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
25 #include "midgard_ops.h"
26 #include "util/u_memory.h"
28 /* Create a mask of accessed components from a swizzle to figure out vector
32 swizzle_to_access_mask(unsigned swizzle
)
34 unsigned component_mask
= 0;
36 for (int i
= 0; i
< 4; ++i
) {
37 unsigned c
= (swizzle
>> (2 * i
)) & 3;
38 component_mask
|= (1 << c
);
41 return component_mask
;
44 /* Does the mask cover more than a scalar? */
47 is_single_component_mask(unsigned mask
)
51 for (int c
= 0; c
< 8; ++c
) {
56 return components
== 1;
59 /* Checks for an SSA data hazard between two adjacent instructions, keeping in
60 * mind that we are a vector architecture and we can write to different
61 * components simultaneously */
64 can_run_concurrent_ssa(midgard_instruction
*first
, midgard_instruction
*second
)
66 /* Each instruction reads some registers and writes to a register. See
67 * where the first writes */
69 /* Figure out where exactly we wrote to */
70 int source
= first
->ssa_args
.dest
;
71 int source_mask
= first
->mask
;
73 /* As long as the second doesn't read from the first, we're okay */
74 if (second
->ssa_args
.src0
== source
) {
75 if (first
->type
== TAG_ALU_4
) {
76 /* Figure out which components we just read from */
78 int q
= second
->alu
.src1
;
79 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
81 /* Check if there are components in common, and fail if so */
82 if (swizzle_to_access_mask(m
->swizzle
) & source_mask
)
89 if (second
->ssa_args
.src1
== source
)
92 /* Otherwise, it's safe in that regard. Another data hazard is both
93 * writing to the same place, of course */
95 if (second
->ssa_args
.dest
== source
) {
96 /* ...but only if the components overlap */
98 if (second
->mask
& source_mask
)
108 midgard_instruction
**segment
, unsigned segment_size
,
109 midgard_instruction
*ains
)
111 for (int s
= 0; s
< segment_size
; ++s
)
112 if (!can_run_concurrent_ssa(segment
[s
], ains
))
120 /* Schedules, but does not emit, a single basic block. After scheduling, the
121 * final tag and size of the block are known, which are necessary for branching
124 static midgard_bundle
125 schedule_bundle(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*ins
, int *skip
)
127 int instructions_emitted
= 0, packed_idx
= 0;
128 midgard_bundle bundle
= { 0 };
130 uint8_t tag
= ins
->type
;
132 /* Default to the instruction's tag */
137 uint32_t control
= 0;
138 size_t bytes_emitted
= sizeof(control
);
140 /* TODO: Constant combining */
141 int index
= 0, last_unit
= 0;
143 /* Previous instructions, for the purpose of parallelism */
144 midgard_instruction
*segment
[4] = {0};
145 int segment_size
= 0;
147 instructions_emitted
= -1;
148 midgard_instruction
*pins
= ins
;
150 unsigned constant_count
= 0;
153 midgard_instruction
*ains
= pins
;
155 /* Advance instruction pointer */
157 ains
= mir_next_op(pins
);
161 /* Out-of-work condition */
162 if ((struct list_head
*) ains
== &block
->instructions
)
165 /* Ensure that the chain can continue */
166 if (ains
->type
!= TAG_ALU_4
) break;
168 /* If there's already something in the bundle and we
169 * have weird scheduler constraints, break now */
170 if (ains
->precede_break
&& index
) break;
172 /* According to the presentation "The ARM
173 * Mali-T880 Mobile GPU" from HotChips 27,
174 * there are two pipeline stages. Branching
175 * position determined experimentally. Lines
176 * are executed in parallel:
179 * [ VADD ] [ SMUL ] [ LUT ] [ BRANCH ]
181 * Verify that there are no ordering dependencies here.
183 * TODO: Allow for parallelism!!!
