• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * © Copyright 2017-2018 Alyssa Rosenzweig
3  * © Copyright 2017-2018 Connor Abbott
4  * © Copyright 2017-2018 Lyude Paul
5  *
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:
12  *
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
15  * Software.
16  *
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
23  * SOFTWARE.
24  *
25  */
26 
27 #ifndef __PANFROST_JOB_H__
28 #define __PANFROST_JOB_H__
29 
30 #include <stdint.h>
31 #include <panfrost-misc.h>
32 
33 #define MALI_SHORT_PTR_BITS (sizeof(uintptr_t)*8)
34 
35 #define MALI_FBD_HIERARCHY_WEIGHTS 8
36 
37 #define MALI_PAYLOAD_SIZE 256
38 
39 typedef u32 mali_jd_core_req;
40 
41 enum mali_job_type {
42         JOB_NOT_STARTED	= 0,
43         JOB_TYPE_NULL = 1,
44         JOB_TYPE_SET_VALUE = 2,
45         JOB_TYPE_CACHE_FLUSH = 3,
46         JOB_TYPE_COMPUTE = 4,
47         JOB_TYPE_VERTEX = 5,
48         JOB_TYPE_GEOMETRY = 6,
49         JOB_TYPE_TILER = 7,
50         JOB_TYPE_FUSED = 8,
51         JOB_TYPE_FRAGMENT = 9,
52 };
53 
54 enum mali_draw_mode {
55         MALI_DRAW_NONE      = 0x0,
56         MALI_POINTS         = 0x1,
57         MALI_LINES          = 0x2,
58         MALI_LINE_STRIP     = 0x4,
59         MALI_LINE_LOOP      = 0x6,
60         MALI_TRIANGLES      = 0x8,
61         MALI_TRIANGLE_STRIP = 0xA,
62         MALI_TRIANGLE_FAN   = 0xC,
63         MALI_POLYGON        = 0xD,
64         MALI_QUADS          = 0xE,
65         MALI_QUAD_STRIP     = 0xF,
66 
67         /* All other modes invalid */
68 };
69 
70 /* Applies to tiler_gl_enables */
71 
72 
73 #define MALI_OCCLUSION_QUERY    (1 << 3)
74 #define MALI_OCCLUSION_PRECISE  (1 << 4)
75 
76 #define MALI_FRONT_FACE(v)      (v << 5)
77 #define MALI_CCW (0)
78 #define MALI_CW  (1)
79 
80 #define MALI_CULL_FACE_FRONT    (1 << 6)
81 #define MALI_CULL_FACE_BACK     (1 << 7)
82 
83 /* TODO: Might this actually be a finer bitfield? */
84 #define MALI_DEPTH_STENCIL_ENABLE 0x6400
85 
86 #define DS_ENABLE(field) \
87 	(field == MALI_DEPTH_STENCIL_ENABLE) \
88 	? "MALI_DEPTH_STENCIL_ENABLE" \
89 	: (field == 0) ? "0" \
90 	: "0 /* XXX: Unknown, check hexdump */"
91 
92 /* Used in stencil and depth tests */
93 
94 enum mali_func {
95         MALI_FUNC_NEVER    = 0,
96         MALI_FUNC_LESS     = 1,
97         MALI_FUNC_EQUAL    = 2,
98         MALI_FUNC_LEQUAL   = 3,
99         MALI_FUNC_GREATER  = 4,
100         MALI_FUNC_NOTEQUAL = 5,
101         MALI_FUNC_GEQUAL   = 6,
102         MALI_FUNC_ALWAYS   = 7
103 };
104 
105 /* Same OpenGL, but mixed up. Why? Because forget me, that's why! */
106 
107 enum mali_alt_func {
108         MALI_ALT_FUNC_NEVER    = 0,
109         MALI_ALT_FUNC_GREATER  = 1,
110         MALI_ALT_FUNC_EQUAL    = 2,
111         MALI_ALT_FUNC_GEQUAL   = 3,
112         MALI_ALT_FUNC_LESS     = 4,
113         MALI_ALT_FUNC_NOTEQUAL = 5,
114         MALI_ALT_FUNC_LEQUAL   = 6,
115         MALI_ALT_FUNC_ALWAYS   = 7
116 };
117 
118 /* Flags apply to unknown2_3? */
119 
120 #define MALI_HAS_MSAA		(1 << 0)
121 #define MALI_CAN_DISCARD 	(1 << 5)
122 
123 /* Applies on SFBD systems, specifying that programmable blending is in use */
124 #define MALI_HAS_BLEND_SHADER 	(1 << 6)
125 
126 /* func is mali_func */
127 #define MALI_DEPTH_FUNC(func)	   (func << 8)
128 #define MALI_GET_DEPTH_FUNC(flags) ((flags >> 8) & 0x7)
129 #define MALI_DEPTH_FUNC_MASK	   MALI_DEPTH_FUNC(0x7)
130 
131 #define MALI_DEPTH_TEST		(1 << 11)
132 
133 /* Next flags to unknown2_4 */
134 #define MALI_STENCIL_TEST      	(1 << 0)
135 
136 /* What?! */
137 #define MALI_SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER (1 << 1)
138 
139 #define MALI_NO_DITHER		(1 << 9)
140 #define MALI_DEPTH_RANGE_A	(1 << 12)
141 #define MALI_DEPTH_RANGE_B	(1 << 13)
142 #define MALI_NO_MSAA		(1 << 14)
143 
144 /* Stencil test state is all encoded in a single u32, just with a lot of
145  * enums... */
146 
147 enum mali_stencil_op {
148         MALI_STENCIL_KEEP 	= 0,
149         MALI_STENCIL_REPLACE 	= 1,
150         MALI_STENCIL_ZERO 	= 2,
151         MALI_STENCIL_INVERT 	= 3,
152         MALI_STENCIL_INCR_WRAP 	= 4,
153         MALI_STENCIL_DECR_WRAP 	= 5,
154         MALI_STENCIL_INCR 	= 6,
155         MALI_STENCIL_DECR 	= 7
156 };
157 
158 struct mali_stencil_test {
159         unsigned ref  			: 8;
160         unsigned mask 			: 8;
161         enum mali_func func 		: 3;
162         enum mali_stencil_op sfail 	: 3;
163         enum mali_stencil_op dpfail 	: 3;
164         enum mali_stencil_op dppass 	: 3;
165         unsigned zero			: 4;
166 } __attribute__((packed));
167 
168 /* Blending is a mess, since anything fancy triggers a blend shader, and
169  * -those- are not understood whatsover yet */
170 
171 #define MALI_MASK_R (1 << 0)
172 #define MALI_MASK_G (1 << 1)
173 #define MALI_MASK_B (1 << 2)
174 #define MALI_MASK_A (1 << 3)
175 
176 enum mali_nondominant_mode {
177         MALI_BLEND_NON_MIRROR = 0,
178         MALI_BLEND_NON_ZERO = 1
179 };
180 
181 enum mali_dominant_blend {
182         MALI_BLEND_DOM_SOURCE = 0,
183         MALI_BLEND_DOM_DESTINATION  = 1
184 };
185 
186 enum mali_dominant_factor {
187         MALI_DOMINANT_UNK0 = 0,
188         MALI_DOMINANT_ZERO = 1,
189         MALI_DOMINANT_SRC_COLOR = 2,
190         MALI_DOMINANT_DST_COLOR = 3,
191         MALI_DOMINANT_UNK4 = 4,
192         MALI_DOMINANT_SRC_ALPHA = 5,
193         MALI_DOMINANT_DST_ALPHA = 6,
194         MALI_DOMINANT_CONSTANT = 7,
195 };
196 
197 enum mali_blend_modifier {
198         MALI_BLEND_MOD_UNK0 = 0,
199         MALI_BLEND_MOD_NORMAL = 1,
200         MALI_BLEND_MOD_SOURCE_ONE = 2,
201         MALI_BLEND_MOD_DEST_ONE = 3,
202 };
203 
204 struct mali_blend_mode {
205         enum mali_blend_modifier clip_modifier : 2;
206         unsigned unused_0 : 1;
207         unsigned negate_source : 1;
208 
209         enum mali_dominant_blend dominant : 1;
210 
211         enum mali_nondominant_mode nondominant_mode : 1;
212 
213         unsigned unused_1 : 1;
214 
215         unsigned negate_dest : 1;
216 
217         enum mali_dominant_factor dominant_factor : 3;
218         unsigned complement_dominant : 1;
219 } __attribute__((packed));
220 
221 struct mali_blend_equation {
222         /* Of type mali_blend_mode */
223         unsigned rgb_mode : 12;
224         unsigned alpha_mode : 12;
225 
226         unsigned zero1 : 4;
227 
228         /* Corresponds to MALI_MASK_* above and glColorMask arguments */
229 
230         unsigned color_mask : 4;
231 
232         /* Attached constant for CONSTANT_ALPHA, etc */
233 
234 #ifndef BIFROST
235         float constant;
236 #endif
237 } __attribute__((packed));
238 
239 /* Used with channel swizzling */
240 enum mali_channel {
241 	MALI_CHANNEL_RED = 0,
242 	MALI_CHANNEL_GREEN = 1,
243 	MALI_CHANNEL_BLUE = 2,
244 	MALI_CHANNEL_ALPHA = 3,
245 	MALI_CHANNEL_ZERO = 4,
246 	MALI_CHANNEL_ONE = 5,
247 	MALI_CHANNEL_RESERVED_0 = 6,
248 	MALI_CHANNEL_RESERVED_1 = 7,
249 };
250 
251 struct mali_channel_swizzle {
252 	enum mali_channel r : 3;
253 	enum mali_channel g : 3;
254 	enum mali_channel b : 3;
255 	enum mali_channel a : 3;
256 } __attribute__((packed));
257 
258 /* Compressed per-pixel formats. Each of these formats expands to one to four
259  * floating-point or integer numbers, as defined by the OpenGL specification.
