2 * Copyright (C) 2020 Collabora Ltd.
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
23 * Authors (Collabora):
24 * Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
29 #include "util/half_float.h"
30 #include "bifrost/disassemble.h"
32 /* Instruction packing tests */
35 bit_test_single(struct panfrost_device
*dev
,
38 bool fma
, enum bit_debug debug
)
40 /* First, simulate the instruction */
41 struct bit_state s
= { 0 };
42 memcpy(s
.r
, input
, 16);
43 bit_step(&s
, ins
, fma
);
45 /* Next, wrap it up and pack it */
47 bi_instruction ldubo
= {
48 .type
= BI_LOAD_UNIFORM
,
57 .dest
= BIR_INDEX_REGISTER
| 0,
58 .dest_type
= nir_type_uint32
,
62 bi_instruction ldva
= {
63 .type
= BI_LOAD_VAR_ADDRESS
,
65 .dest
= BIR_INDEX_REGISTER
| 32,
66 .dest_type
= nir_type_uint32
,
69 BIR_INDEX_REGISTER
| 61,
70 BIR_INDEX_REGISTER
| 62,
84 BIR_INDEX_REGISTER
| 0,
85 ldva
.dest
, ldva
.dest
+ 1, ldva
.dest
+ 2,
89 nir_type_uint32
, nir_type_uint32
, nir_type_uint32
,
94 bi_context
*ctx
= rzalloc(NULL
, bi_context
);
95 ctx
->stage
= MESA_SHADER_VERTEX
;
97 bi_block
*blk
= rzalloc(ctx
, bi_block
);
98 blk
->scheduled
= true;
100 blk
->base
.predecessors
= _mesa_set_create(blk
,
102 _mesa_key_pointer_equal
);
104 list_inithead(&ctx
->blocks
);
105 list_addtail(&blk
->base
.link
, &ctx
->blocks
);
106 list_inithead(&blk
->clauses
);
108 bi_clause
*clauses
[4] = {
109 rzalloc(ctx
, bi_clause
),
110 rzalloc(ctx
, bi_clause
),
111 rzalloc(ctx
, bi_clause
),
112 rzalloc(ctx
, bi_clause
)
115 for (unsigned i
= 0; i
< 4; ++i
) {
116 clauses
[i
]->bundle_count
= 1;
117 list_addtail(&clauses
[i
]->link
, &blk
->clauses
);
118 clauses
[i
]->scoreboard_id
= (i
& 1);
121 clauses
[i
]->dependencies
= 1 << (~i
& 1);
122 clauses
[i
]->data_register_write_barrier
= true;
126 clauses
[0]->bundles
[0].add
= &ldubo
;
127 clauses
[0]->clause_type
= BIFROST_CLAUSE_UBO
;
130 clauses
[1]->bundles
[0].fma
= ins
;
132 clauses
[1]->bundles
[0].add
= ins
;
134 clauses
[0]->constant_count
= 1;
135 clauses
[1]->constant_count
= 1;
136 clauses
[1]->constants
[0] = ins
->constant
.u64
;
138 clauses
[2]->bundles
[0].add
= &ldva
;
139 clauses
[3]->bundles
[0].add
= &st
;
141 clauses
[2]->clause_type
= BIFROST_CLAUSE_UBO
;
142 clauses
[3]->clause_type
= BIFROST_CLAUSE_SSBO_STORE
;
144 panfrost_program prog
;
145 bi_pack(ctx
, &prog
.compiled
);
147 bool succ
= bit_vertex(dev
, prog
, input
, 16, NULL
, 0,
150 if (debug
>= BIT_DEBUG_ALL
|| (!succ
&& debug
>= BIT_DEBUG_FAIL
)) {
151 bi_print_shader(ctx
, stderr
);
152 disassemble_bifrost(stderr
, prog
.compiled
.data
, prog
.compiled
.size
, true);
156 fprintf(stderr
, "FAIL\n");
159 /* Utilities for generating tests */
162 bit_generate_float4(float *mem
)
164 for (unsigned i
= 0; i
< 4; ++i
)
165 mem
[i
] = (float) ((rand() & 255) - 127) / 16.0;
169 bit_generate_half8(uint16_t *mem
)
171 for (unsigned i
= 0; i
< 8; ++i
)
172 mem
[i
] = _mesa_float_to_half(((float) (rand() & 255) - 127) / 16.0);
175 static bi_instruction
176 bit_ins(enum bi_class C
, unsigned argc
, nir_alu_type base
, unsigned size
)
