2 * © Copyright 2017-2018 Alyssa Rosenzweig
3 * © Copyright 2017-2018 Connor Abbott
4 * © Copyright 2017-2018 Lyude Paul
5 * © Copyright2019 Collabora
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
28 #ifndef __PANFROST_JOB_H__
29 #define __PANFROST_JOB_H__
32 #include <panfrost-misc.h>
34 #define MALI_SHORT_PTR_BITS (sizeof(uintptr_t)*8)
36 #define MALI_FBD_HIERARCHY_WEIGHTS 8
38 #define MALI_PAYLOAD_SIZE 256
40 typedef u32 mali_jd_core_req
;
45 JOB_TYPE_SET_VALUE
= 2,
46 JOB_TYPE_CACHE_FLUSH
= 3,
49 JOB_TYPE_GEOMETRY
= 6,
52 JOB_TYPE_FRAGMENT
= 9,
59 MALI_LINE_STRIP
= 0x4,
62 MALI_TRIANGLE_STRIP
= 0xA,
63 MALI_TRIANGLE_FAN
= 0xC,
66 MALI_QUAD_STRIP
= 0xF,
68 /* All other modes invalid */
71 /* Applies to tiler_gl_enables */
74 #define MALI_OCCLUSION_QUERY (1 << 3)
75 #define MALI_OCCLUSION_PRECISE (1 << 4)
77 /* Set for a glFrontFace(GL_CCW) in a Y=0=TOP coordinate system (like Gallium).
78 * In OpenGL, this would corresponds to glFrontFace(GL_CW). Mesa and the blob
79 * disagree about how to do viewport flipping, so the blob actually sets this
80 * for GL_CW but then has a negative viewport stride */
81 #define MALI_FRONT_CCW_TOP (1 << 5)
83 #define MALI_CULL_FACE_FRONT (1 << 6)
84 #define MALI_CULL_FACE_BACK (1 << 7)
86 /* TODO: Might this actually be a finer bitfield? */
87 #define MALI_DEPTH_STENCIL_ENABLE 0x6400
89 #define DS_ENABLE(field) \
90 (field == MALI_DEPTH_STENCIL_ENABLE) \
91 ? "MALI_DEPTH_STENCIL_ENABLE" \
92 : (field == 0) ? "0" \
93 : "0 /* XXX: Unknown, check hexdump */"
95 /* Used in stencil and depth tests */
101 MALI_FUNC_LEQUAL
= 3,
102 MALI_FUNC_GREATER
= 4,
103 MALI_FUNC_NOTEQUAL
= 5,
104 MALI_FUNC_GEQUAL
= 6,
108 /* Same OpenGL, but mixed up. Why? Because forget me, that's why! */
111 MALI_ALT_FUNC_NEVER
= 0,
112 MALI_ALT_FUNC_GREATER
= 1,
113 MALI_ALT_FUNC_EQUAL
= 2,
114 MALI_ALT_FUNC_GEQUAL
= 3,
115 MALI_ALT_FUNC_LESS
= 4,
116 MALI_ALT_FUNC_NOTEQUAL
= 5,
117 MALI_ALT_FUNC_LEQUAL
= 6,
118 MALI_ALT_FUNC_ALWAYS
= 7
121 /* Flags apply to unknown2_3? */
123 #define MALI_HAS_MSAA (1 << 0)
124 #define MALI_CAN_DISCARD (1 << 5)
126 /* Applies on SFBD systems, specifying that programmable blending is in use */
127 #define MALI_HAS_BLEND_SHADER (1 << 6)
129 /* func is mali_func */
130 #define MALI_DEPTH_FUNC(func) (func << 8)
131 #define MALI_GET_DEPTH_FUNC(flags) ((flags >> 8) & 0x7)
132 #define MALI_DEPTH_FUNC_MASK MALI_DEPTH_FUNC(0x7)
134 #define MALI_DEPTH_TEST (1 << 11)
136 /* Next flags to unknown2_4 */
137 #define MALI_STENCIL_TEST (1 << 0)
140 #define MALI_SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER (1 << 1)
142 #define MALI_NO_DITHER (1 << 9)
143 #define MALI_DEPTH_RANGE_A (1 << 12)
144 #define MALI_DEPTH_RANGE_B (1 << 13)
145 #define MALI_NO_MSAA (1 << 14)
147 /* Stencil test state is all encoded in a single u32, just with a lot of
150 enum mali_stencil_op
{
151 MALI_STENCIL_KEEP
= 0,
152 MALI_STENCIL_REPLACE
= 1,
153 MALI_STENCIL_ZERO
= 2,
154 MALI_STENCIL_INVERT
= 3,
155 MALI_STENCIL_INCR_WRAP
= 4,
156 MALI_STENCIL_DECR_WRAP
= 5,
157 MALI_STENCIL_INCR
= 6,
158 MALI_STENCIL_DECR
= 7
161 struct mali_stencil_test
{
164 enum mali_func func
: 3;
165 enum mali_stencil_op sfail
: 3;
166 enum mali_stencil_op dpfail
: 3;
167 enum mali_stencil_op dppass
: 3;
169 } __attribute__((packed
));
171 /* Blending is a mess, since anything fancy triggers a blend shader, and
172 * -those- are not understood whatsover yet */
174 #define MALI_MASK_R (1 << 0)
175 #define MALI_MASK_G (1 << 1)
176 #define MALI_MASK_B (1 << 2)
177 #define MALI_MASK_A (1 << 3)
179 enum mali_nondominant_mode
{
180 MALI_BLEND_NON_MIRROR
= 0,
181 MALI_BLEND_NON_ZERO
= 1
184 enum mali_dominant_blend
{
185 MALI_BLEND_DOM_SOURCE
= 0,
186 MALI_BLEND_DOM_DESTINATION
= 1
189 enum mali_dominant_factor
{
190 MALI_DOMINANT_UNK0
= 0,
191 MALI_DOMINANT_ZERO
= 1,
192 MALI_DOMINANT_SRC_COLOR
= 2,
193 MALI_DOMINANT_DST_COLOR
= 3,
194 MALI_DOMINANT_UNK4
= 4,
195 MALI_DOMINANT_SRC_ALPHA
= 5,
196 MALI_DOMINANT_DST_ALPHA
= 6,
197 MALI_DOMINANT_CONSTANT
= 7,
200 enum mali_blend_modifier
{
201 MALI_BLEND_MOD_UNK0
= 0,
202 MALI_BLEND_MOD_NORMAL
= 1,
203 MALI_BLEND_MOD_SOURCE_ONE
= 2,
204 MALI_BLEND_MOD_DEST_ONE
= 3,
207 struct mali_blend_mode
{
208 enum mali_blend_modifier clip_modifier
: 2;
209 unsigned unused_0
: 1;
210 unsigned negate_source
: 1;
212 enum mali_dominant_blend dominant
: 1;
214 enum mali_nondominant_mode nondominant_mode
: 1;
216 unsigned unused_1
: 1;
218 unsigned negate_dest
: 1;
220 enum mali_dominant_factor dominant_factor
: 3;
221 unsigned complement_dominant
: 1;
222 } __attribute__((packed
));
224 struct mali_blend_equation
{
225 /* Of type mali_blend_mode */
226 unsigned rgb_mode
: 12;
227 unsigned alpha_mode
: 12;
231 /* Corresponds to MALI_MASK_* above and glColorMask arguments */
233 unsigned color_mask
: 4;
234 } __attribute__((packed
));
236 /* Used with channel swizzling */
238 MALI_CHANNEL_RED
= 0,
239 MALI_CHANNEL_GREEN
= 1,
240 MALI_CHANNEL_BLUE
= 2,
241 MALI_CHANNEL_ALPHA
= 3,
242 MALI_CHANNEL_ZERO
= 4,
243 MALI_CHANNEL_ONE
= 5,
244 MALI_CHANNEL_RESERVED_0
= 6,
245 MALI_CHANNEL_RESERVED_1
= 7,
248 struct mali_channel_swizzle
{
249 enum mali_channel r
: 3;
250 enum mali_channel g
: 3;
251 enum mali_channel b
: 3;
252 enum mali_channel a
: 3;
253 } __attribute__((packed
));
255 /* Compressed per-pixel formats. Each of these formats expands to one to four
256 * floating-point or integer numbers, as defined by the OpenGL specification.
