2 * © Copyright 2017-2018 Alyssa Rosenzweig
3 * © Copyright 2017-2018 Connor Abbott
4 * © Copyright 2017-2018 Lyude Paul
5 * © Copyright2019 Collabora, Ltd.
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__
33 #include <panfrost-misc.h>
38 JOB_TYPE_WRITE_VALUE
= 2,
39 JOB_TYPE_CACHE_FLUSH
= 3,
42 JOB_TYPE_GEOMETRY
= 6,
45 JOB_TYPE_FRAGMENT
= 9,
52 MALI_LINE_STRIP
= 0x4,
55 MALI_TRIANGLE_STRIP
= 0xA,
56 MALI_TRIANGLE_FAN
= 0xC,
59 MALI_QUAD_STRIP
= 0xF,
61 /* All other modes invalid */
64 /* Applies to tiler_gl_enables */
66 #define MALI_OCCLUSION_QUERY (1 << 3)
67 #define MALI_OCCLUSION_PRECISE (1 << 4)
69 /* Set for a glFrontFace(GL_CCW) in a Y=0=TOP coordinate system (like Gallium).
70 * In OpenGL, this would corresponds to glFrontFace(GL_CW). Mesa and the blob
71 * disagree about how to do viewport flipping, so the blob actually sets this
72 * for GL_CW but then has a negative viewport stride */
74 #define MALI_FRONT_CCW_TOP (1 << 5)
76 #define MALI_CULL_FACE_FRONT (1 << 6)
77 #define MALI_CULL_FACE_BACK (1 << 7)
79 /* Used in stencil and depth tests */
86 MALI_FUNC_GREATER
= 4,
87 MALI_FUNC_NOTEQUAL
= 5,
92 /* Flags apply to unknown2_3? */
94 #define MALI_HAS_MSAA (1 << 0)
95 #define MALI_CAN_DISCARD (1 << 5)
97 /* Applies on SFBD systems, specifying that programmable blending is in use */
98 #define MALI_HAS_BLEND_SHADER (1 << 6)
100 /* func is mali_func */
101 #define MALI_DEPTH_FUNC(func) (func << 8)
102 #define MALI_GET_DEPTH_FUNC(flags) ((flags >> 8) & 0x7)
103 #define MALI_DEPTH_FUNC_MASK MALI_DEPTH_FUNC(0x7)
105 #define MALI_DEPTH_WRITEMASK (1 << 11)
107 /* Next flags to unknown2_4 */
108 #define MALI_STENCIL_TEST (1 << 0)
111 #define MALI_SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER (1 << 1)
113 #define MALI_NO_DITHER (1 << 9)
114 #define MALI_DEPTH_RANGE_A (1 << 12)
115 #define MALI_DEPTH_RANGE_B (1 << 13)
116 #define MALI_NO_MSAA (1 << 14)
118 /* Stencil test state is all encoded in a single u32, just with a lot of
121 enum mali_stencil_op
{
122 MALI_STENCIL_KEEP
= 0,
123 MALI_STENCIL_REPLACE
= 1,
124 MALI_STENCIL_ZERO
= 2,
125 MALI_STENCIL_INVERT
= 3,
126 MALI_STENCIL_INCR_WRAP
= 4,
127 MALI_STENCIL_DECR_WRAP
= 5,
128 MALI_STENCIL_INCR
= 6,
129 MALI_STENCIL_DECR
= 7
132 struct mali_stencil_test
{
135 enum mali_func func
: 3;
136 enum mali_stencil_op sfail
: 3;
137 enum mali_stencil_op dpfail
: 3;
138 enum mali_stencil_op dppass
: 3;
140 } __attribute__((packed
));
142 #define MALI_MASK_R (1 << 0)
143 #define MALI_MASK_G (1 << 1)
144 #define MALI_MASK_B (1 << 2)
145 #define MALI_MASK_A (1 << 3)
147 enum mali_nondominant_mode
{
148 MALI_BLEND_NON_MIRROR
= 0,
149 MALI_BLEND_NON_ZERO
= 1
152 enum mali_dominant_blend
{
153 MALI_BLEND_DOM_SOURCE
= 0,
154 MALI_BLEND_DOM_DESTINATION
= 1
157 enum mali_dominant_factor
{
158 MALI_DOMINANT_UNK0
= 0,
159 MALI_DOMINANT_ZERO
= 1,
160 MALI_DOMINANT_SRC_COLOR
= 2,
161 MALI_DOMINANT_DST_COLOR
= 3,
162 MALI_DOMINANT_UNK4
= 4,
163 MALI_DOMINANT_SRC_ALPHA
= 5,
164 MALI_DOMINANT_DST_ALPHA
= 6,
165 MALI_DOMINANT_CONSTANT
= 7,
168 enum mali_blend_modifier
{
169 MALI_BLEND_MOD_UNK0
= 0,
170 MALI_BLEND_MOD_NORMAL
= 1,
171 MALI_BLEND_MOD_SOURCE_ONE
= 2,
172 MALI_BLEND_MOD_DEST_ONE
= 3,
175 struct mali_blend_mode
{
176 enum mali_blend_modifier clip_modifier
: 2;
177 unsigned unused_0
: 1;
178 unsigned negate_source
: 1;
180 enum mali_dominant_blend dominant
: 1;
182 enum mali_nondominant_mode nondominant_mode
: 1;
184 unsigned unused_1
: 1;
186 unsigned negate_dest
: 1;
188 enum mali_dominant_factor dominant_factor
: 3;
189 unsigned complement_dominant
: 1;
190 } __attribute__((packed
));
192 struct mali_blend_equation
{
193 /* Of type mali_blend_mode */
194 unsigned rgb_mode
: 12;
195 unsigned alpha_mode
: 12;
199 /* Corresponds to MALI_MASK_* above and glColorMask arguments */
201 unsigned color_mask
: 4;
202 } __attribute__((packed
));
204 /* Used with channel swizzling */
206 MALI_CHANNEL_RED
= 0,
207 MALI_CHANNEL_GREEN
= 1,
208 MALI_CHANNEL_BLUE
= 2,
209 MALI_CHANNEL_ALPHA
= 3,
210 MALI_CHANNEL_ZERO
= 4,
211 MALI_CHANNEL_ONE
= 5,
212 MALI_CHANNEL_RESERVED_0
= 6,
213 MALI_CHANNEL_RESERVED_1
= 7,
216 struct mali_channel_swizzle
{
217 enum mali_channel r
: 3;
218 enum mali_channel g
: 3;
219 enum mali_channel b
: 3;
220 enum mali_channel a
: 3;
221 } __attribute__((packed
));
223 /* Compressed per-pixel formats. Each of these formats expands to one to four
224 * floating-point or integer numbers, as defined by the OpenGL specification.
