1407 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1409 bool direct
= nir_src_is_const(*src_offset
);
1410 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1413 offset
+= nir_src_as_uint(*src_offset
);
1415 /* We may need to apply a fractional offset */
1416 int component
= (is_flat
|| is_interp
) ?
1417 nir_intrinsic_component(instr
) : 0;
1418 reg
= nir_dest_index(&instr
->dest
);
1420 if (is_uniform
&& !ctx
->is_blend
) {
1421 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1422 } else if (is_ubo
) {
1423 nir_src index
= instr
->src
[0];
1425 /* TODO: Is indirect block number possible? */
1426 assert(nir_src_is_const(index
));
1428 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1429 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1430 } else if (is_global
|| is_shared
) {
1431 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1432 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1433 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
, is_flat
);
1434 } else if (ctx
->is_blend
) {
1435 /* For blend shaders, load the input color, which is
1436 * preloaded to r0 */
1438 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), reg
);
1439 emit_mir_instruction(ctx
, move
);
1440 schedule_barrier(ctx
);
1441 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1442 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1444 DBG("Unknown load\n");
1451 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1452 case nir_intrinsic_load_barycentric_pixel
:
1453 case nir_intrinsic_load_barycentric_centroid
:
1456 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1458 case nir_intrinsic_load_raw_output_pan
:
1459 case nir_intrinsic_load_output_u8_as_fp16_pan
:
1460 reg
= nir_dest_index(&instr
->dest
);
1461 assert(ctx
->is_blend
);
1463 /* T720 and below use different blend opcodes with slightly
1464 * different semantics than T760 and up */
1466 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1467 bool old_blend
= ctx
->quirks
& MIDGARD_OLD_BLEND
;
1469 if (instr
->intrinsic
== nir_intrinsic_load_output_u8_as_fp16_pan
) {
1470 ld
.load_store
.op
= old_blend
?
1471 midgard_op_ld_color_buffer_u8_as_fp16_old
:
1472 midgard_op_ld_color_buffer_u8_as_fp16
;
1475 ld
.load_store
.address
= 1;
1476 ld
.load_store
.arg_2
= 0x1E;
1479 for (unsigned c
= 2; c
< 16; ++c
)
1480 ld
.swizzle
[0][c
] = 0;
1483 emit_mir_instruction(ctx
, ld
);
1486 case nir_intrinsic_load_blend_const_color_rgba
: {
1487 assert(ctx
->is_blend
);
1488 reg
= nir_dest_index(&instr
->dest
);
1490 /* Blend constants are embedded directly in the shader and
1491 * patched in, so we use some magic routing */
1493 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1494 ins
.has_constants
= true;
1495 ins
.has_blend_constant
= true;
1496 emit_mir_instruction(ctx
, ins
);
1500 case nir_intrinsic_store_zs_output_pan
: {
1501 assert(ctx
->stage
== MESA_SHADER_FRAGMENT
);
1502 emit_fragment_store(ctx
, nir_src_index(ctx
, &instr
->src
[0]),
1505 midgard_instruction
*br
= ctx
->writeout_branch
[MIDGARD_ZS_RT
];
1507 if (!nir_intrinsic_component(instr
))
1508 br
->writeout_depth
= true;
1509 if (nir_intrinsic_component(instr
) ||
1510 instr
->num_components
)
1511 br
->writeout_stencil
= true;
1512 assert(br
->writeout_depth
| br
->writeout_stencil
);
1516 case nir_intrinsic_store_output
:
1517 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1519 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1521 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1523 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1524 const nir_variable
*var
;
1525 enum midgard_rt_id rt
;
1527 var
= search_var(&ctx
->nir
->outputs
,
1528 nir_intrinsic_base(instr
));
1530 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1531 rt
= MIDGARD_COLOR_RT0
;
1532 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1533 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1538 emit_fragment_store(ctx
, reg
, rt
);
1539 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1540 /* We should have been vectorized, though we don't
1541 * currently check that st_vary is emitted only once
1542 * per slot (this is relevant, since there's not a mask
1543 * parameter available on the store [set to 0 by the
1544 * blob]). We do respect the component by adjusting the
1545 * swizzle. If this is a constant source, we'll need to
1546 * emit that explicitly. */
1548 emit_explicit_constant(ctx
, reg
, reg
);
1550 unsigned dst_component
= nir_intrinsic_component(instr
);
1551 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1553 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1554 st
.load_store
.arg_1
= 0x9E;
1555 st
.load_store
.arg_2
= 0x1E;
1557 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1560 st
