5 * Copyright (c) 2013 Connor Abbott (connor@abbott.cx)
6 * Copyright (c) 2018 Alyssa Rosenzweig (alyssa@rosenzweig.io)
8 * Permission is hereby granted, free of charge, to any person obtaining a copy
9 * of this software and associated documentation files (the "Software"), to deal
10 * in the Software without restriction, including without limitation the rights
11 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
12 * copies of the Software, and to permit persons to whom the Software is
13 * furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice shall be included in
16 * all copies or substantial portions of the Software.
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
21 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
33 #define MIDGARD_DBG_MSGS 0x0001
34 #define MIDGARD_DBG_SHADERS 0x0002
36 extern int midgard_debug
;
39 midgard_word_type_alu
,
40 midgard_word_type_load_store
,
41 midgard_word_type_texture
,
42 midgard_word_type_unknown
58 midgard_alu_op_fadd
= 0x10,
59 midgard_alu_op_fmul
= 0x14,
61 midgard_alu_op_fmin
= 0x28,
62 midgard_alu_op_fmax
= 0x2C,
64 midgard_alu_op_fmov
= 0x30,
65 midgard_alu_op_froundeven
= 0x34,
66 midgard_alu_op_ftrunc
= 0x35,
67 midgard_alu_op_ffloor
= 0x36,
68 midgard_alu_op_fceil
= 0x37,
69 midgard_alu_op_ffma
= 0x38,
70 midgard_alu_op_fdot3
= 0x3C,
71 midgard_alu_op_fdot3r
= 0x3D,
72 midgard_alu_op_fdot4
= 0x3E,
73 midgard_alu_op_freduce
= 0x3F,
75 midgard_alu_op_iadd
= 0x40,
76 midgard_alu_op_ishladd
= 0x41,
77 midgard_alu_op_isub
= 0x46,
79 midgard_alu_op_imul
= 0x58,
81 midgard_alu_op_imin
= 0x60,
82 midgard_alu_op_umin
= 0x61,
83 midgard_alu_op_imax
= 0x62,
84 midgard_alu_op_umax
= 0x63,
85 midgard_alu_op_iasr
= 0x68,
86 midgard_alu_op_ilsr
= 0x69,
87 midgard_alu_op_ishl
= 0x6E,
89 midgard_alu_op_iand
= 0x70,
90 midgard_alu_op_ior
= 0x71,
91 midgard_alu_op_inand
= 0x72, /* ~(a & b), for inot let a = b */
92 midgard_alu_op_inor
= 0x73, /* ~(a | b) */
93 midgard_alu_op_iandnot
= 0x74, /* (a & ~b), used for not/b2f */
94 midgard_alu_op_iornot
= 0x75, /* (a | ~b) */
95 midgard_alu_op_ixor
= 0x76,
96 midgard_alu_op_inxor
= 0x77, /* ~(a & b) */
97 midgard_alu_op_iclz
= 0x78, /* Number of zeroes on left */
98 midgard_alu_op_ibitcount8
= 0x7A, /* Counts bits in 8-bit increments */
99 midgard_alu_op_imov
= 0x7B,
100 midgard_alu_op_iabs
= 0x7C,
102 midgard_alu_op_feq
= 0x80,
103 midgard_alu_op_fne
= 0x81,
104 midgard_alu_op_flt
= 0x82,
105 midgard_alu_op_fle
= 0x83,
106 midgard_alu_op_fball_eq
= 0x88,
107 midgard_alu_op_bball_eq
= 0x89,
108 midgard_alu_op_fball_lt
= 0x8A, /* all(lessThan(.., ..)) */
109 midgard_alu_op_fball_lte
= 0x8B, /* all(lessThanEqual(.., ..)) */
111 midgard_alu_op_bbany_neq
= 0x90, /* used for bvec4(1) */
112 midgard_alu_op_fbany_neq
= 0x91, /* bvec4(0) also */
113 midgard_alu_op_fbany_lt
= 0x92, /* any(lessThan(.., ..)) */
114 midgard_alu_op_fbany_lte
= 0x93, /* any(lessThanEqual(.., ..)) */
115 midgard_alu_op_f2i
= 0x99,
116 midgard_alu_op_f2u8
= 0x9C,
117 midgard_alu_op_f2u
= 0x9D,
119 midgard_alu_op_ieq
= 0xA0,
120 midgard_alu_op_ine
= 0xA1,
121 midgard_alu_op_ult
= 0xA2,
122 midgard_alu_op_ule
= 0xA3,
123 midgard_alu_op_ilt
= 0xA4,
124 midgard_alu_op_ile
= 0xA5,
125 midgard_alu_op_iball_eq
= 0xA8,
126 midgard_alu_op_iball_neq
= 0xA9,
127 midgard_alu_op_uball_lt
