2 * Copyright (C) 2019 Connor Abbott <cwabbott0@gmail.com>
3 * Copyright (C) 2019 Lyude Paul <thatslyude@gmail.com>
4 * Copyright (C) 2019 Ryan Houdek <Sonicadvance1@gmail.com>
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 enum bifrost_clause_type
{
33 BIFROST_CLAUSE_NONE
= 0,
34 BIFROST_CLAUSE_LOAD_VARY
= 1,
35 BIFROST_CLAUSE_UBO
= 2,
36 BIFROST_CLAUSE_TEX
= 3,
37 BIFROST_CLAUSE_SSBO_LOAD
= 5,
38 BIFROST_CLAUSE_SSBO_STORE
= 6,
39 BIFROST_CLAUSE_BLEND
= 9,
40 BIFROST_CLAUSE_ATEST
= 13,
41 BIFROST_CLAUSE_64BIT
= 15
44 struct bifrost_header
{
46 // If true, convert any infinite result of any floating-point operation to
47 // the biggest representable number.
48 unsigned suppress_inf
: 1;
49 // Convert any NaN results to 0.
50 unsigned suppress_nan
: 1;
52 // true if the execution mask of the next clause is the same as the mask of
53 // the current clause.
54 unsigned back_to_back
: 1;
55 unsigned no_end_of_shader
: 1;
57 // Set to true for fragment shaders, to implement this bit of spec text
58 // from section 7.1.5 of the GLSL ES spec:
60 // "Stores to image and buffer variables performed by helper invocations
61 // have no effect on the underlying image or buffer memory."
63 // Helper invocations are threads (invocations) corresponding to pixels in
64 // a quad that aren't actually part of the triangle, but are included to
65 // make derivatives work correctly. They're usually turned on, but they
66 // need to be masked off for GLSL-level stores. This bit seems to be the
67 // only bit that's actually different between fragment shaders and other
68 // shaders, so this is probably what it's doing.
69 unsigned elide_writes
: 1;
70 // If backToBack is off:
71 // - true for conditional branches and fallthrough
72 // - false for unconditional branches
73 // The blob seems to always set it to true if back-to-back is on.
74 unsigned branch_cond
: 1;
75 // This bit is set when the next clause writes to the data register of some
77 unsigned datareg_writebarrier
: 1;
79 unsigned scoreboard_deps
: 8;
80 unsigned scoreboard_index
: 3;
81 enum bifrost_clause_type clause_type
: 4;
82 unsigned unk3
: 1; // part of clauseType?
83 enum bifrost_clause_type next_clause_type
: 4;
84 unsigned unk4
: 1; // part of nextClauseType?
85 } __attribute__((packed
));
87 enum bifrost_packed_src
{
88 BIFROST_SRC_PORT0
= 0,
89 BIFROST_SRC_PORT1
= 1,
90 BIFROST_SRC_PORT3
= 2,
91 BIFROST_SRC_STAGE
= 3,
92 BIFROST_SRC_CONST_LO
= 4,
93 BIFROST_SRC_CONST_HI
= 5,
94 BIFROST_SRC_PASS_FMA
= 6,
95 BIFROST_SRC_PASS_ADD
= 7,
98 #define BIFROST_FMA_EXT (0xe0000)
99 #define BIFROST_FMA_OP_MOV BIFROST_FMA_EXT | (0x32d)
101 struct bifrost_fma_inst
{
104 } __attribute__((packed
));
106 struct bifrost_fma_2src
{
110 } __attribute__((packed
));
112 #define BIFROST_ADD_OP_BLEND (0x1952c)
113 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
114 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
116 struct bifrost_add_inst
{
119 } __attribute__((packed
));
121 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
122 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
123 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
124 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
126 struct bifrost_add_2src
{
130 } __attribute__((packed
));
132 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
134 struct bifrost_st_vary
{
138 unsigned channels
: 2;
140 } __attribute__((packed
));
142 #define BIFROST_ADD_OP_ATEST (0xc8f)
144 struct bifrost_add_atest
{
145 /* gl_SampleMask (R60) */
151 /* If half, X/Y select. If !half, always set */
152 unsigned component
: 1;
156 } __attribute__((packed
));
158 enum bifrost_outmod
{
161 BIFROST_SAT_SIGNED
= 0x2,
165 enum bifrost_roundmode
{
166 BIFROST_RTE
= 0x0, /* round to even */
167 BIFROST_RTP
= 0x1, /* round to positive */
168 BIFROST_RTN
= 0x2, /* round to negative */
169 BIFROST_RTZ
= 0x3 /* round to zero */
172 /* NONE: Same as fmax() and fmin() -- return the other
173 * number if any number is NaN. Also always return +0 if
174 * one argument is +0 and the other is -0.