225 * There are various places in OpenGL where the user can specify a compressed
226 * format in memory, which all use the same 8-bit enum in the various
227 * descriptors, although different hardware units support different formats.
230 /* The top 3 bits specify how the bits of each component are interpreted. */
233 #define MALI_FORMAT_COMPRESSED (0 << 5)
235 /* e.g. R11F_G11F_B10F */
236 #define MALI_FORMAT_SPECIAL (2 << 5)
238 /* signed normalized, e.g. RGBA8_SNORM */
239 #define MALI_FORMAT_SNORM (3 << 5)
242 #define MALI_FORMAT_UINT (4 << 5)
244 /* e.g. RGBA8 and RGBA32F */
245 #define MALI_FORMAT_UNORM (5 << 5)
247 /* e.g. RGBA8I and RGBA16F */
248 #define MALI_FORMAT_SINT (6 << 5)
250 /* These formats seem to largely duplicate the others. They're used at least
251 * for Bifrost framebuffer output.
253 #define MALI_FORMAT_SPECIAL2 (7 << 5)
255 /* If the high 3 bits are 3 to 6 these two bits say how many components
258 #define MALI_NR_CHANNELS(n) ((n - 1) << 3)
260 /* If the high 3 bits are 3 to 6, then the low 3 bits say how big each
261 * component is, except the special MALI_CHANNEL_FLOAT which overrides what the
265 #define MALI_CHANNEL_4 2
267 #define MALI_CHANNEL_8 3
269 #define MALI_CHANNEL_16 4
271 #define MALI_CHANNEL_32 5
273 /* For MALI_FORMAT_SINT it means a half-float (e.g. RG16F). For
274 * MALI_FORMAT_UNORM, it means a 32-bit float.
276 #define MALI_CHANNEL_FLOAT 7
279 MALI_ETC2_RGB8
= MALI_FORMAT_COMPRESSED
| 0x1,
280 MALI_ETC2_R11_UNORM
= MALI_FORMAT_COMPRESSED
| 0x2,
281 MALI_ETC2_RGBA8
= MALI_FORMAT_COMPRESSED
| 0x3,
282 MALI_ETC2_RG11_UNORM
= MALI_FORMAT_COMPRESSED
| 0x4,
283 MALI_ETC2_R11_SNORM
= MALI_FORMAT_COMPRESSED
| 0x11,
284 MALI_ETC2_RG11_SNORM
= MALI_FORMAT_COMPRESSED
| 0x12,
285 MALI_ETC2_RGB8A1
= MALI_FORMAT_COMPRESSED
| 0x13,
286 MALI_ASTC_SRGB_SUPP
= MALI_FORMAT_COMPRESSED
| 0x16,
287 MALI_ASTC_HDR_SUPP
= MALI_FORMAT_COMPRESSED
| 0x17,
289 MALI_RGB565
= MALI_FORMAT_SPECIAL
| 0x0,
290 MALI_RGB5_X1_UNORM
= MALI_FORMAT_SPECIAL
| 0x1,
291 MALI_RGB5_A1_UNORM
= MALI_FORMAT_SPECIAL
| 0x2,
292 MALI_RGB10_A2_UNORM
= MALI_FORMAT_SPECIAL
| 0x3,
293 MALI_RGB10_A2_SNORM
= MALI_FORMAT_SPECIAL
| 0x5,
294 MALI_RGB10_A2UI
= MALI_FORMAT_SPECIAL
| 0x7,
295 MALI_RGB10_A2I
= MALI_FORMAT_SPECIAL
| 0x9,
297 MALI_RGB332_UNORM
= MALI_FORMAT_SPECIAL
| 0xb,
298 MALI_RGB233_UNORM
= MALI_FORMAT_SPECIAL
| 0xc,
300 MALI_Z24X8_UNORM
= MALI_FORMAT_SPECIAL
| 0xd,
301 MALI_R32_FIXED
= MALI_FORMAT_SPECIAL
| 0x11,
302 MALI_RG32_FIXED
= MALI_FORMAT_SPECIAL
| 0x12,
303 MALI_RGB32_FIXED
= MALI_FORMAT_SPECIAL
| 0x13,
304 MALI_RGBA32_FIXED
= MALI_FORMAT_SPECIAL
| 0x14,
305 MALI_R11F_G11F_B10F
= MALI_FORMAT_SPECIAL
| 0x19,
306 MALI_R9F_G9F_B9F_E5F
= MALI_FORMAT_SPECIAL
| 0x1b,
307 /* Only used for varyings, to indicate the transformed gl_Position */
308 MALI_VARYING_POS
= MALI_FORMAT_SPECIAL
| 0x1e,
309 /* Only used for varyings, to indicate that the write should be
312 MALI_VARYING_DISCARD
= MALI_FORMAT_SPECIAL
| 0x1f,
314 MALI_R8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
315 MALI_R16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
316 MALI_R32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
317 MALI_RG8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
318 MALI_RG16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
319 MALI_RG32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
320 MALI_RGB8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
321 MALI_RGB16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
322 MALI_RGB32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
323 MALI_RGBA8_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
324 MALI_RGBA16_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
325 MALI_RGBA32_SNORM
= MALI_FORMAT_SNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
327 MALI_R8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
328 MALI_R16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
329 MALI_R32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
330 MALI_RG8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
331 MALI_RG16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
332 MALI_RG32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
333 MALI_RGB8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
334 MALI_RGB16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
335 MALI_RGB32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
336 MALI_RGBA8UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
337 MALI_RGBA16UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
338 MALI_RGBA32UI
= MALI_FORMAT_UINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
340 MALI_R8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
341 MALI_R16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
342 MALI_R32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
343 MALI_R32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT
,
344 MALI_RG8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
345 MALI_RG16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
346 MALI_RG32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
347 MALI_RG32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT
,
348 MALI_RGB8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
349 MALI_RGB16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
350 MALI_RGB32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
351 MALI_RGB32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT
,
352 MALI_RGBA4_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_4
,
353 MALI_RGBA8_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
354 MALI_RGBA16_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
355 MALI_RGBA32_UNORM
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
356 MALI_RGBA32F
= MALI_FORMAT_UNORM
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT
,
358 MALI_R8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_8
,
359 MALI_R16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_16
,
360 MALI_R32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_32
,
361 MALI_R16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT
,
362 MALI_RG8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_8
,
363 MALI_RG16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_16
,
364 MALI_RG32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_32
,
365 MALI_RG16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT
,
366 MALI_RGB8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_8
,
367 MALI_RGB16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_16
,
368 MALI_RGB32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_32
,
369 MALI_RGB16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT
,
370 MALI_RGBA8I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_8
,
371 MALI_RGBA16I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_16
,
372 MALI_RGBA32I
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_32
,
373 MALI_RGBA16F
= MALI_FORMAT_SINT
| MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT
,
375 MALI_RGBA4
= MALI_FORMAT_SPECIAL2
| 0x8,
376 MALI_RGBA8_2
= MALI_FORMAT_SPECIAL2
| 0xd,
377 MALI_RGB10_A2_2
= MALI_FORMAT_SPECIAL2
| 0xe,
381 /* Alpha coverage is encoded as 4-bits (from a clampf), with inversion
382 * literally performing a bitwise invert. This function produces slightly wrong
383 * results and I'm not sure why; some rounding issue I suppose... */
385 #define MALI_ALPHA_COVERAGE(clampf) ((uint16_t) (int) (clampf * 15.0f))
386 #define MALI_GET_ALPHA_COVERAGE(nibble) ((float) nibble / 15.0f)
388 /* Applies to midgard1.flags_lo */
390 /* Should be set when the fragment shader updates the depth value. */
391 #define MALI_WRITES_Z (1 << 4)
393 /* Should the hardware perform early-Z testing? Normally should be set
394 * for performance reasons. Clear if you use: discard,
395 * alpha-to-coverage... * It's also possible this disables
396 * forward-pixel kill; we're not quite sure which bit is which yet.