257 * There are various places in OpenGL where the user can specify a compressed
258 * format in memory, which all use the same 8-bit enum in the various
259 * descriptors, although different hardware units support different formats.
262 /* The top 3 bits specify how the bits of each component are interpreted. */
264 /* e.g. R11F_G11F_B10F */
265 #define MALI_FORMAT_SPECIAL (2 << 5)
267 /* signed normalized, e.g. RGBA8_SNORM */
268 #define MALI_FORMAT_SNORM (3 << 5)
271 #define MALI_FORMAT_UINT (4 << 5)
273 /* e.g. RGBA8 and RGBA32F */
274 #define MALI_FORMAT_UNORM (5 << 5)
276 /* e.g. RGBA8I and RGBA16F */
277 #define MALI_FORMAT_SINT (6 << 5)
279 /* These formats seem to largely duplicate the others. They're used at least
280 * for Bifrost framebuffer output.
282 #define MALI_FORMAT_SPECIAL2 (7 << 5)
284 /* If the high 3 bits are 3 to 6 these two bits say how many components
287 #define MALI_NR_CHANNELS(n) ((n - 1) << 3)
289 /* If the high 3 bits are 3 to 6, then the low 3 bits say how big each
290 * component is, except the special MALI_CHANNEL_FLOAT which overrides what the
294 #define MALI_CHANNEL_4 2
296 #define MALI_CHANNEL_8 3
298 #define MALI_CHANNEL_16 4
300 #define MALI_CHANNEL_32 5
302 /* For MALI_FORMAT_SINT it means a half-float (e.g. RG16F). For
303 * MALI_FORMAT_UNORM, it means a 32-bit float.
305 #define MALI_CHANNEL_FLOAT 7
308 MALI_RGB565
= MALI_FORMAT_SPECIAL
| 0x0,
309 MALI_RGB5_A1_UNORM
= MALI_FORMAT_SPECIAL
| 0x2,
310 MALI_RGB10_A2_UNORM
= MALI_FORMAT_SPECIAL
| 0x3,
311 MALI_RGB10_A2_SNORM
= MALI_FORMAT_SPECIAL
| 0x5,
312 MALI_RGB10_A2UI
= MALI_FORMAT_SPECIAL
| 0x7,
313 MALI_RGB10_A2I
= MALI_FORMAT_SPECIAL
| 0x9,
316 MALI_NV12
= MALI_FORMAT_SPECIAL
| 0xc,
318 MALI_Z32_UNORM
= MALI_FORMAT_SPECIAL
| 0xD,
319 MALI_R32_FIXED
= MALI_FORMAT_SPECIAL
| 0x11,
320 MALI_RG32_FIXED
= MALI_FORMAT_SPECIAL
| 0x12,
321 MALI_RGB32_FIXED
= MALI_FORMAT_SPECIAL
| 0x13,
322 MALI_RGBA32_FIXED
= MALI_FORMAT_SPECIAL
| 0x14,
323 MALI_R11F_G11F_B10F
= MALI_FORMAT_SPECIAL
| 0x19,
324 /* Only used for varyings, to indicate the transformed gl_Position */
325 MALI_VARYING_POS
= MALI_FORMAT_SPECIAL
| 0x1e,
326 /* Only used for varyings, to indicate that the write should be
329 MALI_VARYING_DISCARD
= MALI_FORMAT_SPECIAL
| 0x1f,
331 MALI_R8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
332 MALI_R16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
333 MALI_R32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
334 MALI_RG8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
335 MALI_RG16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
336 MALI_RG32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
337 MALI_RGB8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
338 MALI_RGB16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
339 MALI_RGB32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
340 MALI_RGBA8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
341 MALI_RGBA16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
342 MALI_RGBA32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
344 MALI_R8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
345 MALI_R16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
346 MALI_R32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
347 MALI_RG8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
348 MALI_RG16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
349 MALI_RG32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
350 MALI_RGB8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
351 MALI_RGB16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
352 MALI_RGB32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
353 MALI_RGBA8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
354 MALI_RGBA16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
355 MALI_RGBA32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
357 MALI_R8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
358 MALI_R16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
359 MALI_R32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
360 MALI_R32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT
,
361 MALI_RG8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
362 MALI_RG16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
363 MALI_RG32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
364 MALI_RG32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT
,
365 MALI_RGB8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
366 MALI_RGB16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
367 MALI_RGB32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
368 MALI_RGB32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT
,
369 MALI_RGBA4_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_4
,
370 MALI_RGBA8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
371 MALI_RGBA16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
372 MALI_RGBA32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
373 MALI_RGBA32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT
,
375 MALI_R8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
376 MALI_R16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
377 MALI_R32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
378 MALI_R16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT
,
379 MALI_RG8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
380 MALI_RG16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
381 MALI_RG32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
382 MALI_RG16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT
,
383 MALI_RGB8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
384 MALI_RGB16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
385 MALI_RGB32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
386 MALI_RGB16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT
,
387 MALI_RGBA8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
388 MALI_RGBA16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
389 MALI_RGBA32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
390 MALI_RGBA16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT
,
392 MALI_RGBA4
= MALI_FORMAT_SPECIAL2
| 0x8,
393 MALI_RGBA8_2
= MALI_FORMAT_SPECIAL2
| 0xd,
394 MALI_RGB10_A2_2
= MALI_FORMAT_SPECIAL2
| 0xe,
398 /* Alpha coverage is encoded as 4-bits (from a clampf), with inversion
399 * literally performing a bitwise invert. This function produces slightly wrong
400 * results and I'm not sure why; some rounding issue I suppose... */
402 #define MALI_ALPHA_COVERAGE(clampf) ((uint16_t) (int) (clampf * 15.0f))
403 #define MALI_GET_ALPHA_COVERAGE(nibble) ((float) nibble / 15.0f)
405 /* Applies to unknown1 */
407 /* Should the hardware perform early-Z testing? Normally should be set
408 * for performance reasons. Clear if you use: discard,
409 * alpha-to-coverage... * It's also possible this disables
410 * forward-pixel kill; we're not quite sure which bit is which yet.
