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
);
492 NIR_PASS(progress
, nir
, nir_opt_algebraic_distribute_src_mods
);
494 /* We implement booleans as 32-bit 0/~0 */
495 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
497 /* Now that booleans are lowered, we can run out late opts */
498 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
500 /* Lower mods for float ops only. Integer ops don't support modifiers
501 * (saturate doesn't make sense on integers, neg/abs require dedicated
504 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
505 NIR_PASS(progress
, nir
, nir_copy_prop
);
506 NIR_PASS(progress
, nir
, nir_opt_dce
);
508 /* Take us out of SSA */
509 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
510 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
512 /* We are a vector architecture; write combine where possible */
513 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
514 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
516 NIR_PASS(progress
, nir
, nir_opt_dce
);
519 /* Do not actually emit a load; instead, cache the constant for inlining */
522 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
524 nir_ssa_def def
= instr
->def
;
526 midgard_constants
*consts
= rzalloc(NULL
, midgard_constants
);
528 assert(instr
->def
.num_components
* instr
->def
.bit_size
<= sizeof(*consts
) * 8);
530 #define RAW_CONST_COPY(bits) \
531 nir_const_value_to_array(consts->u##bits, instr->value, \
532 instr->def.num_components, u##bits)
534 switch (instr
->def
.bit_size
) {
548 unreachable("Invalid bit_size for load_const instruction\n");
551 /* Shifted for SSA, +1 for off-by-one */
552 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, consts
);
555 /* Normally constants are embedded implicitly, but for I/O and such we have to
556 * explicitly emit a move with the constant source */
559 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
561 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
563 if (constant_value
) {
564 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), to
);
565 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
566 emit_mir_instruction(ctx
, ins
);
571 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
573 unsigned comp
= src
->swizzle
[0];
575 for (unsigned c
= 1; c
< nr_components
; ++c
) {
576 if (src
->swizzle
[c
] != comp
)
583 #define ALU_CASE(nir, _op) \
585 op = midgard_alu_op_##_op; \
586 assert(src_bitsize == dst_bitsize); \
589 #define ALU_CASE_BCAST(nir, _op, count) \
591 op = midgard_alu_op_##_op; \
592 broadcast_swizzle = count; \
593 assert(src_bitsize == dst_bitsize); \
595 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
596 * special treatment override this anyway. */
598 static midgard_reg_mode
599 reg_mode_for_nir(nir_alu_instr
*instr
)
601 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
603 switch (src_bitsize
) {
605 return midgard_reg_mode_8
;
607 return midgard_reg_mode_16
;
609 return midgard_reg_mode_32
;
611 return midgard_reg_mode_64
;
613 unreachable("Invalid bit size");
618 mir_copy_src(midgard_instruction
*ins
, nir_alu_instr
*instr
, unsigned i
, unsigned to
)
620 unsigned bits
= nir_src_bit_size(instr
->src
[i
].src
);
622 ins
->src
[to
] = nir_src_index(NULL
, &instr
->src
[i
].src
);
623 ins
->src_types
[to
] = nir_op_infos
[instr
->op
].input_types
[i
] | bits
;
627 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
629 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
630 * is handled elsewhere */
632 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
633 midgard_emit_derivatives(ctx
, instr
);
637 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
639 unsigned nr_components
= nir_dest_num_components(instr
->dest
.dest
);
640 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
643 /* Number of components valid to check for the instruction (the rest
644 * will be forced to the last), or 0 to use as-is. Relevant as
645 * ball-type instructions have a channel count in NIR but are all vec4
648 unsigned broadcast_swizzle
= 0;
650 /* What register mode should we operate in? */
651 midgard_reg_mode reg_mode
=
652 reg_mode_for_nir(instr
);
654 /* Do we need a destination override? Used for inline
657 midgard_dest_override dest_override
=
658 midgard_dest_override_none
;
660 /* Should we use a smaller respective source and sign-extend? */
662 bool half_1
= false, sext_1
= false;
663 bool half_2
= false, sext_2
= false;
665 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
666 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
669 ALU_CASE(fadd
, fadd
);
670 ALU_CASE(fmul
, fmul
);
671 ALU_CASE(fmin
, fmin
);
672 ALU_CASE(fmax
, fmax
);
673 ALU_CASE(imin
, imin
);
674 ALU_CASE(imax
, imax
);
675 ALU_CASE(umin
, umin
);
676 ALU_CASE(umax
, umax
);
677 ALU_CASE(ffloor
, ffloor
);
678 ALU_CASE(fround_even
, froundeven
);
679 ALU_CASE(ftrunc
, ftrunc
);
680 ALU_CASE(fceil
, fceil
);
681 ALU_CASE(fdot3
, fdot3
);
682 ALU_CASE(fdot4
, fdot4
);
683 ALU_CASE(iadd
, iadd
);
684 ALU_CASE(isub
, isub
);
685 ALU_CASE(imul
, imul
);
687 /* Zero shoved as second-arg */
688 ALU_CASE(iabs
, iabsdiff
);
692 ALU_CASE(feq32
, feq
);
693 ALU_CASE(fne32
, fne
);
694 ALU_CASE(flt32
, flt
);
695 ALU_CASE(ieq32
, ieq
);
696 ALU_CASE(ine32
, ine
);
697 ALU_CASE(ilt32
, ilt
);
698 ALU_CASE(ult32
, ult
);
700 /* We don't have a native b2f32 instruction. Instead, like many
701 * GPUs, we exploit booleans as 0/~0 for false/true, and
702 * correspondingly AND
703 * by 1.0 to do the type conversion. For the moment, prime us
706 * iand [whatever], #0
708 * At the end of emit_alu (as MIR), we'll fix-up the constant
711 ALU_CASE(b2f32
, iand
);
712 ALU_CASE(b2i32
, iand
);
714 /* Likewise, we don't have a dedicated f2b32 instruction, but
715 * we can do a "not equal to 0.0" test. */
717 ALU_CASE(f2b32
, fne
);
718 ALU_CASE(i2b32
, ine
);
720 ALU_CASE(frcp
, frcp
);
721 ALU_CASE(frsq
, frsqrt
);
722 ALU_CASE(fsqrt
, fsqrt
);
723 ALU_CASE(fexp2
, fexp2
);
724 ALU_CASE(flog2
, flog2
);
726 ALU_CASE(f2i64
, f2i_rtz
);
727 ALU_CASE(f2u64
, f2u_rtz
);
728 ALU_CASE(i2f64
, i2f_rtz
);
729 ALU_CASE(u2f64
, u2f_rtz
);
731 ALU_CASE(f2i32
, f2i_rtz
);
732 ALU_CASE(f2u32
, f2u_rtz
);
733 ALU_CASE(i2f32
, i2f_rtz
);
734 ALU_CASE(u2f32
, u2f_rtz
);
736 ALU_CASE(f2i16
, f2i_rtz
);
737 ALU_CASE(f2u16
, f2u_rtz
);
738 ALU_CASE(i2f16
, i2f_rtz
);
739 ALU_CASE(u2f16
, u2f_rtz
);
741 ALU_CASE(fsin
, fsin
);
742 ALU_CASE(fcos
, fcos
);
744 /* We'll set invert */
745 ALU_CASE(inot
, imov
);
746 ALU_CASE(iand
, iand
);
748 ALU_CASE(ixor
, ixor
);
749 ALU_CASE(ishl
, ishl
);
750 ALU_CASE(ishr
, iasr
);
751 ALU_CASE(ushr
, ilsr
);
753 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
754 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
755 ALU_CASE(b32all_fequal4
, fball_eq
);
757 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
758 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
759 ALU_CASE(b32any_fnequal4
, fbany_neq
);
761 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
762 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
763 ALU_CASE(b32all_iequal4
, iball_eq
);
765 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
766 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
767 ALU_CASE(b32any_inequal4
, ibany_neq
);
769 /* Source mods will be shoved in later */
770 ALU_CASE(fabs
, fmov
);
771 ALU_CASE(fneg
, fmov
);
772 ALU_CASE(fsat