397 * TODO: How does this interact with blending?*/
399 #define MALI_EARLY_Z (1 << 6)
401 /* Should the hardware calculate derivatives (via helper invocations)? Set in a
402 * fragment shader that uses texturing or derivative functions */
404 #define MALI_HELPER_INVOCATIONS (1 << 7)
406 /* Flags denoting the fragment shader's use of tilebuffer readback. If the
407 * shader might read any part of the tilebuffer, set MALI_READS_TILEBUFFER. If
408 * it might read depth/stencil in particular, also set MALI_READS_ZS */
410 #define MALI_READS_ZS (1 << 8)
411 #define MALI_READS_TILEBUFFER (1 << 12)
413 /* Applies to midgard1.flags_hi */
415 /* Should be set when the fragment shader updates the stencil value. */
416 #define MALI_WRITES_S (1 << 2)
418 /* Mode to suppress generation of Infinity and NaN values by clamping inf
419 * (-inf) to MAX_FLOAT (-MIN_FLOAT) and flushing NaN to 0.0
421 * Compare suppress_inf/suppress_nan flags on the Bifrost clause header for the
422 * same functionality.
424 * This is not conformant on GLES3 or OpenCL, but is optional on GLES2, where
425 * it works around app bugs (e.g. in glmark2-es2 -bterrain with FP16).
427 #define MALI_SUPPRESS_INF_NAN (1 << 3)
429 /* Flags for bifrost1.unk1 */
431 /* Shader uses less than 32 registers, partitioned as [R0, R15] U [R48, R63],
432 * allowing for full thread count. If clear, the full [R0, R63] register set is
433 * available at half thread count */
434 #define MALI_BIFROST_FULL_THREAD (1 << 9)
436 /* Enable early-z testing (presumably). This flag may not be set if the shader:
440 * - Writes gl_FragDepth
442 * This differs from Midgard which sets the MALI_EARLY_Z flag even with
443 * blending, although I've begun to suspect that flag does not in fact enable
445 #define MALI_BIFROST_EARLY_Z (1 << 15)
447 /* First clause type is ATEST */
448 #define MALI_BIFROST_FIRST_ATEST (1 << 26)
450 /* The raw Midgard blend payload can either be an equation or a shader
451 * address, depending on the context */
453 union midgard_blend
{
457 struct mali_blend_equation equation
;
462 /* We need to load the tilebuffer to blend (i.e. the destination factor is not
465 #define MALI_BLEND_LOAD_TIB (0x1)
467 /* A blend shader is used to blend this render target */
468 #define MALI_BLEND_MRT_SHADER (0x2)
470 /* On MRT Midgard systems (using an MFBD), each render target gets its own
471 * blend descriptor */
473 #define MALI_BLEND_SRGB (0x400)
475 /* Dithering is specified here for MFBD, otherwise NO_DITHER for SFBD */
476 #define MALI_BLEND_NO_DITHER (0x800)
478 struct midgard_blend_rt
{
479 /* Flags base value of 0x200 to enable the render target.
480 * OR with 0x1 for blending (anything other than REPLACE).
481 * OR with 0x2 for programmable blending
482 * OR with MALI_BLEND_SRGB for implicit sRGB
486 union midgard_blend blend
;
487 } __attribute__((packed
));
489 /* On Bifrost systems (all MRT), each render target gets one of these
492 enum bifrost_shader_type
{
493 BIFROST_BLEND_F16
= 0,
494 BIFROST_BLEND_F32
= 1,
495 BIFROST_BLEND_I32
= 2,
496 BIFROST_BLEND_U32
= 3,
497 BIFROST_BLEND_I16
= 4,
498 BIFROST_BLEND_U16
= 5,
501 #define BIFROST_MAX_RENDER_TARGET_COUNT 8
503 struct bifrost_blend_rt
{
504 /* This is likely an analogue of the flags on
505 * midgard_blend_rt */
507 u16 flags
; // = 0x200
509 /* Single-channel blend constants are encoded in a sort of
510 * fixed-point. Basically, the float is mapped to a byte, becoming
511 * a high byte, and then the lower-byte is added for precision.
512 * For the original float f:
514 * f = (constant_hi / 255) + (constant_lo / 65535)
516 * constant_hi = int(f / 255)
517 * constant_lo = 65535*f - (65535/255) * constant_hi
521 struct mali_blend_equation equation
;
525 * - 0x3 when this slot is unused (everything else is 0 except the index)
526 * - 0x11 when this is the fourth slot (and it's used)
527 * - 0 when there is a blend shader
531 /* increments from 0 to 3 */
536 /* So far, I've only seen:
537 * - R001 for 1-component formats
538 * - RG01 for 2-component formats
539 * - RGB1 for 3-component formats
540 * - RGBA for 4-component formats
543 enum mali_format format
: 8;
545 /* Type of the shader output variable. Note, this can
546 * be different from the format.
547 * enum bifrost_shader_type
554 /* Only the low 32 bits of the blend shader are stored, the
555 * high 32 bits are implicitly the same as the original shader.
556 * According to the kernel driver, the program counter for
557 * shaders is actually only 24 bits, so shaders cannot cross
558 * the 2^24-byte boundary, and neither can the blend shader.
559 * The blob handles this by allocating a 2^24 byte pool for
560 * shaders, and making sure that any blend shaders are stored
561 * in the same pool as the original shader. The kernel will
562 * make sure this allocation is aligned to 2^24 bytes.