411 * TODO: How does this interact with blending?*/
413 #define MALI_EARLY_Z (1 << 10)
415 /* Should the hardware calculate derivatives (via helper invocations)? Set in a
416 * fragment shader that uses texturing or derivative functions */
418 #define MALI_HELPER_INVOCATIONS (1 << 11)
420 /* Flags denoting the fragment shader's use of tilebuffer readback. If the
421 * shader might read any part of the tilebuffer, set MALI_READS_TILEBUFFER. If
422 * it might read depth/stencil in particular, also set MALI_READS_ZS */
424 #define MALI_READS_ZS (1 << 12)
425 #define MALI_READS_TILEBUFFER (1 << 16)
427 /* The raw Midgard blend payload can either be an equation or a shader
428 * address, depending on the context */
430 union midgard_blend
{
434 struct mali_blend_equation equation
;
439 /* On MRT Midgard systems (using an MFBD), each render target gets its own
440 * blend descriptor */
442 struct midgard_blend_rt
{
443 /* Flags base value of 0x200 to enable the render target.
444 * OR with 0x1 for blending (anything other than REPLACE).
445 * OR with 0x2 for programmable blending with 0-2 registers
446 * OR with 0x3 for programmable blending with 2+ registers
450 union midgard_blend blend
;
451 } __attribute__((packed
));
453 /* On Bifrost systems (all MRT), each render target gets one of these
456 struct bifrost_blend_rt
{
457 /* This is likely an analogue of the flags on
458 * midgard_blend_rt */
460 u16 flags
; // = 0x200
462 /* Single-channel blend constants are encoded in a sort of
463 * fixed-point. Basically, the float is mapped to a byte, becoming
464 * a high byte, and then the lower-byte is added for precision.
465 * For the original float f:
467 * f = (constant_hi / 255) + (constant_lo / 65535)
469 * constant_hi = int(f / 255)
470 * constant_lo = 65535*f - (65535/255) * constant_hi
475 struct mali_blend_equation equation
;
478 * - 0x3 when this slot is unused (everything else is 0 except the index)
479 * - 0x11 when this is the fourth slot (and it's used)
480 + * - 0 when there is a blend shader
483 /* increments from 0 to 3 */
488 /* So far, I've only seen:
489 * - R001 for 1-component formats
490 * - RG01 for 2-component formats
491 * - RGB1 for 3-component formats
492 * - RGBA for 4-component formats
495 enum mali_format format
: 8;
497 /* Type of the shader output variable. Note, this can
498 * be different from the format.
500 * 0: f16 (mediump float)
501 * 1: f32 (highp float)
503 * 3: u32 (highp uint)
504 * 4: i16 (mediump int)
505 * 5: u16 (mediump uint)
511 /* Only the low 32 bits of the blend shader are stored, the
512 * high 32 bits are implicitly the same as the original shader.
513 * According to the kernel driver, the program counter for
514 * shaders is actually only 24 bits, so shaders cannot cross
515 * the 2^24-byte boundary, and neither can the blend shader.
516 * The blob handles this by allocating a 2^24 byte pool for
517 * shaders, and making sure that any blend shaders are stored
518 * in the same pool as the original shader. The kernel will
519 * make sure this allocation is aligned to 2^24 bytes.
523 } __attribute__((packed
));
525 /* Descriptor for the shader. Following this is at least one, up to four blend
526 * descriptors for each active render target */
528 struct mali_shader_meta
{
537 u32 uniform_buffer_count
: 4;
538 u32 unk1
: 28; // = 0x800000 for vertex, 0x958020 for tiler
541 /* 0x200 except MALI_NO_ALPHA_TO_COVERAGE. Mysterious 1
542 * other times. Who knows really? */
545 /* Whole number of uniform registers used, times two;
546 * whole number of work registers used (no scale).
548 unsigned work_count
: 5;
549 unsigned uniform_count
: 5;
550 unsigned unknown2
: 6;
554 /* On bifrost: Exactly the same as glPolygonOffset() for both.
555 * On midgard: Depth factor is exactly as passed to glPolygonOffset.
556 * Depth units is equal to the value passed to glDeptOhffset + 1.0f
557 * (use MALI_NEGATIVE)
567 u8 stencil_mask_front
;
568 u8 stencil_mask_back
;
571 struct mali_stencil_test stencil_front
;
572 struct mali_stencil_test stencil_back
;
577 /* On Bifrost, some system values are preloaded in
578 * registers R55-R62 by the thread dispatcher prior to
579 * the start of shader execution. This is a bitfield
580 * with one entry for each register saying which
581 * registers need to be preloaded. Right now, the known
585 * - R55 : gl_LocalInvocationID.xy
586 * - R56 : gl_LocalInvocationID.z + unknown in high 16 bits
587 * - R57 : gl_WorkGroupID.x
588 * - R58 : gl_WorkGroupID.y
589 * - R59 : gl_WorkGroupID.z
590 * - R60 : gl_GlobalInvocationID.x
591 * - R61 : gl_GlobalInvocationID.y/gl_VertexID (without base)
592 * - R62 : gl_GlobalInvocationID.z/gl_InstanceID (without base)
595 * - R55 : unknown, never seen (but the bit for this is
597 * - R56 : unknown (bit always unset)
598 * - R57 : gl_PrimitiveID
599 * - R58 : gl_FrontFacing in low bit, potentially other stuff
600 * - R59 : u16 fragment coordinates (used to compute
601 * gl_FragCoord.xy, together with sample positions)
602 * - R60 : gl_SampleMask (used in epilog, so pretty
603 * much always used, but the bit is always 0 -- is
604 * this just always pushed?)