, fmov
);
774 /* For size conversion, we use a move. Ideally though we would squash
775 * these ops together; maybe that has to happen after in NIR as part of
776 * propagation...? An earlier algebraic pass ensured we step down by
777 * only / exactly one size. If stepping down, we use a dest override to
778 * reduce the size; if stepping up, we use a larger-sized move with a
779 * half source and a sign/zero-extension modifier */
785 /* If we end up upscale, we'll need a sign-extend on the
786 * operand (the second argument) */
797 if (instr
->op
== nir_op_f2f16
|| instr
->op
== nir_op_f2f32
||
798 instr
->op
== nir_op_f2f64
)
799 op
= midgard_alu_op_fmov
;
801 op
= midgard_alu_op_imov
;
803 if (dst_bitsize
== (src_bitsize
* 2)) {
807 /* Use a greater register mode */
809 } else if (src_bitsize
== (dst_bitsize
* 2)) {
810 /* Converting down */
811 dest_override
= midgard_dest_override_lower
;
817 /* For greater-or-equal, we lower to less-or-equal and flip the
825 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
826 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
827 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
828 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
831 /* Swap via temporary */
832 nir_alu_src temp
= instr
->src
[1];
833 instr
->src
[1] = instr
->src
[0];
834 instr
->src
[0] = temp
;
839 case nir_op_b32csel
: {
840 /* Midgard features both fcsel and icsel, depending on
841 * the type of the arguments/output. However, as long
842 * as we're careful we can _always_ use icsel and
843 * _never_ need fcsel, since the latter does additional
844 * floating-point-specific processing whereas the
845 * former just moves bits on the wire. It's not obvious
846 * why these are separate opcodes, save for the ability
847 * to do things like sat/pos/abs/neg for free */
849 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
850 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
852 /* The condition is the first argument; move the other
853 * arguments up one to be a binary instruction for
854 * Midgard with the condition last */
856 nir_alu_src temp
= instr
->src
[2];
858 instr
->src
[2] = instr
->src
[0];
859 instr
->src
[0] = instr
->src
[1];
860 instr
->src
[1] = temp
;
866 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
871 /* Midgard can perform certain modifiers on output of an ALU op */
874 if (midgard_is_integer_out_op(op
)) {
875 outmod
= midgard_outmod_int_wrap
;
877 bool sat
= instr
->dest
.saturate
|| instr
->op
== nir_op_fsat
;
878 outmod
= sat
? midgard_outmod_sat
: midgard_outmod_none
;
882 /* Fetch unit, quirks, etc information */
883 unsigned opcode_props
= alu_opcode_props
[op
].props
;
884 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
886 midgard_instruction ins
= {
888 .dest
= nir_dest_index(&instr
->dest
.dest
),
889 .dest_type
= nir_op_infos
[instr
->op
].output_type
890 | nir_dest_bit_size(instr
->dest
.dest
),
893 for (unsigned i
= nr_inputs
; i
< ARRAY_SIZE(ins
.src
); ++i
)
896 if (quirk_flipped_r24
) {
898 mir_copy_src(&ins
, instr
, 0, 1);
900 for (unsigned i
= 0; i
< nr_inputs
; ++i
)
901 mir_copy_src(&ins
, instr
, i
, quirk_flipped_r24
? 1 : i
);
904 nir_alu_src
*nirmods
[3] = { NULL
};
906 if (nr_inputs
>= 2) {
907 nirmods
[0] = &instr
->src
[0];
908 nirmods
[1] = &instr
->src
[1];
909 } else if (nr_inputs
== 1) {
910 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
916 nirmods
[2] = &instr
->src
[2];
918 /* These were lowered to a move, so apply the corresponding mod */
920 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
921 nir_alu_src
*s
= nirmods
[quirk_flipped_r24
];
923 if (instr
->op
== nir_op_fneg
)
924 s
->negate
= !s
->negate
;
926 if (instr
->op
== nir_op_fabs
)
930 bool is_int
= midgard_is_integer_op(op
);
932 ins
.mask
= mask_of(nr_components
);
934 midgard_vector_alu alu
= {
936 .reg_mode
= reg_mode
,
937 .dest_override
= dest_override
,
940 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, sext_1
)),
941 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
, broadcast_swizzle
, half_2
, sext_2
)),
944 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
947 ins
.mask
&= instr
->dest
.write_mask
;
949 for (unsigned m
= 0; m
< 3; ++m
) {
953 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; ++c
)
954 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[c
];
956 /* Replicate. TODO: remove when vec16 lands */
957 for (unsigned c
= NIR_MAX_VEC_COMPONENTS
; c
< MIR_VEC_COMPONENTS
; ++c
)
958 ins
.swizzle
[m
][c
] = nirmods
[m
]->swizzle
[NIR_MAX_VEC_COMPONENTS
- 1];
961 if (nr_inputs
== 3) {
962 /* Conditions can't have mods */
963 assert(!nirmods
[2]->abs
);
964 assert(!nirmods
[2]->negate
);
969 /* Late fixup for emulated instructions */
971 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
972 /* Presently, our second argument is an inline #0 constant.
973 * Switch over to an embedded 1.0 constant (that can't fit
974 * inline, since we're 32-bit, not 16-bit like the inline
977 ins
.has_inline_constant
= false;
978 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
979 ins
.src_types
[1] = nir_type_float32
;
980 ins
.has_constants
= true;
982 if (instr
->op
== nir_op_b2f32
)
983 ins
.constants
.f32
[0] = 1.0f
;
985 ins
.constants
.i32
[0] = 1;
987 for (unsigned c
= 0; c
< 16; ++c
)
988 ins
.swizzle
[1][c
] = 0;
989 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
990 /* Lots of instructions need a 0 plonked in */
991 ins
.has_inline_constant
= false;
992 ins
.src
[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
993 ins
.src_types
[1] = nir_type_uint32
;
994 ins
.has_constants
= true;
995 ins
.constants
.u32
[0] = 0;
997 for (unsigned c
= 0; c
< 16; ++c
)
998 ins
.swizzle
[1][c
] = 0;
999 } else if (instr
->op
== nir_op_inot
) {
1003 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1004 /* To avoid duplicating the lookup tables (probably), true LUT
1005 * instructions can only operate as if they were scalars. Lower
1006 * them here by changing the component. */
1008 unsigned orig_mask
= ins
.mask
;
1010 for (int i
= 0; i
< nr_components
; ++i
) {
1011 /* Mask the associated component, dropping the
1012 * instruction if needed */
1015 ins
.mask
&= orig_mask
;
1020 for (unsigned j
= 0; j
< MIR_VEC_COMPONENTS
; ++j
)
1021 ins
.swizzle
[0][j
] = nirmods
[0]->swizzle
[i
]; /* Pull from the correct component */
1023 emit_mir_instruction(ctx
, ins
);
1026 emit_mir_instruction(ctx
, ins
);
1033 mir_set_intr_mask(nir_instr
*instr
, midgard_instruction
*ins
, bool is_read
)
1035 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
1036 unsigned nir_mask
= 0;
1040 nir_mask
= mask_of(nir_intrinsic_dest_components(intr
));
1041 dsize
= nir_dest_bit_size(intr
->dest
);
1043 nir_mask
= nir_intrinsic_write_mask(intr
);
1047 /* Once we have the NIR mask, we need to normalize to work in 32-bit space */
1048 unsigned bytemask
= pan_to_bytemask(dsize
, nir_mask
);
1049 mir_set_bytemask(ins
, bytemask
);
1052 ins
->load_64
= true;
1055 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1056 * optimized) versions of UBO #0 */
1058 static midgard_instruction
*
1060 compiler_context
*ctx
,
1064 nir_src
*indirect_offset
,
1065 unsigned indirect_shift
,
1068 /* TODO: half-floats */
1070 midgard_instruction ins
= m_ld_ubo_int4(dest
, 0);
1071 ins
.constants
.u32
[0] = offset
;
1073 if (instr
->type
== nir_instr_type_intrinsic
)
1074 mir_set_intr_mask(instr