566 } __attribute__((packed
));
568 /* Descriptor for the shader. Following this is at least one, up to four blend
569 * descriptors for each active render target */
571 struct mali_shader_meta
{
580 u32 uniform_buffer_count
: 4;
581 u32 unk1
: 28; // = 0x800000 for vertex, 0x958020 for tiler
584 unsigned uniform_buffer_count
: 4;
585 unsigned flags_lo
: 12;
588 unsigned work_count
: 5;
589 unsigned uniform_count
: 5;
590 unsigned flags_hi
: 6;
594 /* Same as glPolygoOffset() arguments */
603 u8 stencil_mask_front
;
604 u8 stencil_mask_back
;
607 struct mali_stencil_test stencil_front
;
608 struct mali_stencil_test stencil_back
;
613 /* On Bifrost, some system values are preloaded in
614 * registers R55-R62 by the thread dispatcher prior to
615 * the start of shader execution. This is a bitfield
616 * with one entry for each register saying which
617 * registers need to be preloaded. Right now, the known
621 * - R55 : gl_LocalInvocationID.xy
622 * - R56 : gl_LocalInvocationID.z + unknown in high 16 bits
623 * - R57 : gl_WorkGroupID.x
624 * - R58 : gl_WorkGroupID.y
625 * - R59 : gl_WorkGroupID.z
626 * - R60 : gl_GlobalInvocationID.x
627 * - R61 : gl_GlobalInvocationID.y/gl_VertexID (without base)
628 * - R62 : gl_GlobalInvocationID.z/gl_InstanceID (without base)
631 * - R55 : unknown, never seen (but the bit for this is
633 * - R56 : unknown (bit always unset)
634 * - R57 : gl_PrimitiveID
635 * - R58 : gl_FrontFacing in low bit, potentially other stuff
636 * - R59 : u16 fragment coordinates (used to compute
637 * gl_FragCoord.xy, together with sample positions)
638 * - R60 : gl_SampleMask (used in epilog, so pretty
639 * much always used, but the bit is always 0 -- is
640 * this just always pushed?)
641 * - R61 : gl_SampleMaskIn and gl_SampleID, used by
642 * varying interpolation.
643 * - R62 : unknown (bit always unset).
645 * Later GPUs (starting with Mali-G52?) support
646 * preloading float varyings into r0-r7. This is
647 * indicated by setting 0x40. There is no distinction
648 * here between 1 varying and 2.
650 u32 preload_regs
: 8;
651 /* In units of 8 bytes or 64 bits, since the
652 * uniform/const port loads 64 bits at a time.
654 u32 uniform_count
: 7;
655 u32 unk4
: 10; // = 2
664 /* Blending information for the older non-MRT Midgard HW. Check for
665 * MALI_HAS_BLEND_SHADER to decide how to interpret.
668 union midgard_blend blend
;
669 } __attribute__((packed
));
671 /* This only concerns hardware jobs */
673 /* Possible values for job_descriptor_size */
675 #define MALI_JOB_32 0
676 #define MALI_JOB_64 1
678 struct mali_job_descriptor_header
{
679 u32 exception_status
;
680 u32 first_incomplete_task
;
682 u8 job_descriptor_size
: 1;
683 enum mali_job_type job_type
: 7;
685 u8 unknown_flags
: 7;
687 u16 job_dependency_index_1
;
688 u16 job_dependency_index_2
;
690 } __attribute__((packed
));
692 /* These concern exception_status */
694 /* Access type causing a fault, paralleling AS_FAULTSTATUS_* entries in the
697 enum mali_exception_access
{
698 /* Atomic in the kernel for MMU, but that doesn't make sense for a job
699 * fault so it's just unused */
700 MALI_EXCEPTION_ACCESS_NONE
= 0,
702 MALI_EXCEPTION_ACCESS_EXECUTE
= 1,
703 MALI_EXCEPTION_ACCESS_READ
= 2,
704 MALI_EXCEPTION_ACCESS_WRITE
= 3
707 /* Details about write_value from panfrost igt tests which use it as a generic
708 * dword write primitive */
710 #define MALI_WRITE_VALUE_ZERO 3
712 struct mali_payload_write_value
{
714 u32 value_descriptor
;
717 } __attribute__((packed
));
722 * This structure lets the attribute unit compute the address of an attribute
723 * given the vertex and instance ID. Unfortunately, the way this works is
724 * rather complicated when instancing is enabled.
726 * To explain this, first we need to explain how compute and vertex threads are
727 * dispatched. This is a guess (although a pretty firm guess!) since the
728 * details are mostly hidden from the driver, except for attribute instancing.
729 * When a quad is dispatched, it receives a single, linear index. However, we
730 * need to translate that index into a (vertex id, instance id) pair, or a
731 * (local id x, local id y, local id z) triple for compute shaders (although
732 * vertex shaders and compute shaders are handled almost identically).
733 * Focusing on vertex shaders, one option would be to do:
735 * vertex_id = linear_id % num_vertices
736 * instance_id = linear_id / num_vertices
738 * but this involves a costly division and modulus by an arbitrary number.
739 * Instead, we could pad num_vertices. We dispatch padded_num_vertices *
740 * num_instances threads instead of num_vertices * num_instances, which results
741 * in some "extra" threads with vertex_id >= num_vertices, which we have to
742 * discard. The more we pad num_vertices, the more "wasted" threads we
743 * dispatch, but the division is potentially easier.
745 * One straightforward choice is to pad num_vertices to the next power of two,
746 * which means that the division and modulus are just simple bit shifts and
747 * masking. But the actual algorithm is a bit more complicated. The thread
748 * dispatcher has special support for dividing by 3, 5, 7, and 9, in addition
749 * to dividing by a power of two. This is possibly using the technique
750 * described in patent US20170010862A1. As a result, padded_num_vertices can be
751 * 1, 3, 5, 7, or 9 times a power of two. This results in less wasted threads,
752 * since we need less padding.
754 * padded_num_vertices is picked by the hardware. The driver just specifies the
755 * actual number of vertices. At least for Mali G71, the first few cases are
758 * num_vertices | padded_num_vertices
765 * Note that padded_num_vertices is a multiple of four (presumably because
766 * threads are dispatched in groups of 4). Also, padded_num_vertices is always
767 * at least one more than num_vertices, which seems like a quirk of the
768 * hardware. For larger num_vertices, the hardware uses the following
769 * algorithm: using the binary representation of num_vertices, we look at the
770 * most significant set bit as well as the following 3 bits. Let n be the
771 * number of bits after those 4 bits. Then we set padded_num_vertices according
772 * to the following table:
774 * high bits | padded_num_vertices
781 * For example, if num_vertices = 70 is passed to glDraw(), its binary
782 * representation is 1000110, so n = 3 and the high bits are 1000, and
783 * therefore padded_num_vertices = 9 * 2^3 = 72.