178 nir_alu_type T
= base
| size
;
180 bi_instruction ins
= {
182 .dest
= BIR_INDEX_REGISTER
| 0,
186 for (unsigned i
= 0; i
< argc
; ++i
) {
187 ins
.src
[i
] = BIR_INDEX_REGISTER
| i
;
188 ins
.src_types
[i
] = T
;
194 #define BIT_FOREACH_SWIZZLE(swz, args, sz) \
195 for (unsigned swz = 0; swz < ((sz == 16) ? (1 << (2 * args)) : 1); ++swz)
198 bit_apply_swizzle(bi_instruction
*ins
, unsigned swz
, unsigned args
, unsigned sz
)
200 unsigned slots_per_arg
= (sz
== 16) ? 4 : 1;
201 unsigned slots_per_chan
= (sz
== 16) ? 1 : 0;
202 unsigned mask
= (sz
== 16) ? 1 : 0;
204 for (unsigned i
= 0; i
< args
; ++i
) {
205 for (unsigned j
= 0; j
< (32 / sz
); ++j
) {
206 ins
->swizzle
[i
][j
] = ((swz
>> (slots_per_arg
* i
)) >> (slots_per_chan
* j
)) & mask
;
211 /* Tests all 64 combinations of floating point modifiers for a given
212 * instruction / floating-type / test type */
215 bit_fmod_helper(struct panfrost_device
*dev
,
216 enum bi_class c
, unsigned size
, bool fma
,
217 uint32_t *input
, enum bit_debug debug
, unsigned op
)
219 bi_instruction ins
= bit_ins(c
, 2, nir_type_float
, size
);
221 bool fp16
= (size
== 16);
222 bool has_outmods
= fma
|| !fp16
;
224 for (unsigned outmod
= 0; outmod
< (has_outmods
? 4 : 1); ++outmod
) {
225 BIT_FOREACH_SWIZZLE(swz
, 2, size
) {
226 for (unsigned inmod
= 0; inmod
< 16; ++inmod
) {
229 ins
.src_abs
[0] = (inmod
& 0x1);
230 ins
.src_abs
[1] = (inmod
& 0x2);
231 ins
.src_neg
[0] = (inmod
& 0x4);
232 ins
.src_neg
[1] = (inmod
& 0x8);
233 bit_apply_swizzle(&ins
, swz
, 2, size
);
234 bit_test_single(dev
, &ins
, input
, fma
, debug
);
241 bit_fma_helper(struct panfrost_device
*dev
,
242 unsigned size
, uint32_t *input
, enum bit_debug debug
)
244 bi_instruction ins
= bit_ins(BI_FMA
, 3, nir_type_float
, size
);
246 for (unsigned outmod
= 0; outmod
< 4; ++outmod
) {
247 for (unsigned inmod
= 0; inmod
< 8; ++inmod
) {
249 ins
.src_neg
[0] = (inmod
& 0x1);
250 ins
.src_neg
[1] = (inmod
& 0x2);
251 ins
.src_neg
[2] = (inmod
& 0x4);
252 bit_test_single(dev
, &ins
, input
, true, debug
);
258 bit_fma_mscale_helper(struct panfrost_device
*dev
, uint32_t *input
, enum bit_debug debug
)
260 bi_instruction ins
= bit_ins(BI_FMA
, 4, nir_type_float
, 32);
261 ins
.op
.mscale
= true;
262 ins
.src_types
[3] = nir_type_int32
;
263 ins
.src
[2] = ins
.src
[3]; /* Not enough ports! */
265 for (unsigned outmod
= 0; outmod
< 4; ++outmod
) {
266 for (unsigned inmod
= 0; inmod
< 8; ++inmod
) {
268 ins
.src_abs
[0] = (inmod
& 0x1);
269 ins
.src_neg
[1] = (inmod
& 0x2);
270 ins
.src_neg
[2] = (inmod
& 0x4);
271 bit_test_single(dev
, &ins
, input
, true, debug
);
277 bit_csel_helper(struct panfrost_device
*dev
,
278 unsigned size
, uint32_t *input
, enum bit_debug debug
)
280 bi_instruction ins
= bit_ins(BI_CSEL
, 4, nir_type_uint
, size
);
282 /* SCHEDULER: We can only read 3 registers at once. */
283 ins
.src
[2] = ins
.src
[0];
285 for (enum bi_cond cond
= BI_COND_LT
; cond
<= BI_COND_NE
; ++cond
) {
287 bit_test_single(dev
, &ins
, input
, true, debug
);
292 bit_special_helper(struct panfrost_device
*dev
