2 * © Copyright 2017-2018 Alyssa Rosenzweig
3 * © Copyright 2017-2018 Connor Abbott
4 * © Copyright 2017-2018 Lyude Paul
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
27 #ifndef __PANFROST_JOB_H__
28 #define __PANFROST_JOB_H__
31 #include <panfrost-misc.h>
33 #define MALI_SHORT_PTR_BITS (sizeof(uintptr_t)*8)
35 #define MALI_FBD_HIERARCHY_WEIGHTS 8
37 #define MALI_PAYLOAD_SIZE 256
39 typedef u32 mali_jd_core_req
;
44 JOB_TYPE_SET_VALUE
= 2,
45 JOB_TYPE_CACHE_FLUSH
= 3,
48 JOB_TYPE_GEOMETRY
= 6,
51 JOB_TYPE_FRAGMENT
= 9,
58 MALI_LINE_STRIP
= 0x4,
61 MALI_TRIANGLE_STRIP
= 0xA,
62 MALI_TRIANGLE_FAN
= 0xC,
65 MALI_QUAD_STRIP
= 0xF,
67 /* All other modes invalid */
70 /* Applies to tiler_gl_enables */
73 #define MALI_OCCLUSION_QUERY (1 << 3)
74 #define MALI_OCCLUSION_PRECISE (1 << 4)
76 #define MALI_FRONT_FACE(v) (v << 5)
80 #define MALI_CULL_FACE_FRONT (1 << 6)
81 #define MALI_CULL_FACE_BACK (1 << 7)
83 /* TODO: Might this actually be a finer bitfield? */
84 #define MALI_DEPTH_STENCIL_ENABLE 0x6400
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 */"
92 /* Used in stencil and depth tests */
99 MALI_FUNC_GREATER
= 4,
100 MALI_FUNC_NOTEQUAL
= 5,
101 MALI_FUNC_GEQUAL
= 6,
105 /* Same OpenGL, but mixed up. Why? Because forget me, that's why! */
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
118 /* Flags apply to unknown2_3? */
120 #define MALI_HAS_MSAA (1 << 0)
121 #define MALI_CAN_DISCARD (1 << 5)
123 /* Applies on SFBD systems, specifying that programmable blending is in use */
124 #define MALI_HAS_BLEND_SHADER (1 << 6)
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)
131 #define MALI_DEPTH_TEST (1 << 11)
133 /* Next flags to unknown2_4 */
134 #define MALI_STENCIL_TEST (1 << 0)
137 #define MALI_SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER (1 << 1)
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)
144 /* Stencil test state is all encoded in a single u32, just with a lot of
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
158 struct mali_stencil_test
{
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;
166 } __attribute__((packed
));
168 /* Blending is a mess, since anything fancy triggers a blend shader, and
169 * -those- are not understood whatsover yet */
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)
176 enum mali_nondominant_mode
{
177 MALI_BLEND_NON_MIRROR
= 0,
178 MALI_BLEND_NON_ZERO
= 1
181 enum mali_dominant_blend
{
182 MALI_BLEND_DOM_SOURCE
= 0,
183 MALI_BLEND_DOM_DESTINATION
= 1
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,
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,
204 struct mali_blend_mode
{
205 enum mali_blend_modifier clip_modifier
: 2;
206 unsigned unused_0
: 1;
207 unsigned negate_source
: 1;
209 enum mali_dominant_blend dominant
: 1;
211 enum mali_nondominant_mode nondominant_mode
: 1;
213 unsigned unused_1
: 1;
215 unsigned negate_dest
: 1;
217 enum mali_dominant_factor dominant_factor
: 3;
218 unsigned complement_dominant
: 1;
219 } __attribute__((packed
));
221 struct mali_blend_equation
{
222 /* Of type mali_blend_mode */
223 unsigned rgb_mode
: 12;
224 unsigned alpha_mode
: 12;
228 /* Corresponds to MALI_MASK_* above and glColorMask arguments */
230 unsigned color_mask
: 4;
232 /* Attached constant for CONSTANT_ALPHA, etc */
237 } __attribute__((packed
));
239 /* Used with channel swizzling */
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,
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
));
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.