785 * The attribute unit works in terms of the original linear_id. if
786 * num_instances = 1, then they are the same, and everything is simple.
787 * However, with instancing things get more complicated. There are four
788 * possible modes, two of them we can group together:
790 * 1. Use the linear_id directly. Only used when there is no instancing.
792 * 2. Use the linear_id modulo a constant. This is used for per-vertex
793 * attributes with instancing enabled by making the constant equal
794 * padded_num_vertices. Because the modulus is always padded_num_vertices, this
795 * mode only supports a modulus that is a power of 2 times 1, 3, 5, 7, or 9.
796 * The shift field specifies the power of two, while the extra_flags field
797 * specifies the odd number. If shift = n and extra_flags = m, then the modulus
798 * is (2m + 1) * 2^n. As an example, if num_vertices = 70, then as computed
799 * above, padded_num_vertices = 9 * 2^3, so we should set extra_flags = 4 and
800 * shift = 3. Note that we must exactly follow the hardware algorithm used to
801 * get padded_num_vertices in order to correctly implement per-vertex
804 * 3. Divide the linear_id by a constant. In order to correctly implement
805 * instance divisors, we have to divide linear_id by padded_num_vertices times
806 * to user-specified divisor. So first we compute padded_num_vertices, again
807 * following the exact same algorithm that the hardware uses, then multiply it
808 * by the GL-level divisor to get the hardware-level divisor. This case is
809 * further divided into two more cases. If the hardware-level divisor is a
810 * power of two, then we just need to shift. The shift amount is specified by
811 * the shift field, so that the hardware-level divisor is just 2^shift.
813 * If it isn't a power of two, then we have to divide by an arbitrary integer.
814 * For that, we use the well-known technique of multiplying by an approximation
815 * of the inverse. The driver must compute the magic multiplier and shift
816 * amount, and then the hardware does the multiplication and shift. The
817 * hardware and driver also use the "round-down" optimization as described in
818 * http://ridiculousfish.com/files/faster_unsigned_division_by_constants.pdf.
819 * The hardware further assumes the multiplier is between 2^31 and 2^32, so the
820 * high bit is implicitly set to 1 even though it is set to 0 by the driver --
821 * presumably this simplifies the hardware multiplier a little. The hardware
822 * first multiplies linear_id by the multiplier and takes the high 32 bits,
823 * then applies the round-down correction if extra_flags = 1, then finally
824 * shifts right by the shift field.
826 * There are some differences between ridiculousfish's algorithm and the Mali
827 * hardware algorithm, which means that the reference code from ridiculousfish
828 * doesn't always produce the right constants. Mali does not use the pre-shift
829 * optimization, since that would make a hardware implementation slower (it
830 * would have to always do the pre-shift, multiply, and post-shift operations).
831 * It also forces the multplier to be at least 2^31, which means that the
832 * exponent is entirely fixed, so there is no trial-and-error. Altogether,
833 * given the divisor d, the algorithm the driver must follow is:
835 * 1. Set shift = floor(log2(d)).
836 * 2. Compute m = ceil(2^(shift + 32) / d) and e = 2^(shift + 32) % d.
837 * 3. If e <= 2^shift, then we need to use the round-down algorithm. Set
838 * magic_divisor = m - 1 and extra_flags = 1.
839 * 4. Otherwise, set magic_divisor = m and extra_flags = 0.
841 * Unrelated to instancing/actual attributes, images (the OpenCL kind) are
842 * implemented as special attributes, denoted by MALI_ATTR_IMAGE. For images,
843 * let shift=extra_flags=0. Stride is set to the image format's bytes-per-pixel
844 * (*NOT the row stride*). Size is set to the size of the image itself.
846 * Special internal attribtues and varyings (gl_VertexID, gl_FrontFacing, etc)
847 * use particular fixed addresses with modified structures.
850 enum mali_attr_mode
{
851 MALI_ATTR_UNUSED
= 0,
852 MALI_ATTR_LINEAR
= 1,
853 MALI_ATTR_POT_DIVIDE
= 2,
854 MALI_ATTR_MODULO
= 3,
855 MALI_ATTR_NPOT_DIVIDE
= 4,
859 /* Pseudo-address for gl_VertexID, gl_FragCoord, gl_FrontFacing */
861 #define MALI_ATTR_VERTEXID (0x22)
862 #define MALI_ATTR_INSTANCEID (0x24)
863 #define MALI_VARYING_FRAG_COORD (0x25)
864 #define MALI_VARYING_FRONT_FACING (0x26)
866 /* This magic "pseudo-address" is used as `elements` to implement
867 * gl_PointCoord. When read from a fragment shader, it generates a point
868 * coordinate per the OpenGL ES 2.0 specification. Flipped coordinate spaces
869 * require an affine transformation in the shader. */
871 #define MALI_VARYING_POINT_COORD (0x61)
873 /* Used for comparison to check if an address is special. Mostly a guess, but
874 * it doesn't really matter. */
876 #define MALI_RECORD_SPECIAL (0x100)
879 /* This is used for actual attributes. */
881 /* The bottom 3 bits are the mode */
882 mali_ptr elements
: 64 - 8;
888 /* The entry after an NPOT_DIVIDE entry has this format. It stores
889 * extra information that wouldn't fit in a normal entry.
892 u32 unk
; /* = 0x20 */
895 /* This is the original, GL-level divisor. */
898 } __attribute__((packed
));
900 struct mali_attr_meta
{
901 /* Vertex buffer index */
904 unsigned unknown1
: 2;
905 unsigned swizzle
: 12;
906 enum mali_format format
: 8;
908 /* Always observed to be zero at the moment */
909 unsigned unknown3
: 2;
911 /* When packing multiple attributes in a buffer, offset addresses by
912 * this value. Obscurely, this is signed. */
914 } __attribute__((packed
));
916 #define FBD_MASK (~0x3f)
918 /* MFBD, rather than SFBD */
919 #define MALI_MFBD (0x1)
921 /* ORed into an MFBD address to specify the fbx section is included */
922 #define MALI_MFBD_TAG_EXTRA (0x2)
924 /* Uniform buffer objects are 64-bit fields divided as:
927 * mali_ptr ptr : 64 - 10;
929 * The size is actually the size minus 1 (MALI_POSITIVE), in units of 16 bytes.
930 * This gives a maximum of 2^14 bytes, which just so happens to be the GL
931 * minimum-maximum for GL_MAX_UNIFORM_BLOCK_SIZE.
933 * The pointer is missing the bottom 2 bits and top 8 bits. The top 8 bits
934 * should be 0 for userspace pointers, according to
935 * https://lwn.net/Articles/718895/. By reusing these bits, we can make each
936 * entry in the table only 64 bits.