605 * - R61 : gl_SampleMaskIn and gl_SampleID, used by
606 * varying interpolation.
607 * - R62 : unknown (bit always unset).
609 u32 preload_regs
: 8;
610 /* In units of 8 bytes or 64 bits, since the
611 * uniform/const port loads 64 bits at a time.
613 u32 uniform_count
: 7;
614 u32 unk4
: 10; // = 2
621 /* zero on bifrost */
624 /* Blending information for the older non-MRT Midgard HW. Check for
625 * MALI_HAS_BLEND_SHADER to decide how to interpret.
628 union midgard_blend blend
;
629 } __attribute__((packed
));
631 /* This only concerns hardware jobs */
633 /* Possible values for job_descriptor_size */
635 #define MALI_JOB_32 0
636 #define MALI_JOB_64 1
638 struct mali_job_descriptor_header
{
639 u32 exception_status
;
640 u32 first_incomplete_task
;
642 u8 job_descriptor_size
: 1;
643 enum mali_job_type job_type
: 7;
645 u8 unknown_flags
: 7;
647 u16 job_dependency_index_1
;
648 u16 job_dependency_index_2
;
654 } __attribute__((packed
));
656 struct mali_payload_set_value
{
659 } __attribute__((packed
));
661 /* Special attributes have a fixed index */
662 #define MALI_SPECIAL_ATTRIBUTE_BASE 16
663 #define MALI_VERTEX_ID (MALI_SPECIAL_ATTRIBUTE_BASE + 0)
664 #define MALI_INSTANCE_ID (MALI_SPECIAL_ATTRIBUTE_BASE + 1)
669 * This structure lets the attribute unit compute the address of an attribute
670 * given the vertex and instance ID. Unfortunately, the way this works is
671 * rather complicated when instancing is enabled.
673 * To explain this, first we need to explain how compute and vertex threads are
674 * dispatched. This is a guess (although a pretty firm guess!) since the
675 * details are mostly hidden from the driver, except for attribute instancing.
676 * When a quad is dispatched, it receives a single, linear index. However, we
677 * need to translate that index into a (vertex id, instance id) pair, or a
678 * (local id x, local id y, local id z) triple for compute shaders (although
679 * vertex shaders and compute shaders are handled almost identically).
680 * Focusing on vertex shaders, one option would be to do:
682 * vertex_id = linear_id % num_vertices
683 * instance_id = linear_id / num_vertices
685 * but this involves a costly division and modulus by an arbitrary number.
686 * Instead, we could pad num_vertices. We dispatch padded_num_vertices *
687 * num_instances threads instead of num_vertices * num_instances, which results
688 * in some "extra" threads with vertex_id >= num_vertices, which we have to
689 * discard. The more we pad num_vertices, the more "wasted" threads we
690 * dispatch, but the division is potentially easier.
692 * One straightforward choice is to pad num_vertices to the next power of two,
693 * which means that the division and modulus are just simple bit shifts and
694 * masking. But the actual algorithm is a bit more complicated. The thread
695 * dispatcher has special support for dividing by 3, 5, 7, and 9, in addition
696 * to dividing by a power of two. This is possibly using the technique
697 * described in patent US20170010862A1. As a result, padded_num_vertices can be
698 * 1, 3, 5, 7, or 9 times a power of two. This results in less wasted threads,
699 * since we need less padding.
701 * padded_num_vertices is picked by the hardware. The driver just specifies the
702 * actual number of vertices. At least for Mali G71, the first few cases are
705 * num_vertices | padded_num_vertices
712 * Note that padded_num_vertices is a multiple of four (presumably because
713 * threads are dispatched in groups of 4). Also, padded_num_vertices is always
714 * at least one more than num_vertices, which seems like a quirk of the
715 * hardware. For larger num_vertices, the hardware uses the following
716 * algorithm: using the binary representation of num_vertices, we look at the
717 * most significant set bit as well as the following 3 bits. Let n be the
718 * number of bits after those 4 bits. Then we set padded_num_vertices according
719 * to the following table:
721 * high bits | padded_num_vertices
728 * For example, if num_vertices = 70 is passed to glDraw(), its binary
729 * representation is 1000110, so n = 3 and the high bits are 1000, and
730 * therefore padded_num_vertices = 9 * 2^3 = 72.
732 * The attribute unit works in terms of the original linear_id. if
733 * num_instances = 1, then they are the same, and everything is simple.
734 * However, with instancing things get more complicated. There are four
735 * possible modes, two of them we can group together:
737 * 1. Use the linear_id directly. Only used when there is no instancing.
739 * 2. Use the linear_id modulo a constant. This is used for per-vertex
740 * attributes with instancing enabled by making the constant equal
741 * padded_num_vertices. Because the modulus is always padded_num_vertices, this
742 * mode only supports a modulus that is a power of 2 times 1, 3, 5, 7, or 9.
743 * The shift field specifies the power of two, while the extra_flags field
744 * specifies the odd number. If shift = n and extra_flags = m, then the modulus
745 * is (2m + 1) * 2^n. As an example, if num_vertices = 70, then as computed
746 * above, padded_num_vertices = 9 * 2^3, so we should set extra_flags = 4 and
747 * shift = 3. Note that we must exactly follow the hardware algorithm used to
748 * get padded_num_vertices in order to correctly implement per-vertex
751 * 3. Divide the linear_id by a constant. In order to correctly implement
752 * instance divisors, we have to divide linear_id by padded_num_vertices times
753 * to user-specified divisor. So first we compute padded_num_vertices, again
754 * following the exact same algorithm that the hardware uses, then multiply it
755 * by the GL-level divisor to get the hardware-level divisor. This case is
756 * further divided into two more cases. If the hardware-level divisor is a
757 * power of two, then we just need to shift. The shift amount is specified by
758 * the shift field, so that the hardware-level divisor is just 2^shift.
760 * If it isn't a power of two, then we have to divide by an arbitrary integer.
761 * For that, we use the well-known technique of multiplying by an approximation
762 * of the inverse. The driver must compute the magic multiplier and shift
763 * amount, and then the hardware does the multiplication and shift. The
764 * hardware and driver also use the "round-down" optimization as described in
765 * http://ridiculousfish.com/files/faster_unsigned_division_by_constants.pdf.
766 * The hardware further assumes the multiplier is between 2^31 and 2^32, so the
767 * high bit is implicitly set to 1 even though it is set to 0 by the driver --
768 * presumably this simplifies the hardware multiplier a little. The hardware
769 * first multiplies linear_id by the multiplier and takes the high 32 bits,
770 * then applies the round-down correction if extra_flags = 1, then finally
771 * shifts right by the shift field.