265 /* The top 3 bits specify how the bits of each component are interpreted. */
267 /* e.g. R11F_G11F_B10F */
268 #define MALI_FORMAT_SPECIAL (2 << 5)
270 /* signed normalized, e.g. RGBA8_SNORM */
271 #define MALI_FORMAT_SNORM (3 << 5)
274 #define MALI_FORMAT_UINT (4 << 5)
276 /* e.g. RGBA8 and RGBA32F */
277 #define MALI_FORMAT_UNORM (5 << 5)
279 /* e.g. RGBA8I and RGBA16F */
280 #define MALI_FORMAT_SINT (6 << 5)
282 /* These formats seem to largely duplicate the others. They're used at least
283 * for Bifrost framebuffer output.
285 #define MALI_FORMAT_SPECIAL2 (7 << 5)
287 /* If the high 3 bits are 3 to 6 these two bits say how many components
290 #define MALI_NR_CHANNELS(n) ((n - 1) << 3)
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
297 #define MALI_CHANNEL_4 2
299 #define MALI_CHANNEL_8 3
301 #define MALI_CHANNEL_16 4
303 #define MALI_CHANNEL_32 5
305 /* For MALI_FORMAT_SINT it means a half-float (e.g. RG16F). For
306 * MALI_FORMAT_UNORM, it means a 32-bit float.
308 #define MALI_CHANNEL_FLOAT 7
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,
319 MALI_NV12
= MALI_FORMAT_SPECIAL
| 0xc,
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
332 MALI_VARYING_DISCARD
= MALI_FORMAT_SPECIAL
| 0x1f,
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
,
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
,
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
,
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
,
395 MALI_RGBA4
= MALI_FORMAT_SPECIAL2
| 0x8,
396 MALI_RGBA8_2
= MALI_FORMAT_SPECIAL2
| 0xd,
397 MALI_RGB10_A2_2
= MALI_FORMAT_SPECIAL2
| 0xe,
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... */
405 #define MALI_ALPHA_COVERAGE(clampf) ((uint16_t) (int) (clampf * 15.0f))
406 #define MALI_GET_ALPHA_COVERAGE(nibble) ((float) nibble / 15.0f)
408 /* Applies to unknown1 */
409 #define MALI_NO_ALPHA_TO_COVERAGE (1 << 10)
411 struct mali_blend_meta
{
413 /* Base value of 0x200.
414 * OR with 0x1 for blending (anything other than REPLACE).
415 * OR with 0x2 for programmable blending
420 /* For programmable blending, these turn into the blend_shader address */
421 struct mali_blend_equation blend_equation_1
;
424 struct mali_blend_equation blend_equation_2
;
427 struct mali_blend_equation blend_equation
;
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
435 /* increments from 0 to 3 */
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
447 enum mali_format format
: 8;
449 /* Type of the shader output variable. Note, this can
450 * be different from the format.
452 * 0: f16 (mediump float)
453 * 1: f32 (highp float)
455 * 3: u32 (highp uint)
456 * 4: i16 (mediump int)
457 * 5: u16 (mediump uint)
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.
476 } __attribute__((packed
));
478 struct mali_shader_meta
{
487 u32 uniform_buffer_count
: 4;
488 u32 unk1
: 28; // = 0x800000 for vertex, 0x958020 for tiler
491 /* 0x200 except MALI_NO_ALPHA_TO_COVERAGE. Mysterious 1
492 * other times. Who knows really? */
495 /* Whole number of uniform registers used, times two;
496 * whole number of work registers used (no scale).
498 unsigned work_count
: 5;
499 unsigned uniform_count
: 5;
500 unsigned unknown2
: 6;
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)
517 u8 stencil_mask_front
;
518 u8 stencil_mask_back
;
521 struct mali_stencil_test stencil_front
;
522 struct mali_stencil_test stencil_back
;
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
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)
545 * - R55 : unknown, never seen (but the bit for this is
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).
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.
563 u32 uniform_count
: 7;
564 u32 unk4
: 10; // = 2
571 /* zero on bifrost */
574 /* Blending information for the older non-MRT Midgard HW. Check for
575 * MALI_HAS_BLEND_SHADER to decide how to interpret.
579 mali_ptr blend_shader
;
580 struct mali_blend_equation blend_equation
;
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).
