2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 #include <sys/types.h>
33 #include "main/mtypes.h"
34 #include "compiler/glsl/glsl_to_nir.h"
35 #include "compiler/nir_types.h"
36 #include "main/imports.h"
37 #include "compiler/nir/nir_builder.h"
38 #include "util/half_float.h"
39 #include "util/u_math.h"
40 #include "util/u_debug.h"
41 #include "util/u_dynarray.h"
42 #include "util/list.h"
43 #include "main/mtypes.h"
46 #include "midgard_nir.h"
47 #include "midgard_compile.h"
48 #include "midgard_ops.h"
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)
73 midgard_is_branch_unit(unsigned unit
)
75 return (unit
== ALU_ENAB_BRANCH
) || (unit
== ALU_ENAB_BR_COMPACT
);
79 midgard_block_add_successor(midgard_block
*block
, midgard_block
*successor
)
81 block
->successors
[block
->nr_successors
++] = successor
;
82 assert(block
->nr_successors
<= ARRAY_SIZE(block
->successors
));
85 /* Helpers to generate midgard_instruction's using macro magic, since every
86 * driver seems to do it that way */
88 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
90 #define M_LOAD_STORE(name, rname, uname) \
91 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
92 midgard_instruction i = { \
93 .type = TAG_LOAD_STORE_4, \
101 .op = midgard_op_##name, \
102 .swizzle = SWIZZLE_XYZW, \
110 #define M_LOAD(name) M_LOAD_STORE(name, dest, src0)
111 #define M_STORE(name) M_LOAD_STORE(name, src0, dest)
113 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
114 * the corresponding Midgard source */
116 static midgard_vector_alu_src
117 vector_alu_modifiers(nir_alu_src
*src
, bool is_int
, unsigned broadcast_count
,
118 bool half
, bool sext
)
120 if (!src
) return blank_alu_src
;
122 /* Figure out how many components there are so we can adjust the
123 * swizzle. Specifically we want to broadcast the last channel so
124 * things like ball2/3 work
127 if (broadcast_count
) {
128 uint8_t last_component
= src
->swizzle
[broadcast_count
- 1];
130 for (unsigned c
= broadcast_count
; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
131 src
->swizzle
[c
] = last_component
;
135 midgard_vector_alu_src alu_src
= {
139 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
)
143 alu_src
.mod
= midgard_int_normal
;
145 /* Sign/zero-extend if needed */
149 midgard_int_sign_extend
150 : midgard_int_zero_extend
;
153 /* These should have been lowered away */
154 assert(!(src
->abs
|| src
->negate
));
156 alu_src
.mod
= (src
->abs
<< 0) | (src
->negate
<< 1);
162 /* load/store instructions have both 32-bit and 16-bit variants, depending on
163 * whether we are using vectors composed of highp or mediump. At the moment, we
164 * don't support half-floats -- this requires changes in other parts of the
165 * compiler -- therefore the 16-bit versions are commented out. */
167 //M_LOAD(ld_attr_16);
169 //M_LOAD(ld_vary_16);
171 //M_LOAD(ld_uniform_16);
172 M_LOAD(ld_uniform_32
);
173 M_LOAD(ld_color_buffer_8
);
174 //M_STORE(st_vary_16);
176 M_LOAD(st_cubemap_coords
);
178 static midgard_instruction
179 v_alu_br_compact_cond(midgard_jmp_writeout_op op
, unsigned tag
, signed offset
, unsigned cond
)
181 midgard_branch_cond branch
= {
189 memcpy(&compact
, &branch
, sizeof(branch
));
191 midgard_instruction ins
= {
193 .unit
= ALU_ENAB_BR_COMPACT
,
194 .prepacked_branch
= true,
195 .compact_branch
= true,
196 .br_compact
= compact
199 if (op
== midgard_jmp_writeout_op_writeout
)
205 static midgard_instruction
206 v_branch(bool conditional
, bool invert
)
208 midgard_instruction ins
= {
210 .unit
= ALU_ENAB_BRANCH
,
211 .compact_branch
= true,
213 .conditional
= conditional
,
214 .invert_conditional
= invert
221 static midgard_branch_extended
222 midgard_create_branch_extended( midgard_condition cond
,
223 midgard_jmp_writeout_op op
,
225 signed quadword_offset
)
227 /* For unclear reasons, the condition code is repeated 8 times */
228 uint16_t duplicated_cond
=
238 midgard_branch_extended branch
= {
240 .dest_tag
= dest_tag
,
241 .offset
= quadword_offset
,
242 .cond
= duplicated_cond
249 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
251 ins
->has_constants
= true;
252 memcpy(&ins
->constants
, constants
, 16);
256 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
258 return glsl_count_attribute_slots(type
, false);
261 /* Lower fdot2 to a vector multiplication followed by channel addition */
263 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
265 if (alu
->op
!= nir_op_fdot2
)
268 b
->cursor
= nir_before_instr(&alu
->instr
);
270 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
271 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
273 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
275 nir_ssa_def
*sum
= nir_fadd(b
,
276 nir_channel(b
, product
, 0),
277 nir_channel(b
, product
, 1));
279 /* Replace the fdot2 with this sum */
280 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
284 midgard_nir_sysval_for_intrinsic(nir_intrinsic_instr
*instr
)
286 switch (instr
->intrinsic
) {
287 case nir_intrinsic_load_viewport_scale
:
288 return PAN_SYSVAL_VIEWPORT_SCALE
;
289 case nir_intrinsic_load_viewport_offset
:
290 return PAN_SYSVAL_VIEWPORT_OFFSET
;
296 static int sysval_for_instr(compiler_context
*ctx
, nir_instr
*instr
,
299 nir_intrinsic_instr
*intr
;
300 nir_dest
*dst
= NULL
;
304 switch (instr
->type
) {
305 case nir_instr_type_intrinsic
:
306 intr
= nir_instr_as_intrinsic(instr
);
307 sysval
= midgard_nir_sysval_for_intrinsic(intr
);
310 case nir_instr_type_tex
:
311 tex
= nir_instr_as_tex(instr
);
312 if (tex
->op
!= nir_texop_txs
)
315 sysval
= PAN_SYSVAL(TEXTURE_SIZE
,
316 PAN_TXS_SYSVAL_ID(tex
->texture_index
,
317 nir_tex_instr_dest_size(tex
) -
318 (tex
->is_array
? 1 : 0),
327 *dest
= nir_dest_index(ctx
, dst
);
333 midgard_nir_assign_sysval_body(compiler_context
*ctx
, nir_instr
*instr
)
337 sysval
= sysval_for_instr(ctx
, instr
, NULL
);
341 /* We have a sysval load; check if it's already been assigned */
343 if (_mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
))
346 /* It hasn't -- so assign it now! */
348 unsigned id
= ctx
->sysval_count
++;
349 _mesa_hash_table_u64_insert(ctx
->sysval_to_id
, sysval
, (void *) ((uintptr_t) id
+ 1));
350 ctx
->sysvals
[id
] = sysval
;
354 midgard_nir_assign_sysvals(compiler_context
*ctx
, nir_shader
*shader
)
356 ctx
->sysval_count
= 0;
358 nir_foreach_function(function
, shader
) {
359 if (!function
->impl
) continue;
361 nir_foreach_block(block
, function
->impl
) {
362 nir_foreach_instr_safe(instr
, block
) {
363 midgard_nir_assign_sysval_body(ctx
, instr
);
370 midgard_nir_lower_fdot2(nir_shader
*shader
)
372 bool progress
= false;
374 nir_foreach_function(function
, shader
) {
375 if (!function
->impl
) continue;
378 nir_builder
*b
= &_b
;
379 nir_builder_init(b
, function
->impl
);
381 nir_foreach_block(block
, function
->impl
) {
382 nir_foreach_instr_safe(instr
, block
) {
383 if (instr
->type
!= nir_instr_type_alu
) continue;
385 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
386 midgard_nir_lower_fdot2_body(b
, alu
);
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
)
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
, midgard_nir_lower_fdot2
);
412 NIR_PASS(progress
, nir
, nir_lower_idiv
);
414 nir_lower_tex_options lower_tex_1st_pass_options
= {
419 nir_lower_tex_options lower_tex_2nd_pass_options
= {
420 .lower_txs_lod
= true,
423 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_1st_pass_options
);
424 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_2nd_pass_options
);
429 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
430 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
432 NIR_PASS(progress
, nir
, nir_copy_prop
);
433 NIR_PASS(progress
, nir
, nir_opt_dce
);
434 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
435 NIR_PASS(progress
, nir
, nir_opt_cse
);
436 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
437 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
438 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
440 if (lower_flrp
!= 0) {
441 bool lower_flrp_progress
= false;
442 NIR_PASS(lower_flrp_progress
,
446 false /* always_precise */,
447 nir
->options
->lower_ffma
);
448 if (lower_flrp_progress
) {
449 NIR_PASS(progress
, nir
,
450 nir_opt_constant_folding
);
454 /* Nothing should rematerialize any flrps, so we only
455 * need to do this lowering once.