.load_store
.op
= midgard_op_st_vary_32u
;
1563 st
.load_store
.op
= midgard_op_st_vary_32i
;
1565 case nir_type_float
:
1566 st
.load_store
.op
= midgard_op_st_vary_32
;
1569 unreachable("Attempted to store unknown type");
1573 /* nir_intrinsic_component(store_intr) encodes the
1574 * destination component start. Source component offset
1575 * adjustment is taken care of in
1576 * install_registers_instr(), when offset_swizzle() is
1579 unsigned src_component
= COMPONENT_X
;
1581 assert(nr_comp
> 0);
1582 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1583 st
.swizzle
[0][i
] = src_component
;
1584 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1588 emit_mir_instruction(ctx
, st
);
1590 DBG("Unknown store\n");
1596 /* Special case of store_output for lowered blend shaders */
1597 case nir_intrinsic_store_raw_output_pan
:
1598 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1599 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1601 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1602 /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
1603 * reg = r0.x. We want to splatter. So we can do a 32-bit move
1606 * imov r0.xyzw, r0.xxxx
1609 unsigned expanded
= make_compiler_temp(ctx
);
1611 midgard_instruction splatter
= v_mov(reg
, expanded
);
1613 for (unsigned c
= 0; c
< 16; ++c
)
1614 splatter
.swizzle
[1][c
] = 0;
1616 emit_mir_instruction(ctx
, splatter
);
1617 emit_fragment_store(ctx
, expanded
, ctx
->blend_rt
);
1619 emit_fragment_store(ctx
, reg
, ctx
->blend_rt
);
1623 case nir_intrinsic_store_global
:
1624 case nir_intrinsic_store_shared
:
1625 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1626 emit_explicit_constant(ctx
, reg
, reg
);
1628 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1631 case nir_intrinsic_load_ssbo_address
:
1632 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1635 case nir_intrinsic_get_buffer_size
:
1636 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1639 case nir_intrinsic_load_viewport_scale
:
1640 case nir_intrinsic_load_viewport_offset
:
1641 case nir_intrinsic_load_num_work_groups
:
1642 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1643 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1646 case nir_intrinsic_load_work_group_id
:
1647 case nir_intrinsic_load_local_invocation_id
:
1648 emit_compute_builtin(ctx
, instr
);
1651 case nir_intrinsic_load_vertex_id
:
1652 case nir_intrinsic_load_instance_id
:
1653 emit_vertex_builtin(ctx
, instr
);
1656 case nir_intrinsic_memory_barrier_buffer
:
1657 case nir_intrinsic_memory_barrier_shared
:
1660 case nir_intrinsic_control_barrier
:
1661 schedule_barrier(ctx
);
1662 emit_control_barrier(ctx
);
1663 schedule_barrier(ctx
);
1667 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1674 midgard_tex_format(enum glsl_sampler_dim dim
)
1677 case GLSL_SAMPLER_DIM_1D
:
1678 case GLSL_SAMPLER_DIM_BUF
:
1681 case GLSL_SAMPLER_DIM_2D
:
1682 case GLSL_SAMPLER_DIM_EXTERNAL
:
1683 case GLSL_SAMPLER_DIM_RECT
:
1686 case GLSL_SAMPLER_DIM_3D
:
1689 case GLSL_SAMPLER_DIM_CUBE
:
1690 return MALI_TEX_CUBE
;
1693 DBG("Unknown sampler dim type\n");
1699 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1703 pan_attach_constant_bias(
1704 compiler_context
*ctx
,
1706 midgard_texture_word
*word
)
1708 /* To attach as constant, it has to *be* constant */
1710 if (!nir_src_is_const(lod
))
1713 float f
= nir_src_as_float(lod
);
1715 /* Break into fixed-point */
1717 float lod_frac
= f
- lod_int
;
1719 /* Carry over negative fractions */
1720 if (lod_frac
< 0.0) {
1726 word
->bias
= float_to_ubyte(lod_frac
);
1727 word
->bias_int
= lod_int
;
1733 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1734 unsigned midgard_texop
)
1737 //assert (!instr->sampler);
1739 int texture_index
= instr
->texture_index
;
1740 int sampler_index
= texture_index
;
1742 nir_alu_type dest_base
= nir_alu_type_get_base_type(instr
->dest_type
);
1743 nir_alu_type dest_type
= dest_base
| nir_dest_bit_size(instr
->dest
);
1745 midgard_instruction ins
= {
1746 .type
= TAG_TEXTURE_4
,
1748 .dest
= nir_dest_index(&instr
->dest
),
1749 .src
= { ~0, ~0, ~0, ~0 },
1750 .dest_type
= dest_type
,
1751 .swizzle
= SWIZZLE_IDENTITY_4
,
1753 .op
= midgard_texop
,
1754 .format
= midgard_tex_format(instr
->sampler_dim
),
1755 .texture_handle
= texture_index
,
1756 .sampler_handle
= sampler_index
,
1757 .shadow
= instr
->is_shadow
,
1761 /* We may need a temporary for the coordinate */
1763 bool needs_temp_coord
=
1764 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
1765 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
1768 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
1770 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1771 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1772 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1773 unsigned sz