, &ins
, true);
1076 if (indirect_offset
) {
1077 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1078 ins
.src_types
[2] = nir_type_uint32
;
1079 ins
.load_store
.arg_2
= (indirect_shift
<< 5);
1081 ins
.load_store
.arg_2
= 0x1E;
1084 ins
.load_store
.arg_1
= index
;
1086 return emit_mir_instruction(ctx
, ins
);
1089 /* Globals are like UBOs if you squint. And shared memory is like globals if
1090 * you squint even harder */
1094 compiler_context
*ctx
,
1103 midgard_instruction ins
;
1106 ins
= m_ld_int4(srcdest
, 0);
1108 ins
= m_st_int4(srcdest
, 0);
1110 mir_set_offset(ctx
, &ins
, offset
, is_shared
);
1111 mir_set_intr_mask(instr
, &ins
, is_read
);
1113 emit_mir_instruction(ctx
, ins
);
1118 compiler_context
*ctx
,
1119 unsigned dest
, unsigned offset
,
1120 unsigned nr_comp
, unsigned component
,
1121 nir_src
*indirect_offset
, nir_alu_type type
, bool flat
)
1123 /* XXX: Half-floats? */
1124 /* TODO: swizzle, mask */
1126 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1127 ins
.mask
= mask_of(nr_comp
);
1129 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
.swizzle
[0]); ++i
)
1130 ins
.swizzle
[0][i
] = MIN2(i
+ component
, COMPONENT_W
);
1132 midgard_varying_parameter p
= {
1134 .interpolation
= midgard_interp_default
,
1139 memcpy(&u
, &p
, sizeof(p
));
1140 ins
.load_store
.varying_parameters
= u
;
1142 if (indirect_offset
) {
1143 ins
.src
[2] = nir_src_index(ctx
, indirect_offset
);
1144 ins
.src_types
[2] = nir_type_uint32
;
1146 ins
.load_store
.arg_2
= 0x1E;
1148 ins
.load_store
.arg_1
= 0x9E;
1150 /* Use the type appropriate load */
1154 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1157 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1159 case nir_type_float
:
1160 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1163 unreachable("Attempted to load unknown type");
1167 emit_mir_instruction(ctx
, ins
);
1172 compiler_context
*ctx
,
1173 unsigned dest
, unsigned offset
,
1174 unsigned nr_comp
, nir_alu_type t
)
1176 midgard_instruction ins
= m_ld_attr_32(dest
, offset
);
1177 ins
.load_store
.arg_1
= 0x1E;
1178 ins
.load_store
.arg_2
= 0x1E;
1179 ins
.mask
= mask_of(nr_comp
);
1181 /* Use the type appropriate load */
1185 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1188 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1190 case nir_type_float
:
1191 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1194 unreachable("Attempted to load unknown type");
1198 emit_mir_instruction(ctx
, ins
);
1202 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
,
1203 unsigned nr_components
, unsigned offset
)
1207 /* Figure out which uniform this is */
1208 int sysval
= panfrost_sysval_for_instr(instr
, &nir_dest
);
1209 void *val
= _mesa_hash_table_u64_search(ctx
->sysvals
.sysval_to_id
, sysval
);
1211 unsigned dest
= nir_dest_index(&nir_dest
);
1213 /* Sysvals are prefix uniforms */
1214 unsigned uniform
= ((uintptr_t) val
) - 1;
1216 /* Emit the read itself -- this is never indirect */
1217 midgard_instruction
*ins
=
1218 emit_ubo_read(ctx
, instr
, dest
, (uniform
* 16) + offset
, NULL
, 0, 0);
1220 ins
->mask
= mask_of(nr_components
);
1224 compute_builtin_arg(nir_op op
)
1227 case nir_intrinsic_load_work_group_id
:
1229 case nir_intrinsic_load_local_invocation_id
:
1232 unreachable("Invalid compute paramater loaded");
1237 emit_fragment_store(compiler_context
*ctx
, unsigned src
, enum midgard_rt_id rt
)
1239 assert(rt
< ARRAY_SIZE(ctx
->writeout_branch
));
1241 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
1245 emit_explicit_constant(ctx
, src
, src
);
1247 struct midgard_instruction ins
=
1248 v_branch(false, false);
1250 ins
.writeout
= true;
1252 /* Add dependencies */
1254 ins
.src_types
[0] = nir_type_uint32
;
1255 ins
.constants
.u32
[0] = rt
== MIDGARD_ZS_RT
?
1256 0xFF : (rt
- MIDGARD_COLOR_RT0
) * 0x100;
1258 /* Emit the branch */
1259 br
= emit_mir_instruction(ctx
, ins
);
1260 schedule_barrier(ctx
);
1261 ctx
->writeout_branch
[rt
] = br
;
1263 /* Push our current location = current block count - 1 = where we'll
1264 * jump to. Maybe a bit too clever for my own good */
1266 br
->branch
.target_block
= ctx
->block_count
- 1;
1270 emit_compute_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1272 unsigned reg
= nir_dest_index(&instr
->dest
);
1273 midgard_instruction ins
= m_ld_compute_id(reg
, 0);
1274 ins
.mask
= mask_of(3);
1275 ins
.swizzle
[0][3] = COMPONENT_X
; /* xyzx */
1276 ins
.load_store
.arg_1
= compute_builtin_arg(instr
->intrinsic
);
1277 emit_mir_instruction(ctx
, ins
);
1281 vertex_builtin_arg(nir_op op
)
1284 case nir_intrinsic_load_vertex_id
:
1285 return PAN_VERTEX_ID
;
1286 case nir_intrinsic_load_instance_id
:
1287 return PAN_INSTANCE_ID
;
1289 unreachable("Invalid vertex builtin");
1294 emit_vertex_builtin(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1296 unsigned reg
= nir_dest_index(&instr
->dest
);
1297 emit_attr_read(ctx
, reg
, vertex_builtin_arg(instr
->intrinsic
), 1, nir_type_int
);
1301 emit_control_barrier(compiler_context
*ctx
)
1303 midgard_instruction ins
= {
1304 .type
= TAG_TEXTURE_4
,
1305 .src
= { ~0, ~0, ~0, ~0 },
1307 .op
= TEXTURE_OP_BARRIER
,
1309 /* TODO: optimize */
1310 .barrier_buffer
= 1,
1315 emit_mir_instruction(ctx
, ins
);
1318 static const nir_variable
*
1319 search_var(struct exec_list
*vars
, unsigned driver_loc
)
1321 nir_foreach_variable(var
, vars
) {
1322 if (var
->data
.driver_location
== driver_loc
)
1330 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1332 unsigned offset
= 0, reg
;
1334 switch (instr
->intrinsic
) {
1335 case nir_intrinsic_discard_if
:
1336 case nir_intrinsic_discard
: {
1337 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1338 struct midgard_instruction discard
= v_branch(conditional
, false);
1339 discard
.branch
.target_type
= TARGET_DISCARD
;
1342 discard
.src
[0] = nir_src_index(ctx
, &instr
->src
[0]);
1343 discard
.src_types
[0] = nir_type_uint32
;
1346 emit_mir_instruction(ctx
, discard
);
1347 schedule_barrier(ctx
);
1352 case nir_intrinsic_load_uniform
:
1353 case nir_intrinsic_load_ubo
:
1354 case nir_intrinsic_load_global
:
1355 case nir_intrinsic_load_shared
:
1356 case nir_intrinsic_load_input
:
1357 case nir_intrinsic_load_interpolated_input
: {
1358 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1359 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1360 bool is_global
= instr
->intrinsic
== nir_intrinsic_load_global
;
1361 bool is_shared
= instr
->intrinsic
== nir_intrinsic_load_shared
;
1362 bool is_flat
= instr
->intrinsic
== nir_intrinsic_load_input
;
1363 bool is_interp
= instr
->intrinsic
== nir_intrinsic_load_interpolated_input
;
1365 /* Get the base type of the intrinsic */
1366 /* TODO: Infer type? Does it matter? */
1368 (is_ubo
|| is_global
|| is_shared
) ? nir_type_uint
:
1369 (is_interp
) ? nir_type_float
:
1370 nir_intrinsic_type(instr
);
1372 t
= nir_alu_type_get_base_type(t
);
1374 if (!(is_ubo
|| is_global
)) {
1375 offset
= nir_intrinsic_base(instr
);
1378 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1380 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1382 bool direct
= nir_src_is_const(*src_offset
);
1383 nir_src
*indirect_offset
= direct
? NULL
: src_offset
;
1386 offset
+= nir_src_as_uint(*src_offset
);
1388 /* We may need to apply a fractional offset */
1389 int component
= (is_flat
|| is_interp
) ?