939 #define MALI_MAKE_UBO(elements, ptr) \
940 (MALI_POSITIVE((elements)) | (((ptr) >> 2) << 10))
942 /* On Bifrost, these fields are the same between the vertex and tiler payloads.
943 * They also seem to be the same between Bifrost and Midgard. They're shared in
947 /* Applies to unknown_draw */
949 #define MALI_DRAW_INDEXED_UINT8 (0x10)
950 #define MALI_DRAW_INDEXED_UINT16 (0x20)
951 #define MALI_DRAW_INDEXED_UINT32 (0x30)
952 #define MALI_DRAW_INDEXED_SIZE (0x30)
953 #define MALI_DRAW_INDEXED_SHIFT (4)
955 #define MALI_DRAW_VARYING_SIZE (0x100)
957 /* Set to use first vertex as the provoking vertex for flatshading. Clear to
958 * use the last vertex. This is the default in DX and VK, but not in GL. */
960 #define MALI_DRAW_FLATSHADE_FIRST (0x800)
962 #define MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX (0x10000)
964 struct mali_vertex_tiler_prefix
{
965 /* This is a dynamic bitfield containing the following things in this order:
967 * - gl_WorkGroupSize.x
968 * - gl_WorkGroupSize.y
969 * - gl_WorkGroupSize.z
970 * - gl_NumWorkGroups.x
971 * - gl_NumWorkGroups.y
972 * - gl_NumWorkGroups.z
974 * The number of bits allocated for each number is based on the *_shift
975 * fields below. For example, workgroups_y_shift gives the bit that
976 * gl_NumWorkGroups.y starts at, and workgroups_z_shift gives the bit
977 * that gl_NumWorkGroups.z starts at (and therefore one after the bit
978 * that gl_NumWorkGroups.y ends at). The actual value for each gl_*
979 * value is one more than the stored value, since if any of the values
980 * are zero, then there would be no invocations (and hence no job). If
981 * there were 0 bits allocated to a given field, then it must be zero,
982 * and hence the real value is one.
984 * Vertex jobs reuse the same job dispatch mechanism as compute jobs,
985 * effectively doing glDispatchCompute(1, vertex_count, instance_count)
986 * where vertex count is the number of vertices.
988 u32 invocation_count
;
990 /* Bitfield for shifts:
994 * workgroups_x_shift : 6
995 * workgroups_y_shift : 6
996 * workgroups_z_shift : 6
997 * workgroups_x_shift_2 : 4
999 u32 invocation_shifts
;
1002 u32 unknown_draw
: 22;
1004 /* This is the the same as workgroups_x_shift_2 in compute shaders, but
1005 * always 5 for vertex jobs and 6 for tiler jobs. I suspect this has
1006 * something to do with how many quads get put in the same execution
1007 * engine, which is a balance (you don't want to starve the engine, but
1008 * you also want to distribute work evenly).
1010 u32 workgroups_x_shift_3
: 6;
1013 /* Negative of min_index. This is used to compute
1014 * the unbiased index in tiler/fragment shader runs.
1016 * The hardware adds offset_bias_correction in each run,
1017 * so that absent an index bias, the first vertex processed is
1018 * genuinely the first vertex (0). But with an index bias,
1019 * the first vertex process is numbered the same as the bias.
1021 * To represent this more conviniently:
1022 * unbiased_index = lower_bound_index +
1024 * offset_bias_correction
1026 * This is done since the hardware doesn't accept a index_bias
1027 * and this allows it to recover the unbiased index.
1029 int32_t offset_bias_correction
;
1032 /* Like many other strictly nonzero quantities, index_count is
1033 * subtracted by one. For an indexed cube, this is equal to 35 = 6
1034 * faces * 2 triangles/per face * 3 vertices/per triangle - 1. That is,
1035 * for an indexed draw, index_count is the number of actual vertices
1036 * rendered whereas invocation_count is the number of unique vertices
1037 * rendered (the number of times the vertex shader must be invoked).
1038 * For non-indexed draws, this is just equal to invocation_count. */
1042 /* No hidden structure; literally just a pointer to an array of uint
1043 * indices (width depends on flags). Thanks, guys, for not making my
1044 * life insane for once! NULL for non-indexed draws. */
1047 } __attribute__((packed
));
1049 /* Point size / line width can either be specified as a 32-bit float (for
1050 * constant size) or as a [machine word size]-bit GPU pointer (for varying size). If a pointer
1051 * is selected, by setting the appropriate MALI_DRAW_VARYING_SIZE bit in the tiler
1052 * payload, the contents of varying_pointer will be intepreted as an array of
1053 * fp16 sizes, one for each vertex. gl_PointSize is therefore implemented by
1054 * creating a special MALI_R16F varying writing to varying_pointer. */
1056 union midgard_primitive_size
{
1061 struct bifrost_tiler_heap_meta
{
1064 /* note: these are just guesses! */
1065 mali_ptr tiler_heap_start
;
1066 mali_ptr tiler_heap_free
;
1067 mali_ptr tiler_heap_end
;
1069 /* hierarchy weights? but they're still 0 after the job has run... */
1073 } __attribute__((packed
));
1075 struct bifrost_tiler_meta
{
1076 u32 tiler_heap_next_start
; /* To be written by the GPU */
1077 u32 used_hierarchy_mask
; /* To be written by the GPU */
1078 u16 hierarchy_mask
; /* Five values observed: 0xa, 0x14, 0x28, 0x50, 0xa0 */
1083 mali_ptr tiler_heap_meta
;
1084 /* TODO what is this used for? */
1086 } __attribute__((packed
));
1088 struct bifrost_tiler_only
{
1090 union midgard_primitive_size primitive_size
;
1092 mali_ptr tiler_meta
;
1094 u64 zero1
, zero2
, zero3
, zero4
, zero5
, zero6
;
1095 } __attribute__((packed
));
1097 struct mali_vertex_tiler_postfix
{
1098 u16 gl_enables
; // 0x6 on Midgard, 0x2 on Bifrost
1100 /* Both zero for non-instanced draws. For instanced draws, a
1101 * decomposition of padded_num_vertices. See the comments about the
1102 * corresponding fields in mali_attr for context. */
1104 unsigned instance_shift
: 5;
1105 unsigned instance_odd
: 3;
1109 /* Offset for first vertex in buffer */
1114 /* Zero for vertex jobs. Pointer to the position (gl_Position) varying
1115 * output from the vertex shader for tiler jobs.
1118 u64 position_varying
;
1120 /* An array of mali_uniform_buffer_meta's. The size is given by the
1123 u64 uniform_buffers
;
1125 /* On Bifrost, this is a pointer to an array of bifrost_texture_descriptor.