= 0xAA,
128 midgard_alu_op_uball_lte
= 0xAB,
129 midgard_alu_op_iball_lt
= 0xAC,
130 midgard_alu_op_iball_lte
= 0xAD,
132 midgard_alu_op_ibany_eq
= 0xB0,
133 midgard_alu_op_ibany_neq
= 0xB1,
134 midgard_alu_op_ubany_lt
= 0xB2,
135 midgard_alu_op_ubany_lte
= 0xB3,
136 midgard_alu_op_ibany_lt
= 0xB4, /* any(lessThan(.., ..)) */
137 midgard_alu_op_ibany_lte
= 0xB5, /* any(lessThanEqual(.., ..)) */
138 midgard_alu_op_i2f
= 0xB8,
139 midgard_alu_op_u2f
= 0xBC,
141 midgard_alu_op_icsel_v
= 0xC0, /* condition code r31 */
142 midgard_alu_op_icsel
= 0xC1, /* condition code r31.w */
143 midgard_alu_op_fcsel_v
= 0xC4,
144 midgard_alu_op_fcsel
= 0xC5,
145 midgard_alu_op_fround
= 0xC6,
147 midgard_alu_op_fatan_pt2
= 0xE8,
148 midgard_alu_op_fpow_pt1
= 0xEC,
150 midgard_alu_op_frcp
= 0xF0,
151 midgard_alu_op_frsqrt
= 0xF2,
152 midgard_alu_op_fsqrt
= 0xF3,
153 midgard_alu_op_fexp2
= 0xF4,
154 midgard_alu_op_flog2
= 0xF5,
155 midgard_alu_op_fsin
= 0xF6,
156 midgard_alu_op_fcos
= 0xF7,
157 midgard_alu_op_fatan2_pt1
= 0xF9,
161 midgard_outmod_none
= 0,
162 midgard_outmod_pos
= 1,
163 midgard_outmod_int
= 2,
164 midgard_outmod_sat
= 3
168 midgard_reg_mode_8
= 0,
169 midgard_reg_mode_16
= 1,
170 midgard_reg_mode_32
= 2,
171 midgard_reg_mode_64
= 3 /* TODO: verify */
175 midgard_dest_override_lower
= 0,
176 midgard_dest_override_upper
= 1,
177 midgard_dest_override_none
= 2
178 } midgard_dest_override
;
181 midgard_int_sign_extend
= 0,
182 midgard_int_zero_extend
= 1,
183 midgard_int_normal
= 2,
184 midgard_int_reserved
= 3
187 #define MIDGARD_FLOAT_MOD_ABS (1 << 0)
188 #define MIDGARD_FLOAT_MOD_NEG (1 << 1)
191 __attribute__((__packed__
))
193 /* Either midgard_int_mod or from midgard_float_mod_*, depending on the
197 /* replicate lower half if dest = half, or low/high half selection if
201 bool rep_high
: 1; /* unused if dest = full */
202 bool half
: 1; /* only matters if dest = full */
203 unsigned swizzle
: 8;
205 midgard_vector_alu_src
;
208 __attribute__((__packed__
))
210 midgard_alu_op op
: 8;
211 midgard_reg_mode reg_mode
: 2;
214 midgard_dest_override dest_override
: 2;
215 midgard_outmod outmod
: 2;
221 __attribute__((__packed__
))
225 bool full
: 1; /* 0 = half, 1 = full */
226 unsigned component
: 3;
228 midgard_scalar_alu_src
;
231 __attribute__((__packed__
))
233 midgard_alu_op op
: 8;
236 unsigned unknown
: 1;
237 midgard_outmod outmod
: 2;
238 bool output_full
: 1;
239 unsigned output_component
: 3;
244 __attribute__((__packed__
))
246 unsigned src1_reg
: 5;
247 unsigned src2_reg
: 5;
248 unsigned out_reg
: 5;
253 /* In addition to conditional branches and jumps (unconditional branches),
254 * Midgard implements a bit of fixed function functionality used in fragment
255 * shaders via specially crafted branches. These have special branch opcodes,
256 * which perform a fixed-function operation and/or use the results of a
257 * fixed-function operation as the branch condition. */
260 /* Regular branches */
261 midgard_jmp_writeout_op_branch_uncond
= 1,
262 midgard_jmp_writeout_op_branch_cond
= 2,
264 /* In a fragment shader, execute a discard_if instruction, with the
265 * corresponding condition code. Terminates the shader, so generally
266 * set the branch target to out of the shader */
267 midgard_jmp_writeout_op_discard
= 4,