460 NIR_PASS(progress
, nir
, nir_opt_undef
);
461 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
463 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
466 nir_var_function_temp
);
468 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
471 /* Must be run at the end to prevent creation of fsin/fcos ops */
472 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
477 NIR_PASS(progress
, nir
, nir_opt_dce
);
478 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
479 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
480 NIR_PASS(progress
, nir
, nir_copy_prop
);
483 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
485 /* We implement booleans as 32-bit 0/~0 */
486 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
488 /* Now that booleans are lowered, we can run out late opts */
489 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
491 /* Lower mods for float ops only. Integer ops don't support modifiers
492 * (saturate doesn't make sense on integers, neg/abs require dedicated
495 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
496 NIR_PASS(progress
, nir
, nir_copy_prop
);
497 NIR_PASS(progress
, nir
, nir_opt_dce
);
499 /* Take us out of SSA */
500 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
501 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
503 /* We are a vector architecture; write combine where possible */
504 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
505 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
507 NIR_PASS(progress
, nir
, nir_opt_dce
);
510 /* Do not actually emit a load; instead, cache the constant for inlining */
513 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
515 nir_ssa_def def
= instr
->def
;
517 float *v
= rzalloc_array(NULL
, float, 4);
518 nir_const_load_to_arr(v
, instr
, f32
);
520 /* Shifted for SSA, +1 for off-by-one */
521 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (def
.index
<< 1) + 1, v
);
524 /* Normally constants are embedded implicitly, but for I/O and such we have to
525 * explicitly emit a move with the constant source */
528 emit_explicit_constant(compiler_context
*ctx
, unsigned node
, unsigned to
)
530 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, node
+ 1);
532 if (constant_value
) {
533 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, to
);
534 attach_constants(ctx
, &ins
, constant_value
, node
+ 1);
535 emit_mir_instruction(ctx
, ins
);
540 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
542 unsigned comp
= src
->swizzle
[0];
544 for (unsigned c
= 1; c
< nr_components
; ++c
) {
545 if (src
->swizzle
[c
] != comp
)
552 /* Midgard puts scalar conditionals in r31.w; move an arbitrary source (the
553 * output of a conditional test) into that register */
556 emit_condition(compiler_context
*ctx
, nir_src
*src
, bool for_branch
, unsigned component
)
558 int condition
= nir_src_index(ctx
, src
);
560 /* Source to swizzle the desired component into w */
562 const midgard_vector_alu_src alu_src
= {
563 .swizzle
= SWIZZLE(component
, component
, component
, component
),
566 /* There is no boolean move instruction. Instead, we simulate a move by
567 * ANDing the condition with itself to get it into r31.w */
569 midgard_instruction ins
= {
572 /* We need to set the conditional as close as possible */
573 .precede_break
= true,
574 .unit
= for_branch
? UNIT_SMUL
: UNIT_SADD
,
575 .mask
= 1 << COMPONENT_W
,
580 .dest
= SSA_FIXED_REGISTER(31),
584 .op
= midgard_alu_op_iand
,
585 .outmod
= midgard_outmod_int_wrap
,
586 .reg_mode
= midgard_reg_mode_32
,
587 .dest_override
= midgard_dest_override_none
,
588 .src1
= vector_alu_srco_unsigned(alu_src
),
589 .src2
= vector_alu_srco_unsigned(alu_src
)
593 emit_mir_instruction(ctx
, ins
);
596 /* Or, for mixed conditions (with csel_v), here's a vector version using all of
600 emit_condition_mixed(compiler_context
*ctx
, nir_alu_src
*src
, unsigned nr_comp
)
602 int condition
= nir_src_index(ctx
, &src
->src
);
604 /* Source to swizzle the desired component into w */
606 const midgard_vector_alu_src alu_src
= {
607 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
),
610 /* There is no boolean move instruction. Instead, we simulate a move by
611 * ANDing the condition with itself to get it into r31.w */
613 midgard_instruction ins
= {
615 .precede_break
= true,
616 .mask
= mask_of(nr_comp
),
620 .dest
= SSA_FIXED_REGISTER(31),
623 .op
= midgard_alu_op_iand
,
624 .outmod
= midgard_outmod_int_wrap
,
625 .reg_mode
= midgard_reg_mode_32
,
626 .dest_override
= midgard_dest_override_none
,
627 .src1
= vector_alu_srco_unsigned(alu_src
),
628 .src2
= vector_alu_srco_unsigned(alu_src
)
632 emit_mir_instruction(ctx
, ins
);
637 /* Likewise, indirect offsets are put in r27.w. TODO: Allow componentwise
638 * pinning to eliminate this move in all known cases */
641 emit_indirect_offset(compiler_context
*ctx
, nir_src
*src
)
643 int offset
= nir_src_index(ctx
, src
);
645 midgard_instruction ins
= {
647 .mask
= 1 << COMPONENT_W
,
649 .src0
= SSA_UNUSED_1
,
651 .dest
= SSA_FIXED_REGISTER(REGISTER_OFFSET
),
654 .op
= midgard_alu_op_imov
,
655 .outmod
= midgard_outmod_int_wrap
,
656 .reg_mode
= midgard_reg_mode_32
,
657 .dest_override
= midgard_dest_override_none
,
658 .src1
= vector_alu_srco_unsigned(zero_alu_src
),
659 .src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
)
663 emit_mir_instruction(ctx
, ins
);
666 #define ALU_CASE(nir, _op) \
668 op = midgard_alu_op_##_op; \
669 assert(src_bitsize == dst_bitsize); \
672 #define ALU_CASE_BCAST(nir, _op, count) \
674 op = midgard_alu_op_##_op; \
675 broadcast_swizzle = count; \
676 assert(src_bitsize == dst_bitsize); \
679 nir_is_fzero_constant(nir_src src
)
681 if (!nir_src_is_const(src
))
684 for (unsigned c
= 0; c
< nir_src_num_components(src
); ++c
) {
685 if (nir_src_comp_as_float(src
, c
) != 0.0)
692 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
693 * special treatment override this anyway. */
695 static midgard_reg_mode
696 reg_mode_for_nir(nir_alu_instr
*instr
)
698 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
700 switch (src_bitsize
) {
702 return midgard_reg_mode_8
;
704 return midgard_reg_mode_16
;
706 return midgard_reg_mode_32
;
708 return midgard_reg_mode_64
;
710 unreachable("Invalid bit size");
715 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
717 /* Derivatives end up emitted on the texture pipe, not the ALUs. This
718 * is handled elsewhere */
720 if (instr
->op
== nir_op_fddx
|| instr
->op
== nir_op_fddy
) {
721 midgard_emit_derivatives(ctx
, instr
);
725 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
727 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
.dest
);
728 unsigned nr_components
= nir_dest_num_components(instr
->dest
.dest
);
729 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
731 /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
732 * supported. A few do not and are commented for now. Also, there are a
733 * number of NIR ops which Midgard does not support and need to be
734 * lowered, also TODO. This switch block emits the opcode and calling
735 * convention of the Midgard instruction; actual packing is done in
740 /* Number of components valid to check for the instruction (the rest
741 * will be forced to the last), or 0 to use as-is. Relevant as
742 * ball-type instructions have a channel count in NIR but are all vec4
745 unsigned broadcast_swizzle
= 0;
747 /* What register mode should we operate in? */
748 midgard_reg_mode reg_mode
=
749 reg_mode_for_nir(instr
);
751 /* Do we need a destination override? Used for inline
754 midgard_dest_override dest_override
=
755 midgard_dest_override_none
;
757 /* Should we use a smaller respective source and sign-extend? */
759 bool half_1
= false, sext_1
= false;
760 bool half_2
= false, sext_2
= false;
762 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