= nir_src_bit_size(instr
->src
[i
].src
);
1774 nir_alu_type T
= nir_tex_instr_src_type(instr
, i
) | sz
;
1776 switch (instr
->src
[i
].src_type
) {
1777 case nir_tex_src_coord
: {
1778 emit_explicit_constant(ctx
, index
, index
);
1780 unsigned coord_mask
= mask_of(instr
->coord_components
);
1782 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
1785 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
1787 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1788 /* texelFetch is undefined on samplerCube */
1789 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1791 /* For cubemaps, we use a special ld/st op to
1792 * select the face and copy the xy into the
1793 * texture register */
1795 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
1797 ld
.src_types
[1] = T
;
1798 ld
.mask
= 0x3; /* xy */
1799 ld
.load_store
.arg_1
= 0x20;
1800 ld
.swizzle
[1][3] = COMPONENT_X
;
1801 emit_mir_instruction(ctx
, ld
);
1804 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1805 ins
.swizzle
[1][3] = COMPONENT_X
;
1806 } else if (needs_temp_coord
) {
1807 /* mov coord_temp, coords */
1808 midgard_instruction mov
= v_mov(index
, coords
);
1809 mov
.mask
= coord_mask
;
1812 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
1814 emit_mir_instruction(ctx
, mov
);
1819 ins
.src
[1] = coords
;
1820 ins
.src_types
[1] = T
;
1822 /* Texelfetch coordinates uses all four elements
1823 * (xyz/index) regardless of texture dimensionality,
1824 * which means it's necessary to zero the unused
1825 * components to keep everything happy */
1827 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1828 /* mov index.zw, #0, or generalized */
1829 midgard_instruction mov
=
1830 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
1831 mov
.has_constants
= true;
1832 mov
.mask
= coord_mask
^ 0xF;
1833 emit_mir_instruction(ctx
, mov
);
1836 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1837 /* Array component in w but NIR wants it in z,
1838 * but if we have a temp coord we already fixed
1841 if (nr_components
== 3) {
1842 ins
.swizzle
[1][2] = COMPONENT_Z
;
1843 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
1844 } else if (nr_components
== 2) {
1846 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1847 ins
.swizzle
[1][3] = COMPONENT_X
;
1849 unreachable("Invalid texture 2D components");
1852 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1854 ins
.swizzle
[1][2] = COMPONENT_Z
;
1855 ins
.swizzle
[1][3] = COMPONENT_W
;
1861 case nir_tex_src_bias
:
1862 case nir_tex_src_lod
: {
1863 /* Try as a constant if we can */
1865 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1866 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1869 ins
.texture
.lod_register
= true;
1871 ins
.src_types
[2] = T
;
1873 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1874 ins
.swizzle
[2][c
] = COMPONENT_X
;
1876 emit_explicit_constant(ctx
, index
, index
);
1881 case nir_tex_src_offset
: {
1882 ins
.texture
.offset_register
= true;
1884 ins
.src_types
[3] = T
;
1886 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1887 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
1889 emit_explicit_constant(ctx
, index
, index
);
1893 case nir_tex_src_comparator
: {
1894 unsigned comp
= COMPONENT_Z
;
1896 /* mov coord_temp.foo, coords */
1897 midgard_instruction mov
= v_mov(index
, coords
);
1898 mov
.mask
= 1 << comp
;
1900 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
1901 mov
.swizzle
[1][i
] = COMPONENT_X
;
1903 emit_mir_instruction(ctx
, mov
);
1908 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
1914 emit_mir_instruction(ctx
, ins
);
1916 /* Used for .cont and .last hinting */
1917 ctx
->texture_op_count
++;
1921 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1923 switch (instr
->op
) {
1926 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1929 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1932 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1935 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
1938 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
1945 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1947 switch (instr
->type
) {
1948 case nir_jump_break
: {
1949 /* Emit a branch out of the loop */
1950 struct midgard_instruction br
= v_branch(false, false);
1951 br
.branch
.target_type
= TARGET_BREAK
;
1952 br
.branch
.target_break
= ctx
->current_loop_depth
;
1953 emit_mir_instruction(ctx
, br
);
1958 DBG("Unknown jump type %d\n", instr
->type
);
1964 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1966 switch (instr
->type
) {
1967 case nir_instr_type_load_const
:
1968 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1971 case nir_instr_type_intrinsic