1126 * On Midgard, this is a pointer to an array of pointers to the texture
1127 * descriptors, number of pointers bounded by number of textures. The
1128 * indirection is needed to accomodate varying numbers and sizes of
1129 * texture descriptors */
1132 /* For OpenGL, from what I've seen, this is intimately connected to
1133 * texture_meta. cwabbott says this is not the case under Vulkan, hence
1134 * why this field is seperate (Midgard is Vulkan capable). Pointer to
1135 * array of sampler descriptors (which are uniform in size) */
1136 u64 sampler_descriptor
;
1140 u64 attributes
; /* struct attribute_buffer[] */
1141 u64 attribute_meta
; /* attribute_meta[] */
1142 u64 varyings
; /* struct attr */
1143 u64 varying_meta
; /* pointer */
1145 u64 occlusion_counter
; /* A single bit as far as I can tell */
1147 /* On Bifrost, this points directly to a mali_shared_memory structure.
1148 * On Midgard, this points to a framebuffer (either SFBD or MFBD as
1149 * tagged), which embeds a mali_shared_memory structure */
1150 mali_ptr shared_memory
;
1151 } __attribute__((packed
));
1153 struct midgard_payload_vertex_tiler
{
1154 struct mali_vertex_tiler_prefix prefix
;
1155 struct mali_vertex_tiler_postfix postfix
;
1157 union midgard_primitive_size primitive_size
;
1158 } __attribute__((packed
));
1160 struct bifrost_payload_vertex
{
1161 struct mali_vertex_tiler_prefix prefix
;
1162 struct mali_vertex_tiler_postfix postfix
;
1163 } __attribute__((packed
));
1165 struct bifrost_payload_tiler
{
1166 struct mali_vertex_tiler_prefix prefix
;
1167 struct bifrost_tiler_only tiler
;
1168 struct mali_vertex_tiler_postfix postfix
;
1169 } __attribute__((packed
));
1171 struct bifrost_payload_fused
{
1172 struct mali_vertex_tiler_prefix prefix
;
1173 struct bifrost_tiler_only tiler
;
1174 struct mali_vertex_tiler_postfix tiler_postfix
;
1175 u64 padding
; /* zero */
1176 struct mali_vertex_tiler_postfix vertex_postfix
;
1177 } __attribute__((packed
));
1179 /* Purposeful off-by-one in width, height fields. For example, a (64, 64)
1180 * texture is stored as (63, 63) in these fields. This adjusts for that.
1181 * There's an identical pattern in the framebuffer descriptor. Even vertex
1182 * count fields work this way, hence the generic name -- integral fields that
1183 * are strictly positive generally need this adjustment. */
1185 #define MALI_POSITIVE(dim) (dim - 1)
1187 /* Used with wrapping. Unclear what top bit conveys */
1189 enum mali_wrap_mode
{
1190 MALI_WRAP_REPEAT
= 0x8 | 0x0,
1191 MALI_WRAP_CLAMP_TO_EDGE
= 0x8 | 0x1,
1192 MALI_WRAP_CLAMP
= 0x8 | 0x2,
1193 MALI_WRAP_CLAMP_TO_BORDER
= 0x8 | 0x3,
1194 MALI_WRAP_MIRRORED_REPEAT
= 0x8 | 0x4 | 0x0,
1195 MALI_WRAP_MIRRORED_CLAMP_TO_EDGE
= 0x8 | 0x4 | 0x1,
1196 MALI_WRAP_MIRRORED_CLAMP
= 0x8 | 0x4 | 0x2,
1197 MALI_WRAP_MIRRORED_CLAMP_TO_BORDER
= 0x8 | 0x4 | 0x3,
1200 /* Shared across both command stream and Midgard, and even with Bifrost */
1202 enum mali_texture_type
{
1203 MALI_TEX_CUBE
= 0x0,
1210 #define MAX_MIP_LEVELS (13)
1212 /* Cubemap bloats everything up */
1213 #define MAX_CUBE_FACES (6)
1215 /* For each pointer, there is an address and optionally also a stride */
1216 #define MAX_ELEMENTS (2)
1218 /* It's not known why there are 4-bits allocated -- this enum is almost
1219 * certainly incomplete */
1221 enum mali_texture_layout
{
1222 /* For a Z/S texture, this is linear */
1223 MALI_TEXTURE_TILED
= 0x1,
1225 /* Z/S textures cannot be tiled */
1226 MALI_TEXTURE_LINEAR
= 0x2,
1229 MALI_TEXTURE_AFBC
= 0xC
1232 /* Corresponds to the type passed to glTexImage2D and so forth */
1234 struct mali_texture_format
{
1235 unsigned swizzle
: 12;
1236 enum mali_format format
: 8;
1239 unsigned unknown1
: 1;
1241 enum mali_texture_type type
: 2;
1242 enum mali_texture_layout layout
: 4;
1245 unsigned unknown2
: 1;
1247 /* Set to allow packing an explicit stride */
1248 unsigned manual_stride
: 1;
1251 } __attribute__((packed
));
1253 struct mali_texture_descriptor
{
1257 uint16_t array_size
;
1259 struct mali_texture_format format
;
1263 /* One for non-mipmapped, zero for mipmapped */
1266 /* Zero for non-mipmapped, (number of levels - 1) for mipmapped */
1269 /* Swizzling is a single 32-bit word, broken up here for convenience.
1270 * Here, swizzling refers to the ES 3.0 texture parameters for channel
1271 * level swizzling, not the internal pixel-level swizzling which is
1272 * below OpenGL's reach */
1274 unsigned swizzle
: 12;
1275 unsigned swizzle_zero
: 20;
1280 } __attribute__((packed
));
1282 /* While Midgard texture descriptors are variable length, Bifrost descriptors
1283 * are fixed like samplers with more pointers to expand if necessary */
1285 struct bifrost_texture_descriptor
{
1286 unsigned format_unk
: 4; /* 2 */
1287 enum mali_texture_type type
: 2;
1289 unsigned format_swizzle
: 12;
1290 enum mali_format format
: 8;
1292 unsigned format_unk3
: 1; /* 0 */
1294 uint16_t width
; /* MALI_POSITIVE */
1295 uint16_t height
; /* MALI_POSITIVE */
1297 /* OpenGL swizzle */
1298 unsigned swizzle
: 12;
1299 enum mali_texture_layout layout
: 4;
1300 uint8_t levels
: 8; /* Number of levels-1 if mipmapped, 0 if not */
1303 unsigned levels_unk
: 24; /* 0 */
1304 unsigned level_2
: 8; /* Number of levels, again? */
1308 uint16_t array_size
;
1313 } __attribute__((packed
));
1317 #define MALI_SAMP_MAG_NEAREST (1 << 0)
1318 #define MALI_SAMP_MIN_NEAREST (1 << 1)
1320 /* TODO: What do these bits mean individually? Only seen set together */
1322 #define MALI_SAMP_MIP_LINEAR_1 (1 << 3)
1323 #define MALI_SAMP_MIP_LINEAR_2 (1 << 4)
1325 /* Flag in filter_mode, corresponding to OpenCL's NORMALIZED_COORDS_TRUE
1326 * sampler_t flag. For typical OpenGL textures, this is always set. */
1328 #define MALI_SAMP_NORM_COORDS (1 << 5)
1330 /* Used for lod encoding. Thanks @urjaman for pointing out these routines can
1331 * be cleaned up a lot. */
1333 #define DECODE_FIXED_16(x) ((float) (x / 256.0))
1335 static inline int16_t
1336 FIXED_16(float x
, bool allow_negative
)
1338 /* Clamp inputs, accounting for float error */
1339 float max_lod
= (32.0 - (1.0 / 512.0));
1340 float min_lod
= allow_negative
? -max_lod
: 0.0;
1342 x
= ((x
> max_lod
) ? max_lod
: ((x
< min_lod
) ? min_lod
: x
));
1344 return (int) (x
* 256.0);
1347 struct mali_sampler_descriptor
{
1348 uint16_t filter_mode
;
1350 /* Fixed point, signed.