773 * There are some differences between ridiculousfish's algorithm and the Mali
774 * hardware algorithm, which means that the reference code from ridiculousfish
775 * doesn't always produce the right constants. Mali does not use the pre-shift
776 * optimization, since that would make a hardware implementation slower (it
777 * would have to always do the pre-shift, multiply, and post-shift operations).
778 * It also forces the multplier to be at least 2^31, which means that the
779 * exponent is entirely fixed, so there is no trial-and-error. Altogether,
780 * given the divisor d, the algorithm the driver must follow is:
782 * 1. Set shift = floor(log2(d)).
783 * 2. Compute m = ceil(2^(shift + 32) / d) and e = 2^(shift + 32) % d.
784 * 3. If e <= 2^shift, then we need to use the round-down algorithm. Set
785 * magic_divisor = m - 1 and extra_flags = 1.
786 * 4. Otherwise, set magic_divisor = m and extra_flags = 0.
789 enum mali_attr_mode
{
790 MALI_ATTR_UNUSED
= 0,
791 MALI_ATTR_LINEAR
= 1,
792 MALI_ATTR_POT_DIVIDE
= 2,
793 MALI_ATTR_MODULO
= 3,
794 MALI_ATTR_NPOT_DIVIDE
= 4,
797 /* This magic "pseudo-address" is used as `elements` to implement
798 * gl_PointCoord. When read from a fragment shader, it generates a point
799 * coordinate per the OpenGL ES 2.0 specification. Flipped coordinate spaces
800 * require an affine transformation in the shader. */
802 #define MALI_VARYING_POINT_COORD (0x60)
805 /* This is used for actual attributes. */
807 /* The bottom 3 bits are the mode */
808 mali_ptr elements
: 64 - 8;
814 /* The entry after an NPOT_DIVIDE entry has this format. It stores
815 * extra information that wouldn't fit in a normal entry.
818 u32 unk
; /* = 0x20 */
821 /* This is the original, GL-level divisor. */
824 } __attribute__((packed
));
826 struct mali_attr_meta
{
827 /* Vertex buffer index */
830 unsigned unknown1
: 2;
831 unsigned swizzle
: 12;
832 enum mali_format format
: 8;
834 /* Always observed to be zero at the moment */
835 unsigned unknown3
: 2;
837 /* When packing multiple attributes in a buffer, offset addresses by this value */
839 } __attribute__((packed
));
847 #define FBD_MASK (~0x3f)
849 struct mali_uniform_buffer_meta
{
850 /* This is actually the size minus 1 (MALI_POSITIVE), in units of 16
851 * bytes. This gives a maximum of 2^14 bytes, which just so happens to
852 * be the GL minimum-maximum for GL_MAX_UNIFORM_BLOCK_SIZE.
856 /* This is missing the bottom 2 bits and top 8 bits. The top 8 bits
857 * should be 0 for userspace pointers, according to
858 * https://lwn.net/Articles/718895/. By reusing these bits, we can make
859 * each entry in the table only 64 bits.
861 mali_ptr ptr
: 64 - 10;
864 /* On Bifrost, these fields are the same between the vertex and tiler payloads.
865 * They also seem to be the same between Bifrost and Midgard. They're shared in
869 /* Applies to unknown_draw */
871 #define MALI_DRAW_INDEXED_UINT8 (0x10)
872 #define MALI_DRAW_INDEXED_UINT16 (0x20)
873 #define MALI_DRAW_INDEXED_UINT32 (0x30)
874 #define MALI_DRAW_VARYING_SIZE (0x100)
875 #define MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX (0x10000)
877 struct mali_vertex_tiler_prefix
{
878 /* This is a dynamic bitfield containing the following things in this order:
880 * - gl_WorkGroupSize.x
881 * - gl_WorkGroupSize.y
882 * - gl_WorkGroupSize.z
883 * - gl_NumWorkGroups.x
884 * - gl_NumWorkGroups.y
885 * - gl_NumWorkGroups.z
887 * The number of bits allocated for each number is based on the *_shift
888 * fields below. For example, workgroups_y_shift gives the bit that
889 * gl_NumWorkGroups.y starts at, and workgroups_z_shift gives the bit
890 * that gl_NumWorkGroups.z starts at (and therefore one after the bit
891 * that gl_NumWorkGroups.y ends at). The actual value for each gl_*
892 * value is one more than the stored value, since if any of the values
893 * are zero, then there would be no invocations (and hence no job). If
894 * there were 0 bits allocated to a given field, then it must be zero,
895 * and hence the real value is one.
897 * Vertex jobs reuse the same job dispatch mechanism as compute jobs,
898 * effectively doing glDispatchCompute(1, vertex_count, instance_count)
899 * where vertex count is the number of vertices.
901 u32 invocation_count
;
903 u32 size_y_shift
: 5;
904 u32 size_z_shift
: 5;
905 u32 workgroups_x_shift
: 6;
906 u32 workgroups_y_shift
: 6;
907 u32 workgroups_z_shift
: 6;
908 /* This is max(workgroups_x_shift, 2) in all the cases I've seen. */
909 u32 workgroups_x_shift_2
: 4;
912 u32 unknown_draw
: 22;
914 /* This is the the same as workgroups_x_shift_2 in compute shaders, but
915 * always 5 for vertex jobs and 6 for tiler jobs. I suspect this has
916 * something to do with how many quads get put in the same execution
917 * engine, which is a balance (you don't want to starve the engine, but
918 * you also want to distribute work evenly).
920 u32 workgroups_x_shift_3
: 6;
923 /* Negative of draw_start for TILER jobs from what I've seen */
924 int32_t negative_start
;
927 /* Like many other strictly nonzero quantities, index_count is
928 * subtracted by one. For an indexed cube, this is equal to 35 = 6
929 * faces * 2 triangles/per face * 3 vertices/per triangle - 1. That is,
930 * for an indexed draw, index_count is the number of actual vertices
931 * rendered whereas invocation_count is the number of unique vertices
932 * rendered (the number of times the vertex shader must be invoked).