587 struct mali_blend_meta blend_meta
[];
589 } __attribute__((packed
));
591 /* This only concerns hardware jobs */
593 /* Possible values for job_descriptor_size */
595 #define MALI_JOB_32 0
596 #define MALI_JOB_64 1
598 struct mali_job_descriptor_header
{
599 u32 exception_status
;
600 u32 first_incomplete_task
;
602 u8 job_descriptor_size
: 1;
603 enum mali_job_type job_type
: 7;
605 u8 unknown_flags
: 7;
607 u16 job_dependency_index_1
;
608 u16 job_dependency_index_2
;
614 } __attribute__((packed
));
616 struct mali_payload_set_value
{
619 } __attribute__((packed
));
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)
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.
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:
642 * vertex_id = linear_id % num_vertices
643 * instance_id = linear_id / num_vertices
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.
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.
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
665 * num_vertices | padded_num_vertices
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:
681 * high bits | padded_num_vertices
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.
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:
697 * 1. Use the linear_id directly. Only used when there is no instancing.
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
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.
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.
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:
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.
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,
757 /* This magic "pseudo-address" is used as `elements` to implement
758 * gl_PointCoord. When read from a fragment shader, it generates a point
759 * coordinate per the OpenGL ES 2.0 specification. Flipped coordinate spaces
760 * require an affine transformation in the shader. */
762 #define MALI_VARYING_POINT_COORD (0x60)
765 /* This is used for actual attributes. */
767 /* The bottom 3 bits are the mode */
768 mali_ptr elements
: 64 - 8;
774 /* The entry after an NPOT_DIVIDE entry has this format. It stores
775 * extra information that wouldn't fit in a normal entry.
778 u32 unk
; /* = 0x20 */
781 /* This is the original, GL-level divisor. */
784 } __attribute__((packed
));
786 struct mali_attr_meta
{
787 /* Vertex buffer index */
790 unsigned unknown1
: 2;
791 unsigned swizzle
: 12;
792 enum mali_format format
: 8;
794 /* Always observed to be zero at the moment */
795 unsigned unknown3
: 2;
797 /* When packing multiple attributes in a buffer, offset addresses by this value */
799 } __attribute__((packed
));
807 #define FBD_MASK (~0x3f)
809 struct mali_uniform_buffer_meta
{
810 /* This is actually the size minus 1 (MALI_POSITIVE), in units of 16
811 * bytes. This gives a maximum of 2^14 bytes, which just so happens to
812 * be the GL minimum-maximum for GL_MAX_UNIFORM_BLOCK_SIZE.
816 /* This is missing the bottom 2 bits and top 8 bits. The top 8 bits
817 * should be 0 for userspace pointers, according to
818 * https://lwn.net/Articles/718895/. By reusing these bits, we can make
819 * each entry in the table only 64 bits.
821 mali_ptr ptr
: 64 - 10;
824 /* On Bifrost, these fields are the same between the vertex and tiler payloads.
825 * They also seem to be the same between Bifrost and Midgard. They're shared in
829 /* Applies to unknown_draw */
831 #define MALI_DRAW_INDEXED_UINT8 (0x10)
832 #define MALI_DRAW_INDEXED_UINT16 (0x20)
833 #define MALI_DRAW_INDEXED_UINT32 (0x30)
834 #define MALI_DRAW_VARYING_SIZE (0x100)
835 #define MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX (0x10000)
837 struct mali_vertex_tiler_prefix
{
838 /* This is a dynamic bitfield containing the following things in this order:
840 * - gl_WorkGroupSize.x
841 * - gl_WorkGroupSize.y
842 * - gl_WorkGroupSize.z
843 * - gl_NumWorkGroups.x
844 * - gl_NumWorkGroups.y
845 * - gl_NumWorkGroups.z
847 * The number of bits allocated for each number is based on the *_shift
848 * fields below. For example, workgroups_y_shift gives the bit that
849 * gl_NumWorkGroups.y starts at, and workgroups_z_shift gives the bit
850 * that gl_NumWorkGroups.z starts at (and therefore one after the bit
851 * that gl_NumWorkGroups.y ends at). The actual value for each gl_*
852 * value is one more than the stored value, since if any of the values
853 * are zero, then there would be no invocations (and hence no job). If
854 * there were 0 bits allocated to a given field, then it must be zero,
855 * and hence the real value is one.