176 * NAN_WINS: Instead of never returning a NaN, always return
177 * one. The "greater"/"lesser" NaN is always returned, first
178 * by checking the sign and then the mantissa bits.
180 * SRC1_WINS: For max, implement src0 > src1 ? src0 : src1.
181 * For min, implement src0 < src1 ? src0 : src1. This
182 * includes handling NaN's and signedness of 0 differently
183 * from above, since +0 and -0 compare equal and comparisons
184 * always return false for NaN's. As a result, this mode is
187 * SRC0_WINS: For max, implement src0 < src1 ? src1 : src0
188 * For min, implement src0 > src1 ? src1 : src0
192 enum bifrost_minmax_mode
{
193 BIFROST_MINMAX_NONE
= 0x0,
194 BIFROST_NAN_WINS
= 0x1,
195 BIFROST_SRC1_WINS
= 0x2,
196 BIFROST_SRC0_WINS
= 0x3,
199 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
201 struct bifrost_fma_add
{
204 unsigned src1_abs
: 1;
205 unsigned src0_neg
: 1;
206 unsigned src1_neg
: 1;
208 unsigned src0_abs
: 1;
209 enum bifrost_roundmode roundmode
: 2;
210 enum bifrost_outmod outmod
: 2;
212 } __attribute__((packed
));
214 #define BIFROST_FMA_OP_FMAX16 (0xC0 >> 2)
215 #define BIFROST_FMA_OP_FMIN16 (0xCC >> 2)
216 #define BIFROST_FMA_OP_FADD16 (0xD8 >> 2)
218 struct bifrost_fma_add_minmax16
{
221 /* abs2 inferred as (src1 < src0) */
223 unsigned src0_neg
: 1;
224 unsigned src1_neg
: 1;
225 unsigned src0_swizzle
: 2;
226 unsigned src1_swizzle
: 2;
228 enum bifrost_outmod outmod
: 2;
229 /* roundmode for add, min/max mode for min/max */
231 } __attribute__((packed
));
233 #define BIFROST_FMA_OP_FMA (0x00)
235 struct bifrost_fma_fma
{
239 unsigned src_expand
: 3;
240 unsigned src0_abs
: 1;
241 enum bifrost_roundmode roundmode
: 2;
242 enum bifrost_outmod outmod
: 2;
243 unsigned src0_neg
: 1; /* 14 */
244 unsigned src2_neg
: 1;
245 unsigned src1_abs
: 1;
246 unsigned src2_abs
: 1; /* 17 */
248 } __attribute__((packed
));
250 #define BIFROST_FMA_OP_FMA16 (0x2)
252 struct bifrost_fma_fma16
{
256 unsigned swizzle_0
: 2;
257 unsigned swizzle_1
: 2;
258 enum bifrost_roundmode roundmode
: 2;
259 enum bifrost_outmod outmod
: 2;
260 unsigned src0_neg
: 1;
261 unsigned src2_neg
: 1;
262 unsigned swizzle_2
: 2;
264 } __attribute__((packed
));
266 enum bifrost_csel_cond
{
277 #define BIFROST_FMA_OP_CSEL4 (0x5c)
278 #define BIFROST_FMA_OP_CSEL4_V16 (0xdc)
280 struct bifrost_csel4
{
285 enum bifrost_csel_cond cond
: 3;
287 } __attribute__((packed
));
289 struct bifrost_shift_fma
{
293 unsigned half
: 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
294 unsigned unk
: 1; /* always set? */
295 unsigned invert_1
: 1; /* Inverts sources to combining op */
296 /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
297 unsigned invert_2
: 1;
299 } __attribute__((packed
));
301 struct bifrost_shift_add