260  * There are various places in OpenGL where the user can specify a compressed
261  * format in memory, which all use the same 8-bit enum in the various
262  * descriptors, although different hardware units support different formats.
263  */
264 
265 /* The top 3 bits specify how the bits of each component are interpreted. */
266 
267 /* e.g. R11F_G11F_B10F */
268 #define MALI_FORMAT_SPECIAL (2 << 5)
269 
270 /* signed normalized, e.g. RGBA8_SNORM */
271 #define MALI_FORMAT_SNORM (3 << 5)
272 
273 /* e.g. RGBA8UI */
274 #define MALI_FORMAT_UINT (4 << 5)
275 
276 /* e.g. RGBA8 and RGBA32F */
277 #define MALI_FORMAT_UNORM (5 << 5)
278 
279 /* e.g. RGBA8I and RGBA16F */
280 #define MALI_FORMAT_SINT (6 << 5)
281 
282 /* These formats seem to largely duplicate the others. They're used at least
283  * for Bifrost framebuffer output.
284  */
285 #define MALI_FORMAT_SPECIAL2 (7 << 5)
286 
287 /* If the high 3 bits are 3 to 6 these two bits say how many components
288  * there are.
289  */
290 #define MALI_NR_CHANNELS(n) ((n - 1) << 3)
291 
292 /* If the high 3 bits are 3 to 6, then the low 3 bits say how big each
293  * component is, except the special MALI_CHANNEL_FLOAT which overrides what the
294  * bits mean.
295  */
296 
297 #define MALI_CHANNEL_4 2
298 
299 #define MALI_CHANNEL_8 3
300 
301 #define MALI_CHANNEL_16 4
302 
303 #define MALI_CHANNEL_32 5
304 
305 /* For MALI_FORMAT_SINT it means a half-float (e.g. RG16F). For
306  * MALI_FORMAT_UNORM, it means a 32-bit float.
307  */
308 #define MALI_CHANNEL_FLOAT 7
309 
310 enum mali_format {
311 	MALI_RGB565         = MALI_FORMAT_SPECIAL | 0x0,
312 	MALI_RGB5_A1_UNORM  = MALI_FORMAT_SPECIAL | 0x2,
313 	MALI_RGB10_A2_UNORM = MALI_FORMAT_SPECIAL | 0x3,
314 	MALI_RGB10_A2_SNORM = MALI_FORMAT_SPECIAL | 0x5,
315 	MALI_RGB10_A2UI     = MALI_FORMAT_SPECIAL | 0x7,
316 	MALI_RGB10_A2I      = MALI_FORMAT_SPECIAL | 0x9,
317 
318 	/* YUV formats */
319 	MALI_NV12           = MALI_FORMAT_SPECIAL | 0xc,
320 
321 	MALI_Z32_UNORM      = MALI_FORMAT_SPECIAL | 0xD,
322 	MALI_R32_FIXED      = MALI_FORMAT_SPECIAL | 0x11,
323 	MALI_RG32_FIXED     = MALI_FORMAT_SPECIAL | 0x12,
324 	MALI_RGB32_FIXED    = MALI_FORMAT_SPECIAL | 0x13,
325 	MALI_RGBA32_FIXED   = MALI_FORMAT_SPECIAL | 0x14,
326 	MALI_R11F_G11F_B10F = MALI_FORMAT_SPECIAL | 0x19,
327 	/* Only used for varyings, to indicate the transformed gl_Position */
328 	MALI_VARYING_POS    = MALI_FORMAT_SPECIAL | 0x1e,
329 	/* Only used for varyings, to indicate that the write should be
330 	 * discarded.
331 	 */
332 	MALI_VARYING_DISCARD = MALI_FORMAT_SPECIAL | 0x1f,
333 
334 	MALI_R8_SNORM     = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
335 	MALI_R16_SNORM    = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
336 	MALI_R32_SNORM    = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
337 	MALI_RG8_SNORM    = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
338 	MALI_RG16_SNORM   = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
339 	MALI_RG32_SNORM   = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
340 	MALI_RGB8_SNORM   = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
341 	MALI_RGB16_SNORM  = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
342 	MALI_RGB32_SNORM  = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
343 	MALI_RGBA8_SNORM  = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
344 	MALI_RGBA16_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
345 	MALI_RGBA32_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
346 
347 	MALI_R8UI     = MALI_FORMAT_UINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
348 	MALI_R16UI    = MALI_FORMAT_UINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
349 	MALI_R32UI    = MALI_FORMAT_UINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
350 	MALI_RG8UI    = MALI_FORMAT_UINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
351 	MALI_RG16UI   = MALI_FORMAT_UINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
352 	MALI_RG32UI   = MALI_FORMAT_UINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
353 	MALI_RGB8UI   = MALI_FORMAT_UINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
354 	MALI_RGB16UI  = MALI_FORMAT_UINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
355 	MALI_RGB32UI  = MALI_FORMAT_UINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
356 	MALI_RGBA8UI  = MALI_FORMAT_UINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
357 	MALI_RGBA16UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
358 	MALI_RGBA32UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
359 
360 	MALI_R8_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
361 	MALI_R16_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
362 	MALI_R32_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
363 	MALI_R32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT,
364 	MALI_RG8_UNORM    = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
365 	MALI_RG16_UNORM   = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
366 	MALI_RG32_UNORM   = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
367 	MALI_RG32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT,
368 	MALI_RGB8_UNORM   = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
369 	MALI_RGB16_UNORM  = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
370 	MALI_RGB32_UNORM  = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
371 	MALI_RGB32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT,
372 	MALI_RGBA4_UNORM  = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_4,
373 	MALI_RGBA8_UNORM  = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
374 	MALI_RGBA16_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
375 	MALI_RGBA32_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
376 	MALI_RGBA32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT,
377 
378 	MALI_R8I     = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
379 	MALI_R16I    = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
380 	MALI_R32I    = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
381 	MALI_R16F    = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT,
382 	MALI_RG8I    = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
383 	MALI_RG16I   = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
384 	MALI_RG32I   = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
385 	MALI_RG16F   = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT,
386 	MALI_RGB8I   = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
387 	MALI_RGB16I  = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
388 	MALI_RGB32I  = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
389 	MALI_RGB16F  = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT,
390 	MALI_RGBA8I  = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
391 	MALI_RGBA16I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
392 	MALI_RGBA32I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
393 	MALI_RGBA16F = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT,
394 
395 	MALI_RGBA4      = MALI_FORMAT_SPECIAL2 | 0x8,
396 	MALI_RGBA8_2    = MALI_FORMAT_SPECIAL2 | 0xd,
397 	MALI_RGB10_A2_2 = MALI_FORMAT_SPECIAL2 | 0xe,
398 };
399 
400 
401 /* Alpha coverage is encoded as 4-bits (from a clampf), with inversion
402  * literally performing a bitwise invert. This function produces slightly wrong
403  * results and I'm not sure why; some rounding issue I suppose... */
404 
405 #define MALI_ALPHA_COVERAGE(clampf) ((uint16_t) (int) (clampf * 15.0f))
406 #define MALI_GET_ALPHA_COVERAGE(nibble) ((float) nibble / 15.0f)
407 
408 /* Applies to unknown1 */
409 #define MALI_NO_ALPHA_TO_COVERAGE (1 << 10)
410 
411 struct mali_blend_meta {
412 #ifndef BIFROST
413         /* Base value of 0x200.