763 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
766 ALU_CASE(fadd
, fadd
);
767 ALU_CASE(fmul
, fmul
);
768 ALU_CASE(fmin
, fmin
);
769 ALU_CASE(fmax
, fmax
);
770 ALU_CASE(imin
, imin
);
771 ALU_CASE(imax
, imax
);
772 ALU_CASE(umin
, umin
);
773 ALU_CASE(umax
, umax
);
774 ALU_CASE(ffloor
, ffloor
);
775 ALU_CASE(fround_even
, froundeven
);
776 ALU_CASE(ftrunc
, ftrunc
);
777 ALU_CASE(fceil
, fceil
);
778 ALU_CASE(fdot3
, fdot3
);
779 ALU_CASE(fdot4
, fdot4
);
780 ALU_CASE(iadd
, iadd
);
781 ALU_CASE(isub
, isub
);
782 ALU_CASE(imul
, imul
);
784 /* Zero shoved as second-arg */
785 ALU_CASE(iabs
, iabsdiff
);
789 ALU_CASE(feq32
, feq
);
790 ALU_CASE(fne32
, fne
);
791 ALU_CASE(flt32
, flt
);
792 ALU_CASE(ieq32
, ieq
);
793 ALU_CASE(ine32
, ine
);
794 ALU_CASE(ilt32
, ilt
);
795 ALU_CASE(ult32
, ult
);
797 /* We don't have a native b2f32 instruction. Instead, like many
798 * GPUs, we exploit booleans as 0/~0 for false/true, and
799 * correspondingly AND
800 * by 1.0 to do the type conversion. For the moment, prime us
803 * iand [whatever], #0
805 * At the end of emit_alu (as MIR), we'll fix-up the constant
808 ALU_CASE(b2f32
, iand
);
809 ALU_CASE(b2i32
, iand
);
811 /* Likewise, we don't have a dedicated f2b32 instruction, but
812 * we can do a "not equal to 0.0" test. */
814 ALU_CASE(f2b32
, fne
);
815 ALU_CASE(i2b32
, ine
);
817 ALU_CASE(frcp
, frcp
);
818 ALU_CASE(frsq
, frsqrt
);
819 ALU_CASE(fsqrt
, fsqrt
);
820 ALU_CASE(fexp2
, fexp2
);
821 ALU_CASE(flog2
, flog2
);
823 ALU_CASE(f2i32
, f2i_rtz
);
824 ALU_CASE(f2u32
, f2u_rtz
);
825 ALU_CASE(i2f32
, i2f_rtz
);
826 ALU_CASE(u2f32
, u2f_rtz
);
828 ALU_CASE(f2i16
, f2i_rtz
);
829 ALU_CASE(f2u16
, f2u_rtz
);
830 ALU_CASE(i2f16
, i2f_rtz
);
831 ALU_CASE(u2f16
, u2f_rtz
);
833 ALU_CASE(fsin
, fsin
);
834 ALU_CASE(fcos
, fcos
);
836 /* We'll set invert */
837 ALU_CASE(inot
, imov
);
838 ALU_CASE(iand
, iand
);
840 ALU_CASE(ixor
, ixor
);
841 ALU_CASE(ishl
, ishl
);
842 ALU_CASE(ishr
, iasr
);
843 ALU_CASE(ushr
, ilsr
);
845 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
846 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
847 ALU_CASE(b32all_fequal4
, fball_eq
);
849 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
850 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
851 ALU_CASE(b32any_fnequal4
, fbany_neq
);
853 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
854 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
855 ALU_CASE(b32all_iequal4
, iball_eq
);
857 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
858 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
859 ALU_CASE(b32any_inequal4
, ibany_neq
);
861 /* Source mods will be shoved in later */
862 ALU_CASE(fabs
, fmov
);
863 ALU_CASE(fneg
, fmov
);
864 ALU_CASE(fsat
, fmov
);
866 /* For size conversion, we use a move. Ideally though we would squash
867 * these ops together; maybe that has to happen after in NIR as part of
868 * propagation...? An earlier algebraic pass ensured we step down by
869 * only / exactly one size. If stepping down, we use a dest override to
870 * reduce the size; if stepping up, we use a larger-sized move with a
871 * half source and a sign/zero-extension modifier */
876 /* If we end up upscale, we'll need a sign-extend on the
877 * operand (the second argument) */
883 op
= midgard_alu_op_imov
;
885 if (dst_bitsize
== (src_bitsize
* 2)) {
889 /* Use a greater register mode */
891 } else if (src_bitsize
== (dst_bitsize
* 2)) {
892 /* Converting down */
893 dest_override
= midgard_dest_override_lower
;
900 assert(src_bitsize
== 32);
902 op
= midgard_alu_op_fmov
;
903 dest_override
= midgard_dest_override_lower
;
908 assert(src_bitsize
== 16);
910 op
= midgard_alu_op_fmov
;
917 /* For greater-or-equal, we lower to less-or-equal and flip the
925 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
926 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
927 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
928 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
931 /* Swap via temporary */
932 nir_alu_src temp
= instr
->src
[1];
933 instr
->src
[1] = instr
->src
[0];
934 instr
->src
[0] = temp
;
939 case nir_op_b32csel
: {
940 /* Midgard features both fcsel and icsel, depending on
941 * the type of the arguments/output. However, as long
942 * as we're careful we can _always_ use icsel and
943 * _never_ need fcsel, since the latter does additional
944 * floating-point-specific processing whereas the
945 * former just moves bits on the wire. It's not obvious
946 * why these are separate opcodes, save for the ability
947 * to do things like sat/pos/abs/neg for free */
949 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
950 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
952 /* csel works as a two-arg in Midgard, since the condition is hardcoded in r31.w */
955 /* Emit the condition into r31 */
958 emit_condition_mixed(ctx
, &instr
->src
[0], nr_components
);
960 emit_condition(ctx
, &instr
->src
[0].src
, false, instr
->src
[0].swizzle
[0]);
962 /* The condition is the first argument; move the other
963 * arguments up one to be a binary instruction for
966 memmove(instr
->src
, instr
->src
+ 1, 2 * sizeof(nir_alu_src
));
971 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
976 /* Midgard can perform certain modifiers on output of an ALU op */
979 if (midgard_is_integer_out_op(op
)) {
980 outmod
= midgard_outmod_int_wrap
;
982 bool sat
= instr
->dest
.saturate
|| instr
->op
== nir_op_fsat
;
983 outmod
= sat
? midgard_outmod_sat
: midgard_outmod_none
;
986 /* fmax(a, 0.0) can turn into a .pos modifier as an optimization */
988 if (instr
->op
== nir_op_fmax
) {
989 if (nir_is_fzero_constant(instr
->src
[0].src
)) {
990 op
= midgard_alu_op_fmov
;
992 outmod
= midgard_outmod_pos
;
993 instr
->src
[0] = instr
->src
[1];
994 } else if (nir_is_fzero_constant(instr
->src
[1].src
)) {
995 op
= midgard_alu_op_fmov
;
997 outmod
= midgard_outmod_pos
;
1001 /* Fetch unit, quirks, etc information */
1002 unsigned opcode_props
= alu_opcode_props
[op
].props
;
1003 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
1005 /* src0 will always exist afaik, but src1 will not for 1-argument
1006 * instructions. The latter can only be fetched if the instruction
1007 * needs it, or else we may segfault. */
1009 unsigned src0
= nir_alu_src_index(ctx
, &instr
->src
[0]);
1010 unsigned src1
= nr_inputs
== 2 ? nir_alu_src_index(ctx
, &instr
->src
[1]) : SSA_UNUSED_0
;
1012 /* Rather than use the instruction generation helpers, we do it
1013 * ourselves here to avoid the mess */
1015 midgard_instruction ins
= {
1018 .src0
= quirk_flipped_r24
? SSA_UNUSED_1
: src0
,
1019 .src1
= quirk_flipped_r24
? src0
: src1
,
1024 nir_alu_src
*nirmods
[2] = { NULL
};
1026 if (nr_inputs
== 2) {
1027 nirmods
[0] = &instr
->src
[0];
1028 nirmods
[1] = &instr
->src
[1];
1029 } else if (nr_inputs
== 1) {
1030 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
1035 /* These were lowered to a move, so apply the corresponding mod */
1037 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
1038 nir_alu_src
*s
= nirmods
[quirk_flipped_r24
];
1040 if (instr
->op
== nir_op_fneg
)
1041 s
->negate
= !s
->negate
;
1043 if (instr
->op
== nir_op_fabs
)
1047 bool is_int
= midgard_is_integer_op(op
);
1049 ins
.mask
= mask_of(nr_components
);
1051 midgard_vector_alu alu
= {
1053 .reg_mode
= reg_mode
,
1054 .dest_override
= dest_override
,
1057 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, sext_1
)),
1058 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
, broadcast_swizzle
, half_2
, sext_2
)),
1061 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
1064 ins
.mask
&= instr
->dest
.write_mask
;
1068 /* Late fixup for emulated instructions */
1070 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1071 /* Presently, our second argument is an inline #0 constant.