{
307 unsigned invert_1
: 1;
308 unsigned invert_2
: 1;
311 } __attribute__((packed
));
313 #define BIFROST_FMA_INT16_TO_32 (0xe0198 >> 3)
315 struct bifrost_fma_int16_to_32
{
317 unsigned is_unsigned
: 1;
318 unsigned swizzle
: 1;
319 unsigned to_float
: 1;
321 } __attribute__((packed
));
323 /* We could combine but it's easier to just use FMA_ONE_SRC */
324 #define BIFROST_FMA_FLOAT16_TO_32(y) (0xe01a2 | ((y) ? 1 : 0))
326 /* Two sources for vectorization */
327 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
329 /* Again we could combine but easier to just ONE_SRC with an
330 * argumnt for unsignedness */
331 #define BIFROST_FMA_FLOAT32_TO_INT(u) (0xe0136 | ((u) ? 1 : 0))
332 #define BIFROST_FMA_INT_TO_FLOAT32(u) (0xe0178 | ((u) ? 1 : 0))
334 /* Used for f2i16 and i2f16 */
335 #define BIFROST_FMA_F2I16 (0xe00)
337 struct bifrost_fma_f2i_i2f16
{
339 unsigned is_unsigned
: 1;
340 unsigned direction
: 2; /* 00 for i2f, 11 for f2i */
341 unsigned swizzle
: 2;
342 unsigned unk
: 2; /* always 10 */
343 unsigned direction_2
: 1; /* 0 for f2i, 1 for i2f */
345 } __attribute__((packed
));
347 enum bifrost_ldst_type
{
348 BIFROST_LDST_F16
= 0,
349 BIFROST_LDST_F32
= 1,
350 BIFROST_LDST_I32
= 2,
354 #define BIFROST_ADD_OP_LD_VAR_ADDR (0x18000 >> 10)
356 struct bifrost_ld_var_addr
{
359 unsigned location
: 5;
360 enum bifrost_ldst_type type
: 2;
362 } __attribute__((packed
));
364 #define BIFROST_ADD_OP_LD_ATTR (0x08000 >> 12)
366 struct bifrost_ld_attr
{
369 unsigned location
: 5;
370 unsigned channels
: 2; /* MALI_POSITIVE */
371 enum bifrost_ldst_type type
: 2;
373 } __attribute__((packed
));
375 enum bifrost_interp_mode
{
376 BIFROST_INTERP_PER_FRAG
= 0x0,
377 BIFROST_INTERP_CENTROID
= 0x1,
378 BIFROST_INTERP_DEFAULT
= 0x2,
379 BIFROST_INTERP_EXPLICIT
= 0x3
382 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
383 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
385 struct bifrost_ld_var
{
388 /* If top two bits set, indirect with src in bottom three */
391 unsigned channels
: 2; /* MALI_POSITIVE */
392 enum bifrost_interp_mode interp_mode
: 2;
396 } __attribute__((packed
));
398 struct bifrost_tex_ctrl
{
399 unsigned sampler_index
: 4; // also used to signal indirects
400 unsigned tex_index
: 7;
401 bool no_merge_index
: 1; // whether to merge (direct) sampler & texture indices
402 bool filter
: 1; // use the usual filtering pipeline (0 for texelFetch & textureGather)
404 bool texel_offset
: 1; // *Offset()
407 unsigned tex_type
: 2; // 2D, 3D, Cube, Buffer
408 bool compute_lod
: 1; // 0 for *Lod()
409 bool not_supply_lod
: 1; // 0 for *Lod() or when a bias is applied
410 bool calc_gradients
: 1; // 0 for *Grad()
412 unsigned result_type
: 4; // integer, unsigned, float TODO: why is this 4 bits?