1390 nir_intrinsic_component(instr
) : 0;
1391 reg
= nir_dest_index(&instr
->dest
);
1393 if (is_uniform
&& !ctx
->is_blend
) {
1394 emit_ubo_read(ctx
, &instr
->instr
, reg
, (ctx
->sysvals
.sysval_count
+ offset
) * 16, indirect_offset
, 4, 0);
1395 } else if (is_ubo
) {
1396 nir_src index
= instr
->src
[0];
1398 /* TODO: Is indirect block number possible? */
1399 assert(nir_src_is_const(index
));
1401 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1402 emit_ubo_read(ctx
, &instr
->instr
, reg
, offset
, indirect_offset
, 0, uindex
);
1403 } else if (is_global
|| is_shared
) {
1404 emit_global(ctx
, &instr
->instr
, true, reg
, src_offset
, is_shared
);
1405 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1406 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, indirect_offset
, t
, is_flat
);
1407 } else if (ctx
->is_blend
) {
1408 /* For blend shaders, load the input color, which is
1409 * preloaded to r0 */
1411 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), reg
);
1412 emit_mir_instruction(ctx
, move
);
1413 schedule_barrier(ctx
);
1414 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1415 emit_attr_read(ctx
, reg
, offset
, nr_comp
, t
);
1417 DBG("Unknown load\n");
1424 /* Artefact of load_interpolated_input. TODO: other barycentric modes */
1425 case nir_intrinsic_load_barycentric_pixel
:
1426 case nir_intrinsic_load_barycentric_centroid
:
1429 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1431 case nir_intrinsic_load_raw_output_pan
:
1432 case nir_intrinsic_load_output_u8_as_fp16_pan
:
1433 reg
= nir_dest_index(&instr
->dest
);
1434 assert(ctx
->is_blend
);
1436 /* T720 and below use different blend opcodes with slightly
1437 * different semantics than T760 and up */
1439 midgard_instruction ld
= m_ld_color_buffer_32u(reg
, 0);
1440 bool old_blend
= ctx
->quirks
& MIDGARD_OLD_BLEND
;
1442 if (instr
->intrinsic
== nir_intrinsic_load_output_u8_as_fp16_pan
) {
1443 ld
.load_store
.op
= old_blend
?
1444 midgard_op_ld_color_buffer_u8_as_fp16_old
:
1445 midgard_op_ld_color_buffer_u8_as_fp16
;
1448 ld
.load_store
.address
= 1;
1449 ld
.load_store
.arg_2
= 0x1E;
1452 for (unsigned c
= 2; c
< 16; ++c
)
1453 ld
.swizzle
[0][c
] = 0;
1456 emit_mir_instruction(ctx
, ld
);
1459 case nir_intrinsic_load_blend_const_color_rgba
: {
1460 assert(ctx
->is_blend
);
1461 reg
= nir_dest_index(&instr
->dest
);
1463 /* Blend constants are embedded directly in the shader and
1464 * patched in, so we use some magic routing */
1466 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), reg
);
1467 ins
.has_constants
= true;
1468 ins
.has_blend_constant
= true;
1469 emit_mir_instruction(ctx
, ins
);
1473 case nir_intrinsic_store_zs_output_pan
: {
1474 assert(ctx
->stage
== MESA_SHADER_FRAGMENT
);
1475 emit_fragment_store(ctx
, nir_src_index(ctx
, &instr
->src
[0]),
1478 midgard_instruction
*br
= ctx
->writeout_branch
[MIDGARD_ZS_RT
];
1480 if (!nir_intrinsic_component(instr
))
1481 br
->writeout_depth
= true;
1482 if (nir_intrinsic_component(instr
) ||
1483 instr
->num_components
)
1484 br
->writeout_stencil
= true;
1485 assert(br
->writeout_depth
| br
->writeout_stencil
);
1489 case nir_intrinsic_store_output
:
1490 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1492 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1494 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1496 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1497 const nir_variable
*var
;
1498 enum midgard_rt_id rt
;
1500 var
= search_var(&ctx
->nir
->outputs
,
1501 nir_intrinsic_base(instr
));
1503 if (var
->data
.location
== FRAG_RESULT_COLOR
)
1504 rt
= MIDGARD_COLOR_RT0
;
1505 else if (var
->data
.location
>= FRAG_RESULT_DATA0
)
1506 rt
= MIDGARD_COLOR_RT0
+ var
->data
.location
-
1511 emit_fragment_store(ctx
, reg
, rt
);
1512 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1513 /* We should have been vectorized, though we don't
1514 * currently check that st_vary is emitted only once
1515 * per slot (this is relevant, since there's not a mask
1516 * parameter available on the store [set to 0 by the
1517 * blob]). We do respect the component by adjusting the
1518 * swizzle. If this is a constant source, we'll need to
1519 * emit that explicitly. */
1521 emit_explicit_constant(ctx
, reg
, reg
);
1523 unsigned dst_component
= nir_intrinsic_component(instr
);
1524 unsigned nr_comp
= nir_src_num_components(instr
->src
[0]);
1526 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1527 st
.load_store
.arg_1
= 0x9E;
1528 st
.load_store
.arg_2
= 0x1E;
1530 switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr
))) {
1533 st
.load_store
.op
= midgard_op_st_vary_32u
;
1536 st
.load_store
.op
= midgard_op_st_vary_32i
;
1538 case nir_type_float
:
1539 st
.load_store
.op
= midgard_op_st_vary_32
;
1542 unreachable("Attempted to store unknown type");
1546 /* nir_intrinsic_component(store_intr) encodes the
1547 * destination component start. Source component offset
1548 * adjustment is taken care of in
1549 * install_registers_instr(), when offset_swizzle() is
1552 unsigned src_component
= COMPONENT_X
;
1554 assert(nr_comp
> 0);
1555 for (unsigned i
= 0; i
< ARRAY_SIZE(st
.swizzle
); ++i
) {
1556 st
.swizzle
[0][i
] = src_component
;
1557 if (i
>= dst_component
&& i
< dst_component
+ nr_comp
- 1)
1561 emit_mir_instruction(ctx
, st
);
1563 DBG("Unknown store\n");
1569 /* Special case of store_output for lowered blend shaders */
1570 case nir_intrinsic_store_raw_output_pan
:
1571 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1572 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1574 if (ctx
->quirks
& MIDGARD_OLD_BLEND
) {
1575 /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
1576 * reg = r0.x. We want to splatter. So we can do a 32-bit move
1579 * imov r0.xyzw, r0.xxxx
1582 unsigned expanded
= make_compiler_temp(ctx
);
1584 midgard_instruction splatter
= v_mov(reg
, expanded
);
1586 for (unsigned c
= 0; c
< 16; ++c
)
1587 splatter
.swizzle
[1][c
] = 0;
1589 emit_mir_instruction(ctx
, splatter
);
1590 emit_fragment_store(ctx
, expanded
, ctx
->blend_rt
);
1592 emit_fragment_store(ctx
, reg
, ctx
->blend_rt
);
1596 case nir_intrinsic_store_global
:
1597 case nir_intrinsic_store_shared
:
1598 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1599 emit_explicit_constant(ctx
, reg
, reg
);
1601 emit_global(ctx
, &instr
->instr
, false, reg
, &instr
->src
[1], instr
->intrinsic
== nir_intrinsic_store_shared
);
1604 case nir_intrinsic_load_ssbo_address
:
1605 emit_sysval_read(ctx
, &instr
->instr
, 1, 0);
1608 case nir_intrinsic_get_buffer_size
:
1609 emit_sysval_read(ctx
, &instr
->instr
, 1, 8);
1612 case nir_intrinsic_load_viewport_scale
:
1613 case nir_intrinsic_load_viewport_offset
:
1614 case nir_intrinsic_load_num_work_groups
:
1615 case nir_intrinsic_load_sampler_lod_parameters_pan
:
1616 emit_sysval_read(ctx
, &instr
->instr
, 3, 0);
1619 case nir_intrinsic_load_work_group_id
:
1620 case nir_intrinsic_load_local_invocation_id
:
1621 emit_compute_builtin(ctx
, instr
);
1624 case nir_intrinsic_load_vertex_id
:
1625 case nir_intrinsic_load_instance_id
:
1626 emit_vertex_builtin(ctx
, instr
);
1629 case nir_intrinsic_memory_barrier_buffer
:
1630 case nir_intrinsic_memory_barrier_shared
:
1633 case nir_intrinsic_control_barrier
:
1634 schedule_barrier(ctx
);
1635 emit_control_barrier(ctx
);
1636 schedule_barrier(ctx
);
1640 fprintf(stderr
, "Unhandled intrinsic %s\n", nir_intrinsic_infos
[instr
->intrinsic
].name
);
1647 midgard_tex_format(enum glsl_sampler_dim dim
)
1650 case GLSL_SAMPLER_DIM_1D
:
1651 case GLSL_SAMPLER_DIM_BUF
:
1654 case GLSL_SAMPLER_DIM_2D