1072 * Switch over to an embedded 1.0 constant (that can't fit
1073 * inline, since we're 32-bit, not 16-bit like the inline
1076 ins
.ssa_args
.inline_constant
= false;
1077 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1078 ins
.has_constants
= true;
1080 if (instr
->op
== nir_op_b2f32
) {
1081 ins
.constants
[0] = 1.0f
;
1083 /* Type pun it into place */
1085 memcpy(&ins
.constants
[0], &one
, sizeof(uint32_t));
1088 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1089 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1090 /* Lots of instructions need a 0 plonked in */
1091 ins
.ssa_args
.inline_constant
= false;
1092 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1093 ins
.has_constants
= true;
1094 ins
.constants
[0] = 0.0f
;
1095 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1096 } else if (instr
->op
== nir_op_inot
) {
1100 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1101 /* To avoid duplicating the lookup tables (probably), true LUT
1102 * instructions can only operate as if they were scalars. Lower
1103 * them here by changing the component. */
1105 uint8_t original_swizzle
[4];
1106 memcpy(original_swizzle
, nirmods
[0]->swizzle
, sizeof(nirmods
[0]->swizzle
));
1107 unsigned orig_mask
= ins
.mask
;
1109 for (int i
= 0; i
< nr_components
; ++i
) {
1110 /* Mask the associated component, dropping the
1111 * instruction if needed */
1114 ins
.mask
&= orig_mask
;
1119 for (int j
= 0; j
< 4; ++j
)
1120 nirmods
[0]->swizzle
[j
] = original_swizzle
[i
]; /* Pull from the correct component */
1122 ins
.alu
.src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, false));
1123 emit_mir_instruction(ctx
, ins
);
1126 emit_mir_instruction(ctx
, ins
);
1132 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1133 * optimized) versions of UBO #0 */
1137 compiler_context
*ctx
,
1140 nir_src
*indirect_offset
,
1143 /* TODO: half-floats */
1145 midgard_instruction ins
= m_ld_uniform_32(dest
, offset
);
1147 /* TODO: Don't split */
1148 ins
.load_store
.varying_parameters
= (offset
& 7) << 7;
1149 ins
.load_store
.address
= offset
>> 3;
1151 if (indirect_offset
) {
1152 emit_indirect_offset(ctx
, indirect_offset
);
1153 ins
.load_store
.unknown
= 0x8700 | index
; /* xxx: what is this? */
1155 ins
.load_store
.unknown
= 0x1E00 | index
; /* xxx: what is this? */
1158 emit_mir_instruction(ctx
, ins
);
1163 compiler_context
*ctx
,
1164 unsigned dest
, unsigned offset
,
1165 unsigned nr_comp
, unsigned component
,
1166 nir_src
*indirect_offset
, nir_alu_type type
)
1168 /* XXX: Half-floats? */
1169 /* TODO: swizzle, mask */
1171 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1172 ins
.mask
= mask_of(nr_comp
);
1173 ins
.load_store
.swizzle
= SWIZZLE_XYZW
>> (2 * component
);
1175 midgard_varying_parameter p
= {
1177 .interpolation
= midgard_interp_default
,
1178 .flat
= /*var->data.interpolation == INTERP_MODE_FLAT*/ 0
1182 memcpy(&u
, &p
, sizeof(p
));
1183 ins
.load_store
.varying_parameters
= u
;
1185 if (indirect_offset
) {
1186 /* We need to add in the dynamic index, moved to r27.w */
1187 emit_indirect_offset(ctx
, indirect_offset
);
1188 ins
.load_store
.unknown
= 0x79e; /* xxx: what is this? */
1190 /* Just a direct load */
1191 ins
.load_store
.unknown
= 0x1e9e; /* xxx: what is this? */
1194 /* Use the type appropriate load */
1198 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1201 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1203 case nir_type_float
:
1204 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1207 unreachable("Attempted to load unknown type");
1211 emit_mir_instruction(ctx
, ins
);
1215 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
)
1219 /* Figure out which uniform this is */
1220 int sysval
= sysval_for_instr(ctx
, instr
, &dest
);
1221 void *val
= _mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
);
1223 /* Sysvals are prefix uniforms */
1224 unsigned uniform
= ((uintptr_t) val
) - 1;
1226 /* Emit the read itself -- this is never indirect */
1227 emit_ubo_read(ctx
, dest
, uniform
, NULL
, 0);
1231 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1233 unsigned offset
= 0, reg
;
1235 switch (instr
->intrinsic
) {
1236 case nir_intrinsic_discard_if
:
1237 emit_condition(ctx
, &instr
->src
[0], true, COMPONENT_X
);
1241 case nir_intrinsic_discard
: {
1242 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1243 struct midgard_instruction discard
= v_branch(conditional
, false);
1244 discard
.branch
.target_type
= TARGET_DISCARD
;
1245 emit_mir_instruction(ctx
, discard
);
1249 case nir_intrinsic_load_uniform
:
1250 case nir_intrinsic_load_ubo
:
1251 case nir_intrinsic_load_input
: {
1252 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1253 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1255 /* Get the base type of the intrinsic */
1256 /* TODO: Infer type? Does it matter? */
1258 is_ubo
? nir_type_uint
: nir_intrinsic_type(instr
);
1259 t
= nir_alu_type_get_base_type(t
);
1262 offset
= nir_intrinsic_base(instr
);
1265 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1267 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1269 bool direct
= nir_src_is_const(*src_offset
);
1272 offset
+= nir_src_as_uint(*src_offset
);
1274 /* We may need to apply a fractional offset */
1275 int component
= instr
->intrinsic
== nir_intrinsic_load_input
?
1276 nir_intrinsic_component(instr
) : 0;
1277 reg
= nir_dest_index(ctx
, &instr
->dest
);
1279 if (is_uniform
&& !ctx
->is_blend
) {
1280 emit_ubo_read(ctx
, reg
, ctx
->sysval_count
+ offset
, !direct
? &instr
->src
[0] : NULL
, 0);
1281 } else if (is_ubo
) {
1282 nir_src index
= instr
->src
[0];
1284 /* We don't yet support indirect UBOs. For indirect
1285 * block numbers (if that's possible), we don't know
1286 * enough about the hardware yet. For indirect sources,
1287 * we know what we need but we need to add some NIR
1288 * support for lowering correctly with respect to
1291 assert(nir_src_is_const(index
));
1292 assert(nir_src_is_const(*src_offset
));
1294 /* TODO: Alignment */
1295 assert((offset
& 0xF) == 0);
1297 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1298 emit_ubo_read(ctx
, reg
, offset
/ 16, NULL
, uindex
);
1299 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1300 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, !direct
? &instr
->src
[0] : NULL
, t
);
1301 } else if (ctx
->is_blend
) {
1302 /* For blend shaders, load the input color, which is
1303 * preloaded to r0 */
1305 midgard_instruction move
= v_mov(SSA_FIXED_REGISTER(0), blank_alu_src
, reg
);
1306 emit_mir_instruction(ctx
, move
);
1307 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1308 midgard_instruction ins
= m_ld_attr_32(reg
, offset
);
1309 ins
.load_store
.unknown
= 0x1E1E; /* XXX: What is this? */
1310 ins
.mask
= mask_of(nr_comp
);
1312 /* Use the type appropriate load */
1316 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1319 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1321 case nir_type_float
:
1322 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1325 unreachable("Attempted to load unknown type");
1329 emit_mir_instruction(ctx
, ins
);
1331 DBG("Unknown load\n");
1338 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1340 case nir_intrinsic_load_raw_output_pan
:
1341 reg
= nir_dest_index(ctx
, &instr
->dest
);
1342 assert(ctx
->is_blend
);
1344 midgard_instruction ins
= m_ld_color_buffer_8(reg
, 0);
1345 emit_mir_instruction(ctx
, ins
);
1348 case nir_intrinsic_load_blend_const_color_rgba
: {
1349 assert(ctx
->is_blend
);
1350 reg
= nir_dest_index(ctx
, &instr
->dest
);
1352 /* Blend constants are embedded directly in the shader and
1353 * patched in, so we use some magic routing */
1355 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, reg
);
1356 ins
.has_constants
= true;
1357 ins
.has_blend_constant
= true;
1358 emit_mir_instruction(ctx
, ins
);
1362 case nir_intrinsic_store_output
:
1363 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1365 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1367 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1369 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1370 /* gl_FragColor is not emitted with load/store
1371 * instructions. Instead, it gets plonked into
1372 * r0 at the end of the shader and we do the
1373 * framebuffer writeout dance. TODO: Defer
1376 midgard_instruction move
= v_mov(reg
, blank_alu_src
, SSA_FIXED_REGISTER(0));
1377 emit_mir_instruction(ctx
, move
);
1379 /* Save the index we're writing to for later reference
1380 * in the epilogue */
1382 ctx
->fragment_output
= reg
;
1383 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1384 /* We should have been vectorized, though we don't
1385 * currently check that st_vary is emitted only once