933 * For non-indexed draws, this is just equal to invocation_count. */
937 /* No hidden structure; literally just a pointer to an array of uint
938 * indices (width depends on flags). Thanks, guys, for not making my
939 * life insane for once! NULL for non-indexed draws. */
942 } __attribute__((packed
));
944 /* Point size / line width can either be specified as a 32-bit float (for
945 * constant size) or as a [machine word size]-bit GPU pointer (for varying size). If a pointer
946 * is selected, by setting the appropriate MALI_DRAW_VARYING_SIZE bit in the tiler
947 * payload, the contents of varying_pointer will be intepreted as an array of
948 * fp16 sizes, one for each vertex. gl_PointSize is therefore implemented by
949 * creating a special MALI_R16F varying writing to varying_pointer. */
951 union midgard_primitive_size
{
956 struct bifrost_vertex_only
{
962 } __attribute__((packed
));
964 struct bifrost_tiler_heap_meta
{
967 /* note: these are just guesses! */
968 mali_ptr tiler_heap_start
;
969 mali_ptr tiler_heap_free
;
970 mali_ptr tiler_heap_end
;
972 /* hierarchy weights? but they're still 0 after the job has run... */
974 } __attribute__((packed
));
976 struct bifrost_tiler_meta
{
983 mali_ptr tiler_heap_meta
;
984 /* TODO what is this used for? */
986 } __attribute__((packed
));
988 struct bifrost_tiler_only
{
990 union midgard_primitive_size primitive_size
;
994 u64 zero1
, zero2
, zero3
, zero4
, zero5
, zero6
;
999 } __attribute__((packed
));
1001 struct bifrost_scratchpad
{
1003 u32 flags
; // = 0x1f
1004 /* This is a pointer to a CPU-inaccessible buffer, 16 pages, allocated
1005 * during startup. It seems to serve the same purpose as the
1006 * gpu_scratchpad in the SFBD for Midgard, although it's slightly
1009 mali_ptr gpu_scratchpad
;
1010 } __attribute__((packed
));
1012 struct mali_vertex_tiler_postfix
{
1013 /* Zero for vertex jobs. Pointer to the position (gl_Position) varying
1014 * output from the vertex shader for tiler jobs.
1017 uintptr_t position_varying
;
1019 /* An array of mali_uniform_buffer_meta's. The size is given by the
1022 uintptr_t uniform_buffers
;
1024 /* This is a pointer to an array of pointers to the texture
1025 * descriptors, number of pointers bounded by number of textures. The
1026 * indirection is needed to accomodate varying numbers and sizes of
1027 * texture descriptors */
1028 uintptr_t texture_trampoline
;
1030 /* For OpenGL, from what I've seen, this is intimately connected to
1031 * texture_meta. cwabbott says this is not the case under Vulkan, hence
1032 * why this field is seperate (Midgard is Vulkan capable). Pointer to
1033 * array of sampler descriptors (which are uniform in size) */
1034 uintptr_t sampler_descriptor
;
1038 uintptr_t _shader_upper
: MALI_SHORT_PTR_BITS
- 4; /* struct shader_meta */
1039 uintptr_t attributes
; /* struct attribute_buffer[] */
1040 uintptr_t attribute_meta
; /* attribute_meta[] */
1041 uintptr_t varyings
; /* struct attr */
1042 uintptr_t varying_meta
; /* pointer */
1044 uintptr_t occlusion_counter
; /* A single bit as far as I can tell */
1046 /* Note: on Bifrost, this isn't actually the FBD. It points to
1047 * bifrost_scratchpad instead. However, it does point to the same thing
1048 * in vertex and tiler jobs.
1050 mali_ptr framebuffer
;
1051 } __attribute__((packed
));
1053 struct midgard_payload_vertex_tiler
{
1055 union midgard_primitive_size primitive_size
;
1058 struct mali_vertex_tiler_prefix prefix
;
1064 u32 gl_enables
; // 0x5
1066 /* Offset for first vertex in buffer */
1071 struct mali_vertex_tiler_postfix postfix
;
1074 union midgard_primitive_size primitive_size
;
1076 } __attribute__((packed
));
1078 struct bifrost_payload_vertex
{
1079 struct mali_vertex_tiler_prefix prefix
;
1080 struct bifrost_vertex_only vertex
;
1081 struct mali_vertex_tiler_postfix postfix
;
1082 } __attribute__((packed
));
1084 struct bifrost_payload_tiler
{
1085 struct mali_vertex_tiler_prefix prefix
;
1086 struct bifrost_tiler_only tiler
;
1087 struct mali_vertex_tiler_postfix postfix
;
1088 } __attribute__((packed
));
1090 struct bifrost_payload_fused
{
1091 struct mali_vertex_tiler_prefix prefix
;
1092 struct bifrost_tiler_only tiler
;
1093 struct mali_vertex_tiler_postfix tiler_postfix
;
1094 u64 padding
; /* zero */
1095 struct bifrost_vertex_only vertex
;
1096 struct mali_vertex_tiler_postfix vertex_postfix
;
1097 } __attribute__((packed
));
1099 /* Pointed to from texture_trampoline, mostly unknown still, haven't
1100 * managed to replay successfully */
1102 /* Purposeful off-by-one in width, height fields. For example, a (64, 64)
1103 * texture is stored as (63, 63) in these fields. This adjusts for that.
1104 * There's an identical pattern in the framebuffer descriptor. Even vertex
1105 * count fields work this way, hence the generic name -- integral fields that
1106 * are strictly positive generally need this adjustment. */
1108 #define MALI_POSITIVE(dim) (dim - 1)
1110 /* Opposite of MALI_POSITIVE, found in the depth_units field */
1112 #define MALI_NEGATIVE(dim) (dim + 1)
1114 /* Used with wrapping. Incomplete (this is a 4-bit field...) */
1116 enum mali_wrap_mode
{
1117 MALI_WRAP_REPEAT
= 0x8,
1118 MALI_WRAP_CLAMP_TO_EDGE
= 0x9,
1119 MALI_WRAP_CLAMP_TO_BORDER
= 0xB,
1120 MALI_WRAP_MIRRORED_REPEAT
= 0xC
1124 #define MAX_MIP_LEVELS (13)
1126 /* Cubemap bloats everything up */
1127 #define MAX_FACES (6)
1129 /* For each pointer, there is an address and optionally also a stride */
1130 #define MAX_ELEMENTS (2)
1132 /* Corresponds to the type passed to glTexImage2D and so forth */
1135 #define MALI_TEX_3D (0x04)
1137 /* Flags for usage2 */
1138 #define MALI_TEX_MANUAL_STRIDE (0x20)
1140 struct mali_texture_format
{
1141 unsigned swizzle
: 12;
1142 enum mali_format format
: 8;
1144 unsigned usage1
: 3;
1145 unsigned is_not_cubemap
: 1;
1146 unsigned usage2
: 8;
1147 } __attribute__((packed
));
1149 struct mali_texture_descriptor
{
1153 uint16_t array_size
;
1155 struct mali_texture_format format
;
1159 /* One for non-mipmapped, zero for mipmapped */
1162 /* Zero for non-mipmapped, (number of levels - 1) for mipmapped */
1163 uint8_t nr_mipmap_levels
;
1165 /* Swizzling is a single 32-bit word, broken up here for convenience.