:
1655 case GLSL_SAMPLER_DIM_EXTERNAL
:
1656 case GLSL_SAMPLER_DIM_RECT
:
1659 case GLSL_SAMPLER_DIM_3D
:
1662 case GLSL_SAMPLER_DIM_CUBE
:
1663 return MALI_TEX_CUBE
;
1666 DBG("Unknown sampler dim type\n");
1672 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1676 pan_attach_constant_bias(
1677 compiler_context
*ctx
,
1679 midgard_texture_word
*word
)
1681 /* To attach as constant, it has to *be* constant */
1683 if (!nir_src_is_const(lod
))
1686 float f
= nir_src_as_float(lod
);
1688 /* Break into fixed-point */
1690 float lod_frac
= f
- lod_int
;
1692 /* Carry over negative fractions */
1693 if (lod_frac
< 0.0) {
1699 word
->bias
= float_to_ubyte(lod_frac
);
1700 word
->bias_int
= lod_int
;
1706 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1707 unsigned midgard_texop
)
1710 //assert (!instr->sampler);
1712 int texture_index
= instr
->texture_index
;
1713 int sampler_index
= texture_index
;
1715 nir_alu_type dest_base
= nir_alu_type_get_base_type(instr
->dest_type
);
1716 nir_alu_type dest_type
= dest_base
| nir_dest_bit_size(instr
->dest
);
1718 midgard_instruction ins
= {
1719 .type
= TAG_TEXTURE_4
,
1721 .dest
= nir_dest_index(&instr
->dest
),
1722 .src
= { ~0, ~0, ~0, ~0 },
1723 .dest_type
= dest_type
,
1724 .swizzle
= SWIZZLE_IDENTITY_4
,
1726 .op
= midgard_texop
,
1727 .format
= midgard_tex_format(instr
->sampler_dim
),
1728 .texture_handle
= texture_index
,
1729 .sampler_handle
= sampler_index
,
1730 .shadow
= instr
->is_shadow
,
1734 if (instr
->is_shadow
&& !instr
->is_new_style_shadow
)
1735 for (int i
= 0; i
< 4; ++i
)
1736 ins
.swizzle
[0][i
] = COMPONENT_X
;
1738 /* We may need a temporary for the coordinate */
1740 bool needs_temp_coord
=
1741 (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) ||
1742 (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) ||
1745 unsigned coords
= needs_temp_coord
? make_compiler_temp_reg(ctx
) : 0;
1747 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1748 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1749 unsigned nr_components
= nir_src_num_components(instr
->src
[i
].src
);
1750 unsigned sz
= nir_src_bit_size(instr
->src
[i
].src
);
1751 nir_alu_type T
= nir_tex_instr_src_type(instr
, i
) | sz
;
1753 switch (instr
->src
[i
].src_type
) {
1754 case nir_tex_src_coord
: {
1755 emit_explicit_constant(ctx
, index
, index
);
1757 unsigned coord_mask
= mask_of(instr
->coord_components
);
1759 bool flip_zw
= (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) && (coord_mask
& (1 << COMPONENT_Z
));
1762 coord_mask
^= ((1 << COMPONENT_Z
) | (1 << COMPONENT_W
));
1764 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1765 /* texelFetch is undefined on samplerCube */
1766 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1768 /* For cubemaps, we use a special ld/st op to
1769 * select the face and copy the xy into the
1770 * texture register */
1772 midgard_instruction ld
= m_ld_cubemap_coords(coords
, 0);
1774 ld
.src_types
[1] = T
;
1775 ld
.mask
= 0x3; /* xy */
1776 ld
.load_store
.arg_1
= 0x20;
1777 ld
.swizzle
[1][3] = COMPONENT_X
;
1778 emit_mir_instruction(ctx
, ld
);
1781 ins
.swizzle
[1][2] = instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1782 ins
.swizzle
[1][3] = COMPONENT_X
;
1783 } else if (needs_temp_coord
) {
1784 /* mov coord_temp, coords */
1785 midgard_instruction mov
= v_mov(index
, coords
);
1786 mov
.mask
= coord_mask
;
1789 mov
.swizzle
[1][COMPONENT_W
] = COMPONENT_Z
;
1791 emit_mir_instruction(ctx
, mov
);
1796 ins
.src
[1] = coords
;
1797 ins
.src_types
[1] = T
;
1799 /* Texelfetch coordinates uses all four elements
1800 * (xyz/index) regardless of texture dimensionality,
1801 * which means it's necessary to zero the unused
1802 * components to keep everything happy */
1804 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1805 /* mov index.zw, #0, or generalized */
1806 midgard_instruction mov
=
1807 v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), coords
);
1808 mov
.has_constants
= true;
1809 mov
.mask
= coord_mask
^ 0xF;
1810 emit_mir_instruction(ctx
, mov
);
1813 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1814 /* Array component in w but NIR wants it in z,
1815 * but if we have a temp coord we already fixed
1818 if (nr_components
== 3) {
1819 ins
.swizzle
[1][2] = COMPONENT_Z
;
1820 ins
.swizzle
[1][3] = needs_temp_coord
? COMPONENT_W
: COMPONENT_Z
;
1821 } else if (nr_components
== 2) {
1823 instr
->is_shadow
? COMPONENT_Z
: COMPONENT_X
;
1824 ins
.swizzle
[1][3] = COMPONENT_X
;
1826 unreachable("Invalid texture 2D components");
1829 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1831 ins
.swizzle
[1][2] = COMPONENT_Z
;
1832 ins
.swizzle
[1][3] = COMPONENT_W
;
1838 case nir_tex_src_bias
:
1839 case nir_tex_src_lod
: {
1840 /* Try as a constant if we can */
1842 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1843 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1846 ins
.texture
.lod_register
= true;
1848 ins
.src_types
[2] = T
;
1850 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1851 ins
.swizzle
[2][c
] = COMPONENT_X
;
1853 emit_explicit_constant(ctx
, index
, index
);
1858 case nir_tex_src_offset
: {
1859 ins
.texture
.offset_register
= true;
1861 ins
.src_types
[3] = T
;
1863 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
)
1864 ins
.swizzle
[3][c
] = (c
> COMPONENT_Z
) ? 0 : c
;
1866 emit_explicit_constant(ctx
, index
, index
);
1870 case nir_tex_src_comparator
: {
1871 unsigned comp
= COMPONENT_Z
;
1873 /* mov coord_temp.foo, coords */
1874 midgard_instruction mov
= v_mov(index
, coords
);
1875 mov
.mask
= 1 << comp
;
1877 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
)
1878 mov
.swizzle
[1][i
] = COMPONENT_X
;
1880 emit_mir_instruction(ctx
, mov
);
1885 fprintf(stderr
, "Unknown texture source type: %d\n", instr
->src
[i
].src_type
);
1891 emit_mir_instruction(ctx
, ins
);
1895 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1897 switch (instr
->op
) {
1900 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1903 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1906 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1909 emit_sysval_read(ctx
, &instr
->instr
, 4, 0);
1912 fprintf(stderr
, "Unhandled texture op: %d\n", instr
->op
);
1919 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1921 switch (instr
->type
) {
1922 case nir_jump_break
: {
1923 /* Emit a branch out of the loop */
1924 struct midgard_instruction br
= v_branch(false, false);
1925 br
.branch
.target_type
= TARGET_BREAK
;
1926 br
.branch
.target_break
= ctx
->current_loop_depth
;
1927 emit_mir_instruction(ctx
, br
);
1932 DBG("Unknown jump type %d\n", instr
->type
);
1938 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1940 switch (instr
->type
) {
1941 case nir_instr_type_load_const
:
1942 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1945 case nir_instr_type_intrinsic
:
1946 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1949 case nir_instr_type_alu
:
1950 emit_alu(ctx
, nir_instr_as_alu(instr
));
1953 case nir_instr_type_tex
:
1954 emit_tex(ctx
, nir_instr_as_tex(instr
));
1957 case nir_instr_type_jump
:
1958 emit_jump(ctx
, nir_instr_as_jump(instr
));
1961 case nir_instr_type_ssa_undef
:
1966 DBG("Unhandled instruction type\n");
1972 /* ALU instructions can inline or embed constants, which decreases register
1973 * pressure and saves space. */
1975 #define CONDITIONAL_ATTACH(idx) { \
1976 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->src[idx] + 1); \
1979 attach_constants(ctx, alu, entry, alu->src[idx] + 1); \
1980 alu->src[idx] = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