:
1972 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1975 case nir_instr_type_alu
:
1976 emit_alu(ctx
, nir_instr_as_alu(instr
));
1979 case nir_instr_type_tex
:
1980 emit_tex(ctx
, nir_instr_as_tex(instr
));
1983 case nir_instr_type_jump
:
1984 emit_jump(ctx
, nir_instr_as_jump(instr
));
1987 case nir_instr_type_ssa_undef
:
1992 DBG("Unhandled instruction type\n");
1998 /* ALU instructions can inline or embed constants, which decreases register
1999 * pressure and saves space. */
2001 #define CONDITIONAL_ATTACH(idx) { \
2002 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
2005 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
2006 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2011 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
2013 mir_foreach_instr_in_block(block
, alu
) {
2014 /* Other instructions cannot inline constants */
2015 if (alu
->type
!= TAG_ALU_4
) continue;
2016 if (alu
->compact_branch
) continue;
2018 /* If there is already a constant here, we can do nothing */
2019 if (alu
->has_constants
) continue;
2021 CONDITIONAL_ATTACH(0);
2023 if (!alu
->has_constants
) {
2024 CONDITIONAL_ATTACH(1)
2025 } else if (!alu
->inline_constant
) {
2026 /* Corner case: _two_ vec4 constants, for instance with a
2027 * csel. For this case, we can only use a constant
2028 * register for one, we'll have to emit a move for the
2031 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2032 unsigned scratch
= make_compiler_temp(ctx
);
2035 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2036 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2038 /* Set the source */
2039 alu
->src
[1] = scratch
;
2041 /* Inject us -before- the last instruction which set r31 */
2042 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2048 /* Midgard supports two types of constants, embedded constants (128-bit) and
2049 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2050 * constants can be demoted to inline constants, for space savings and
2051 * sometimes a performance boost */
2054 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2056 mir_foreach_instr_in_block(block
, ins
) {
2057 if (!ins
->has_constants
) continue;
2058 if (ins
->has_inline_constant
) continue;
2060 /* Blend constants must not be inlined by definition */
2061 if (ins
->has_blend_constant
) continue;
2063 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2064 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2065 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2067 if (!(is_16
|| is_32
))
2070 /* src1 cannot be an inline constant due to encoding
2071 * restrictions. So, if possible we try to flip the arguments
2074 int op
= ins
->alu
.op
;
2076 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
) &&
2077 alu_opcode_props
[op
].props
& OP_COMMUTES
) {
2081 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2082 /* Extract the source information */
2084 midgard_vector_alu_src
*src
;
2085 int q
= ins
->alu
.src2
;
2086 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2089 /* Component is from the swizzle. Take a nonzero component */
2091 unsigned first_comp
= ffs(ins
->mask
) - 1;
2092 unsigned component
= ins
->swizzle
[1][first_comp
];
2094 /* Scale constant appropriately, if we can legally */
2095 uint16_t scaled_constant
= 0;
2098 scaled_constant
= ins
->constants
.u16
[component
];
2099 } else if (midgard_is_integer_op(op
)) {
2100 scaled_constant
= ins
->constants
.u32
[component
];
2102 /* Constant overflow after resize */
2103 if (scaled_constant
!= ins
->constants
.u32
[component
])
2106 float original
= ins
->constants
.f32
[component
];
2107 scaled_constant
= _mesa_float_to_half(original
);
2109 /* Check for loss of precision. If this is
2110 * mediump, we don't care, but for a highp
2111 * shader, we need to pay attention. NIR
2112 * doesn't yet tell us which mode we're in!
2113 * Practically this prevents most constants
2114 * from being inlined, sadly. */
2116 float fp32
= _mesa_half_to_float(scaled_constant
);
2118 if (fp32
!= original
)
2122 /* We don't know how to handle these with a constant */
2124 if (mir_nontrivial_source2_mod_simple(ins
) || src
->rep_low
|| src
->rep_high
) {
2125 DBG("Bailing inline constant...\n");
2129 /* Make sure that the constant is not itself a vector
2130 * by checking if all accessed values are the same. */
2132 const midgard_constants
*cons
= &ins
->constants
;
2133 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2135 bool is_vector
= false;
2136 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2138 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2139 /* We only care if this component is actually used */
2140 if (!(mask
& (1 << c
)))
2143 uint32_t test
= is_16
?