269 /* Branch if the tilebuffer is not yet ready. At the beginning of a
270 * fragment shader that reads from the tile buffer, for instance via
271 * ARM_shader_framebuffer_fetch or EXT_pixel_local_storage, this branch
272 * operation should be used as a loop. An instruction like
273 * "br.tilebuffer.always -1" does the trick, corresponding to
274 * "while(!is_tilebuffer_ready) */
275 midgard_jmp_writeout_op_tilebuffer_pending
= 6,
277 /* In a fragment shader, try to write out the value pushed to r0 to the
278 * tilebuffer, subject to unknown state in r1.z and r1.w. If this
279 * succeeds, the shader terminates. If it fails, it branches to the
280 * specified branch target. Generally, this should be used in a loop to
281 * itself, acting as "do { write(r0); } while(!write_successful);" */
282 midgard_jmp_writeout_op_writeout
= 7,
283 } midgard_jmp_writeout_op
;
286 midgard_condition_write0
= 0,
288 /* These condition codes denote a conditional branch on FALSE and on
289 * TRUE respectively */
290 midgard_condition_false
= 1,
291 midgard_condition_true
= 2,
293 /* This condition code always branches. For a pure branch, the
294 * unconditional branch coding should be used instead, but for
295 * fixed-function branch opcodes, this is still useful */
296 midgard_condition_always
= 3,
300 __attribute__((__packed__
))
302 midgard_jmp_writeout_op op
: 3; /* == branch_uncond */
303 unsigned dest_tag
: 4; /* tag of branch destination */
304 unsigned unknown
: 2;
307 midgard_branch_uncond
;
310 __attribute__((__packed__
))
312 midgard_jmp_writeout_op op
: 3; /* == branch_cond */
313 unsigned dest_tag
: 4; /* tag of branch destination */
315 midgard_condition cond
: 2;
320 __attribute__((__packed__
))
322 midgard_jmp_writeout_op op
: 3; /* == branch_cond */
323 unsigned dest_tag
: 4; /* tag of branch destination */
324 unsigned unknown
: 2;
328 midgard_branch_extended
;
331 __attribute__((__packed__
))
333 midgard_jmp_writeout_op op
: 3; /* == writeout */
334 unsigned unknown
: 13;
343 midgard_op_ld_st_noop
= 0x03,
345 /* Unclear why this is on the L/S unit, but (with an address of 0,
346 * appropriate swizzle, magic constant 0x24, and xy mask?) moves fp32 cube
347 * map coordinates in r27 to its cube map texture coordinate
348 * destination (e.g r29). 0x4 magic for lding from fp16 instead */
350 midgard_op_st_cubemap_coords
= 0x0E,
352 /* Used in OpenCL. Probably can ld other things as well */
353 midgard_op_ld_global_id
= 0x10,
355 /* val in r27.y, address embedded, outputs result to argument. Invert val for sub. Let val = +-1 for inc/dec. */
356 midgard_op_atomic_add
= 0x40,
357 midgard_op_atomic_and
= 0x44,
358 midgard_op_atomic_or
= 0x48,
359 midgard_op_atomic_xor
= 0x4C,
361 midgard_op_atomic_imin
= 0x50,
362 midgard_op_atomic_umin
= 0x54,
363 midgard_op_atomic_imax
= 0x58,
364 midgard_op_atomic_umax
= 0x5C,
366 midgard_op_atomic_xchg
= 0x60,
368 /* Used for compute shader's __global arguments, __local variables (or
369 * for register spilling) */
371 midgard_op_ld_char
= 0x81,
372 midgard_op_ld_char2
= 0x84,
373 midgard_op_ld_short
= 0x85,
374 midgard_op_ld_char4
= 0x88, /* short2, int, float */
375 midgard_op_ld_short4
= 0x8C, /* int2, float2, long */
376 midgard_op_ld_int4
= 0x90, /* float4, long2 */
378 midgard_op_ld_attr_32
= 0x94,
379 midgard_op_ld_attr_16
= 0x95,