1166 * Here, swizzling refers to the ES 3.0 texture parameters for channel
1167 * level swizzling, not the internal pixel-level swizzling which is
1168 * below OpenGL's reach */
1170 unsigned swizzle
: 12;
1171 unsigned swizzle_zero
: 20;
1177 mali_ptr payload
[MAX_MIP_LEVELS
* MAX_FACES
* MAX_ELEMENTS
];
1178 } __attribute__((packed
));
1180 /* Used as part of filter_mode */
1182 #define MALI_LINEAR 0
1183 #define MALI_NEAREST 1
1184 #define MALI_MIP_LINEAR (0x18)
1186 /* Used to construct low bits of filter_mode */
1188 #define MALI_TEX_MAG(mode) (((mode) & 1) << 0)
1189 #define MALI_TEX_MIN(mode) (((mode) & 1) << 1)
1191 #define MALI_TEX_MAG_MASK (1)
1192 #define MALI_TEX_MIN_MASK (2)
1194 #define MALI_FILTER_NAME(filter) (filter ? "MALI_NEAREST" : "MALI_LINEAR")
1196 /* Used for lod encoding. Thanks @urjaman for pointing out these routines can
1197 * be cleaned up a lot. */
1199 #define DECODE_FIXED_16(x) ((float) (x / 256.0))
1201 static inline uint16_t
1204 /* Clamp inputs, accounting for float error */
1205 float max_lod
= (32.0 - (1.0 / 512.0));
1207 x
= ((x
> max_lod
) ? max_lod
: ((x
< 0.0) ? 0.0 : x
));
1209 return (int) (x
* 256.0);
1212 struct mali_sampler_descriptor
{
1213 uint32_t filter_mode
;
1215 /* Fixed point. Upper 8-bits is before the decimal point, although it
1216 * caps [0-31]. Lower 8-bits is after the decimal point: int(round(x *
1222 /* All one word in reality, but packed a bit */
1224 enum mali_wrap_mode wrap_s
: 4;
1225 enum mali_wrap_mode wrap_t
: 4;
1226 enum mali_wrap_mode wrap_r
: 4;
1227 enum mali_alt_func compare_func
: 3;
1229 /* A single set bit of unknown, ha! */
1230 unsigned unknown2
: 1;
1235 float border_color
[4];
1236 } __attribute__((packed
));
1238 /* viewport0/viewport1 form the arguments to glViewport. viewport1 is
1239 * modified by MALI_POSITIVE; viewport0 is as-is.
1242 struct mali_viewport
{
1243 /* XY clipping planes */
1249 /* Depth clipping planes */
1255 } __attribute__((packed
));
1257 /* From presentations, 16x16 tiles externally. Use shift for fast computation
1258 * of tile numbers. */
1260 #define MALI_TILE_SHIFT 4
1261 #define MALI_TILE_LENGTH (1 << MALI_TILE_SHIFT)
1263 /* Tile coordinates are stored as a compact u32, as only 12 bits are needed to
1264 * each component. Notice that this provides a theoretical upper bound of (1 <<
1265 * 12) = 4096 tiles in each direction, addressing a maximum framebuffer of size
1266 * 65536x65536. Multiplying that together, times another four given that Mali
1267 * framebuffers are 32-bit ARGB8888, means that this upper bound would take 16
1268 * gigabytes of RAM just to store the uncompressed framebuffer itself, let
1269 * alone rendering in real-time to such a buffer.
1273 /* From mali_kbase_10969_workaround.c */
1274 #define MALI_X_COORD_MASK 0x00000FFF
1275 #define MALI_Y_COORD_MASK 0x0FFF0000
1277 /* Extract parts of a tile coordinate */
1279 #define MALI_TILE_COORD_X(coord) ((coord) & MALI_X_COORD_MASK)
1280 #define MALI_TILE_COORD_Y(coord) (((coord) & MALI_Y_COORD_MASK) >> 16)
1281 #define MALI_TILE_COORD_FLAGS(coord) ((coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK))
1283 /* No known flags yet, but just in case...? */
1285 #define MALI_TILE_NO_FLAG (0)
1287 /* Helpers to generate tile coordinates based on the boundary coordinates in
1288 * screen space. So, with the bounds (0, 0) to (128, 128) for the screen, these
1289 * functions would convert it to the bounding tiles (0, 0) to (7, 7).
1290 * Intentional "off-by-one"; finding the tile number is a form of fencepost
1293 #define MALI_MAKE_TILE_COORDS(X, Y) ((X) | ((Y) << 16))
1294 #define MALI_BOUND_TO_TILE(B, bias) ((B - bias) >> MALI_TILE_SHIFT)
1295 #define MALI_COORDINATE_TO_TILE(W, H, bias) MALI_MAKE_TILE_COORDS(MALI_BOUND_TO_TILE(W, bias), MALI_BOUND_TO_TILE(H, bias))
1296 #define MALI_COORDINATE_TO_TILE_MIN(W, H) MALI_COORDINATE_TO_TILE(W, H, 0)
1297 #define MALI_COORDINATE_TO_TILE_MAX(W, H) MALI_COORDINATE_TO_TILE(W, H, 1)
1299 struct mali_payload_fragment
{
1302 mali_ptr framebuffer
;
1303 } __attribute__((packed
));
1305 /* Single Framebuffer Descriptor */
1307 /* Flags apply to format. With just MSAA_A and MSAA_B, the framebuffer is
1308 * configured for 4x. With MSAA_8, it is configured for 8x. */
1310 #define MALI_FRAMEBUFFER_MSAA_8 (1 << 3)
1311 #define MALI_FRAMEBUFFER_MSAA_A (1 << 4)
1312 #define MALI_FRAMEBUFFER_MSAA_B (1 << 23)
1314 /* Fast/slow based on whether all three buffers are cleared at once */
1316 #define MALI_CLEAR_FAST (1 << 18)
1317 #define MALI_CLEAR_SLOW (1 << 28)
1318 #define MALI_CLEAR_SLOW_STENCIL (1 << 31)
1320 struct mali_single_framebuffer
{
1323 u64 unknown_address_0
;
1327 /* Exact format is ironically not known, since EGL is finnicky with the
1328 * blob. MSAA, colourspace, etc are configured here. */
1335 /* Purposeful off-by-one in these fields should be accounted for by the
1336 * MALI_DIMENSION macro */
1343 /* By default, the framebuffer is upside down from OpenGL's
1344 * perspective. Set framebuffer to the end and negate the stride to
1345 * flip in the Y direction */
1347 mali_ptr framebuffer
;
1352 /* Depth and stencil buffers are interleaved, it appears, as they are
1353 * set to the same address in captures. Both fields set to zero if the
1354 * buffer is not being cleared. Depending on GL_ENABLE magic, you might
1355 * get a zero enable despite the buffer being present; that still is
1358 mali_ptr depth_buffer
; // not SAME_VA
1359 u64 depth_buffer_enable
;
1361 mali_ptr stencil_buffer
; // not SAME_VA
1362 u64 stencil_buffer_enable
;
1364 u32 clear_color_1
; // RGBA8888 from glClear, actually used by hardware
1365 u32 clear_color_2
; // always equal, but unclear function?