2144 cons
->u16
[ins
->swizzle
[1][c
]] :
2145 cons
->u32
[ins
->swizzle
[1][c
]];
2147 if (test
!= value
) {
2156 /* Get rid of the embedded constant */
2157 ins
->has_constants
= false;
2159 ins
->has_inline_constant
= true;
2160 ins
->inline_constant
= scaled_constant
;
2165 /* Dead code elimination for branches at the end of a block - only one branch
2166 * per block is legal semantically */
2169 midgard_opt_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2171 bool branched
= false;
2173 mir_foreach_instr_in_block_safe(block
, ins
) {
2174 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2177 mir_remove_instruction(ins
);
2183 /* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
2184 * the move can be propagated away entirely */
2187 mir_compose_float_outmod(midgard_outmod_float
*outmod
, midgard_outmod_float comp
)
2190 if (comp
== midgard_outmod_none
)
2193 if (*outmod
== midgard_outmod_none
) {
2198 /* TODO: Compose rules */
2203 midgard_opt_pos_propagate(compiler_context
*ctx
, midgard_block
*block
)
2205 bool progress
= false;
2207 mir_foreach_instr_in_block_safe(block
, ins
) {
2208 if (ins
->type
!= TAG_ALU_4
) continue;
2209 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2210 if (ins
->alu
.outmod
!= midgard_outmod_pos
) continue;
2212 /* TODO: Registers? */
2213 unsigned src
= ins
->src
[1];
2214 if (src
& PAN_IS_REG
) continue;
2216 /* There might be a source modifier, too */
2217 if (mir_nontrivial_source2_mod(ins
)) continue;
2219 /* Backpropagate the modifier */
2220 mir_foreach_instr_in_block_from_rev(block
, v
, mir_prev_op(ins
)) {
2221 if (v
->type
!= TAG_ALU_4
) continue;
2222 if (v
->dest
!= src
) continue;
2224 /* Can we even take a float outmod? */
2225 if (midgard_is_integer_out_op(v
->alu
.op
)) continue;
2227 midgard_outmod_float temp
= v
->alu
.outmod
;
2228 progress
|= mir_compose_float_outmod(&temp
, ins
->alu
.outmod
);
2230 /* Throw in the towel.. */
2231 if (!progress
) break;
2233 /* Otherwise, transfer the modifier */
2234 v
->alu
.outmod
= temp
;
2235 ins
->alu
.outmod
= midgard_outmod_none
;
2245 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2247 /* Loop to ourselves */
2248 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2249 struct midgard_instruction ins
= v_branch(false, false);
2250 ins
.writeout
= true;
2251 ins
.writeout_depth
= br
->writeout_depth
;
2252 ins
.writeout_stencil
= br
->writeout_stencil
;
2253 ins
.branch
.target_block
= ctx
->block_count
- 1;
2254 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2255 emit_mir_instruction(ctx
, ins
);
2257 ctx
->current_block
->epilogue
= true;
2258 schedule_barrier(ctx
);
2259 return ins
.branch
.target_block
;
2262 static midgard_block
*
2263 emit_block(compiler_context
*ctx
, nir_block
*block
)
2265 midgard_block
*this_block
= ctx
->after_block
;
2266 ctx
->after_block
= NULL
;
2269 this_block
= create_empty_block(ctx
);
2271 list_addtail(&this_block
->base
.link
, &ctx
->blocks
);
2273 this_block
->scheduled
= false;
2276 /* Set up current block */
2277 list_inithead(&this_block
->base
.instructions
);
2278 ctx
->current_block
= this_block
;
2280 nir_foreach_instr(instr
, block
) {
2281 emit_instr(ctx
, instr
);
2282 ++ctx
->instruction_count
;
2288 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2291 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2293 midgard_block
*before_block
= ctx
->current_block
;
2295 /* Speculatively emit the branch, but we can't fill it in until later */
2296 EMIT(branch
, true, true);
2297 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2298 then_branch
->src
[0] = nir_src_index(ctx
, &nif
->condition
);
2299 then_branch
->src_types
[0] = nir_type_uint32
;
2301 /* Emit the two subblocks. */
2302 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2303 midgard_block
*end_then_block
= ctx
->current_block
;
2305 /* Emit a jump from the end of the then block to the end of the else */
2306 EMIT(branch
, false, false);
2307 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2309 /* Emit second block, and check if it's empty */
2311 int else_idx
= ctx
->block_count
;
2312 int count_in
= ctx
->instruction_count
;
2313 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2314 midgard_block
*end_else_block
= ctx
->current_block
;
2315 int after_else_idx
= ctx
->block_count
;
2317 /* Now that we have the subblocks emitted, fix up the branches */
2322 if (ctx
->instruction_count
== count_in
) {
2323 /* The else block is empty, so don't emit an exit jump */
2324 mir_remove_instruction(then_exit
);
2325 then_branch
->branch
.target_block
= after_else_idx
;
2327 then_branch
->branch
.target_block
= else_idx
;
2328 then_exit
->branch
.target_block
= after_else_idx
;
2331 /* Wire up the successors */
2333 ctx
->after_block
= create_empty_block(ctx
);
2335 pan_block_add_successor(&before_block
->base
, &then_block
->base
);
2336 pan_block_add_successor(&before_block