857 * Vertex jobs reuse the same job dispatch mechanism as compute jobs,
858 * effectively doing glDispatchCompute(1, vertex_count, instance_count)
859 * where vertex count is the number of vertices.
861 u32 invocation_count
;
863 u32 size_y_shift
: 5;
864 u32 size_z_shift
: 5;
865 u32 workgroups_x_shift
: 6;
866 u32 workgroups_y_shift
: 6;
867 u32 workgroups_z_shift
: 6;
868 /* This is max(workgroups_x_shift, 2) in all the cases I've seen. */
869 u32 workgroups_x_shift_2
: 4;
872 u32 unknown_draw
: 22;
874 /* This is the the same as workgroups_x_shift_2 in compute shaders, but
875 * always 5 for vertex jobs and 6 for tiler jobs. I suspect this has
876 * something to do with how many quads get put in the same execution
877 * engine, which is a balance (you don't want to starve the engine, but
878 * you also want to distribute work evenly).
880 u32 workgroups_x_shift_3
: 6;
883 /* Negative of draw_start for TILER jobs from what I've seen */
884 int32_t negative_start
;
887 /* Like many other strictly nonzero quantities, index_count is
888 * subtracted by one. For an indexed cube, this is equal to 35 = 6
889 * faces * 2 triangles/per face * 3 vertices/per triangle - 1. That is,
890 * for an indexed draw, index_count is the number of actual vertices
891 * rendered whereas invocation_count is the number of unique vertices
892 * rendered (the number of times the vertex shader must be invoked).
893 * For non-indexed draws, this is just equal to invocation_count. */
897 /* No hidden structure; literally just a pointer to an array of uint
898 * indices (width depends on flags). Thanks, guys, for not making my
899 * life insane for once! NULL for non-indexed draws. */
902 } __attribute__((packed
));
904 /* Point size / line width can either be specified as a 32-bit float (for
905 * constant size) or as a [machine word size]-bit GPU pointer (for varying size). If a pointer
906 * is selected, by setting the appropriate MALI_DRAW_VARYING_SIZE bit in the tiler
907 * payload, the contents of varying_pointer will be intepreted as an array of
908 * fp16 sizes, one for each vertex. gl_PointSize is therefore implemented by
909 * creating a special MALI_R16F varying writing to varying_pointer. */
911 union midgard_primitive_size
{
916 struct bifrost_vertex_only
{
922 } __attribute__((packed
));
924 struct bifrost_tiler_heap_meta
{
927 /* note: these are just guesses! */
928 mali_ptr tiler_heap_start
;
929 mali_ptr tiler_heap_free
;
930 mali_ptr tiler_heap_end
;
932 /* hierarchy weights? but they're still 0 after the job has run... */
934 } __attribute__((packed
));
936 struct bifrost_tiler_meta
{
942 mali_ptr tiler_heap_meta
;
943 /* TODO what is this used for? */
945 } __attribute__((packed
));
947 struct bifrost_tiler_only
{
949 union midgard_primitive_size primitive_size
;
953 u64 zero1
, zero2
, zero3
, zero4
, zero5
, zero6
;
958 } __attribute__((packed
));
960 struct bifrost_scratchpad
{
963 /* This is a pointer to a CPU-inaccessible buffer, 16 pages, allocated
964 * during startup. It seems to serve the same purpose as the
965 * gpu_scratchpad in the SFBD for Midgard, although it's slightly
968 mali_ptr gpu_scratchpad
;
969 } __attribute__((packed
));
971 struct mali_vertex_tiler_postfix
{
972 /* Zero for vertex jobs. Pointer to the position (gl_Position) varying
973 * output from the vertex shader for tiler jobs.