,
293 unsigned size
, uint32_t *input
, enum bit_debug debug
)
295 bi_instruction ins
= bit_ins(BI_SPECIAL
, 2, nir_type_float
, size
);
296 uint32_t exp_input
[4];
298 for (enum bi_special_op op
= BI_SPECIAL_FRCP
; op
<= BI_SPECIAL_EXP2_LOW
; ++op
) {
299 if (op
== BI_SPECIAL_EXP2_LOW
) {
300 /* exp2 only supported in fp32 mode */
304 /* Give expected input */
305 exp_input
[1] = input
[0];
306 float *ff
= (float *) input
;
307 exp_input
[0] = (int) (ff
[0] * (1 << 24));
310 for (unsigned c
= 0; c
< ((size
== 16) ? 2 : 1); ++c
) {
312 ins
.swizzle
[0][0] = c
;
313 bit_test_single(dev
, &ins
,
314 op
== BI_SPECIAL_EXP2_LOW
? exp_input
: input
,
321 bit_table_helper(struct panfrost_device
*dev
, uint32_t *input
, enum bit_debug debug
)
323 bi_instruction ins
= bit_ins(BI_TABLE
, 1, nir_type_float
, 32);
325 for (enum bi_table_op op
= 0; op
<= BI_TABLE_LOG2_U_OVER_U_1_LOW
; ++op
) {
327 bit_test_single(dev
, &ins
, input
, false, debug
);
332 bit_frexp_helper(struct panfrost_device
*dev
, uint32_t *input
, enum bit_debug debug
)
334 bi_instruction ins
= bit_ins(BI_FREXP
, 1, nir_type_float
, 32);
335 ins
.dest_type
= nir_type_int32
;
337 for (enum bi_frexp_op op
= 0; op
<= BI_FREXPE_LOG
; ++op
) {
339 bit_test_single(dev
, &ins
, input
, true, debug
);
344 bit_round_helper(struct panfrost_device
*dev
, uint32_t *input
, unsigned sz
, bool FMA
, enum bit_debug debug
)
346 bi_instruction ins
= bit_ins(BI_ROUND
, 1, nir_type_float
, sz
);
348 for (enum bifrost_roundmode mode
= 0; mode
<= 3; ++mode
) {
349 BIT_FOREACH_SWIZZLE(swz
, 1, sz
) {
350 bit_apply_swizzle(&ins
, swz
, 1, sz
);
351 ins
.roundmode
= mode
;
352 bit_test_single(dev
, &ins
, input
, FMA
, debug
);
358 bit_reduce_helper(struct panfrost_device
*dev
, uint32_t *input
, enum bit_debug debug
)
360 bi_instruction ins
= bit_ins(BI_REDUCE_FMA
, 2, nir_type_float
, 32);
362 for (enum bi_reduce_op op
= 0; op
<= BI_REDUCE_ADD_FREXPM
; ++op
) {
364 bit_test_single(dev
, &ins
, input
, true, debug
);
369 bit_select_helper(struct panfrost_device
*dev
, uint32_t *input
, unsigned size
, enum bit_debug debug
)
371 unsigned C
= 32 / size
;
372 bi_instruction ins
= bit_ins(BI_SELECT
, C
, nir_type_uint
, 32);
374 for (unsigned c
= 0; c
< C
; ++c
)
375 ins
.src_types
[c
] = nir_type_uint
| size
;
378 /* SCHEDULER: We can only read 3 registers at once. */
379 ins
.src
[2] = ins
.src
[0];
382 /* Each argument has swizzle {lo, hi} so 2^C options */
383 unsigned hi
= (size
== 16) ? 1 : 2;
385 for (unsigned add
= 0; add
< ((size
== 16) ? 2 : 1); ++add
) {
386 for (unsigned swizzle
= 0; swizzle
< (1 << C
); ++swizzle
) {
387 for (unsigned i
= 0; i
< C
; ++i
)
388 ins
.swizzle
[i
][0] = ((swizzle
>> i
) & 1) ? hi
: 0;
390 bit_test_single(dev
, &ins
, input
, !add
, debug
);
396 bit_fcmp_helper(struct panfrost_device
*dev
, uint32_t *input
, unsigned size
, enum bit_debug debug
, bool FMA
)
398 bi_instruction ins
= bit_ins(BI_CMP
, 2, nir_type_float
, size
);
399 ins
.dest_type
= nir_type_uint
| size
;
401 /* 16-bit has swizzles and abs. 32-bit has abs/neg mods. */
402 unsigned max_mods