414          * OR with 0x1 for blending (anything other than REPLACE).
415          * OR with 0x2 for programmable blending
416          */
417 
418         u64 unk1;
419 
420         /* For programmable blending, these turn into the blend_shader address */
421         struct mali_blend_equation blend_equation_1;
422 
423         u64 zero2;
424         struct mali_blend_equation blend_equation_2;
425 #else
426         u32 unk1; // = 0x200
427         struct mali_blend_equation blend_equation;
428         /*
429          * - 0x19 normally
430          * - 0x3 when this slot is unused (everything else is 0 except the index)
431          * - 0x11 when this is the fourth slot (and it's used)
432 +	 * - 0 when there is a blend shader
433          */
434         u16 unk2;
435         /* increments from 0 to 3 */
436         u16 index;
437 
438 	union {
439 		struct {
440 			/* So far, I've only seen:
441 			 * - R001 for 1-component formats
442 			 * - RG01 for 2-component formats
443 			 * - RGB1 for 3-component formats
444 			 * - RGBA for 4-component formats
445 			 */
446 			u32 swizzle : 12;
447 			enum mali_format format : 8;
448 
449 			/* Type of the shader output variable. Note, this can
450 			 * be different from the format.
451 			 *
452 			 * 0: f16 (mediump float)
453 			 * 1: f32 (highp float)
454 			 * 2: i32 (highp int)
455 			 * 3: u32 (highp uint)
456 			 * 4: i16 (mediump int)
457 			 * 5: u16 (mediump uint)
458 			 */
459 			u32 shader_type : 3;
460 			u32 zero : 9;
461 		};
462 
463 		/* Only the low 32 bits of the blend shader are stored, the
464 		 * high 32 bits are implicitly the same as the original shader.
465 		 * According to the kernel driver, the program counter for
466 		 * shaders is actually only 24 bits, so shaders cannot cross
467 		 * the 2^24-byte boundary, and neither can the blend shader.
468 		 * The blob handles this by allocating a 2^24 byte pool for
469 		 * shaders, and making sure that any blend shaders are stored
470 		 * in the same pool as the original shader. The kernel will
471 		 * make sure this allocation is aligned to 2^24 bytes.
472 		 */
473 		u32 blend_shader;
474 	};
475 #endif
476 } __attribute__((packed));
477 
478 struct mali_shader_meta {
479         mali_ptr shader;
480         u16 texture_count;
481         u16 sampler_count;
482         u16 attribute_count;
483         u16 varying_count;
484 
485         union {
486                 struct {
487                         u32 uniform_buffer_count : 4;
488                         u32 unk1 : 28; // = 0x800000 for vertex, 0x958020 for tiler
489                 } bifrost1;
490                 struct {
491                         /* 0x200 except MALI_NO_ALPHA_TO_COVERAGE. Mysterious 1
492                          * other times. Who knows really? */
493                         u16 unknown1;
494 
495                         /* Whole number of uniform registers used, times two;
496                          * whole number of work registers used (no scale).
497                          */
498                         unsigned work_count : 5;
499                         unsigned uniform_count : 5;
500                         unsigned unknown2 : 6;
501                 } midgard1;
502         };
503 
504         /* On bifrost: Exactly the same as glPolygonOffset() for both.
505          * On midgard: Depth factor is exactly as passed to glPolygonOffset.
506          * Depth units is equal to the value passed to glDeptOhffset + 1.0f
507          * (use MALI_NEGATIVE)
508          */
509         float depth_units;
510         float depth_factor;
511 
512         u32 unknown2_2;
513 
514         u16 alpha_coverage;
515         u16 unknown2_3;
516 
517         u8 stencil_mask_front;
518         u8 stencil_mask_back;
519         u16 unknown2_4;
520 
521         struct mali_stencil_test stencil_front;
522         struct mali_stencil_test stencil_back;
523 
524         union {
525                 struct {
526                         u32 unk3 : 7;
527                         /* On Bifrost, some system values are preloaded in
528                          * registers R55-R62 by the thread dispatcher prior to
529                          * the start of shader execution. This is a bitfield
530                          * with one entry for each register saying which
531                          * registers need to be preloaded. Right now, the known
532                          * values are:
533                          *
534                          * Vertex/compute:
535                          * - R55 : gl_LocalInvocationID.xy
536                          * - R56 : gl_LocalInvocationID.z + unknown in high 16 bits
537                          * - R57 : gl_WorkGroupID.x
538                          * - R58 : gl_WorkGroupID.y
539                          * - R59 : gl_WorkGroupID.z
540                          * - R60 : gl_GlobalInvocationID.x
541                          * - R61 : gl_GlobalInvocationID.y/gl_VertexID (without base)
542                          * - R62 : gl_GlobalInvocationID.z/gl_InstanceID (without base)
543                          *
544                          * Fragment:
545                          * - R55 : unknown, never seen (but the bit for this is
546                          *   always set?)
547                          * - R56 : unknown (bit always unset)
548                          * - R57 : gl_PrimitiveID
549                          * - R58 : gl_FrontFacing in low bit, potentially other stuff
550                          * - R59 : u16 fragment coordinates (used to compute
551                          *   gl_FragCoord.xy, together with sample positions)
552                          * - R60 : gl_SampleMask (used in epilog, so pretty
553                          *   much always used, but the bit is always 0 -- is
554                          *   this just always pushed?)
555                          * - R61 : gl_SampleMaskIn and gl_SampleID, used by
556                          *   varying interpolation.
557                          * - R62 : unknown (bit always unset).
558                          */
559                         u32 preload_regs : 8;
560                         /* In units of 8 bytes or 64 bits, since the
561                          * uniform/const port loads 64 bits at a time.
562                          */
563                         u32 uniform_count : 7;
564                         u32 unk4 : 10; // = 2
565                 } bifrost2;
566                 struct {
567                         u32 unknown2_7;
568                 } midgard2;
569         };
570 
571         /* zero on bifrost */
572         u32 unknown2_8;
573 
574         /* Blending information for the older non-MRT Midgard HW. Check for
575          * MALI_HAS_BLEND_SHADER to decide how to interpret.