1386 * per slot (this is relevant, since there's not a mask
1387 * parameter available on the store [set to 0 by the
1388 * blob]). We do respect the component by adjusting the
1389 * swizzle. If this is a constant source, we'll need to
1390 * emit that explicitly. */
1392 emit_explicit_constant(ctx
, reg
, reg
);
1394 unsigned component
= nir_intrinsic_component(instr
);
1396 midgard_instruction st
= m_st_vary_32(reg
, offset
);
1397 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
1398 st
.load_store
.swizzle
= SWIZZLE_XYZW
<< (2*component
);
1399 emit_mir_instruction(ctx
, st
);
1401 DBG("Unknown store\n");
1407 /* Special case of store_output for lowered blend shaders */
1408 case nir_intrinsic_store_raw_output_pan
:
1409 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1410 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1412 midgard_instruction move
= v_mov(reg
, blank_alu_src
, SSA_FIXED_REGISTER(0));
1413 emit_mir_instruction(ctx
, move
);
1414 ctx
->fragment_output
= reg
;
1418 case nir_intrinsic_load_alpha_ref_float
:
1419 assert(instr
->dest
.is_ssa
);
1421 float ref_value
= ctx
->alpha_ref
;
1423 /* See emit_load_const */
1424 float *v
= ralloc_array(NULL
, float, 4);
1425 memcpy(v
, &ref_value
, sizeof(float));
1426 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, (instr
->dest
.ssa
.index
<< 1) + 1, v
);
1429 case nir_intrinsic_load_viewport_scale
:
1430 case nir_intrinsic_load_viewport_offset
:
1431 emit_sysval_read(ctx
, &instr
->instr
);
1435 printf ("Unhandled intrinsic\n");
1442 midgard_tex_format(enum glsl_sampler_dim dim
)
1445 case GLSL_SAMPLER_DIM_1D
:
1446 case GLSL_SAMPLER_DIM_BUF
:
1449 case GLSL_SAMPLER_DIM_2D
:
1450 case GLSL_SAMPLER_DIM_EXTERNAL
:
1453 case GLSL_SAMPLER_DIM_3D
:
1456 case GLSL_SAMPLER_DIM_CUBE
:
1457 return MALI_TEX_CUBE
;
1460 DBG("Unknown sampler dim type\n");
1466 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1470 pan_attach_constant_bias(
1471 compiler_context
*ctx
,
1473 midgard_texture_word
*word
)
1475 /* To attach as constant, it has to *be* constant */
1477 if (!nir_src_is_const(lod
))
1480 float f
= nir_src_as_float(lod
);
1482 /* Break into fixed-point */
1484 float lod_frac
= f
- lod_int
;
1486 /* Carry over negative fractions */
1487 if (lod_frac
< 0.0) {
1493 word
->bias
= float_to_ubyte(lod_frac
);
1494 word
->bias_int
= lod_int
;
1499 static enum mali_sampler_type
1500 midgard_sampler_type(nir_alu_type t
) {
1501 switch (nir_alu_type_get_base_type(t
))
1503 case nir_type_float
:
1504 return MALI_SAMPLER_FLOAT
;
1506 return MALI_SAMPLER_SIGNED
;
1508 return MALI_SAMPLER_UNSIGNED
;
1510 unreachable("Unknown sampler type");
1515 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1516 unsigned midgard_texop
)
1519 //assert (!instr->sampler);
1520 //assert (!instr->texture_array_size);
1522 int texture_index
= instr
->texture_index
;
1523 int sampler_index
= texture_index
;
1525 /* No helper to build texture words -- we do it all here */
1526 midgard_instruction ins
= {
1527 .type
= TAG_TEXTURE_4
,
1530 .dest
= nir_dest_index(ctx
, &instr
->dest
),
1535 .op
= midgard_texop
,
1536 .format
= midgard_tex_format(instr
->sampler_dim
),
1537 .texture_handle
= texture_index
,
1538 .sampler_handle
= sampler_index
,
1539 .swizzle
= SWIZZLE_XYZW
,
1540 .in_reg_swizzle
= SWIZZLE_XYZW
,
1546 .sampler_type
= midgard_sampler_type(instr
->dest_type
),
1550 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1551 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1552 midgard_vector_alu_src alu_src
= blank_alu_src
;
1554 switch (instr
->src
[i
].src_type
) {
1555 case nir_tex_src_coord
: {
1556 emit_explicit_constant(ctx
, index
, index
);
1558 /* Texelfetch coordinates uses all four elements
1559 * (xyz/index) regardless of texture dimensionality,
1560 * which means it's necessary to zero the unused
1561 * components to keep everything happy */
1563 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1564 unsigned old_index
= index
;
1566 index
= make_compiler_temp(ctx
);
1568 /* mov index, old_index */
1569 midgard_instruction mov
= v_mov(old_index
, blank_alu_src
, index
);
1571 emit_mir_instruction(ctx
, mov
);
1573 /* mov index.zw, #0 */
1574 mov
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
),
1575 blank_alu_src
, index
);
1576 mov
.has_constants
= true;
1577 mov
.mask
= (1 << COMPONENT_Z
) | (1 << COMPONENT_W
);
1578 emit_mir_instruction(ctx
, mov
);
1581 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1582 /* texelFetch is undefined on samplerCube */
1583 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1585 /* For cubemaps, we use a special ld/st op to
1586 * select the face and copy the xy into the
1587 * texture register */
1589 unsigned temp
= make_compiler_temp(ctx
);
1591 midgard_instruction st
= m_st_cubemap_coords(temp
, 0);
1592 st
.ssa_args
.src0
= index
;
1593 st
.load_store
.unknown
= 0x24; /* XXX: What is this? */
1594 st
.mask
= 0x3; /* xy */
1595 st
.load_store
.swizzle
= alu_src
.swizzle
;
1596 emit_mir_instruction(ctx
, st
);
1598 ins
.ssa_args
.src0
= temp
;
1600 ins
.ssa_args
.src0
= index
;
1603 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
) {
1604 /* Array component in w but NIR wants it in z */
1605 ins
.texture
.in_reg_swizzle
= SWIZZLE_XYZZ
;
1611 case nir_tex_src_bias
:
1612 case nir_tex_src_lod
: {
1613 /* Try as a constant if we can */
1615 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1616 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1619 ins
.texture
.lod_register
= true;
1620 ins
.ssa_args
.src1
= index
;
1621 emit_explicit_constant(ctx
, index
, index
);
1627 unreachable("Unknown texture source type\n");
1631 emit_mir_instruction(ctx
, ins
);
1633 /* Used for .cont and .last hinting */
1634 ctx
->texture_op_count
++;
1638 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1640 /* Fixup op, since only textureLod is permitted in VS but NIR can give
1641 * generic tex in some cases (which confuses the hardware) */
1643 bool is_vertex
= ctx
->stage
== MESA_SHADER_VERTEX
;
1645 if (is_vertex
&& instr
->op
== nir_texop_tex
)
1646 instr
->op
= nir_texop_txl
;
1648 switch (instr
->op
) {
1651 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1654 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1657 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1660 emit_sysval_read(ctx
, &instr
->instr
);
1663 unreachable("Unhanlded texture op");
1668 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1670 switch (instr
->type
) {
1671 case nir_jump_break
: {
1672 /* Emit a branch out of the loop */
1673 struct midgard_instruction br
= v_branch(false, false);
1674 br
.branch
.target_type
= TARGET_BREAK
;
1675 br
.branch
.target_break
= ctx
->current_loop_depth
;
1676 emit_mir_instruction(ctx
, br
);
1681 DBG("Unknown jump type %d\n", instr
->type
);
1687 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1689 switch (instr
->type
) {
1690 case nir_instr_type_load_const
:
1691 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1694 case nir_instr_type_intrinsic
:
1695 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1698 case nir_instr_type_alu
:
1699 emit_alu(ctx
, nir_instr_as_alu(instr
));
1702 case nir_instr_type_tex
:
1703 emit_tex(ctx
, nir_instr_as_tex(instr
));
1706 case nir_instr_type_jump
:
1707 emit_jump(ctx
, nir_instr_as_jump(instr
));
1710 case nir_instr_type_ssa_undef
:
1715 DBG("Unhandled instruction type\n");
1721 /* ALU instructions can inline or embed constants, which decreases register
1722 * pressure and saves space. */
1724 #define CONDITIONAL_ATTACH(src) { \
1725 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->ssa_args.src + 1); \
1728 attach_constants(ctx, alu, entry, alu->ssa_args.src + 1); \
1729 alu->ssa_args.src = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
1734 inline_alu_constants(compiler_context
*ctx
)
1736 mir_foreach_instr(ctx
, alu
) {
1737 /* Other instructions cannot inline constants */
1738 if (alu
->type
!= TAG_ALU_4
) continue;
1740 /* If there is already a constant here, we can do nothing */
1741 if (alu
->has_constants
) continue;
1743 /* It makes no sense to inline constants on a branch */
1744 if (alu
->compact_branch
|| alu
->prepacked_branch
) continue;
1746 CONDITIONAL_ATTACH(src0
);
1748 if (!alu
->has_constants
) {
1749 CONDITIONAL_ATTACH(src1
)
1750 } else if (!alu
->inline_constant
) {
1751 /* Corner case: _two_ vec4 constants, for instance with a
1752 * csel. For this case, we can only use a constant
1753 * register for one, we'll have to emit a move for the
1754 * other. Note, if both arguments are constants, then
1755 * necessarily neither argument depends on the value of
1756 * any particular register. As the destination register
1757 * will be wiped, that means we can spill the constant
1758 * to the destination register.