186 /* Pick a unit for it if it doesn't force a particular unit */
188 int unit
= ains
->unit
;
191 int op
= ains
->alu
.op
;
192 int units
= alu_opcode_props
[op
].props
;
194 bool vectorable
= units
& UNITS_ANY_VECTOR
;
195 bool scalarable
= units
& UNITS_SCALAR
;
196 bool could_scalar
= is_single_component_mask(ains
->mask
);
198 /* Only 16/32-bit can run on a scalar unit */
199 could_scalar
&= ains
->alu
.reg_mode
!= midgard_reg_mode_8
;
200 could_scalar
&= ains
->alu
.reg_mode
!= midgard_reg_mode_64
;
201 could_scalar
&= ains
->alu
.dest_override
== midgard_dest_override_none
;
203 bool vector
= vectorable
&& !(could_scalar
&& scalarable
);
205 /* TODO: Check ahead-of-time for other scalar
206 * hazards that otherwise get aborted out */
209 assert(units
& UNITS_SCALAR
);
212 if (last_unit
>= UNIT_VADD
) {
213 if (units
& UNIT_VLUT
)
218 if ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
))
220 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
222 else if (units
& UNIT_VLUT
)
228 if (last_unit
>= UNIT_VADD
) {
229 if ((units
& UNIT_SMUL
) && !(control
& UNIT_SMUL
))
231 else if (units
& UNIT_VLUT
)
236 if ((units
& UNIT_SADD
) && !(control
& UNIT_SADD
) && !midgard_has_hazard(segment
, segment_size
, ains
))
238 else if (units
& UNIT_SMUL
)
239 unit
= ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
)) ? UNIT_VMUL
: UNIT_SMUL
;
240 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
247 assert(unit
& units
);
250 /* Late unit check, this time for encoding (not parallelism) */
251 if (unit
<= last_unit
) break;
253 /* Clear the segment */
254 if (last_unit
< UNIT_VADD
&& unit
>= UNIT_VADD
)
257 if (midgard_has_hazard(segment
, segment_size
, ains
))
260 /* We're good to go -- emit the instruction */
263 segment
[segment_size
++] = ains
;
265 /* We try to reuse constants if possible, by adjusting
268 if (ains
->has_blend_constant
) {
269 /* Everything conflicts with the blend constant */
270 if (bundle
.has_embedded_constants
)
273 bundle
.has_blend_constant
= 1;
274 bundle
.has_embedded_constants
= 1;
275 } else if (ains
->has_constants
) {
276 /* By definition, blend constants conflict with
277 * everything, so if there are already
278 * constants we break the bundle *now* */
280 if (bundle
.has_blend_constant
)
283 /* For anything but blend constants, we can do
284 * proper analysis, however */
286 /* TODO: Mask by which are used */
287 uint32_t *constants
= (uint32_t *) ains
->constants
;
288 uint32_t *bundles
= (uint32_t *) bundle
.constants
;
290 uint32_t indices
[4] = { 0 };
291 bool break_bundle
= false;
293 for (unsigned i
= 0; i
< 4; ++i
) {
294 uint32_t cons
= constants
[i
];
295 bool constant_found
= false;
297 /* Search for the constant */
298 for (unsigned j
= 0; j
< constant_count
; ++j
) {
299 if (bundles
[j
] != cons
)
302 /* We found it, reuse */
304 constant_found
= true;
311 /* We didn't find it, so allocate it */
312 unsigned idx
= constant_count
++;
315 /* Uh-oh, out of space */
320 /* We have space, copy it in! */
328 /* Cool, we have it in. So use indices as a
331 unsigned swizzle
= SWIZZLE_FROM_ARRAY(indices
);
332 unsigned r_constant
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
334 if (ains
->ssa_args
.src0
== r_constant
)
335 ains
->alu
.src1
= vector_alu_apply_swizzle(ains
->alu
.src1
, swizzle
);
337 if (ains
->ssa_args
.src1
== r_constant
)
338 ains
->alu
.src2
= vector_alu_apply_swizzle(ains
->alu
.src2
, swizzle
);
340 bundle
.has_embedded_constants
= true;
343 if (ains
->unit
& UNITS_ANY_VECTOR
) {
344 bytes_emitted
+= sizeof(midgard_reg_info
);
345 bytes_emitted
+= sizeof(midgard_vector_alu
);
346 } else if (ains
->compact_branch
) {
347 /* All of r0 has to be written out along with
348 * the branch writeout */
350 if (ains
->writeout
) {