576          */
577 
578         union {
579                 mali_ptr blend_shader;
580                 struct mali_blend_equation blend_equation;
581         };
582 
583         /* There can be up to 4 blend_meta's. None of them are required for
584          * vertex shaders or the non-MRT case for Midgard (so the blob doesn't
585          * allocate any space).
586          */
587         struct mali_blend_meta blend_meta[];
588 
589 } __attribute__((packed));
590 
591 /* This only concerns hardware jobs */
592 
593 /* Possible values for job_descriptor_size */
594 
595 #define MALI_JOB_32 0
596 #define MALI_JOB_64 1
597 
598 struct mali_job_descriptor_header {
599         u32 exception_status;
600         u32 first_incomplete_task;
601         u64 fault_pointer;
602         u8 job_descriptor_size : 1;
603         enum mali_job_type job_type : 7;
604         u8 job_barrier : 1;
605         u8 unknown_flags : 7;
606         u16 job_index;
607         u16 job_dependency_index_1;
608         u16 job_dependency_index_2;
609 
610         union {
611                 u64 next_job_64;
612                 u32 next_job_32;
613         };
614 } __attribute__((packed));
615 
616 struct mali_payload_set_value {
617         u64 out;
618         u64 unknown;
619 } __attribute__((packed));
620 
621 /* Special attributes have a fixed index */
622 #define MALI_SPECIAL_ATTRIBUTE_BASE 16
623 #define MALI_VERTEX_ID   (MALI_SPECIAL_ATTRIBUTE_BASE + 0)
624 #define MALI_INSTANCE_ID (MALI_SPECIAL_ATTRIBUTE_BASE + 1)
625 
626 /*
627  * Mali Attributes
628  *
629  * This structure lets the attribute unit compute the address of an attribute
630  * given the vertex and instance ID. Unfortunately, the way this works is
631  * rather complicated when instancing is enabled.
632  *
633  * To explain this, first we need to explain how compute and vertex threads are
634  * dispatched. This is a guess (although a pretty firm guess!) since the
635  * details are mostly hidden from the driver, except for attribute instancing.
636  * When a quad is dispatched, it receives a single, linear index. However, we
637  * need to translate that index into a (vertex id, instance id) pair, or a
638  * (local id x, local id y, local id z) triple for compute shaders (although
639  * vertex shaders and compute shaders are handled almost identically).
640  * Focusing on vertex shaders, one option would be to do:
641  *
642  * vertex_id = linear_id % num_vertices
643  * instance_id = linear_id / num_vertices
644  *
645  * but this involves a costly division and modulus by an arbitrary number.
646  * Instead, we could pad num_vertices. We dispatch padded_num_vertices *
647  * num_instances threads instead of num_vertices * num_instances, which results
648  * in some "extra" threads with vertex_id >= num_vertices, which we have to
649  * discard.  The more we pad num_vertices, the more "wasted" threads we
650  * dispatch, but the division is potentially easier.
651  *
652  * One straightforward choice is to pad num_vertices to the next power of two,
653  * which means that the division and modulus are just simple bit shifts and
654  * masking. But the actual algorithm is a bit more complicated. The thread
655  * dispatcher has special support for dividing by 3, 5, 7, and 9, in addition
656  * to dividing by a power of two. This is possibly using the technique
657  * described in patent US20170010862A1. As a result, padded_num_vertices can be
658  * 1, 3, 5, 7, or 9 times a power of two. This results in less wasted threads,
659  * since we need less padding.
660  *
661  * padded_num_vertices is picked by the hardware. The driver just specifies the
662  * actual number of vertices. At least for Mali G71, the first few cases are
663  * given by:
664  *
665  * num_vertices	| padded_num_vertices
666  * 3		| 4
667  * 4-7		| 8
668  * 8-11		| 12 (3 * 4)
669  * 12-15	| 16
670  * 16-19	| 20 (5 * 4)
671  *
672  * Note that padded_num_vertices is a multiple of four (presumably because
673  * threads are dispatched in groups of 4). Also, padded_num_vertices is always
674  * at least one more than num_vertices, which seems like a quirk of the
675  * hardware. For larger num_vertices, the hardware uses the following
676  * algorithm: using the binary representation of num_vertices, we look at the
677  * most significant set bit as well as the following 3 bits. Let n be the
678  * number of bits after those 4 bits. Then we set padded_num_vertices according
679  * to the following table:
680  *
681  * high bits	| padded_num_vertices
682  * 1000		| 9 * 2^n
683  * 1001		| 5 * 2^(n+1)
684  * 101x		| 3 * 2^(n+2)
685  * 110x		| 7 * 2^(n+1)
686  * 111x		| 2^(n+4)
687  *
688  * For example, if num_vertices = 70 is passed to glDraw(), its binary
689  * representation is 1000110, so n = 3 and the high bits are 1000, and
690  * therefore padded_num_vertices = 9 * 2^3 = 72.
691  *
692  * The attribute unit works in terms of the original linear_id. if
693  * num_instances = 1, then they are the same, and everything is simple.
694  * However, with instancing things get more complicated. There are four
695  * possible modes, two of them we can group together:
696  *
697  * 1. Use the linear_id directly. Only used when there is no instancing.
698  *
699  * 2. Use the linear_id modulo a constant. This is used for per-vertex
700  * attributes with instancing enabled by making the constant equal
701  * padded_num_vertices. Because the modulus is always padded_num_vertices, this
702  * mode only supports a modulus that is a power of 2 times 1, 3, 5, 7, or 9.
703  * The shift field specifies the power of two, while the extra_flags field
704  * specifies the odd number. If shift = n and extra_flags = m, then the modulus
705  * is (2m + 1) * 2^n. As an example, if num_vertices = 70, then as computed
706  * above, padded_num_vertices = 9 * 2^3, so we should set extra_flags = 4 and
707  * shift = 3. Note that we must exactly follow the hardware algorithm used to
708  * get padded_num_vertices in order to correctly implement per-vertex
709  * attributes.
710  *
711  * 3. Divide the linear_id by a constant. In order to correctly implement
712  * instance divisors, we have to divide linear_id by padded_num_vertices times
713  * to user-specified divisor. So first we compute padded_num_vertices, again
714  * following the exact same algorithm that the hardware uses, then multiply it
715  * by the GL-level divisor to get the hardware-level divisor. This case is
716  * further divided into two more cases. If the hardware-level divisor is a
717  * power of two, then we just need to shift. The shift amount is specified by
718  * the shift field, so that the hardware-level divisor is just 2^shift.
719  *
720  * If it isn't a power of two, then we have to divide by an arbitrary integer.
721  * For that, we use the well-known technique of multiplying by an approximation
722  * of the inverse. The driver must compute the magic multiplier and shift
723  * amount, and then the hardware does the multiplication and shift. The
724  * hardware and driver also use the "round-down" optimization as described in
725  * http://ridiculousfish.com/files/faster_unsigned_division_by_constants.pdf.
726  * The hardware further assumes the multiplier is between 2^31 and 2^32, so the
727  * high bit is implicitly set to 1 even though it is set to 0 by the driver --
728  * presumably this simplifies the hardware multiplier a little. The hardware
729  * first multiplies linear_id by the multiplier and takes the high 32 bits,
730  * then applies the round-down correction if extra_flags = 1, then finally
731  * shifts right by the shift field.
732  *
733  * There are some differences between ridiculousfish's algorithm and the Mali
734  * hardware algorithm, which means that the reference code from ridiculousfish
735  * doesn't always produce the right constants. Mali does not use the pre-shift
736  * optimization, since that would make a hardware implementation slower (it
737  * would have to always do the pre-shift, multiply, and post-shift operations).