1761 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->ssa_args
.src1
+ 1);
1762 unsigned scratch
= alu
->ssa_args
.dest
;
1765 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, scratch
);
1766 attach_constants(ctx
, &ins
, entry
, alu
->ssa_args
.src1
+ 1);
1768 /* Force a break XXX Defer r31 writes */
1769 ins
.unit
= UNIT_VLUT
;
1771 /* Set the source */
1772 alu
->ssa_args
.src1
= scratch
;
1774 /* Inject us -before- the last instruction which set r31 */
1775 mir_insert_instruction_before(mir_prev_op(alu
), ins
);
1781 /* Midgard supports two types of constants, embedded constants (128-bit) and
1782 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
1783 * constants can be demoted to inline constants, for space savings and
1784 * sometimes a performance boost */
1787 embedded_to_inline_constant(compiler_context
*ctx
)
1789 mir_foreach_instr(ctx
, ins
) {
1790 if (!ins
->has_constants
) continue;
1792 if (ins
->ssa_args
.inline_constant
) continue;
1794 /* Blend constants must not be inlined by definition */
1795 if (ins
->has_blend_constant
) continue;
1797 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
1798 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
1799 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
1801 if (!(is_16
|| is_32
))
1804 /* src1 cannot be an inline constant due to encoding
1805 * restrictions. So, if possible we try to flip the arguments
1808 int op
= ins
->alu
.op
;
1810 if (ins
->ssa_args
.src0
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
1812 /* These ops require an operational change to flip
1813 * their arguments TODO */
1814 case midgard_alu_op_flt
:
1815 case midgard_alu_op_fle
:
1816 case midgard_alu_op_ilt
:
1817 case midgard_alu_op_ile
:
1818 case midgard_alu_op_fcsel
:
1819 case midgard_alu_op_icsel
:
1820 DBG("Missed non-commutative flip (%s)\n", alu_opcode_props
[op
].name
);
1825 if (alu_opcode_props
[op
].props
& OP_COMMUTES
) {
1826 /* Flip the SSA numbers */
1827 ins
->ssa_args
.src0
= ins
->ssa_args
.src1
;
1828 ins
->ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1830 /* And flip the modifiers */
1834 src_temp
= ins
->alu
.src2
;
1835 ins
->alu
.src2
= ins
->alu
.src1
;
1836 ins
->alu
.src1
= src_temp
;
1840 if (ins
->ssa_args
.src1
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
1841 /* Extract the source information */
1843 midgard_vector_alu_src
*src
;
1844 int q
= ins
->alu
.src2
;
1845 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
1848 /* Component is from the swizzle, e.g. r26.w -> w component. TODO: What if x is masked out? */
1849 int component
= src
->swizzle
& 3;
1851 /* Scale constant appropriately, if we can legally */
1852 uint16_t scaled_constant
= 0;
1854 if (midgard_is_integer_op(op
) || is_16
) {
1855 unsigned int *iconstants
= (unsigned int *) ins
->constants
;
1856 scaled_constant
= (uint16_t) iconstants
[component
];
1858 /* Constant overflow after resize */
1859 if (scaled_constant
!= iconstants
[component
])
1862 float original
= (float) ins
->constants
[component
];
1863 scaled_constant
= _mesa_float_to_half(original
);
1865 /* Check for loss of precision. If this is
1866 * mediump, we don't care, but for a highp
1867 * shader, we need to pay attention. NIR
1868 * doesn't yet tell us which mode we're in!
1869 * Practically this prevents most constants
1870 * from being inlined, sadly. */
1872 float fp32
= _mesa_half_to_float(scaled_constant
);
1874 if (fp32
!= original
)
1878 /* We don't know how to handle these with a constant */
1880 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
1881 if (mir_nontrivial_raw_mod(*src
, is_int
) || src
->half
|| src
->rep_low
|| src
->rep_high
) {
1882 DBG("Bailing inline constant...\n");
1886 /* Make sure that the constant is not itself a
1887 * vector by checking if all accessed values
1888 * (by the swizzle) are the same. */
1890 uint32_t *cons
= (uint32_t *) ins
->constants
;
1891 uint32_t value
= cons
[component
];
1893 bool is_vector
= false;
1894 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
1896 for (int c
= 1; c
< 4; ++c
) {
1897 /* We only care if this component is actually used */
1898 if (!(mask
& (1 << c
)))
1901 uint32_t test
= cons
[(src
->swizzle
>> (2 * c
)) & 3];
1903 if (test
!= value
) {
1912 /* Get rid of the embedded constant */
1913 ins
->has_constants
= false;
1914 ins
->ssa_args
.src1
= SSA_UNUSED_0
;
1915 ins
->ssa_args
.inline_constant
= true;
1916 ins
->inline_constant
= scaled_constant
;
1921 /* Dead code elimination for branches at the end of a block - only one branch
1922 * per block is legal semantically */
1925 midgard_opt_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
1927 bool branched
= false;
1929 mir_foreach_instr_in_block_safe(block
, ins
) {
1930 if (!midgard_is_branch_unit(ins
->unit
)) continue;
1932 /* We ignore prepacked branches since the fragment epilogue is
1933 * just generally special */
1934 if (ins
->prepacked_branch
) continue;
1936 /* Discards are similarly special and may not correspond to the
1939 if (ins
->branch
.target_type
== TARGET_DISCARD
) continue;
1942 /* We already branched, so this is dead */
1943 mir_remove_instruction(ins
);
1950 /* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
1951 * the move can be propagated away entirely */
1954 mir_compose_float_outmod(midgard_outmod_float
*outmod
, midgard_outmod_float comp
)
1957 if (comp
== midgard_outmod_none
)
1960 if (*outmod
== midgard_outmod_none
) {
1965 /* TODO: Compose rules */
1970 midgard_opt_pos_propagate(compiler_context
*ctx
, midgard_block
*block
)
1972 bool progress
= false;
1974 mir_foreach_instr_in_block_safe(block
, ins
) {
1975 if (ins
->type
!= TAG_ALU_4
) continue;
1976 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
1977 if (ins
->alu
.outmod
!= midgard_outmod_pos
) continue;
1979 /* TODO: Registers? */
1980 unsigned src
= ins
->ssa_args
.src1
;
1981 if (src
& IS_REG
) continue;
1982 assert(!mir_has_multiple_writes(ctx
, src
));
1984 /* There might be a source modifier, too */
1985 if (mir_nontrivial_source2_mod(ins
)) continue;
1987 /* Backpropagate the modifier */
1988 mir_foreach_instr_in_block_from_rev(block
, v
, mir_prev_op(ins
)) {
1989 if (v
->type
!= TAG_ALU_4
) continue;
1990 if (v
->ssa_args
.dest
!= src
) continue;
1992 /* Can we even take a float outmod? */
1993 if (midgard_is_integer_out_op(v
->alu
.op
)) continue;
1995 midgard_outmod_float temp
= v
->alu
.outmod
;
1996 progress
|= mir_compose_float_outmod(&temp
, ins
->alu
.outmod
);
1998 /* Throw in the towel.. */
1999 if (!progress
) break;
2001 /* Otherwise, transfer the modifier */
2002 v
->alu
.outmod
= temp
;
2003 ins
->alu
.outmod
= midgard_outmod_none
;
2013 emit_fragment_epilogue(compiler_context
*ctx
)
2015 emit_explicit_constant(ctx
, ctx
->fragment_output
, SSA_FIXED_REGISTER(0));
2017 /* Perform the actual fragment writeout. We have two writeout/branch
2018 * instructions, forming a loop until writeout is successful as per the
2019 * docs. TODO: gl_FragDepth */
2021 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
2022 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
2025 static midgard_block
*
2026 emit_block(compiler_context
*ctx
, nir_block
*block
)
2028 midgard_block
*this_block
= calloc(sizeof(midgard_block
), 1);
2029 list_addtail(&this_block
->link
, &ctx
->blocks
);
2031 this_block
->is_scheduled
= false;
2034 ctx
->texture_index
[0] = -1;
2035 ctx
->texture_index
[1] = -1;
2037 /* Add us as a successor to the block we are following */
2038 if (ctx
->current_block
)
2039 midgard_block_add_successor(ctx
->current_block
, this_block
);
2041 /* Set up current block */
2042 list_inithead(&this_block
->instructions
);
2043 ctx
->current_block
= this_block
;
2045 nir_foreach_instr(instr
, block
) {
2046 emit_instr(ctx