976 uintptr_t position_varying
;
978 /* An array of mali_uniform_buffer_meta's. The size is given by the
981 uintptr_t uniform_buffers
;
983 /* This is a pointer to an array of pointers to the texture
984 * descriptors, number of pointers bounded by number of textures. The
985 * indirection is needed to accomodate varying numbers and sizes of
986 * texture descriptors */
987 uintptr_t texture_trampoline
;
989 /* For OpenGL, from what I've seen, this is intimately connected to
990 * texture_meta. cwabbott says this is not the case under Vulkan, hence
991 * why this field is seperate (Midgard is Vulkan capable). Pointer to
992 * array of sampler descriptors (which are uniform in size) */
993 uintptr_t sampler_descriptor
;
997 uintptr_t _shader_upper
: MALI_SHORT_PTR_BITS
- 4; /* struct shader_meta */
998 uintptr_t attributes
; /* struct attribute_buffer[] */
999 uintptr_t attribute_meta
; /* attribute_meta[] */
1000 uintptr_t varyings
; /* struct attr */
1001 uintptr_t varying_meta
; /* pointer */
1003 uintptr_t occlusion_counter
; /* A single bit as far as I can tell */
1005 /* Note: on Bifrost, this isn't actually the FBD. It points to
1006 * bifrost_scratchpad instead. However, it does point to the same thing
1007 * in vertex and tiler jobs.
1009 mali_ptr framebuffer
;
1013 /* most likely padding to make this a multiple of 64 bytes */
1017 } __attribute__((packed
));
1019 struct midgard_payload_vertex_tiler
{
1021 union midgard_primitive_size primitive_size
;
1024 struct mali_vertex_tiler_prefix prefix
;
1030 u32 gl_enables
; // 0x5
1032 /* Offset for first vertex in buffer */
1037 struct mali_vertex_tiler_postfix postfix
;
1040 union midgard_primitive_size primitive_size
;
1042 } __attribute__((packed
));
1044 struct bifrost_payload_vertex
{
1045 struct mali_vertex_tiler_prefix prefix
;
1046 struct bifrost_vertex_only vertex
;
1047 struct mali_vertex_tiler_postfix postfix
;
1048 } __attribute__((packed
));
1050 struct bifrost_payload_tiler
{
1051 struct mali_vertex_tiler_prefix prefix
;
1052 struct bifrost_tiler_only tiler
;
1053 struct mali_vertex_tiler_postfix postfix
;
1054 } __attribute__((packed
));
1056 struct bifrost_payload_fused
{
1057 struct mali_vertex_tiler_prefix prefix
;
1058 struct bifrost_tiler_only tiler
;
1059 struct mali_vertex_tiler_postfix tiler_postfix
;
1060 struct bifrost_vertex_only vertex
;
1061 struct mali_vertex_tiler_postfix vertex_postfix
;
1062 } __attribute__((packed
));
1064 /* Pointed to from texture_trampoline, mostly unknown still, haven't
1065 * managed to replay successfully */
1067 /* Purposeful off-by-one in width, height fields. For example, a (64, 64)
1068 * texture is stored as (63, 63) in these fields. This adjusts for that.
1069 * There's an identical pattern in the framebuffer descriptor. Even vertex
1070 * count fields work this way, hence the generic name -- integral fields that
1071 * are strictly positive generally need this adjustment. */
1073 #define MALI_POSITIVE(dim) (dim - 1)
1075 /* Opposite of MALI_POSITIVE, found in the depth_units field */
1077 #define MALI_NEGATIVE(dim) (dim + 1)
1079 /* Used with wrapping. Incomplete (this is a 4-bit field...) */
1081 enum mali_wrap_mode
{
1082 MALI_WRAP_REPEAT
= 0x8,
1083 MALI_WRAP_CLAMP_TO_EDGE
= 0x9,
1084 MALI_WRAP_CLAMP_TO_BORDER
= 0xB,
1085 MALI_WRAP_MIRRORED_REPEAT
= 0xC
1089 #define MAX_MIP_LEVELS (13)
1091 /* Cubemap bloats everything up */
1092 #define MAX_FACES (6)
1094 /* Corresponds to the type passed to glTexImage2D and so forth */
1096 struct mali_texture_format
{
1097 unsigned swizzle
: 12;
1098 enum mali_format format
: 8;
1100 unsigned usage1
: 3;
1101 unsigned is_not_cubemap
: 1;
1102 unsigned usage2
: 8;
1103 } __attribute__((packed
));
1105 struct mali_texture_descriptor
{
1112 struct mali_texture_format format
;
1116 /* One for non-mipmapped, zero for mipmapped */
1119 /* Zero for non-mipmapped, (number of levels - 1) for mipmapped */
1120 uint8_t nr_mipmap_levels
;
1122 /* Swizzling is a single 32-bit word, broken up here for convenience.