->base
, &else_block
->base
);
2338 pan_block_add_successor(&end_then_block
->base
, &ctx
->after_block
->base
);
2339 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
);
2343 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2345 /* Remember where we are */
2346 midgard_block
*start_block
= ctx
->current_block
;
2348 /* Allocate a loop number, growing the current inner loop depth */
2349 int loop_idx
= ++ctx
->current_loop_depth
;
2351 /* Get index from before the body so we can loop back later */
2352 int start_idx
= ctx
->block_count
;
2354 /* Emit the body itself */
2355 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2357 /* Branch back to loop back */
2358 struct midgard_instruction br_back
= v_branch(false, false);
2359 br_back
.branch
.target_block
= start_idx
;
2360 emit_mir_instruction(ctx
, br_back
);
2362 /* Mark down that branch in the graph. */
2363 pan_block_add_successor(&start_block
->base
, &loop_block
->base
);
2364 pan_block_add_successor(&ctx
->current_block
->base
, &loop_block
->base
);
2366 /* Find the index of the block about to follow us (note: we don't add
2367 * one; blocks are 0-indexed so we get a fencepost problem) */
2368 int break_block_idx
= ctx
->block_count
;
2370 /* Fix up the break statements we emitted to point to the right place,
2371 * now that we can allocate a block number for them */
2372 ctx
->after_block
= create_empty_block(ctx
);
2374 mir_foreach_block_from(ctx
, start_block
, _block
) {
2375 mir_foreach_instr_in_block(((midgard_block
*) _block
), ins
) {
2376 if (ins
->type
!= TAG_ALU_4
) continue;
2377 if (!ins
->compact_branch
) continue;
2379 /* We found a branch -- check the type to see if we need to do anything */
2380 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2382 /* It's a break! Check if it's our break */
2383 if (ins
->branch
.target_break
!= loop_idx
) continue;
2385 /* Okay, cool, we're breaking out of this loop.
2386 * Rewrite from a break to a goto */
2388 ins
->branch
.target_type
= TARGET_GOTO
;
2389 ins
->branch
.target_block
= break_block_idx
;
2391 pan_block_add_successor(_block
, &ctx
->after_block
->base
);
2395 /* Now that we've finished emitting the loop, free up the depth again
2396 * so we play nice with recursion amid nested loops */
2397 --ctx
->current_loop_depth
;
2399 /* Dump loop stats */
2403 static midgard_block
*
2404 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2406 midgard_block
*start_block
= NULL
;
2408 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2409 switch (node
->type
) {
2410 case nir_cf_node_block
: {
2411 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2414 start_block
= block
;
2419 case nir_cf_node_if
:
2420 emit_if(ctx
, nir_cf_node_as_if(node
));
2423 case nir_cf_node_loop
:
2424 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2427 case nir_cf_node_function
:
2436 /* Due to lookahead, we need to report the first tag executed in the command
2437 * stream and in branch targets. An initial block might be empty, so iterate
2438 * until we find one that 'works' */
2441 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2443 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2445 mir_foreach_block_from(ctx
, initial_block
, _v
) {
2446 midgard_block
*v
= (midgard_block
*) _v
;
2447 if (v
->quadword_count
) {
2448 midgard_bundle
*initial_bundle
=
2449 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2451 return initial_bundle
->tag
;
2455 /* Default to a tag 1 which will break from the shader, in case we jump
2456 * to the exit block (i.e. `return` in a compute shader) */
2461 /* For each fragment writeout instruction, generate a writeout loop to
2462 * associate with it */
2465 mir_add_writeout_loops(compiler_context
*ctx
)
2467 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2468 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2471 unsigned popped
= br
->branch
.target_block
;
2472 pan_block_add_successor(&(mir_get_block(ctx
, popped
- 1)->base
), &ctx
->current_block
->base
);
2473 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2474 br
->branch
.target_type
= TARGET_GOTO
;
2476 /* If we have more RTs, we'll need to restore back after our
2477 * loop terminates */
2479 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2480 midgard_instruction uncond
= v_branch(false, false);
2481 uncond
.branch
.target_block
= popped
;
2482 uncond
.branch
.target_type
= TARGET_GOTO
;
2483 emit_mir_instruction(ctx
, uncond
);
2484 pan_block_add_successor(&ctx
->current_block
->base
, &(mir_get_block(ctx
, popped
)->base
));
2485 schedule_barrier(ctx
);
2487 /* We're last, so we can terminate here */
2488 br
->last_writeout
= true;
2494 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2496 struct util_dynarray
*compiled
= &program
->compiled
;
2498 midgard_debug
= debug_get_option_midgard_debug();
2500 /* TODO: Bound against what? */