351 /* The rules for when "bare" writeout
352 * is safe are when all components are
353 * r0 are written out in the final
354 * bundle, earlier than VLUT, where any
355 * register dependencies of r0 are from
356 * an earlier bundle. We can't verify
357 * this before RA, so we don't try. */
363 midgard_instruction ins
= v_mov(0, blank_alu_src
, SSA_FIXED_REGISTER(0));
364 ins
.unit
= UNIT_VMUL
;
367 /* TODO don't leak */
368 midgard_instruction
*move
=
369 mem_dup(&ins
, sizeof(midgard_instruction
));
370 bytes_emitted
+= sizeof(midgard_reg_info
);
371 bytes_emitted
+= sizeof(midgard_vector_alu
);
372 bundle
.instructions
[packed_idx
++] = move
;
375 if (ains
->unit
== ALU_ENAB_BRANCH
) {
376 bytes_emitted
+= sizeof(midgard_branch_extended
);
378 bytes_emitted
+= sizeof(ains
->br_compact
);
381 bytes_emitted
+= sizeof(midgard_reg_info
);
382 bytes_emitted
+= sizeof(midgard_scalar_alu
);
385 /* Defer marking until after writing to allow for break */
386 control
|= ains
->unit
;
387 last_unit
= ains
->unit
;
388 ++instructions_emitted
;
394 /* Pad ALU op to nearest word */
396 if (bytes_emitted
& 15) {
397 padding
= 16 - (bytes_emitted
& 15);
398 bytes_emitted
+= padding
;
401 /* Constants must always be quadwords */
402 if (bundle
.has_embedded_constants
)
405 /* Size ALU instruction for tag */
406 bundle
.tag
= (TAG_ALU_4
) + (bytes_emitted
/ 16) - 1;
407 bundle
.padding
= padding
;
408 bundle
.control
= bundle
.tag
| control
;
413 case TAG_LOAD_STORE_4
: {
414 /* Load store instructions have two words at once. If
415 * we only have one queued up, we need to NOP pad.
416 * Otherwise, we store both in succession to save space
417 * and cycles -- letting them go in parallel -- skip
418 * the next. The usefulness of this optimisation is
419 * greatly dependent on the quality of the instruction
423 midgard_instruction
*next_op
= mir_next_op(ins
);
425 if ((struct list_head
*) next_op
!= &block
->instructions
&& next_op
->type
== TAG_LOAD_STORE_4
) {
426 /* TODO: Concurrency check */
427 instructions_emitted
++;
433 case TAG_TEXTURE_4
: {
434 /* Which tag we use depends on the shader stage */
435 bool in_frag
= ctx
->stage
== MESA_SHADER_FRAGMENT
;
436 bundle
.tag
= in_frag
? TAG_TEXTURE_4
: TAG_TEXTURE_4_VTX
;
441 unreachable("Unknown tag");
445 /* Copy the instructions into the bundle */
446 bundle
.instruction_count
= instructions_emitted
+ 1 + packed_idx
;
448 midgard_instruction
*uins
= ins
;
449 for (; packed_idx
< bundle
.instruction_count
; ++packed_idx
) {
450 bundle
.instructions
[packed_idx
] = uins
;
451 uins
= mir_next_op(uins
);
454 *skip
= instructions_emitted
;
459 /* Schedule a single block by iterating its instruction to create bundles.
460 * While we go, tally about the bundle sizes to compute the block size. */
463 schedule_block(compiler_context
*ctx
, midgard_block
*block
)
465 util_dynarray_init(&block
->bundles
, NULL
);
467 block
->quadword_count
= 0;
469 mir_foreach_instr_in_block(block
, ins
) {
471 midgard_bundle bundle
= schedule_bundle(ctx
, block
, ins
, &skip
);
472 util_dynarray_append(&block
->bundles
, midgard_bundle
, bundle
);
474 if (bundle
.has_blend_constant
) {
475 /* TODO: Multiblock? */
476 int quadwords_within_block
= block
->quadword_count
+ quadword_size(bundle
.tag
) - 1;
477 ctx
->blend_constant_offset
= quadwords_within_block
* 0x10;
481 ins
= mir_next_op(ins
);
483 block
->quadword_count
+= quadword_size(bundle
.tag
);
486 block
->is_scheduled
= true;
490 schedule_program(compiler_context
*ctx
)
492 /* We run RA prior to scheduling */
494 mir_foreach_block(ctx
, block
) {
495 schedule_block(ctx
, block
);
498 /* Pipeline registers creation is a prepass before RA */
499 mir_create_pipeline_registers(ctx
);
501 struct ra_graph
*g
= allocate_registers(ctx
);
502 install_registers(ctx
, g
);