738  * It also forces the multplier to be at least 2^31, which means that the
739  * exponent is entirely fixed, so there is no trial-and-error. Altogether,
740  * given the divisor d, the algorithm the driver must follow is:
741  *
742  * 1. Set shift = floor(log2(d)).
743  * 2. Compute m = ceil(2^(shift + 32) / d) and e = 2^(shift + 32) % d.
744  * 3. If e <= 2^shift, then we need to use the round-down algorithm. Set
745  * magic_divisor = m - 1 and extra_flags = 1.
746  * 4. Otherwise, set magic_divisor = m and extra_flags = 0.
747  */
748 
749 enum mali_attr_mode {
750 	MALI_ATTR_UNUSED = 0,
751 	MALI_ATTR_LINEAR = 1,
752 	MALI_ATTR_POT_DIVIDE = 2,
753 	MALI_ATTR_MODULO = 3,
754 	MALI_ATTR_NPOT_DIVIDE = 4,
755 };
756 
757 union mali_attr {
758 	/* This is used for actual attributes. */
759 	struct {
760 		/* The bottom 3 bits are the mode */
761 		mali_ptr elements : 64 - 8;
762 		u32 shift : 5;
763 		u32 extra_flags : 3;
764 		u32 stride;
765 		u32 size;
766 	};
767 	/* The entry after an NPOT_DIVIDE entry has this format. It stores
768 	 * extra information that wouldn't fit in a normal entry.
769 	 */
770 	struct {
771 		u32 unk; /* = 0x20 */
772 		u32 magic_divisor;
773 		u32 zero;
774 		/* This is the original, GL-level divisor. */
775 		u32 divisor;
776 	};
777 } __attribute__((packed));
778 
779 struct mali_attr_meta {
780         /* Vertex buffer index */
781         u8 index;
782 
783         unsigned unknown1 : 2;
784         unsigned swizzle : 12;
785         enum mali_format format : 8;
786 
787         /* Always observed to be zero at the moment */
788         unsigned unknown3 : 2;
789 
790         /* When packing multiple attributes in a buffer, offset addresses by this value */
791         uint32_t src_offset;
792 } __attribute__((packed));
793 
794 enum mali_fbd_type {
795         MALI_SFBD = 0,
796         MALI_MFBD = 1,
797 };
798 
799 #define FBD_TYPE (1)
800 #define FBD_MASK (~0x3f)
801 
802 struct mali_uniform_buffer_meta {
803         /* This is actually the size minus 1 (MALI_POSITIVE), in units of 16
804          * bytes. This gives a maximum of 2^14 bytes, which just so happens to
805          * be the GL minimum-maximum for GL_MAX_UNIFORM_BLOCK_SIZE.
806          */
807         u64 size : 10;
808 
809         /* This is missing the bottom 2 bits and top 8 bits. The top 8 bits
810          * should be 0 for userspace pointers, according to
811          * https://lwn.net/Articles/718895/. By reusing these bits, we can make
812          * each entry in the table only 64 bits.
813          */
814         mali_ptr ptr : 64 - 10;
815 };
816 
817 /* On Bifrost, these fields are the same between the vertex and tiler payloads.
818  * They also seem to be the same between Bifrost and Midgard. They're shared in
819  * fused payloads.
820  */
821 
822 /* Applies to unknown_draw */
823 
824 #define MALI_DRAW_INDEXED_UINT8  (0x10)
825 #define MALI_DRAW_INDEXED_UINT16 (0x20)
826 #define MALI_DRAW_INDEXED_UINT32 (0x30)
827 #define MALI_DRAW_VARYING_SIZE   (0x100)
828 #define MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX (0x10000)
829 
830 struct mali_vertex_tiler_prefix {
831         /* This is a dynamic bitfield containing the following things in this order:
832          *
833          * - gl_WorkGroupSize.x
834          * - gl_WorkGroupSize.y
835          * - gl_WorkGroupSize.z
836          * - gl_NumWorkGroups.x
837          * - gl_NumWorkGroups.y
838          * - gl_NumWorkGroups.z
839          *
840          * The number of bits allocated for each number is based on the *_shift
841          * fields below. For example, workgroups_y_shift gives the bit that
842          * gl_NumWorkGroups.y starts at, and workgroups_z_shift gives the bit
843          * that gl_NumWorkGroups.z starts at (and therefore one after the bit
844          * that gl_NumWorkGroups.y ends at). The actual value for each gl_*
845          * value is one more than the stored value, since if any of the values
846          * are zero, then there would be no invocations (and hence no job). If
847          * there were 0 bits allocated to a given field, then it must be zero,
848          * and hence the real value is one.
849          *
850          * Vertex jobs reuse the same job dispatch mechanism as compute jobs,
851          * effectively doing glDispatchCompute(1, vertex_count, instance_count)
852          * where vertex count is the number of vertices.
853          */
854         u32 invocation_count;
855 
856         u32 size_y_shift : 5;
857         u32 size_z_shift : 5;
858         u32 workgroups_x_shift : 6;
859         u32 workgroups_y_shift : 6;
860         u32 workgroups_z_shift : 6;
861         /* This is max(workgroups_x_shift, 2) in all the cases I've seen. */
862         u32 workgroups_x_shift_2 : 4;
863 
864         u32 draw_mode : 4;
865         u32 unknown_draw : 22;
866 
867         /* This is the the same as workgroups_x_shift_2 in compute shaders, but
868          * always 5 for vertex jobs and 6 for tiler jobs. I suspect this has
869          * something to do with how many quads get put in the same execution
870          * engine, which is a balance (you don't want to starve the engine, but
871          * you also want to distribute work evenly).
872          */
873         u32 workgroups_x_shift_3 : 6;
874 
875 
876         /* Negative of draw_start for TILER jobs from what I've seen */
877         int32_t negative_start;
878         u32 zero1;
879 
880         /* Like many other strictly nonzero quantities, index_count is
881          * subtracted by one. For an indexed cube, this is equal to 35 = 6
882          * faces * 2 triangles/per face * 3 vertices/per triangle - 1. That is,
883          * for an indexed draw, index_count is the number of actual vertices
884          * rendered whereas invocation_count is the number of unique vertices
885          * rendered (the number of times the vertex shader must be invoked).
886          * For non-indexed draws, this is just equal to invocation_count. */
887 
888         u32 index_count;
889 
890         /* No hidden structure; literally just a pointer to an array of uint
891          * indices (width depends on flags). Thanks, guys, for not making my
892          * life insane for once! NULL for non-indexed draws. */
893 
894         uintptr_t indices;
895 } __attribute__((packed));
896 
897 /* Point size / line width can either be specified as a 32-bit float (for
898  * constant size) or as a [machine word size]-bit GPU pointer (for varying size). If a pointer
899  * is selected, by setting the appropriate MALI_DRAW_VARYING_SIZE bit in the tiler
900  * payload, the contents of varying_pointer will be intepreted as an array of
901  * fp16 sizes, one for each vertex. gl_PointSize is therefore implemented by
902  * creating a special MALI_R16F varying writing to varying_pointer. */
903 
904 union midgard_primitive_size {
905         float constant;
906         uintptr_t pointer;
907 };
908 
909 struct bifrost_vertex_only {
910         u32 unk2; /* =0x2 */
911 
912         u32 zero0;
913 
914         u64 zero1;
915 } __attribute__((packed));
916 
917 struct bifrost_tiler_heap_meta {
918         u32 zero;
919         u32 heap_size;
920         /* note: these are just guesses! */
921         mali_ptr tiler_heap_start;
922         mali_ptr tiler_heap_free;
923         mali_ptr tiler_heap_end;
924 
925         /* hierarchy weights? but they're still 0 after the job has run... */
926         u32 zeros[12];
927 } __attribute__((packed));
928 
929 struct bifrost_tiler_meta {
930         u64 zero0;
931         u32 unk; // = 0xf0
932         u16 width;
933         u16 height;
934         u64 zero1;
935         mali_ptr tiler_heap_meta;
936         /* TODO what is this used for? */
937         u64 zeros[20];
938 } __attribute__((packed));
939 
940 struct bifrost_tiler_only {
941         /* 0x20 */
942         union midgard_primitive_size primitive_size;
943 
944         mali_ptr tiler_meta;
945 
946         u64 zero1, zero2, zero3, zero4, zero5, zero6;
947 
948         u32 gl_enables;
949         u32 zero7;
950         u64 zero8;
951 } __attribute__((packed));
952 
953 struct bifrost_scratchpad {
954         u32 zero;
955         u32 flags; // = 0x1f
956         /* This is a pointer to a CPU-inaccessible buffer, 16 pages, allocated
957          * during startup. It seems to serve the same purpose as the
958          * gpu_scratchpad in the SFBD for Midgard, although it's slightly
959          * larger.