2501 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2504 ctx
->stage
= nir
->info
.stage
;
2505 ctx
->is_blend
= is_blend
;
2506 ctx
->alpha_ref
= program
->alpha_ref
;
2507 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2508 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2510 /* Start off with a safe cutoff, allowing usage of all 16 work
2511 * registers. Later, we'll promote uniform reads to uniform registers
2512 * if we determine it is beneficial to do so */
2513 ctx
->uniform_cutoff
= 8;
2515 /* Initialize at a global (not block) level hash tables */
2517 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2518 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2520 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2521 * (so we don't accidentally duplicate the epilogue since mesa/st has
2522 * messed with our I/O quite a bit already) */
2524 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2526 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2527 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2528 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2531 NIR_PASS_V(nir
, nir_lower_var_copies
);
2532 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2533 NIR_PASS_V(nir
, nir_split_var_copies
);
2534 NIR_PASS_V(nir
, nir_lower_var_copies
);
2535 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2536 NIR_PASS_V(nir
, nir_lower_var_copies
);
2537 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2539 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2540 NIR_PASS_V(nir
, nir_lower_ssbo
);
2541 NIR_PASS_V(nir
, midgard_nir_lower_zs_store
);
2543 /* Optimisation passes */
2545 optimise_nir(nir
, ctx
->quirks
);
2547 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2548 nir_print_shader(nir
, stdout
);
2551 /* Assign sysvals and counts, now that we're sure
2552 * (post-optimisation) */
2554 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2555 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2556 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2558 nir_foreach_function(func
, nir
) {
2562 list_inithead(&ctx
->blocks
);
2563 ctx
->block_count
= 0;
2566 emit_cf_list(ctx
, &func
->impl
->body
);
2567 break; /* TODO: Multi-function shaders */
2570 util_dynarray_init(compiled
, NULL
);
2572 /* Per-block lowering before opts */
2574 mir_foreach_block(ctx
, _block
) {
2575 midgard_block
*block
= (midgard_block
*) _block
;
2576 inline_alu_constants(ctx
, block
);
2577 midgard_opt_promote_fmov(ctx
, block
);
2578 embedded_to_inline_constant(ctx
, block
);
2580 /* MIR-level optimizations */
2582 bool progress
= false;
2587 mir_foreach_block(ctx
, _block
) {
2588 midgard_block
*block
= (midgard_block
*) _block
;
2589 progress
|= midgard_opt_pos_propagate(ctx
, block
);
2590 progress
|= midgard_opt_copy_prop(ctx
, block
);
2591 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2592 progress
|= midgard_opt_combine_projection(ctx
, block
);
2593 progress
|= midgard_opt_varying_projection(ctx
, block
);
2594 progress
|= midgard_opt_not_propagate(ctx
, block
);
2595 progress
|= midgard_opt_fuse_src_invert(ctx
, block
);
2596 progress
|= midgard_opt_fuse_dest_invert(ctx
, block
);
2597 progress
|= midgard_opt_csel_invert(ctx
, block
);
2598 progress
|= midgard_opt_drop_cmp_invert(ctx
, block
);
2599 progress
|= midgard_opt_invert_branch(ctx
, block
);
2603 mir_foreach_block(ctx
, _block
) {
2604 midgard_block
*block
= (midgard_block
*) _block
;
2605 midgard_lower_invert(ctx
, block
);
2606 midgard_lower_derivatives(ctx
, block
);
2609 /* Nested control-flow can result in dead branches at the end of the
2610 * block. This messes with our analysis and is just dead code, so cull
2612 mir_foreach_block(ctx
, _block
) {
2613 midgard_block
*block
= (midgard_block
*) _block
;
2614 midgard_opt_cull_dead_branch(ctx
, block
);
2617 /* Ensure we were lowered */
2618 mir_foreach_instr_global(ctx
, ins
) {
2619 assert(!ins
->invert
);
2622 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2623 mir_add_writeout_loops(ctx
);
2626 midgard_schedule_program(ctx
);
2629 /* Now that all the bundles are scheduled and we can calculate block
2630 * sizes, emit actual branch instructions rather than placeholders */
2632 int br_block_idx
= 0;
2634 mir_foreach_block(ctx
, _block
) {
2635 midgard_block
*block
= (midgard_block
*) _block
;
2636 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2637 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2638 midgard_instruction
*ins
= bundle
->instructions
[c
];
2640 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2642 /* Parse some basic branch info */
2643 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2644 bool is_conditional
= ins
->branch
.conditional
;
2645 bool is_inverted
= ins
->branch
.invert_conditional
;
2646 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2647 bool is_writeout
= ins
->writeout
;