1123 * Here, swizzling refers to the ES 3.0 texture parameters for channel
1124 * level swizzling, not the internal pixel-level swizzling which is
1125 * below OpenGL's reach */
1127 unsigned swizzle
: 12;
1128 unsigned swizzle_zero
: 20;
1134 mali_ptr swizzled_bitmaps
[MAX_MIP_LEVELS
* MAX_FACES
];
1135 } __attribute__((packed
));
1137 /* Used as part of filter_mode */
1139 #define MALI_LINEAR 0
1140 #define MALI_NEAREST 1
1141 #define MALI_MIP_LINEAR (0x18)
1143 /* Used to construct low bits of filter_mode */
1145 #define MALI_TEX_MAG(mode) (((mode) & 1) << 0)
1146 #define MALI_TEX_MIN(mode) (((mode) & 1) << 1)
1148 #define MALI_TEX_MAG_MASK (1)
1149 #define MALI_TEX_MIN_MASK (2)
1151 #define MALI_FILTER_NAME(filter) (filter ? "MALI_NEAREST" : "MALI_LINEAR")
1153 /* Used for lod encoding. Thanks @urjaman for pointing out these routines can
1154 * be cleaned up a lot. */
1156 #define DECODE_FIXED_16(x) ((float) (x / 256.0))
1158 static inline uint16_t
1161 /* Clamp inputs, accounting for float error */
1162 float max_lod
= (32.0 - (1.0 / 512.0));
1164 x
= ((x
> max_lod
) ? max_lod
: ((x
< 0.0) ? 0.0 : x
));
1166 return (int) (x
* 256.0);
1169 struct mali_sampler_descriptor
{
1170 uint32_t filter_mode
;
1172 /* Fixed point. Upper 8-bits is before the decimal point, although it
1173 * caps [0-31]. Lower 8-bits is after the decimal point: int(round(x *
1179 /* All one word in reality, but packed a bit */
1181 enum mali_wrap_mode wrap_s
: 4;
1182 enum mali_wrap_mode wrap_t
: 4;
1183 enum mali_wrap_mode wrap_r
: 4;
1184 enum mali_alt_func compare_func
: 3;
1186 /* A single set bit of unknown, ha! */
1187 unsigned unknown2
: 1;
1192 float border_color
[4];
1193 } __attribute__((packed
));
1195 /* TODO: What are the floats? Apparently always { -inf, -inf, inf, inf },
1196 * unless the scissor test is enabled.
1198 * viewport0/viewport1 form the arguments to glViewport. viewport1 is modified
1199 * by MALI_POSITIVE; viewport0 is as-is.
1202 struct mali_viewport
{
1203 /* XY clipping planes */
1209 /* Depth clipping planes */
1215 } __attribute__((packed
));
1217 /* TODO: Varying meta is symmetrical with attr_meta, but there is some
1218 * weirdness associated. Figure it out. */
1220 struct mali_unknown6
{
1225 /* From presentations, 16x16 tiles externally. Use shift for fast computation
1226 * of tile numbers. */
1228 #define MALI_TILE_SHIFT 4
1229 #define MALI_TILE_LENGTH (1 << MALI_TILE_SHIFT)
1231 /* Tile coordinates are stored as a compact u32, as only 12 bits are needed to
1232 * each component. Notice that this provides a theoretical upper bound of (1 <<
1233 * 12) = 4096 tiles in each direction, addressing a maximum framebuffer of size
1234 * 65536x65536. Multiplying that together, times another four given that Mali
1235 * framebuffers are 32-bit ARGB8888, means that this upper bound would take 16
1236 * gigabytes of RAM just to store the uncompressed framebuffer itself, let
1237 * alone rendering in real-time to such a buffer.
1241 /* From mali_kbase_10969_workaround.c */
1242 #define MALI_X_COORD_MASK 0x00000FFF
1243 #define MALI_Y_COORD_MASK 0x0FFF0000
1245 /* Extract parts of a tile coordinate */
1247 #define MALI_TILE_COORD_X(coord) ((coord) & MALI_X_COORD_MASK)
1248 #define MALI_TILE_COORD_Y(coord) (((coord) & MALI_Y_COORD_MASK) >> 16)
1249 #define MALI_TILE_COORD_FLAGS(coord) ((coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK))
1251 /* No known flags yet, but just in case...? */
1253 #define MALI_TILE_NO_FLAG (0)
1255 /* Helpers to generate tile coordinates based on the boundary coordinates in
1256 * screen space. So, with the bounds (0, 0) to (128, 128) for the screen, these
1257 * functions would convert it to the bounding tiles (0, 0) to (7, 7).