960          */
961         mali_ptr gpu_scratchpad;
962 } __attribute__((packed));
963 
964 struct mali_vertex_tiler_postfix {
965         /* Zero for vertex jobs. Pointer to the position (gl_Position) varying
966          * output from the vertex shader for tiler jobs.
967          */
968 
969         uintptr_t position_varying;
970 
971         /* An array of mali_uniform_buffer_meta's. The size is given by the
972          * shader_meta.
973          */
974         uintptr_t uniform_buffers;
975 
976         /* This is a pointer to an array of pointers to the texture
977          * descriptors, number of pointers bounded by number of textures. The
978          * indirection is needed to accomodate varying numbers and sizes of
979          * texture descriptors */
980         uintptr_t texture_trampoline;
981 
982         /* For OpenGL, from what I've seen, this is intimately connected to
983          * texture_meta. cwabbott says this is not the case under Vulkan, hence
984          * why this field is seperate (Midgard is Vulkan capable). Pointer to
985          * array of sampler descriptors (which are uniform in size) */
986         uintptr_t sampler_descriptor;
987 
988         uintptr_t uniforms;
989         u8 flags : 4;
990         uintptr_t _shader_upper : MALI_SHORT_PTR_BITS - 4; /* struct shader_meta */
991         uintptr_t attributes; /* struct attribute_buffer[] */
992         uintptr_t attribute_meta; /* attribute_meta[] */
993         uintptr_t varyings; /* struct attr */
994         uintptr_t varying_meta; /* pointer */
995         uintptr_t viewport;
996         uintptr_t occlusion_counter; /* A single bit as far as I can tell */
997 
998         /* Note: on Bifrost, this isn't actually the FBD. It points to
999          * bifrost_scratchpad instead. However, it does point to the same thing
1000          * in vertex and tiler jobs.
1001          */
1002         mali_ptr framebuffer;
1003 
1004 #ifdef __LP64__
1005 #ifdef BIFROST
1006         /* most likely padding to make this a multiple of 64 bytes */
1007         u64 zero7;
1008 #endif
1009 #endif
1010 } __attribute__((packed));
1011 
1012 struct midgard_payload_vertex_tiler {
1013 #ifndef __LP64__
1014         union midgard_primitive_size primitive_size;
1015 #endif
1016 
1017         struct mali_vertex_tiler_prefix prefix;
1018 
1019 #ifndef __LP64__
1020         u32 zero3;
1021 #endif
1022 
1023         u32 gl_enables; // 0x5
1024 
1025         /* Offset for first vertex in buffer */
1026         u32 draw_start;
1027 
1028 	uintptr_t zero5;
1029 
1030         struct mali_vertex_tiler_postfix postfix;
1031 
1032 #ifdef __LP64__
1033         union midgard_primitive_size primitive_size;
1034 #endif
1035 } __attribute__((packed));
1036 
1037 struct bifrost_payload_vertex {
1038         struct mali_vertex_tiler_prefix prefix;
1039         struct bifrost_vertex_only vertex;
1040         struct mali_vertex_tiler_postfix postfix;
1041 } __attribute__((packed));
1042 
1043 struct bifrost_payload_tiler {
1044         struct mali_vertex_tiler_prefix prefix;
1045         struct bifrost_tiler_only tiler;
1046         struct mali_vertex_tiler_postfix postfix;
1047 } __attribute__((packed));
1048 
1049 struct bifrost_payload_fused {
1050         struct mali_vertex_tiler_prefix prefix;
1051         struct bifrost_tiler_only tiler;
1052         struct mali_vertex_tiler_postfix tiler_postfix;
1053         struct bifrost_vertex_only vertex;
1054         struct mali_vertex_tiler_postfix vertex_postfix;
1055 } __attribute__((packed));
1056 
1057 /* Pointed to from texture_trampoline, mostly unknown still, haven't
1058  * managed to replay successfully */
1059 
1060 /* Purposeful off-by-one in width, height fields. For example, a (64, 64)
1061  * texture is stored as (63, 63) in these fields. This adjusts for that.
1062  * There's an identical pattern in the framebuffer descriptor. Even vertex
1063  * count fields work this way, hence the generic name -- integral fields that
1064  * are strictly positive generally need this adjustment. */
1065 
1066 #define MALI_POSITIVE(dim) (dim - 1)
1067 
1068 /* Opposite of MALI_POSITIVE, found in the depth_units field */
1069 
1070 #define MALI_NEGATIVE(dim) (dim + 1)
1071 
1072 /* Used with wrapping. Incomplete (this is a 4-bit field...) */
1073 
1074 enum mali_wrap_mode {
1075         MALI_WRAP_REPEAT = 0x8,
1076         MALI_WRAP_CLAMP_TO_EDGE = 0x9,
1077         MALI_WRAP_CLAMP_TO_BORDER = 0xB,
1078         MALI_WRAP_MIRRORED_REPEAT = 0xC
1079 };
1080 
1081 /* 8192x8192 */
1082 #define MAX_MIP_LEVELS (13)
1083 
1084 /* Cubemap bloats everything up */
1085 #define MAX_FACES (6)
1086 
1087 /* Corresponds to the type passed to glTexImage2D and so forth */
1088 
1089 struct mali_texture_format {
1090         unsigned swizzle : 12;
1091         enum mali_format format : 8;
1092 
1093         unsigned usage1 : 3;
1094         unsigned is_not_cubemap : 1;
1095         unsigned usage2 : 8;
1096 } __attribute__((packed));
1097 
1098 struct mali_texture_descriptor {
1099         uint16_t width;
1100         uint16_t height;
1101         uint16_t depth;
1102 
1103         uint16_t unknown1;
1104 
1105         struct mali_texture_format format;
1106 
1107         uint16_t unknown3;
1108 
1109         /* One for non-mipmapped, zero for mipmapped */
1110         uint8_t unknown3A;
1111 
1112         /* Zero for non-mipmapped, (number of levels - 1) for mipmapped */
1113         uint8_t nr_mipmap_levels;
1114 
1115         /* Swizzling is a single 32-bit word, broken up here for convenience.