380 midgard_op_ld_attr_32i
= 0x97,
381 midgard_op_ld_vary_32
= 0x98,
382 midgard_op_ld_vary_16
= 0x99,
383 midgard_op_ld_vary_32i
= 0x9B,
384 midgard_op_ld_color_buffer_16
= 0x9D,
386 midgard_op_ld_uniform_16
= 0xAC,
388 midgard_op_ld_uniform_32
= 0xB0,
389 midgard_op_ld_color_buffer_8
= 0xBA,
391 midgard_op_st_char
= 0xC0,
392 midgard_op_st_char2
= 0xC4, /* short */
393 midgard_op_st_char4
= 0xC8, /* short2, int, float */
394 midgard_op_st_short4
= 0xCC, /* int2, float2, long */
395 midgard_op_st_int4
= 0xD0, /* float4, long2 */
397 midgard_op_st_vary_32
= 0xD4,
398 midgard_op_st_vary_16
= 0xD5,
399 midgard_op_st_vary_32i
= 0xD7,
401 /* Value to st in r27, location r26.w as short2 */
402 midgard_op_st_image_f
= 0xD8,
403 midgard_op_st_image_ui
= 0xDA,
404 midgard_op_st_image_i
= 0xDB,
405 } midgard_load_store_op
;
408 midgard_interp_centroid
= 1,
409 midgard_interp_default
= 2
410 } midgard_interpolation
;
413 __attribute__((__packed__
))
415 unsigned zero1
: 4; /* Always zero */
417 /* Varying qualifiers, zero if not a varying */
419 unsigned is_varying
: 1; /* Always one for varying, but maybe something else? */
420 midgard_interpolation interpolation
: 2;
422 unsigned zero2
: 2; /* Always zero */
424 midgard_varying_parameter
;
427 __attribute__((__packed__
))
429 midgard_load_store_op op
: 8;
432 unsigned swizzle
: 8;
433 unsigned unknown
: 16;
435 unsigned varying_parameters
: 10;
437 unsigned address
: 9;
439 midgard_load_store_word
;
442 __attribute__((__packed__
))
445 unsigned next_type
: 4;
451 /* Texture pipeline results are in r28-r29 */
452 #define REG_TEX_BASE 28
454 /* Texture opcodes... maybe? */
455 #define TEXTURE_OP_NORMAL 0x11
456 #define TEXTURE_OP_TEXEL_FETCH 0x14
458 /* Texture format types, found in format */
459 #define TEXTURE_CUBE 0x00
460 #define TEXTURE_2D 0x02
461 #define TEXTURE_3D 0x03
464 __attribute__((__packed__
))
467 unsigned next_type
: 4;
471 unsigned unknown3
: 1;
473 /* A little obscure, but last is set for the last texture operation in
474 * a shader. cont appears to just be last's opposite (?). Yeah, I know,
475 * kind of funky.. BiOpen thinks it could do with memory hinting, or
482 unsigned has_offset
: 1;
484 /* Like in Bifrost */
487 unsigned in_reg_select
: 1;
488 unsigned in_reg_upper
: 1;
490 unsigned in_reg_swizzle_left
: 2;
491 unsigned in_reg_swizzle_right
: 2;
493 unsigned unknown1
: 2;
495 unsigned unknown8
: 4;
497 unsigned out_full
: 1;
499 /* Always 1 afaict... */
500 unsigned unknown7
: 2;
502 unsigned out_reg_select
: 1;
503 unsigned out_upper
: 1;
507 unsigned unknown2
: 2;
509 unsigned swizzle
: 8;
510 unsigned unknown4
: 8;
512 unsigned unknownA
: 4;
514 unsigned offset_unknown1
: 1;
515 unsigned offset_reg_select
: 1;
516 unsigned offset_reg_upper
: 1;
517 unsigned offset_unknown4
: 1;
518 unsigned offset_unknown5
: 1;
519 unsigned offset_unknown6
: 1;
520 unsigned offset_unknown7
: 1;
521 unsigned offset_unknown8
: 1;
522 unsigned offset_unknown9
: 1;
524 unsigned unknownB
: 3;
526 /* Texture bias or LOD, depending on whether it is executed in a
527 * fragment/vertex shader respectively. Compute as int(2^8 * biasf).
529 * For texel fetch, this is the LOD as is. */
532 unsigned unknown9
: 8;
534 unsigned texture_handle
: 16;
535 unsigned sampler_handle
: 16;
537 midgard_texture_word
;