1258 * Intentional "off-by-one"; finding the tile number is a form of fencepost
1261 #define MALI_MAKE_TILE_COORDS(X, Y) ((X) | ((Y) << 16))
1262 #define MALI_BOUND_TO_TILE(B, bias) ((B - bias) >> MALI_TILE_SHIFT)
1263 #define MALI_COORDINATE_TO_TILE(W, H, bias) MALI_MAKE_TILE_COORDS(MALI_BOUND_TO_TILE(W, bias), MALI_BOUND_TO_TILE(H, bias))
1264 #define MALI_COORDINATE_TO_TILE_MIN(W, H) MALI_COORDINATE_TO_TILE(W, H, 0)
1265 #define MALI_COORDINATE_TO_TILE_MAX(W, H) MALI_COORDINATE_TO_TILE(W, H, 1)
1267 struct mali_payload_fragment
{
1270 mali_ptr framebuffer
;
1271 } __attribute__((packed
));
1273 /* (Single?) Framebuffer Descriptor */
1275 /* Flags apply to format. With just MSAA_A and MSAA_B, the framebuffer is
1276 * configured for 4x. With MSAA_8, it is configured for 8x. */
1278 #define MALI_FRAMEBUFFER_MSAA_8 (1 << 3)
1279 #define MALI_FRAMEBUFFER_MSAA_A (1 << 4)
1280 #define MALI_FRAMEBUFFER_MSAA_B (1 << 23)
1282 /* Fast/slow based on whether all three buffers are cleared at once */
1284 #define MALI_CLEAR_FAST (1 << 18)
1285 #define MALI_CLEAR_SLOW (1 << 28)
1286 #define MALI_CLEAR_SLOW_STENCIL (1 << 31)
1288 struct mali_single_framebuffer
{
1291 u64 unknown_address_0
;
1295 /* Exact format is ironically not known, since EGL is finnicky with the
1296 * blob. MSAA, colourspace, etc are configured here. */
1303 /* Purposeful off-by-one in these fields should be accounted for by the
1304 * MALI_DIMENSION macro */
1311 /* By default, the framebuffer is upside down from OpenGL's
1312 * perspective. Set framebuffer to the end and negate the stride to
1313 * flip in the Y direction */
1315 mali_ptr framebuffer
;
1320 /* Depth and stencil buffers are interleaved, it appears, as they are
1321 * set to the same address in captures. Both fields set to zero if the
1322 * buffer is not being cleared. Depending on GL_ENABLE magic, you might
1323 * get a zero enable despite the buffer being present; that still is
1326 mali_ptr depth_buffer
; // not SAME_VA
1327 u64 depth_buffer_enable
;
1329 mali_ptr stencil_buffer
; // not SAME_VA
1330 u64 stencil_buffer_enable
;
1332 u32 clear_color_1
; // RGBA8888 from glClear, actually used by hardware
1333 u32 clear_color_2
; // always equal, but unclear function?
1334 u32 clear_color_3
; // always equal, but unclear function?
1335 u32 clear_color_4
; // always equal, but unclear function?
1337 /* Set to zero if not cleared */
1339 float clear_depth_1
; // float32, ditto
1340 float clear_depth_2
; // float32, ditto
1341 float clear_depth_3
; // float32, ditto
1342 float clear_depth_4
; // float32, ditto
1344 u32 clear_stencil
; // Exactly as it appears in OpenGL
1348 /* Very weird format, see generation code in trans_builder.c */
1349 u32 resolution_check
;
1353 u64 unknown_address_1
; /* Pointing towards... a zero buffer? */
1354 u64 unknown_address_2
;
1356 /* See mali_kbase_replay.c */
1357 u64 tiler_heap_free
;
1360 /* More below this, maybe */
1361 } __attribute__((packed
));
1363 /* Format bits for the render target flags */
1365 #define MALI_MFBD_FORMAT_AFBC (1 << 5)
1366 #define MALI_MFBD_FORMAT_MSAA (1 << 7)
1368 struct mali_rt_format
{
1372 unsigned nr_channels
: 2; /* MALI_POSITIVE */
1374 unsigned flags
: 11;
1376 unsigned swizzle
: 12;
1379 } __attribute__((packed
));
1381 struct bifrost_render_target
{
1382 struct mali_rt_format format
;
1388 /* Stuff related to ARM Framebuffer Compression. When AFBC is enabled,
1389 * there is an extra metadata buffer that contains 16 bytes per tile.