414 } __attribute__((packed
));
416 struct bifrost_dual_tex_ctrl
{
417 unsigned sampler_index0
: 2;
419 unsigned tex_index0
: 2;
420 unsigned sampler_index1
: 2;
421 unsigned tex_index1
: 2;
423 } __attribute__((packed
));
425 enum branch_bit_size
{
429 // For the above combinations of bitsize and location, an extra bit is
430 // encoded via comparing the sources. The only possible source of ambiguity
431 // would be if the sources were the same, but then the branch condition
432 // would be always true or always false anyways, so we can ignore it. But
433 // this no longer works when comparing the y component to the x component,
434 // since it's valid to compare the y component of a source against its own
435 // x component. Instead, the extra bit is encoded via an extra bitsize.
438 BR_SIZE_32_AND_16X
= 5,
439 BR_SIZE_32_AND_16Y
= 6,
440 // Used for comparisons with zero and always-true, see below. I think this
441 // only works for integer comparisons.
445 enum bifrost_reg_write_unit
{
446 REG_WRITE_NONE
= 0, // don't write
447 REG_WRITE_TWO
, // write using reg2
448 REG_WRITE_THREE
, // write using reg3
451 struct bifrost_regs
{
452 unsigned uniform_const
: 8;
458 } __attribute__((packed
));
460 enum bifrost_branch_cond
{
465 // Equal vs. not-equal determined by src0/src1 comparison
467 // floating-point comparisons
468 // Becomes UNE when you flip the arguments
470 // TODO what happens when you flip the arguments?
475 enum bifrost_branch_code
{
479 struct bifrost_branch
{
482 /* For BR_SIZE_ZERO, upper two bits become ctrl */
485 /* Offset source -- always uniform/const but
486 * theoretically could support indirect jumps? */
489 enum bifrost_branch_cond cond
: 3;
490 enum branch_bit_size size
: 3;
497 #define BIFROST_FMA_NOP (0x701960 | BIFROST_SRC_STAGE)
498 #define BIFROST_ADD_NOP (0x3D960 | BIFROST_SRC_STAGE)
500 struct bifrost_fmt1
{
505 uint64_t header
: 45;
506 } __attribute__((packed
));
508 #define BIFROST_FMT1_INSTRUCTIONS 0b00101
509 #define BIFROST_FMT1_FINAL 0b01001
510 #define BIFROST_FMT1_CONSTANTS 0b00001
512 #define BIFROST_FMTC_CONSTANTS 0b0011
513 #define BIFROST_FMTC_FINAL 0b0111
515 struct bifrost_fmt_constant
{
520 } __attribute__((packed
));
522 enum bifrost_reg_control
{
523 BIFROST_WRITE_FMA_P2
= 1,
524 BIFROST_WRITE_FMA_P2_READ_P3
= 2,
525 BIFROST_FIRST_WRITE_FMA_P2_READ_P3
= 3,
527 BIFROST_WRITE_ADD_P2
= 5,
528 BIFROST_WRITE_ADD_P2_READ_P3
= 6,
529 BIFROST_WRITE_ADD_P2_FMA_P3
= 7,
531 BIFROST_FIRST_NONE
= 8,
532 BIFROST_FIRST_WRITE_FMA_P2
= 9,
533 /* INSTR_INVALID_ENC */
534 BIFROST_REG_NONE
= 11,
535 BIFROST_FIRST_READ_P3
= 12,
536 BIFROST_FIRST_WRITE_ADD_P2
= 13,
537 BIFROST_FIRST_WRITE_ADD_P2_READ_P3
= 14,
538 BIFROST_FIRST_WRITE_ADD_P2_FMA_P3
= 15