= (size
== 16) ? 64 : (size
== 32) ? 16 : 1;
404 for (enum bi_cond cond
= BI_COND_LT
; cond
<= BI_COND_NE
; ++cond
) {
405 for (unsigned mods
= 0; mods
< max_mods
; ++mods
) {
409 for (unsigned i
= 0; i
< 2; ++i
) {
410 ins
.swizzle
[i
][0] = ((mods
>> (i
* 2)) & 1) ? 1 : 0;
411 ins
.swizzle
[i
][1] = ((mods
>> (i
* 2)) & 2) ? 1 : 0;
414 ins
.src_abs
[0] = (mods
& 16) ? true : false;
415 ins
.src_abs
[1] = (mods
& 32) ? true : false;
416 } else if (size
== 8) {
417 for (unsigned i
= 0; i
< 2; ++i
) {
418 for (unsigned j
= 0; j
< 4; ++j
)
419 ins
.swizzle
[i
][j
] = j
;
421 } else if (size
== 32) {
422 ins
.src_abs
[0] = (mods
& 1) ? true : false;
423 ins
.src_abs
[1] = (mods
& 2) ? true : false;
424 ins
.src_neg
[0] = (mods
& 4) ? true : false;
425 ins
.src_neg
[1] = (mods
& 8) ? true : false;
428 bit_test_single(dev
, &ins
, input
, FMA
, debug
);
434 bit_icmp_helper(struct panfrost_device
*dev
, uint32_t *input
, unsigned size
, nir_alu_type T
, enum bit_debug debug
)
436 bi_instruction ins
= bit_ins(BI_CMP
, 2, T
, size
);
437 ins
.dest_type
= nir_type_uint
| size
;
439 for (enum bi_cond cond
= BI_COND_LT
; cond
<= BI_COND_NE
; ++cond
) {
440 BIT_FOREACH_SWIZZLE(swz
, 2, size
) {
442 bit_apply_swizzle(&ins
, swz
, 2, size
);
443 bit_test_single(dev
, &ins
, input
, false, debug
);
451 bit_convert_helper(struct panfrost_device
*dev
, unsigned from_size
,
452 unsigned to_size
, unsigned cx
, unsigned cy
, bool FMA
,
453 enum bifrost_roundmode roundmode
,
454 uint32_t *input
, enum bit_debug debug
)
456 bi_instruction ins
= {
458 .dest
= BIR_INDEX_REGISTER
| 0,
459 .src
= { BIR_INDEX_REGISTER
| 0 }
462 nir_alu_type Ts
[3] = { nir_type_float
, nir_type_uint
, nir_type_int
};
464 for (unsigned from_base
= 0; from_base
< 3; ++from_base
) {
465 for (unsigned to_base
= 0; to_base
< 3; ++to_base
) {
466 /* Discard invalid combinations.. */
467 if ((from_size
== to_size
) && (from_base
== to_base
))
470 /* Can't switch signedness */
471 if (from_base
&& to_base
)
474 /* No F16_TO_I32, etc */
475 if (from_size
!= to_size
&& from_base
== 0 && to_base
)
478 if (from_size
!= to_size
&& from_base
&& to_base
== 0)
481 /* No need, just ignore the upper half */
482 if (from_size
> to_size
&& from_base
== to_base
&& from_base
)
485 ins
.dest_type
= Ts
[to_base
] | to_size
;
486 ins
.src_types
[0] = Ts
[from_base
] | from_size
;
487 ins
.roundmode
= roundmode
;
488 ins
.swizzle
[0][0] = cx
;
489 ins
.swizzle
[0][1] = cy
;
491 bit_test_single(dev
, &ins
, input
, FMA
, debug
);
497 bit_constant_helper(struct panfrost_device
*dev
,
498 uint32_t *input
, enum bit_debug debug
)
500 enum bi_class C
[3] = { BI_MOV
, BI_ADD
, BI_FMA
};
502 for (unsigned doubled
= 0; doubled
< 2; ++doubled
) {
503 for (unsigned count
= 1; count
<= 3; ++count
) {
504 bi_instruction ins
= bit_ins(C
[count
- 1], count
, nir_type_float
, 32);
506 ins
.src
[0] = BIR_INDEX_CONSTANT
| 0;
507 ins
.src
[1] = (count
>= 2) ? BIR_INDEX_CONSTANT
| (doubled
? 32 : 0) : 0;
508 ins
.src
[2] = (count
>= 3) ? BIR_INDEX_ZERO
: 0;
510 ins
.constant
.u64
= doubled
?