1985 inline_alu_constants(compiler_context
*ctx
, midgard_block
*block
)
1987 mir_foreach_instr_in_block(block
, alu
) {
1988 /* Other instructions cannot inline constants */
1989 if (alu
->type
!= TAG_ALU_4
) continue;
1990 if (alu
->compact_branch
) continue;
1992 /* If there is already a constant here, we can do nothing */
1993 if (alu
->has_constants
) continue;
1995 CONDITIONAL_ATTACH(0);
1997 if (!alu
->has_constants
) {
1998 CONDITIONAL_ATTACH(1)
1999 } else if (!alu
->inline_constant
) {
2000 /* Corner case: _two_ vec4 constants, for instance with a
2001 * csel. For this case, we can only use a constant
2002 * register for one, we'll have to emit a move for the
2005 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->src
[1] + 1);
2006 unsigned scratch
= make_compiler_temp(ctx
);
2009 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), scratch
);
2010 attach_constants(ctx
, &ins
, entry
, alu
->src
[1] + 1);
2012 /* Set the source */
2013 alu
->src
[1] = scratch
;
2015 /* Inject us -before- the last instruction which set r31 */
2016 mir_insert_instruction_before(ctx
, mir_prev_op(alu
), ins
);
2022 /* Midgard supports two types of constants, embedded constants (128-bit) and
2023 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2024 * constants can be demoted to inline constants, for space savings and
2025 * sometimes a performance boost */
2028 embedded_to_inline_constant(compiler_context
*ctx
, midgard_block
*block
)
2030 mir_foreach_instr_in_block(block
, ins
) {
2031 if (!ins
->has_constants
) continue;
2032 if (ins
->has_inline_constant
) continue;
2034 /* Blend constants must not be inlined by definition */
2035 if (ins
->has_blend_constant
) continue;
2037 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
2038 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
2039 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
2041 if (!(is_16
|| is_32
))
2044 /* src1 cannot be an inline constant due to encoding
2045 * restrictions. So, if possible we try to flip the arguments
2048 int op
= ins
->alu
.op
;
2050 if (ins
->src
[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
) &&
2051 alu_opcode_props
[op
].props
& OP_COMMUTES
) {
2055 if (ins
->src
[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2056 /* Extract the source information */
2058 midgard_vector_alu_src
*src
;
2059 int q
= ins
->alu
.src2
;
2060 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2063 /* Component is from the swizzle. Take a nonzero component */
2065 unsigned first_comp
= ffs(ins
->mask
) - 1;
2066 unsigned component
= ins
->swizzle
[1][first_comp
];
2068 /* Scale constant appropriately, if we can legally */
2069 uint16_t scaled_constant
= 0;
2072 scaled_constant
= ins
->constants
.u16
[component
];
2073 } else if (midgard_is_integer_op(op
)) {
2074 scaled_constant
= ins
->constants
.u32
[component
];
2076 /* Constant overflow after resize */
2077 if (scaled_constant
!= ins
->constants
.u32
[component
])
2080 float original
= ins
->constants
.f32
[component
];
2081 scaled_constant
= _mesa_float_to_half(original
);
2083 /* Check for loss of precision. If this is
2084 * mediump, we don't care, but for a highp
2085 * shader, we need to pay attention. NIR
2086 * doesn't yet tell us which mode we're in!
2087 * Practically this prevents most constants
2088 * from being inlined, sadly. */
2090 float fp32
= _mesa_half_to_float(scaled_constant
);
2092 if (fp32
!= original
)
2096 /* We don't know how to handle these with a constant */
2098 if (mir_nontrivial_source2_mod_simple(ins
) || src
->rep_low
|| src
->rep_high
) {
2099 DBG("Bailing inline constant...\n");
2103 /* Make sure that the constant is not itself a vector
2104 * by checking if all accessed values are the same. */
2106 const midgard_constants
*cons
= &ins
->constants
;
2107 uint32_t value
= is_16
? cons
->u16
[component
] : cons
->u32
[component
];
2109 bool is_vector
= false;
2110 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
2112 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
2113 /* We only care if this component is actually used */
2114 if (!(mask
& (1 << c
)))
2117 uint32_t test
= is_16
?
2118 cons
->u16
[ins
->swizzle
[1][c
]] :
2119 cons
->u32
[ins
->swizzle
[1][c
]];
2121 if (test
!= value
) {
2130 /* Get rid of the embedded constant */
2131 ins
->has_constants
= false;
2133 ins
->has_inline_constant
= true;
2134 ins
->inline_constant
= scaled_constant
;
2139 /* Dead code elimination for branches at the end of a block - only one branch
2140 * per block is legal semantically */
2143 midgard_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2145 bool branched
= false;
2147 mir_foreach_instr_in_block_safe(block
, ins
) {
2148 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2151 mir_remove_instruction(ins
);
2158 emit_fragment_epilogue(compiler_context
*ctx
, unsigned rt
)
2160 /* Loop to ourselves */
2161 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2162 struct midgard_instruction ins
= v_branch(false, false);
2163 ins
.writeout
= true;
2164 ins
.writeout_depth
= br
->writeout_depth
;
2165 ins
.writeout_stencil
= br
->writeout_stencil
;
2166 ins
.branch
.target_block
= ctx
->block_count
- 1;
2167 ins
.constants
.u32
[0] = br
->constants
.u32
[0];
2168 emit_mir_instruction(ctx
, ins
);
2170 ctx
->current_block
->epilogue
= true;
2171 schedule_barrier(ctx
);
2172 return ins
.branch
.target_block
;
2175 static midgard_block
*
2176 emit_block(compiler_context
*ctx
, nir_block
*block
)
2178 midgard_block
*this_block
= ctx
->after_block
;
2179 ctx
->after_block
= NULL
;
2182 this_block
= create_empty_block(ctx
);
2184 list_addtail(&this_block
->base
.link
, &ctx
->blocks
);
2186 this_block
->scheduled
= false;
2189 /* Set up current block */
2190 list_inithead(&this_block
->base
.instructions
);
2191 ctx
->current_block
= this_block
;
2193 nir_foreach_instr(instr
, block
) {
2194 emit_instr(ctx
, instr
);
2195 ++ctx
->instruction_count
;
2201 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2204 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2206 midgard_block
*before_block
= ctx
->current_block
;
2208 /* Speculatively emit the branch, but we can't fill it in until later */
2209 EMIT(branch
, true, true);
2210 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2211 then_branch
->src
[0] = nir_src_index(ctx
, &nif
->condition
);
2212 then_branch
->src_types
[0] = nir_type_uint32
;
2214 /* Emit the two subblocks. */
2215 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2216 midgard_block
*end_then_block
= ctx
->current_block
;
2218 /* Emit a jump from the end of the then block to the end of the else */
2219 EMIT(branch
, false, false);
2220 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2222 /* Emit second block, and check if it's empty */
2224 int else_idx
= ctx
->block_count
;
2225 int count_in
= ctx
->instruction_count
;
2226 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2227 midgard_block
*end_else_block
= ctx
->current_block
;
2228 int after_else_idx
= ctx
->block_count
;
2230 /* Now that we have the subblocks emitted, fix up the branches */
2235 if (ctx
->instruction_count
== count_in
) {
2236 /* The else block is empty, so don't emit an exit jump */
2237 mir_remove_instruction(then_exit
);
2238 then_branch
->branch
.target_block
= after_else_idx
;
2240 then_branch
->branch
.target_block
= else_idx
;
2241 then_exit
->branch
.target_block
= after_else_idx
;
2244 /* Wire up the successors */
2246 ctx
->after_block
= create_empty_block(ctx
);
2248 pan_block_add_successor(&before_block
->base
, &then_block