, instr
);
2047 ++ctx
->instruction_count
;
2050 inline_alu_constants(ctx
);
2051 embedded_to_inline_constant(ctx
);
2053 /* Append fragment shader epilogue (value writeout) */
2054 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
2055 if (block
== nir_impl_last_block(ctx
->func
->impl
)) {
2056 emit_fragment_epilogue(ctx
);
2060 if (block
== nir_start_block(ctx
->func
->impl
))
2061 ctx
->initial_block
= this_block
;
2063 if (block
== nir_impl_last_block(ctx
->func
->impl
))
2064 ctx
->final_block
= this_block
;
2066 /* Allow the next control flow to access us retroactively, for
2068 ctx
->current_block
= this_block
;
2070 /* Document the fallthrough chain */
2071 ctx
->previous_source_block
= this_block
;
2076 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2079 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2081 /* Conditional branches expect the condition in r31.w; emit a move for
2082 * that in the _previous_ block (which is the current block). */
2083 emit_condition(ctx
, &nif
->condition
, true, COMPONENT_X
);
2085 /* Speculatively emit the branch, but we can't fill it in until later */
2086 EMIT(branch
, true, true);
2087 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2089 /* Emit the two subblocks */
2090 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2092 /* Emit a jump from the end of the then block to the end of the else */
2093 EMIT(branch
, false, false);
2094 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2096 /* Emit second block, and check if it's empty */
2098 int else_idx
= ctx
->block_count
;
2099 int count_in
= ctx
->instruction_count
;
2100 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2101 int after_else_idx
= ctx
->block_count
;
2103 /* Now that we have the subblocks emitted, fix up the branches */
2108 if (ctx
->instruction_count
== count_in
) {
2109 /* The else block is empty, so don't emit an exit jump */
2110 mir_remove_instruction(then_exit
);
2111 then_branch
->branch
.target_block
= after_else_idx
;
2113 then_branch
->branch
.target_block
= else_idx
;
2114 then_exit
->branch
.target_block
= after_else_idx
;
2119 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2121 /* Remember where we are */
2122 midgard_block
*start_block
= ctx
->current_block
;
2124 /* Allocate a loop number, growing the current inner loop depth */
2125 int loop_idx
= ++ctx
->current_loop_depth
;
2127 /* Get index from before the body so we can loop back later */
2128 int start_idx
= ctx
->block_count
;
2130 /* Emit the body itself */
2131 emit_cf_list(ctx
, &nloop
->body
);
2133 /* Branch back to loop back */
2134 struct midgard_instruction br_back
= v_branch(false, false);
2135 br_back
.branch
.target_block
= start_idx
;
2136 emit_mir_instruction(ctx
, br_back
);
2138 /* Mark down that branch in the graph. Note that we're really branching
2139 * to the block *after* we started in. TODO: Why doesn't the branch
2140 * itself have an off-by-one then...? */
2141 midgard_block_add_successor(ctx
->current_block
, start_block
->successors
[0]);
2143 /* Find the index of the block about to follow us (note: we don't add
2144 * one; blocks are 0-indexed so we get a fencepost problem) */
2145 int break_block_idx
= ctx
->block_count
;
2147 /* Fix up the break statements we emitted to point to the right place,
2148 * now that we can allocate a block number for them */
2150 list_for_each_entry_from(struct midgard_block
, block
, start_block
, &ctx
->blocks
, link
) {
2151 mir_foreach_instr_in_block(block
, ins
) {
2152 if (ins
->type
!= TAG_ALU_4
) continue;
2153 if (!ins
->compact_branch
) continue;
2154 if (ins
->prepacked_branch
) continue;
2156 /* We found a branch -- check the type to see if we need to do anything */
2157 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2159 /* It's a break! Check if it's our break */
2160 if (ins
->branch
.target_break
!= loop_idx
) continue;
2162 /* Okay, cool, we're breaking out of this loop.
2163 * Rewrite from a break to a goto */
2165 ins
->branch
.target_type
= TARGET_GOTO
;
2166 ins
->branch
.target_block
= break_block_idx
;
2170 /* Now that we've finished emitting the loop, free up the depth again
2171 * so we play nice with recursion amid nested loops */
2172 --ctx
->current_loop_depth
;
2174 /* Dump loop stats */
2178 static midgard_block
*
2179 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2181 midgard_block
*start_block
= NULL
;
2183 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2184 switch (node
->type
) {
2185 case nir_cf_node_block
: {
2186 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2189 start_block
= block
;
2194 case nir_cf_node_if
:
2195 emit_if(ctx
, nir_cf_node_as_if(node
));
2198 case nir_cf_node_loop
:
2199 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2202 case nir_cf_node_function
:
2211 /* Due to lookahead, we need to report the first tag executed in the command
2212 * stream and in branch targets. An initial block might be empty, so iterate
2213 * until we find one that 'works' */
2216 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2218 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2220 unsigned first_tag
= 0;
2223 midgard_bundle
*initial_bundle
= util_dynarray_element(&initial_block
->bundles
, midgard_bundle
, 0);
2225 if (initial_bundle
) {
2226 first_tag
= initial_bundle
->tag
;
2230 /* Initial block is empty, try the next block */
2231 initial_block
= list_first_entry(&(initial_block
->link
), midgard_block
, link
);
2232 } while(initial_block
!= NULL
);
2239 midgard_compile_shader_nir(struct midgard_screen
*screen
, nir_shader
*nir
, midgard_program
*program
, bool is_blend
)
2241 struct util_dynarray
*compiled
= &program
->compiled
;
2243 midgard_debug
= debug_get_option_midgard_debug();
2245 compiler_context ictx
= {
2248 .stage
= nir
->info
.stage
,
2251 .is_blend
= is_blend
,
2252 .blend_constant_offset
= 0,
2254 .alpha_ref
= program
->alpha_ref
2257 compiler_context
*ctx
= &ictx
;
2259 /* Start off with a safe cutoff, allowing usage of all 16 work
2260 * registers. Later, we'll promote uniform reads to uniform registers
2261 * if we determine it is beneficial to do so */
2262 ctx
->uniform_cutoff
= 8;
2264 /* Initialize at a global (not block) level hash tables */
2266 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2267 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2268 ctx
->sysval_to_id
= _mesa_hash_table_u64_create(NULL
);
2270 /* Record the varying mapping for the command stream's bookkeeping */
2272 struct exec_list
*varyings
=
2273 ctx
->stage
== MESA_SHADER_VERTEX
? &nir
->outputs
: &nir
->inputs
;
2275 unsigned max_varying
= 0;
2276 nir_foreach_variable(var
, varyings
) {
2277 unsigned loc
= var
->data
.driver_location
;
2278 unsigned sz
= glsl_type_size(var
->type
, FALSE
);
2280 for (int c
= 0; c
< sz
; ++c
) {
2281 program
->varyings
[loc
+ c
] = var
->data
.location
+ c
;
2282 max_varying
= MAX2(max_varying
, loc
+ c
);
2286 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2287 * (so we don't accidentally duplicate the epilogue since mesa/st has
2288 * messed with our I/O quite a bit already) */
2290 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2292 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2293 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2294 NIR_PASS_V(nir
, nir_clamp_psiz
, 1.0, 1024.0);
2297 NIR_PASS_V(nir
, nir_lower_var_copies
);
2298 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2299 NIR_PASS_V(nir
, nir_split_var_copies
);
2300 NIR_PASS_V(nir
, nir_lower_var_copies
);
2301 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2302 NIR_PASS_V(nir
, nir_lower_var_copies
);
2303 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2305 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2307 /* Optimisation passes */