511 0x3f800000ull
| (0x3f000000ull
<< 32ull) :
514 bit_test_single(dev
, &ins
, input
, true, debug
);
520 bit_bitwise_helper(struct panfrost_device
*dev
, uint32_t *input
, unsigned size
, enum bit_debug debug
)
522 bi_instruction ins
= bit_ins(BI_BITWISE
, 3, nir_type_uint
, size
);
525 ins
.src
[2] = BIR_INDEX_ZERO
;
527 /* Force identity swizzle -- bitwise is not swizzleable */
528 for (unsigned i
= 0; i
< 2; ++i
) {
529 for (unsigned j
= 0; j
< (32 / size
); ++j
)
530 ins
.swizzle
[i
][j
] = j
;
533 for (unsigned op
= BI_BITWISE_AND
; op
<= BI_BITWISE_XOR
; ++op
) {
536 for (unsigned mods
= 0; mods
< 4; ++mods
) {
537 ins
.bitwise
.src_invert
[0] = mods
& 1;
538 ins
.bitwise
.src_invert
[1] = mods
& 2;
539 bit_test_single(dev
, &ins
, input
, true, debug
);
545 bit_packing(struct panfrost_device
*dev
, enum bit_debug debug
)
550 bit_generate_float4(input32
);
551 bit_generate_half8(input16
);
553 bit_constant_helper(dev
, (uint32_t *) input32
, debug
);
555 for (unsigned sz
= 16; sz
<= 32; sz
*= 2) {
557 (sz
== 16) ? (uint32_t *) input16
:
558 (uint32_t *) input32
;
560 bit_fmod_helper(dev
, BI_ADD
, sz
, true, input
, debug
, 0);
561 bit_fmod_helper(dev
, BI_ADD
, sz
, false, input
, debug
, 0);
562 bit_round_helper(dev
, (uint32_t *) input32
, sz
, true, debug
);
564 bit_fmod_helper(dev
, BI_MINMAX
, sz
, false, input
, debug
, BI_MINMAX_MIN
);
565 bit_fmod_helper(dev
, BI_MINMAX
, sz
, false, input
, debug
, BI_MINMAX_MAX
);
567 bit_fma_helper(dev
, sz
, input
, debug
);
568 bit_icmp_helper(dev
, input
, sz
, nir_type_uint
, debug
);
569 bit_icmp_helper(dev
, input
, sz
, nir_type_int
, debug
);
572 for (unsigned sz
= 32; sz
<= 32; sz
*= 2)
573 bit_csel_helper(dev
, sz
, (uint32_t *) input32
, debug
);
575 float special
[4] = { 0.9 };
576 uint32_t special16
[4] = { _mesa_float_to_half(special
[0]) | (_mesa_float_to_half(0.2) << 16) };
578 bit_table_helper(dev
, (uint32_t *) special
, debug
);
580 for (unsigned sz
= 16; sz
<= 32; sz
*= 2) {
582 (sz
== 16) ? special16
:
583 (uint32_t *) special
;
585 bit_special_helper(dev
, sz
, input
, debug
);
588 for (unsigned rm
= 0; rm
< 4; ++rm
) {
589 bit_convert_helper(dev
, 32, 32, 0, 0, false, rm
, (uint32_t *) input32
, debug
);
591 for (unsigned c
= 0; c
< 2; ++c
)
592 bit_convert_helper(dev
, 32, 16, c
, 0, false, rm
, (uint32_t *) input32
, debug
);
594 bit_convert_helper(dev
, 16, 32, 0, 0, false, rm
, (uint32_t *) input16
, debug
);
596 for (unsigned c
= 0; c
< 4; ++c
)
597 bit_convert_helper(dev
, 16, 16, c
& 1, c
>> 1, false, rm
, (uint32_t *) input16
, debug
);
600 bit_frexp_helper(dev
, (uint32_t *) input32
, debug
);
601 bit_reduce_helper(dev
, (uint32_t *) input32
, debug
);
603 uint32_t mscale_input
[4];
604 memcpy(mscale_input
, input32
, sizeof(input32
));
605 mscale_input
[3] = 0x7;
606 bit_fma_mscale_helper(dev
, mscale_input
, debug
);
608 for (unsigned sz
= 8; sz
<= 16; sz
*= 2) {
609 bit_select_helper(dev
, (uint32_t *) input32
, sz
, debug
);
612 bit_fcmp_helper(dev
, (uint32_t *) input32
, 32, debug
, true);
613 bit_fcmp_helper(dev
, (uint32_t *) input32
, 16, debug
, true);
615 for (unsigned sz
= 8; sz
<= 32; sz
*= 2)
616 bit_bitwise_helper(dev
, (uint32_t *) input32
, sz
, debug
);