1116          * Here, swizzling refers to the ES 3.0 texture parameters for channel
1117          * level swizzling, not the internal pixel-level swizzling which is
1118          * below OpenGL's reach */
1119 
1120         unsigned swizzle : 12;
1121         unsigned swizzle_zero       : 20;
1122 
1123         uint32_t unknown5;
1124         uint32_t unknown6;
1125         uint32_t unknown7;
1126 
1127         mali_ptr swizzled_bitmaps[MAX_MIP_LEVELS * MAX_FACES];
1128 } __attribute__((packed));
1129 
1130 /* Used as part of filter_mode */
1131 
1132 #define MALI_LINEAR 0
1133 #define MALI_NEAREST 1
1134 #define MALI_MIP_LINEAR (0x18)
1135 
1136 /* Used to construct low bits of filter_mode */
1137 
1138 #define MALI_TEX_MAG(mode) (((mode) & 1) << 0)
1139 #define MALI_TEX_MIN(mode) (((mode) & 1) << 1)
1140 
1141 #define MALI_TEX_MAG_MASK (1)
1142 #define MALI_TEX_MIN_MASK (2)
1143 
1144 #define MALI_FILTER_NAME(filter) (filter ? "MALI_NEAREST" : "MALI_LINEAR")
1145 
1146 /* Used for lod encoding. Thanks @urjaman for pointing out these routines can
1147  * be cleaned up a lot. */
1148 
1149 #define DECODE_FIXED_16(x) ((float) (x / 256.0))
1150 
1151 static inline uint16_t
FIXED_16(float x)1152 FIXED_16(float x)
1153 {
1154         /* Clamp inputs, accounting for float error */
1155         float max_lod = (32.0 - (1.0 / 512.0));
1156 
1157         x = ((x > max_lod) ? max_lod : ((x < 0.0) ? 0.0 : x));
1158 
1159         return (int) (x * 256.0);
1160 }
1161 
1162 struct mali_sampler_descriptor {
1163         uint32_t filter_mode;
1164 
1165         /* Fixed point. Upper 8-bits is before the decimal point, although it
1166          * caps [0-31]. Lower 8-bits is after the decimal point: int(round(x *
1167          * 256)) */
1168 
1169         uint16_t min_lod;
1170         uint16_t max_lod;
1171 
1172         /* All one word in reality, but packed a bit */
1173 
1174         enum mali_wrap_mode wrap_s : 4;
1175         enum mali_wrap_mode wrap_t : 4;
1176         enum mali_wrap_mode wrap_r : 4;
1177         enum mali_alt_func compare_func : 3;
1178 
1179         /* A single set bit of unknown, ha! */
1180         unsigned unknown2 : 1;
1181 
1182         unsigned zero : 16;
1183 
1184         uint32_t zero2;
1185         float border_color[4];
1186 } __attribute__((packed));
1187 
1188 /* TODO: What are the floats? Apparently always { -inf, -inf, inf, inf },
1189  * unless the scissor test is enabled.
1190  *
1191  * viewport0/viewport1 form the arguments to glViewport. viewport1 is modified
1192  * by MALI_POSITIVE; viewport0 is as-is.
1193  */
1194 
1195 struct mali_viewport {
1196         /* XY clipping planes */
1197         float clip_minx;
1198         float clip_miny;
1199         float clip_maxx;
1200         float clip_maxy;
1201 
1202         /* Depth clipping planes */
1203         float clip_minz;
1204         float clip_maxz;
1205 
1206         u16 viewport0[2];
1207         u16 viewport1[2];
1208 } __attribute__((packed));
1209 
1210 /* TODO: Varying meta is symmetrical with attr_meta, but there is some
1211  * weirdness associated. Figure it out. */
1212 
1213 struct mali_unknown6 {
1214         u64 unknown0;
1215         u64 unknown1;
1216 };
1217 
1218 /* From presentations, 16x16 tiles externally. Use shift for fast computation
1219  * of tile numbers. */
1220 
1221 #define MALI_TILE_SHIFT 4
1222 #define MALI_TILE_LENGTH (1 << MALI_TILE_SHIFT)
1223 
1224 /* Tile coordinates are stored as a compact u32, as only 12 bits are needed to
1225  * each component. Notice that this provides a theoretical upper bound of (1 <<
1226  * 12) = 4096 tiles in each direction, addressing a maximum framebuffer of size
1227  * 65536x65536. Multiplying that together, times another four given that Mali
1228  * framebuffers are 32-bit ARGB8888, means that this upper bound would take 16
1229  * gigabytes of RAM just to store the uncompressed framebuffer itself, let
1230  * alone rendering in real-time to such a buffer.
1231  *
1232  * Nice job, guys.*/
1233 
1234 /* From mali_kbase_10969_workaround.c */
1235 #define MALI_X_COORD_MASK 0x00000FFF
1236 #define MALI_Y_COORD_MASK 0x0FFF0000
1237 
1238 /* Extract parts of a tile coordinate */
1239 
1240 #define MALI_TILE_COORD_X(coord) ((coord) & MALI_X_COORD_MASK)
1241 #define MALI_TILE_COORD_Y(coord) (((coord) & MALI_Y_COORD_MASK) >> 16)
1242 #define MALI_TILE_COORD_FLAGS(coord) ((coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK))
1243 
1244 /* No known flags yet, but just in case...? */
1245 
1246 #define MALI_TILE_NO_FLAG (0)
1247 
1248 /* Helpers to generate tile coordinates based on the boundary coordinates in
1249  * screen space. So, with the bounds (0, 0) to (128, 128) for the screen, these
1250  * functions would convert it to the bounding tiles (0, 0) to (7, 7).
1251  * Intentional "off-by-one"; finding the tile number is a form of fencepost
1252  * problem. */
1253 
1254 #define MALI_MAKE_TILE_COORDS(X, Y) ((X) | ((Y) << 16))
1255 #define MALI_BOUND_TO_TILE(B, bias) ((B - bias) >> MALI_TILE_SHIFT)
1256 #define MALI_COORDINATE_TO_TILE(W, H, bias) MALI_MAKE_TILE_COORDS(MALI_BOUND_TO_TILE(W, bias), MALI_BOUND_TO_TILE(H, bias))
1257 #define MALI_COORDINATE_TO_TILE_MIN(W, H) MALI_COORDINATE_TO_TILE(W, H, 0)
1258 #define MALI_COORDINATE_TO_TILE_MAX(W, H) MALI_COORDINATE_TO_TILE(W, H, 1)
1259 
1260 struct mali_payload_fragment {
1261         u32 min_tile_coord;
1262         u32 max_tile_coord;
1263         mali_ptr framebuffer;
1264 } __attribute__((packed));
1265 
1266 /* (Single?) Framebuffer Descriptor */
1267 
1268 /* Flags apply to format. With just MSAA_A and MSAA_B, the framebuffer is
1269  * configured for 4x. With MSAA_8, it is configured for 8x. */
1270 
1271 #define MALI_FRAMEBUFFER_MSAA_8 (1 << 3)
1272 #define MALI_FRAMEBUFFER_MSAA_A (1 << 4)
1273 #define MALI_FRAMEBUFFER_MSAA_B (1 << 23)
1274 
1275 /* Fast/slow based on whether all three buffers are cleared at once */
1276 
1277 #define MALI_CLEAR_FAST         (1 << 18)
1278 #define MALI_CLEAR_SLOW         (1 << 28)
1279 #define MALI_CLEAR_SLOW_STENCIL (1 << 31)
1280 
1281 struct mali_single_framebuffer {
1282         u32 unknown1;
1283         u32 unknown2;
1284         u64 unknown_address_0;
1285         u64 zero1;
1286         u64 zero0;
1287 
1288         /* Exact format is ironically not known, since EGL is finnicky with the
1289          * blob. MSAA, colourspace, etc are configured here. */
1290 
1291         u32 format;
1292 
1293         u32 clear_flags;
1294         u32 zero2;
1295 
1296         /* Purposeful off-by-one in these fields should be accounted for by the
1297          * MALI_DIMENSION macro */
1298 
1299         u16 width;
1300         u16 height;
1301 
1302         u32 zero3[8];
1303 
1304         /* By default, the framebuffer is upside down from OpenGL's
1305          * perspective. Set framebuffer to the end and negate the stride to
1306          * flip in the Y direction */
1307 
1308         mali_ptr framebuffer;
1309         int32_t stride;
1310 
1311         u32 zero4;
1312 
1313         /* Depth and stencil buffers are interleaved, it appears, as they are
1314          * set to the same address in captures. Both fields set to zero if the
1315          * buffer is not being cleared. Depending on GL_ENABLE magic, you might
1316          * get a zero enable despite the buffer being present; that still is
1317          * disabled. */
1318 
1319         mali_ptr depth_buffer; // not SAME_VA
1320         u64 depth_buffer_enable;
1321 
1322         mali_ptr stencil_buffer; // not SAME_VA
1323         u64 stencil_buffer_enable;
1324 
1325         u32 clear_color_1; // RGBA8888 from glClear, actually used by hardware
1326         u32 clear_color_2; // always equal, but unclear function?