->base
);
2249 pan_block_add_successor(&before_block
->base
, &else_block
->base
);
2251 pan_block_add_successor(&end_then_block
->base
, &ctx
->after_block
->base
);
2252 pan_block_add_successor(&end_else_block
->base
, &ctx
->after_block
->base
);
2256 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2258 /* Remember where we are */
2259 midgard_block
*start_block
= ctx
->current_block
;
2261 /* Allocate a loop number, growing the current inner loop depth */
2262 int loop_idx
= ++ctx
->current_loop_depth
;
2264 /* Get index from before the body so we can loop back later */
2265 int start_idx
= ctx
->block_count
;
2267 /* Emit the body itself */
2268 midgard_block
*loop_block
= emit_cf_list(ctx
, &nloop
->body
);
2270 /* Branch back to loop back */
2271 struct midgard_instruction br_back
= v_branch(false, false);
2272 br_back
.branch
.target_block
= start_idx
;
2273 emit_mir_instruction(ctx
, br_back
);
2275 /* Mark down that branch in the graph. */
2276 pan_block_add_successor(&start_block
->base
, &loop_block
->base
);
2277 pan_block_add_successor(&ctx
->current_block
->base
, &loop_block
->base
);
2279 /* Find the index of the block about to follow us (note: we don't add
2280 * one; blocks are 0-indexed so we get a fencepost problem) */
2281 int break_block_idx
= ctx
->block_count
;
2283 /* Fix up the break statements we emitted to point to the right place,
2284 * now that we can allocate a block number for them */
2285 ctx
->after_block
= create_empty_block(ctx
);
2287 mir_foreach_block_from(ctx
, start_block
, _block
) {
2288 mir_foreach_instr_in_block(((midgard_block
*) _block
), ins
) {
2289 if (ins
->type
!= TAG_ALU_4
) continue;
2290 if (!ins
->compact_branch
) continue;
2292 /* We found a branch -- check the type to see if we need to do anything */
2293 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2295 /* It's a break! Check if it's our break */
2296 if (ins
->branch
.target_break
!= loop_idx
) continue;
2298 /* Okay, cool, we're breaking out of this loop.
2299 * Rewrite from a break to a goto */
2301 ins
->branch
.target_type
= TARGET_GOTO
;
2302 ins
->branch
.target_block
= break_block_idx
;
2304 pan_block_add_successor(_block
, &ctx
->after_block
->base
);
2308 /* Now that we've finished emitting the loop, free up the depth again
2309 * so we play nice with recursion amid nested loops */
2310 --ctx
->current_loop_depth
;
2312 /* Dump loop stats */
2316 static midgard_block
*
2317 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2319 midgard_block
*start_block
= NULL
;
2321 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2322 switch (node
->type
) {
2323 case nir_cf_node_block
: {
2324 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2327 start_block
= block
;
2332 case nir_cf_node_if
:
2333 emit_if(ctx
, nir_cf_node_as_if(node
));
2336 case nir_cf_node_loop
:
2337 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2340 case nir_cf_node_function
:
2349 /* Due to lookahead, we need to report the first tag executed in the command
2350 * stream and in branch targets. An initial block might be empty, so iterate
2351 * until we find one that 'works' */
2354 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2356 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2358 mir_foreach_block_from(ctx
, initial_block
, _v
) {
2359 midgard_block
*v
= (midgard_block
*) _v
;
2360 if (v
->quadword_count
) {
2361 midgard_bundle
*initial_bundle
=
2362 util_dynarray_element(&v
->bundles
, midgard_bundle
, 0);
2364 return initial_bundle
->tag
;
2368 /* Default to a tag 1 which will break from the shader, in case we jump
2369 * to the exit block (i.e. `return` in a compute shader) */
2374 /* For each fragment writeout instruction, generate a writeout loop to
2375 * associate with it */
2378 mir_add_writeout_loops(compiler_context
*ctx
)
2380 for (unsigned rt
= 0; rt
< ARRAY_SIZE(ctx
->writeout_branch
); ++rt
) {
2381 midgard_instruction
*br
= ctx
->writeout_branch
[rt
];
2384 unsigned popped
= br
->branch
.target_block
;
2385 pan_block_add_successor(&(mir_get_block(ctx
, popped
- 1)->base
), &ctx
->current_block
->base
);
2386 br
->branch
.target_block
= emit_fragment_epilogue(ctx
, rt
);
2387 br
->branch
.target_type
= TARGET_GOTO
;
2389 /* If we have more RTs, we'll need to restore back after our
2390 * loop terminates */
2392 if ((rt
+ 1) < ARRAY_SIZE(ctx
->writeout_branch
) && ctx
->writeout_branch
[rt
+ 1]) {
2393 midgard_instruction uncond
= v_branch(false, false);
2394 uncond
.branch
.target_block
= popped
;
2395 uncond
.branch
.target_type
= TARGET_GOTO
;
2396 emit_mir_instruction(ctx
, uncond
);
2397 pan_block_add_successor(&ctx
->current_block
->base
, &(mir_get_block(ctx
, popped
)->base
));
2398 schedule_barrier(ctx
);
2400 /* We're last, so we can terminate here */
2401 br
->last_writeout
= true;
2407 midgard_compile_shader_nir(nir_shader
*nir
, panfrost_program
*program
, bool is_blend
, unsigned blend_rt
, unsigned gpu_id
, bool shaderdb
)
2409 struct util_dynarray
*compiled
= &program
->compiled
;
2411 midgard_debug
= debug_get_option_midgard_debug();
2413 /* TODO: Bound against what? */
2414 compiler_context
*ctx
= rzalloc(NULL
, compiler_context
);
2417 ctx
->stage
= nir
->info
.stage
;
2418 ctx
->is_blend
= is_blend
;
2419 ctx
->alpha_ref
= program
->alpha_ref
;
2420 ctx
->blend_rt
= MIDGARD_COLOR_RT0
+ blend_rt
;
2421 ctx
->quirks
= midgard_get_quirks(gpu_id
);
2423 /* Start off with a safe cutoff, allowing usage of all 16 work
2424 * registers. Later, we'll promote uniform reads to uniform registers
2425 * if we determine it is beneficial to do so */
2426 ctx
->uniform_cutoff
= 8;
2428 /* Initialize at a global (not block) level hash tables */
2430 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2431 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2433 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2434 * (so we don't accidentally duplicate the epilogue since mesa/st has
2435 * messed with our I/O quite a bit already) */
2437 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2439 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2440 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2441 NIR_PASS_V(nir
, nir_lower_point_size
, 1.0, 1024.0);
2444 NIR_PASS_V(nir
, nir_lower_var_copies
);
2445 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2446 NIR_PASS_V(nir
, nir_split_var_copies
);
2447 NIR_PASS_V(nir
, nir_lower_var_copies
);
2448 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2449 NIR_PASS_V(nir
, nir_lower_var_copies
);
2450 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2452 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2453 NIR_PASS_V(nir
, nir_lower_ssbo
);
2454 NIR_PASS_V(nir
, midgard_nir_lower_zs_store
);
2456 /* Optimisation passes */
2458 optimise_nir(nir
, ctx
->quirks
);
2460 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2461 nir_print_shader(nir
, stdout
);
2464 /* Assign sysvals and counts, now that we're sure
2465 * (post-optimisation) */
2467 panfrost_nir_assign_sysvals(&ctx
->sysvals
, nir
);
2468 program
->sysval_count
= ctx
->sysvals
.sysval_count
;
2469 memcpy(program
->sysvals
, ctx
->sysvals
.sysvals
, sizeof(ctx
->sysvals
.sysvals
[0]) * ctx
->sysvals
.sysval_count
);
2471 nir_foreach_function(func
, nir
) {
2475 list_inithead(&ctx
->blocks
);
2476 ctx
->block_count
= 0;
2479 emit_cf_list(ctx
, &func
->impl
->body
);
2480 break; /* TODO: Multi-function shaders */
2483 util_dynarray_init(compiled
, NULL
);
2485 /* Per-block lowering before opts */
2487 mir_foreach_block(ctx
, _block
) {
2488 midgard_block
*block
= (midgard_block
*) _block
;