2311 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2312 nir_print_shader(nir
, stdout
);
2315 /* Assign sysvals and counts, now that we're sure
2316 * (post-optimisation) */
2318 midgard_nir_assign_sysvals(ctx
, nir
);
2320 program
->uniform_count
= nir
->num_uniforms
;
2321 program
->sysval_count
= ctx
->sysval_count
;
2322 memcpy(program
->sysvals
, ctx
->sysvals
, sizeof(ctx
->sysvals
[0]) * ctx
->sysval_count
);
2324 nir_foreach_function(func
, nir
) {
2328 list_inithead(&ctx
->blocks
);
2329 ctx
->block_count
= 0;
2332 emit_cf_list(ctx
, &func
->impl
->body
);
2333 emit_block(ctx
, func
->impl
->end_block
);
2335 break; /* TODO: Multi-function shaders */
2338 util_dynarray_init(compiled
, NULL
);
2340 /* MIR-level optimizations */
2342 bool progress
= false;
2347 mir_foreach_block(ctx
, block
) {
2348 progress
|= midgard_opt_pos_propagate(ctx
, block
);
2349 progress
|= midgard_opt_copy_prop(ctx
, block
);
2350 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2351 progress
|= midgard_opt_combine_projection(ctx
, block
);
2352 progress
|= midgard_opt_varying_projection(ctx
, block
);
2356 mir_foreach_block(ctx
, block
) {
2357 midgard_lower_invert(ctx
, block
);
2358 midgard_lower_derivatives(ctx
, block
);
2361 /* Nested control-flow can result in dead branches at the end of the
2362 * block. This messes with our analysis and is just dead code, so cull
2364 mir_foreach_block(ctx
, block
) {
2365 midgard_opt_cull_dead_branch(ctx
, block
);
2368 /* Ensure we were lowered */
2369 mir_foreach_instr_global(ctx
, ins
) {
2370 assert(!ins
->invert
);
2374 schedule_program(ctx
);
2376 /* Now that all the bundles are scheduled and we can calculate block
2377 * sizes, emit actual branch instructions rather than placeholders */
2379 int br_block_idx
= 0;
2381 mir_foreach_block(ctx
, block
) {
2382 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2383 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2384 midgard_instruction
*ins
= bundle
->instructions
[c
];
2386 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2388 if (ins
->prepacked_branch
) continue;
2390 /* Parse some basic branch info */
2391 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2392 bool is_conditional
= ins
->branch
.conditional
;
2393 bool is_inverted
= ins
->branch
.invert_conditional
;
2394 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2396 /* Determine the block we're jumping to */
2397 int target_number
= ins
->branch
.target_block
;
2399 /* Report the destination tag */
2400 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2402 /* Count up the number of quadwords we're
2403 * jumping over = number of quadwords until
2404 * (br_block_idx, target_number) */
2406 int quadword_offset
= 0;
2409 /* Jump to the end of the shader. We
2410 * need to include not only the
2411 * following blocks, but also the
2412 * contents of our current block (since
2413 * discard can come in the middle of
2416 midgard_block
*blk
= mir_get_block(ctx
, br_block_idx
+ 1);
2418 for (midgard_bundle
*bun
= bundle
+ 1; bun
< (midgard_bundle
*)((char*) block
->bundles
.data
+ block
->bundles
.size
); ++bun
) {
2419 quadword_offset
+= quadword_size(bun
->tag
);
2422 mir_foreach_block_from(ctx
, blk
, b
) {
2423 quadword_offset
+= b
->quadword_count
;
2426 } else if (target_number
> br_block_idx
) {
2429 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2430 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2433 quadword_offset
+= blk
->quadword_count
;
2436 /* Jump backwards */
2438 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2439 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2442 quadword_offset
-= blk
->quadword_count
;
2446 /* Unconditional extended branches (far jumps)
2447 * have issues, so we always use a conditional
2448 * branch, setting the condition to always for
2449 * unconditional. For compact unconditional
2450 * branches, cond isn't used so it doesn't
2451 * matter what we pick. */
2453 midgard_condition cond
=
2454 !is_conditional
? midgard_condition_always
:
2455 is_inverted
? midgard_condition_false
:
2456 midgard_condition_true
;
2458 midgard_jmp_writeout_op op
=
2459 is_discard
? midgard_jmp_writeout_op_discard
:
2460 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2461 midgard_jmp_writeout_op_branch_cond
;
2464 midgard_branch_extended branch
=
2465 midgard_create_branch_extended(
2470 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2471 } else if (is_conditional
|| is_discard
) {
2472 midgard_branch_cond branch
= {
2474 .dest_tag
= dest_tag
,
2475 .offset
= quadword_offset
,
2479 assert(branch
.offset
== quadword_offset
);
2481 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2483 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2485 midgard_branch_uncond branch
= {
2487 .dest_tag
= dest_tag
,
2488 .offset
= quadword_offset
,
2492 assert(branch
.offset
== quadword_offset
);
2494 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2502 /* Emit flat binary from the instruction arrays. Iterate each block in
2503 * sequence. Save instruction boundaries such that lookahead tags can
2504 * be assigned easily */
2506 /* Cache _all_ bundles in source order for lookahead across failed branches */
2508 int bundle_count
= 0;
2509 mir_foreach_block(ctx
, block
) {
2510 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2512 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2514 mir_foreach_block(ctx
, block
) {
2515 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2516 source_order_bundles
[bundle_idx
++] = bundle
;
2520 int current_bundle
= 0;
2522 /* Midgard prefetches instruction types, so during emission we
2523 * need to lookahead. Unless this is the last instruction, in
2524 * which we return 1. Or if this is the second to last and the
2525 * last is an ALU, then it's also 1... */
2527 mir_foreach_block(ctx
, block
) {
2528 mir_foreach_bundle_in_block(block
, bundle
) {
2531 if (current_bundle
+ 1 < bundle_count
) {
2532 uint8_t next
= source_order_bundles
[current_bundle
+ 1]->tag
;
2534 if (!(current_bundle
+ 2 < bundle_count
) && IS_ALU(next
)) {
2541 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2545 /* TODO: Free deeper */
2546 //util_dynarray_fini(&block->instructions);
2549 free(source_order_bundles
);
2551 /* Report the very first tag executed */
2552 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2554 /* Deal with off-by-one related to the fencepost problem */
2555 program
->work_register_count
= ctx
->work_registers
+ 1;
2556 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2558 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2559 program
->tls_size
= ctx
->tls_size
;
2561 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2562 disassemble_midgard(program
->compiled
.data
, program
->compiled
.size
);
2564 if (midgard_debug
& MIDGARD_DBG_SHADERDB
) {
2565 unsigned nr_bundles
= 0, nr_ins
= 0, nr_quadwords
= 0;
2567 /* Count instructions and bundles */
2569 mir_foreach_instr_global(ctx
, ins
) {
2573 mir_foreach_block(ctx
, block
) {
2574 nr_bundles
+= util_dynarray_num_elements(
2575 &block
->bundles
, midgard_bundle
);
2577 nr_quadwords
+= block
->quadword_count
;
2580 /* Calculate thread count. There are certain cutoffs by
2581 * register count for thread count */
2583 unsigned nr_registers
= program
->work_register_count
;
2585 unsigned nr_threads
=
2586 (nr_registers
<= 4) ? 4 :
2587 (nr_registers
<= 8) ? 2 :
2592 fprintf(stderr
, "shader%d - %s shader: "
2593 "%u inst, %u bundles, %u quadwords, "
2594 "%u registers, %u threads, %u loops, "
2595 "%d:%d spills:fills\n",
2597 gl_shader_stage_name(ctx
->stage
),
2598 nr_ins
, nr_bundles
, nr_quadwords
,
2599 nr_registers
, nr_threads
,
2601 ctx
->spills
, ctx
->fills
);