1366 u32 clear_color_3
; // always equal, but unclear function?
1367 u32 clear_color_4
; // always equal, but unclear function?
1369 /* Set to zero if not cleared */
1371 float clear_depth_1
; // float32, ditto
1372 float clear_depth_2
; // float32, ditto
1373 float clear_depth_3
; // float32, ditto
1374 float clear_depth_4
; // float32, ditto
1376 u32 clear_stencil
; // Exactly as it appears in OpenGL
1380 /* Logically, by symmetry to the MFBD, this ought to be the size of the
1381 * polygon list. But this doesn't quite compute up. More investigation
1384 u32 tiler_resolution_check
;
1386 u16 tiler_hierarchy_mask
;
1389 /* See pan_tiler.c */
1390 mali_ptr tiler_polygon_list
;
1391 mali_ptr tiler_polygon_list_body
;
1393 /* See mali_kbase_replay.c */
1394 mali_ptr tiler_heap_free
;
1395 mali_ptr tiler_heap_end
;
1397 /* More below this, maybe */
1398 } __attribute__((packed
));
1400 /* Format bits for the render target flags */
1402 #define MALI_MFBD_FORMAT_AFBC (1 << 5)
1403 #define MALI_MFBD_FORMAT_MSAA (1 << 7)
1405 struct mali_rt_format
{
1409 unsigned nr_channels
: 2; /* MALI_POSITIVE */
1411 unsigned flags
: 11;
1413 unsigned swizzle
: 12;
1416 } __attribute__((packed
));
1418 struct bifrost_render_target
{
1419 struct mali_rt_format format
;
1425 /* Stuff related to ARM Framebuffer Compression. When AFBC is enabled,
1426 * there is an extra metadata buffer that contains 16 bytes per tile.
1427 * The framebuffer needs to be the same size as before, since we don't
1428 * know ahead of time how much space it will take up. The
1429 * framebuffer_stride is set to 0, since the data isn't stored linearly
1434 u32 stride
; // stride in units of tiles
1435 u32 unk
; // = 0x20000
1439 /* Heck if I know */
1445 mali_ptr framebuffer
;
1448 u32 framebuffer_stride
: 28; // in units of bytes
1451 u32 clear_color_1
; // RGBA8888 from glClear, actually used by hardware
1452 u32 clear_color_2
; // always equal, but unclear function?
1453 u32 clear_color_3
; // always equal, but unclear function?
1454 u32 clear_color_4
; // always equal, but unclear function?
1455 } __attribute__((packed
));
1457 /* An optional part of bifrost_framebuffer. It comes between the main structure
1458 * and the array of render targets. It must be included if any of these are
1461 * - Transaction Elimination
1463 * - TODO: Anything else?
1466 /* Flags field: note, these are guesses */
1468 #define MALI_EXTRA_PRESENT (0x400)
1469 #define MALI_EXTRA_AFBC (0x20)
1470 #define MALI_EXTRA_AFBC_ZS (0x10)
1471 #define MALI_EXTRA_ZS (0x4)
1473 struct bifrost_fb_extra
{
1475 /* Each tile has an 8 byte checksum, so the stride is "width in tiles * 8" */
1476 u32 checksum_stride
;
1481 /* Note: AFBC is only allowed for 24/8 combined depth/stencil. */
1483 mali_ptr depth_stencil_afbc_metadata
;
1484 u32 depth_stencil_afbc_stride
; // in units of tiles
1487 mali_ptr depth_stencil
;
1493 /* Depth becomes depth/stencil in case of combined D/S */
1495 u32 depth_stride_zero
: 4;
1496 u32 depth_stride
: 28;
1500 u32 stencil_stride_zero
: 4;
1501 u32 stencil_stride
: 28;
1508 } __attribute__((packed
));
1510 /* Flags for mfbd_flags */
1512 /* Enables writing depth results back to main memory (rather than keeping them
1513 * on-chip in the tile buffer and then discarding) */
1515 #define MALI_MFBD_DEPTH_WRITE (1 << 10)
1517 /* The MFBD contains the extra bifrost_fb_extra section */
1519 #define MALI_MFBD_EXTRA (1 << 13)
1521 struct bifrost_framebuffer
{
1524 u32 unknown2
; // = 0x1f, same as SFBD
1525 mali_ptr scratchpad
;
1528 mali_ptr sample_locations
;
1531 u16 width1
, height1
;
1533 u16 width2
, height2
;
1534 u32 unk1
: 19; // = 0x01000
1535 u32 rt_count_1
: 2; // off-by-one (use MALI_POSITIVE)
1536 u32 unk2
: 3; // = 0
1537 u32 rt_count_2
: 3; // no off-by-one
1540 u32 clear_stencil
: 8;
1541 u32 mfbd_flags
: 24; // = 0x100
1545 /* Tiler section begins here */
1546 u32 tiler_polygon_list_size
;
1548 /* Name known from the replay workaround in the kernel. What exactly is
1549 * flagged here is less known. We do that (tiler_hierarchy_mask & 0x1ff)
1550 * specifies a mask of hierarchy weights, which explains some of the
1551 * performance mysteries around setting it. We also see the bottom bit
1552 * of tiler_flags set in the kernel, but no comment why. */
1554 u16 tiler_hierarchy_mask
;
1557 /* See mali_tiler.c for an explanation */
1558 mali_ptr tiler_polygon_list
;
1559 mali_ptr tiler_polygon_list_body
;
1561 /* Names based on we see symmetry with replay jobs which name these
1564 mali_ptr tiler_heap_start
; /* tiler heap_free_address */
1565 mali_ptr tiler_heap_end
;
1567 u32 tiler_weights
[8];
1569 /* optional: struct bifrost_fb_extra extra */
1570 /* struct bifrost_render_target rts[] */
1571 } __attribute__((packed
));
1573 #endif /* __PANFROST_JOB_H__ */