1327         u32 clear_color_3; // always equal, but unclear function?
1328         u32 clear_color_4; // always equal, but unclear function?
1329 
1330         /* Set to zero if not cleared */
1331 
1332         float clear_depth_1; // float32, ditto
1333         float clear_depth_2; // float32, ditto
1334         float clear_depth_3; // float32, ditto
1335         float clear_depth_4; // float32, ditto
1336 
1337         u32 clear_stencil; // Exactly as it appears in OpenGL
1338 
1339         u32 zero6[7];
1340 
1341         /* Very weird format, see generation code in trans_builder.c */
1342         u32 resolution_check;
1343 
1344         u32 tiler_flags;
1345 
1346         u64 unknown_address_1; /* Pointing towards... a zero buffer? */
1347         u64 unknown_address_2;
1348 
1349         /* See mali_kbase_replay.c */
1350         u64 tiler_heap_free;
1351         u64 tiler_heap_end;
1352 
1353         /* More below this, maybe */
1354 } __attribute__((packed));
1355 
1356 /* Format bits for the render target flags */
1357 
1358 #define MALI_MFBD_FORMAT_AFBC 	  (1 << 5)
1359 #define MALI_MFBD_FORMAT_MSAA 	  (1 << 7)
1360 
1361 struct mali_rt_format {
1362         unsigned unk1 : 32;
1363         unsigned unk2 : 3;
1364 
1365         unsigned nr_channels : 2; /* MALI_POSITIVE */
1366 
1367         unsigned flags : 11;
1368 
1369         unsigned swizzle : 12;
1370 
1371         unsigned unk4 : 4;
1372 } __attribute__((packed));
1373 
1374 struct bifrost_render_target {
1375         struct mali_rt_format format;
1376 
1377         u64 zero1;
1378 
1379         union {
1380                 struct {
1381                         /* Stuff related to ARM Framebuffer Compression. When AFBC is enabled,
1382                          * there is an extra metadata buffer that contains 16 bytes per tile.
1383                          * The framebuffer needs to be the same size as before, since we don't
1384                          * know ahead of time how much space it will take up. The
1385                          * framebuffer_stride is set to 0, since the data isn't stored linearly
1386                          * anymore.
1387                          */
1388 
1389                         mali_ptr metadata;
1390                         u32 stride; // stride in units of tiles
1391                         u32 unk; // = 0x20000
1392                 } afbc;
1393 
1394                 struct {
1395                         /* Heck if I know */
1396                         u64 unk;
1397                         mali_ptr pointer;
1398                 } chunknown;
1399         };
1400 
1401         mali_ptr framebuffer;
1402 
1403         u32 zero2 : 4;
1404         u32 framebuffer_stride : 28; // in units of bytes
1405         u32 zero3;
1406 
1407         u32 clear_color_1; // RGBA8888 from glClear, actually used by hardware
1408         u32 clear_color_2; // always equal, but unclear function?
1409         u32 clear_color_3; // always equal, but unclear function?
1410         u32 clear_color_4; // always equal, but unclear function?
1411 } __attribute__((packed));
1412 
1413 /* An optional part of bifrost_framebuffer. It comes between the main structure
1414  * and the array of render targets. It must be included if any of these are
1415  * enabled:
1416  *
1417  * - Transaction Elimination
1418  * - Depth/stencil
1419  * - TODO: Anything else?
1420  */
1421 
1422 /* Flags field: note, these are guesses */
1423 
1424 #define MALI_EXTRA_PRESENT      (0x400)
1425 #define MALI_EXTRA_AFBC         (0x20)
1426 #define MALI_EXTRA_AFBC_ZS      (0x10)
1427 #define MALI_EXTRA_ZS           (0x4)
1428 
1429 struct bifrost_fb_extra {
1430         mali_ptr checksum;
1431         /* Each tile has an 8 byte checksum, so the stride is "width in tiles * 8" */
1432         u32 checksum_stride;
1433 
1434         u32 flags;
1435 
1436         union {
1437                 /* Note: AFBC is only allowed for 24/8 combined depth/stencil. */
1438                 struct {
1439                         mali_ptr depth_stencil_afbc_metadata;
1440                         u32 depth_stencil_afbc_stride; // in units of tiles
1441                         u32 zero1;
1442 
1443                         mali_ptr depth_stencil;
1444 
1445                         u64 padding;
1446                 } ds_afbc;
1447 
1448                 struct {
1449                         /* Depth becomes depth/stencil in case of combined D/S */
1450                         mali_ptr depth;
1451                         u32 depth_stride_zero : 4;
1452                         u32 depth_stride : 28;
1453                         u32 zero1;
1454 
1455                         mali_ptr stencil;
1456                         u32 stencil_stride_zero : 4;
1457                         u32 stencil_stride : 28;
1458                         u32 zero2;
1459                 } ds_linear;
1460         };
1461 
1462 
1463         u64 zero3, zero4;
1464 } __attribute__((packed));
1465 
1466 /* flags for unk3 */
1467 
1468 /* Enables writing depth results back to main memory (rather than keeping them
1469  * on-chip in the tile buffer and then discarding) */
1470 
1471 #define MALI_MFBD_DEPTH_WRITE (1 << 10)
1472 
1473 /* The MFBD contains the extra bifrost_fb_extra section */
1474 
1475 #define MALI_MFBD_EXTRA (1 << 13)
1476 
1477 struct bifrost_framebuffer {
1478         u32 unk0; // = 0x10
1479 
1480         u32 unknown2; // = 0x1f, same as SFBD
1481         mali_ptr scratchpad;
1482 
1483         /* 0x10 */
1484         mali_ptr sample_locations;
1485         mali_ptr unknown1;
1486         /* 0x20 */
1487         u16 width1, height1;
1488         u32 zero3;
1489         u16 width2, height2;
1490         u32 unk1 : 19; // = 0x01000
1491         u32 rt_count_1 : 2; // off-by-one (use MALI_POSITIVE)
1492         u32 unk2 : 3; // = 0
1493         u32 rt_count_2 : 3; // no off-by-one
1494         u32 zero4 : 5;
1495         /* 0x30 */
1496         u32 clear_stencil : 8;
1497         u32 unk3 : 24; // = 0x100
1498         float clear_depth;
1499         mali_ptr tiler_meta;
1500         /* 0x40 */
1501 
1502         /* Note: these are guesses! */
1503         mali_ptr tiler_scratch_start;
1504         mali_ptr tiler_scratch_middle;
1505 
1506         /* These are not, since we see symmetry with replay jobs which name these explicitly */
1507         mali_ptr tiler_heap_start;
1508         mali_ptr tiler_heap_end;
1509 
1510         u64 zero9, zero10, zero11, zero12;
1511 
1512         /* optional: struct bifrost_fb_extra extra */
1513         /* struct bifrost_render_target rts[] */
1514 } __attribute__((packed));
1515 
1516 #endif /* __PANFROST_JOB_H__ */
1517