2649 /* Determine the block we're jumping to */
2650 int target_number
= ins
->branch
.target_block
;
2652 /* Report the destination tag */
2653 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2655 /* Count up the number of quadwords we're
2656 * jumping over = number of quadwords until
2657 * (br_block_idx, target_number) */
2659 int quadword_offset
= 0;
2663 } else if (target_number
> br_block_idx
) {
2666 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2667 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2670 quadword_offset
+= blk
->quadword_count
;
2673 /* Jump backwards */
2675 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2676 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2679 quadword_offset
-= blk
->quadword_count
;
2683 /* Unconditional extended branches (far jumps)
2684 * have issues, so we always use a conditional
2685 * branch, setting the condition to always for
2686 * unconditional. For compact unconditional
2687 * branches, cond isn't used so it doesn't
2688 * matter what we pick. */
2690 midgard_condition cond
=
2691 !is_conditional
? midgard_condition_always
:
2692 is_inverted
? midgard_condition_false
:
2693 midgard_condition_true
;
2695 midgard_jmp_writeout_op op
=
2696 is_discard
? midgard_jmp_writeout_op_discard
:
2697 is_writeout
? midgard_jmp_writeout_op_writeout
:
2698 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2699 midgard_jmp_writeout_op_branch_cond
;
2702 midgard_branch_extended branch
=
2703 midgard_create_branch_extended(
2708 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2709 } else if (is_conditional
|| is_discard
) {
2710 midgard_branch_cond branch
= {
2712 .dest_tag
= dest_tag
,
2713 .offset
= quadword_offset
,
2717 assert(branch
.offset
== quadword_offset
);
2719 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2721 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2723 midgard_branch_uncond branch
= {
2725 .dest_tag
= dest_tag
,
2726 .offset
= quadword_offset
,
2730 assert(branch
.offset
== quadword_offset
);
2732 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2740 /* Emit flat binary from the instruction arrays. Iterate each block in
2741 * sequence. Save instruction boundaries such that lookahead tags can
2742 * be assigned easily */
2744 /* Cache _all_ bundles in source order for lookahead across failed branches */
2746 int bundle_count
= 0;
2747 mir_foreach_block(ctx
, _block
) {
2748 midgard_block
*block
= (midgard_block
*) _block
;
2749 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2751 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2753 mir_foreach_block(ctx
, _block
) {
2754 midgard_block
*block
= (midgard_block
*) _block
;
2755 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2756 source_order_bundles
[bundle_idx
++] = bundle
;
2760 int current_bundle
= 0;
2762 /* Midgard prefetches instruction types, so during emission we
2763 * need to lookahead. Unless this is the last instruction, in
2764 * which we return 1. */
2766 mir_foreach_block(ctx
, _block
) {
2767 midgard_block
*block
= (midgard_block
*) _block
;
2768 mir_foreach_bundle_in_block(block
, bundle
) {
2771 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2772 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2774 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2778 /* TODO: Free deeper */
2779 //util_dynarray_fini(&block->instructions);
2782 free(source_order_bundles
);
2784 /* Report the very first tag executed */
2785 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2787 /* Deal with off-by-one related to the fencepost problem */
2788 program
->work_register_count
= ctx
->work_registers
+ 1;
2789 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2791 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2792 program
->tls_size
= ctx
->tls_size
;
2794 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2795 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
2797 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
2798 unsigned nr_bundles
= 0, nr_ins
= 0;
2800 /* Count instructions and bundles */
2802 mir_foreach_block(ctx
, _block
) {
2803 midgard_block
*block
= (midgard_block
*) _block
;
2804 nr_bundles
+= util_dynarray_num_elements(
2805 &block
->bundles
, midgard_bundle
);
2807 mir_foreach_bundle_in_block(block
, bun
)
2808 nr_ins
+= bun
->instruction_count
;
2811 /* Calculate thread count. There are certain cutoffs by
2812 * register count for thread count */
2814 unsigned nr_registers
= program
->work_register_count
;
2816 unsigned nr_threads
=
2817 (nr_registers
<= 4) ? 4 :
2818 (nr_registers
<= 8) ? 2 :
2823 fprintf(stderr
, "shader%d - %s shader: "
2824 "%u inst, %u bundles, %u quadwords, "
2825 "%u registers, %u threads, %u loops, "
2826 "%u:%u spills:fills\n",
2828 gl_shader_stage_name(ctx
->stage
),
2829 nr_ins
, nr_bundles
, ctx
->quadword_count
,
2830 nr_registers
, nr_threads
,
2832 ctx
->spills
, ctx
->fills
);