2489 inline_alu_constants(ctx
, block
);
2490 midgard_opt_promote_fmov(ctx
, block
);
2491 embedded_to_inline_constant(ctx
, block
);
2493 /* MIR-level optimizations */
2495 bool progress
= false;
2500 mir_foreach_block(ctx
, _block
) {
2501 midgard_block
*block
= (midgard_block
*) _block
;
2502 progress
|= midgard_opt_copy_prop(ctx
, block
);
2503 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2504 progress
|= midgard_opt_combine_projection(ctx
, block
);
2505 progress
|= midgard_opt_varying_projection(ctx
, block
);
2506 progress
|= midgard_opt_not_propagate(ctx
, block
);
2507 progress
|= midgard_opt_fuse_src_invert(ctx
, block
);
2508 progress
|= midgard_opt_fuse_dest_invert(ctx
, block
);
2509 progress
|= midgard_opt_csel_invert(ctx
, block
);
2510 progress
|= midgard_opt_drop_cmp_invert(ctx
, block
);
2511 progress
|= midgard_opt_invert_branch(ctx
, block
);
2515 mir_foreach_block(ctx
, _block
) {
2516 midgard_block
*block
= (midgard_block
*) _block
;
2517 midgard_lower_invert(ctx
, block
);
2518 midgard_lower_derivatives(ctx
, block
);
2521 /* Nested control-flow can result in dead branches at the end of the
2522 * block. This messes with our analysis and is just dead code, so cull
2524 mir_foreach_block(ctx
, _block
) {
2525 midgard_block
*block
= (midgard_block
*) _block
;
2526 midgard_cull_dead_branch(ctx
, block
);
2529 /* Ensure we were lowered */
2530 mir_foreach_instr_global(ctx
, ins
) {
2531 assert(!ins
->invert
);
2534 if (ctx
->stage
== MESA_SHADER_FRAGMENT
)
2535 mir_add_writeout_loops(ctx
);
2537 /* Analyze now that the code is known but before scheduling creates
2538 * pipeline registers which are harder to track */
2539 mir_analyze_helper_terminate(ctx
);
2540 mir_analyze_helper_requirements(ctx
);
2543 midgard_schedule_program(ctx
);
2546 /* Now that all the bundles are scheduled and we can calculate block
2547 * sizes, emit actual branch instructions rather than placeholders */
2549 int br_block_idx
= 0;
2551 mir_foreach_block(ctx
, _block
) {
2552 midgard_block
*block
= (midgard_block
*) _block
;
2553 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2554 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2555 midgard_instruction
*ins
= bundle
->instructions
[c
];
2557 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2559 /* Parse some basic branch info */
2560 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2561 bool is_conditional
= ins
->branch
.conditional
;
2562 bool is_inverted
= ins
->branch
.invert_conditional
;
2563 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2564 bool is_writeout
= ins
->writeout
;
2566 /* Determine the block we're jumping to */
2567 int target_number
= ins
->branch
.target_block
;
2569 /* Report the destination tag */
2570 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2572 /* Count up the number of quadwords we're
2573 * jumping over = number of quadwords until
2574 * (br_block_idx, target_number) */
2576 int quadword_offset
= 0;
2580 } else if (target_number
> br_block_idx
) {
2583 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2584 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2587 quadword_offset
+= blk
->quadword_count
;
2590 /* Jump backwards */
2592 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2593 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2596 quadword_offset
-= blk
->quadword_count
;
2600 /* Unconditional extended branches (far jumps)
2601 * have issues, so we always use a conditional
2602 * branch, setting the condition to always for
2603 * unconditional. For compact unconditional
2604 * branches, cond isn't used so it doesn't
2605 * matter what we pick. */
2607 midgard_condition cond
=
2608 !is_conditional
? midgard_condition_always
:
2609 is_inverted
? midgard_condition_false
:
2610 midgard_condition_true
;
2612 midgard_jmp_writeout_op op
=
2613 is_discard
? midgard_jmp_writeout_op_discard
:
2614 is_writeout
? midgard_jmp_writeout_op_writeout
:
2615 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2616 midgard_jmp_writeout_op_branch_cond
;
2619 midgard_branch_extended branch
=
2620 midgard_create_branch_extended(
2625 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2626 } else if (is_conditional
|| is_discard
) {
2627 midgard_branch_cond branch
= {
2629 .dest_tag
= dest_tag
,
2630 .offset
= quadword_offset
,
2634 assert(branch
.offset
== quadword_offset
);
2636 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2638 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2640 midgard_branch_uncond branch
= {
2642 .dest_tag
= dest_tag
,
2643 .offset
= quadword_offset
,
2647 assert(branch
.offset
== quadword_offset
);
2649 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2657 /* Emit flat binary from the instruction arrays. Iterate each block in
2658 * sequence. Save instruction boundaries such that lookahead tags can
2659 * be assigned easily */
2661 /* Cache _all_ bundles in source order for lookahead across failed branches */
2663 int bundle_count
= 0;
2664 mir_foreach_block(ctx
, _block
) {
2665 midgard_block
*block
= (midgard_block
*) _block
;
2666 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2668 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2670 mir_foreach_block(ctx
, _block
) {
2671 midgard_block
*block
= (midgard_block
*) _block
;
2672 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2673 source_order_bundles
[bundle_idx
++] = bundle
;
2677 int current_bundle
= 0;
2679 /* Midgard prefetches instruction types, so during emission we
2680 * need to lookahead. Unless this is the last instruction, in
2681 * which we return 1. */
2683 mir_foreach_block(ctx
, _block
) {
2684 midgard_block
*block
= (midgard_block
*) _block
;
2685 mir_foreach_bundle_in_block(block
, bundle
) {
2688 if (!bundle
->last_writeout
&& (current_bundle
+ 1 < bundle_count
))
2689 lookahead
= source_order_bundles
[current_bundle
+ 1]->tag
;
2691 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2695 /* TODO: Free deeper */
2696 //util_dynarray_fini(&block->instructions);
2699 free(source_order_bundles
);
2701 /* Report the very first tag executed */
2702 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2704 /* Deal with off-by-one related to the fencepost problem */
2705 program
->work_register_count
= ctx
->work_registers
+ 1;
2706 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2708 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2709 program
->tls_size
= ctx
->tls_size
;
2711 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2712 disassemble_midgard(stdout
, program
->compiled
.data
, program
->compiled
.size
, gpu_id
, ctx
->stage
);
2714 if (midgard_debug
& MIDGARD_DBG_SHADERDB
|| shaderdb
) {
2715 unsigned nr_bundles
= 0, nr_ins
= 0;
2717 /* Count instructions and bundles */
2719 mir_foreach_block(ctx
, _block
) {
2720 midgard_block
*block
= (midgard_block
*) _block
;
2721 nr_bundles
+= util_dynarray_num_elements(
2722 &block
->bundles
, midgard_bundle
);
2724 mir_foreach_bundle_in_block(block
, bun
)
2725 nr_ins
+= bun
->instruction_count
;
2728 /* Calculate thread count. There are certain cutoffs by
2729 * register count for thread count */
2731 unsigned nr_registers
= program
->work_register_count
;
2733 unsigned nr_threads
=
2734 (nr_registers
<= 4) ? 4 :
2735 (nr_registers
<= 8) ? 2 :
2740 fprintf(stderr
, "shader%d - %s shader: "
2741 "%u inst, %u bundles, %u quadwords, "
2742 "%u registers, %u threads, %u loops, "
2743 "%u:%u spills:fills\n",
2745 gl_shader_stage_name(ctx
->stage
),
2746 nr_ins
, nr_bundles
, ctx
->quadword_count
,
2747 nr_registers
, nr_threads
,
2749 ctx
->spills
, ctx
->fills
);