1390 * The framebuffer needs to be the same size as before, since we don't
1391 * know ahead of time how much space it will take up. The
1392 * framebuffer_stride is set to 0, since the data isn't stored linearly
1397 u32 stride
; // stride in units of tiles
1398 u32 unk
; // = 0x20000
1402 /* Heck if I know */
1408 mali_ptr framebuffer
;
1411 u32 framebuffer_stride
: 28; // in units of bytes
1414 u32 clear_color_1
; // RGBA8888 from glClear, actually used by hardware
1415 u32 clear_color_2
; // always equal, but unclear function?
1416 u32 clear_color_3
; // always equal, but unclear function?
1417 u32 clear_color_4
; // always equal, but unclear function?
1418 } __attribute__((packed
));
1420 /* An optional part of bifrost_framebuffer. It comes between the main structure
1421 * and the array of render targets. It must be included if any of these are
1424 * - Transaction Elimination
1426 * - TODO: Anything else?
1429 /* Flags field: note, these are guesses */
1431 #define MALI_EXTRA_PRESENT (0x400)
1432 #define MALI_EXTRA_AFBC (0x20)
1433 #define MALI_EXTRA_AFBC_ZS (0x10)
1434 #define MALI_EXTRA_ZS (0x4)
1436 struct bifrost_fb_extra
{
1438 /* Each tile has an 8 byte checksum, so the stride is "width in tiles * 8" */
1439 u32 checksum_stride
;
1444 /* Note: AFBC is only allowed for 24/8 combined depth/stencil. */
1446 mali_ptr depth_stencil_afbc_metadata
;
1447 u32 depth_stencil_afbc_stride
; // in units of tiles
1450 mali_ptr depth_stencil
;
1456 /* Depth becomes depth/stencil in case of combined D/S */
1458 u32 depth_stride_zero
: 4;
1459 u32 depth_stride
: 28;
1463 u32 stencil_stride_zero
: 4;
1464 u32 stencil_stride
: 28;
1471 } __attribute__((packed
));
1473 /* flags for unk3 */
1475 /* Enables writing depth results back to main memory (rather than keeping them
1476 * on-chip in the tile buffer and then discarding) */
1478 #define MALI_MFBD_DEPTH_WRITE (1 << 10)
1480 /* The MFBD contains the extra bifrost_fb_extra section */
1482 #define MALI_MFBD_EXTRA (1 << 13)
1484 struct bifrost_framebuffer
{
1487 u32 unknown2
; // = 0x1f, same as SFBD
1488 mali_ptr scratchpad
;
1491 mali_ptr sample_locations
;
1494 u16 width1
, height1
;
1496 u16 width2
, height2
;
1497 u32 unk1
: 19; // = 0x01000
1498 u32 rt_count_1
: 2; // off-by-one (use MALI_POSITIVE)
1499 u32 unk2
: 3; // = 0
1500 u32 rt_count_2
: 3; // no off-by-one
1503 u32 clear_stencil
: 8;
1504 u32 unk3
: 24; // = 0x100
1506 mali_ptr tiler_meta
;
1509 /* Note: these are guesses! */
1510 mali_ptr tiler_scratch_start
;
1511 mali_ptr tiler_scratch_middle
;
1513 /* These are not, since we see symmetry with replay jobs which name these explicitly */
1514 mali_ptr tiler_heap_start
;
1515 mali_ptr tiler_heap_end
;
1517 u64 zero9
, zero10
, zero11
, zero12
;
1519 /* optional: struct bifrost_fb_extra extra */
1520 /* struct bifrost_render_target rts[] */
1521 } __attribute__((packed
));
1523 #endif /* __PANFROST_JOB_H__ */