1351 * Upper 7 bits before the decimal point, although it caps [0-31].
1352 * Lower 8 bits after the decimal point: int(round(x * 256)) */
1358 /* All one word in reality, but packed a bit. Comparisons are flipped
1361 enum mali_wrap_mode wrap_s
: 4;
1362 enum mali_wrap_mode wrap_t
: 4;
1363 enum mali_wrap_mode wrap_r
: 4;
1364 enum mali_func compare_func
: 3;
1366 /* No effect on 2D textures. For cubemaps, set for ES3 and clear for
1367 * ES2, controlling seamless cubemapping */
1368 unsigned seamless_cube_map
: 1;
1373 float border_color
[4];
1374 } __attribute__((packed
));
1376 /* Bifrost sampler descriptors look pretty similar */
1378 #define BIFROST_SAMP_MIN_NEAREST (1)
1379 #define BIFROST_SAMP_MAG_LINEAR (1)
1381 struct bifrost_sampler_descriptor
{
1384 enum mali_wrap_mode wrap_r
: 4;
1385 enum mali_wrap_mode wrap_t
: 4;
1386 enum mali_wrap_mode wrap_s
: 4;
1390 uint8_t norm_coords
: 1;
1392 uint8_t min_filter
: 1;
1394 uint8_t mag_filter
: 1;
1395 uint8_t mip_filter
: 1;
1403 } __attribute__((packed
));
1405 /* viewport0/viewport1 form the arguments to glViewport. viewport1 is
1406 * modified by MALI_POSITIVE; viewport0 is as-is.
1409 struct mali_viewport
{
1410 /* XY clipping planes */
1416 /* Depth clipping planes */
1422 } __attribute__((packed
));
1424 /* From presentations, 16x16 tiles externally. Use shift for fast computation
1425 * of tile numbers. */
1427 #define MALI_TILE_SHIFT 4
1428 #define MALI_TILE_LENGTH (1 << MALI_TILE_SHIFT)
1430 /* Tile coordinates are stored as a compact u32, as only 12 bits are needed to
1431 * each component. Notice that this provides a theoretical upper bound of (1 <<
1432 * 12) = 4096 tiles in each direction, addressing a maximum framebuffer of size
1433 * 65536x65536. Multiplying that together, times another four given that Mali
1434 * framebuffers are 32-bit ARGB8888, means that this upper bound would take 16
1435 * gigabytes of RAM just to store the uncompressed framebuffer itself, let
1436 * alone rendering in real-time to such a buffer.
1440 /* From mali_kbase_10969_workaround.c */
1441 #define MALI_X_COORD_MASK 0x00000FFF
1442 #define MALI_Y_COORD_MASK 0x0FFF0000
1444 /* Extract parts of a tile coordinate */
1446 #define MALI_TILE_COORD_X(coord) ((coord) & MALI_X_COORD_MASK)
1447 #define MALI_TILE_COORD_Y(coord) (((coord) & MALI_Y_COORD_MASK) >> 16)
1449 /* Helpers to generate tile coordinates based on the boundary coordinates in
1450 * screen space. So, with the bounds (0, 0) to (128, 128) for the screen, these
1451 * functions would convert it to the bounding tiles (0, 0) to (7, 7).
1452 * Intentional "off-by-one"; finding the tile number is a form of fencepost
1455 #define MALI_MAKE_TILE_COORDS(X, Y) ((X) | ((Y) << 16))
1456 #define MALI_BOUND_TO_TILE(B, bias) ((B - bias) >> MALI_TILE_SHIFT)
1457 #define MALI_COORDINATE_TO_TILE(W, H, bias) MALI_MAKE_TILE_COORDS(MALI_BOUND_TO_TILE(W, bias), MALI_BOUND_TO_TILE(H, bias))
1458 #define MALI_COORDINATE_TO_TILE_MIN(W, H) MALI_COORDINATE_TO_TILE(W, H, 0)
1459 #define MALI_COORDINATE_TO_TILE_MAX(W, H) MALI_COORDINATE_TO_TILE(W, H, 1)
1461 struct mali_payload_fragment
{
1464 mali_ptr framebuffer
;
1465 } __attribute__((packed
));
1467 /* Single Framebuffer Descriptor */
1469 /* Flags apply to format. With just MSAA_A and MSAA_B, the framebuffer is
1470 * configured for 4x. With MSAA_8, it is configured for 8x. */
1472 #define MALI_SFBD_FORMAT_MSAA_8 (1 << 3)
1473 #define MALI_SFBD_FORMAT_MSAA_A (1 << 4)
1474 #define MALI_SFBD_FORMAT_MSAA_B (1 << 4)
1475 #define MALI_SFBD_FORMAT_SRGB (1 << 5)
1477 /* Fast/slow based on whether all three buffers are cleared at once */
1479 #define MALI_CLEAR_FAST (1 << 18)
1480 #define MALI_CLEAR_SLOW (1 << 28)
1481 #define MALI_CLEAR_SLOW_STENCIL (1 << 31)
1483 /* Configures hierarchical tiling on Midgard for both SFBD/MFBD (embedded
1484 * within the larget framebuffer descriptor). Analogous to
1485 * bifrost_tiler_heap_meta and bifrost_tiler_meta*/
1487 /* See pan_tiler.c for derivation */
1488 #define MALI_HIERARCHY_MASK ((1 << 9) - 1)
1490 /* Flag disabling the tiler for clear-only jobs, with
1491 hierarchical tiling */
1492 #define MALI_TILER_DISABLED (1 << 12)
1494 /* Flag selecting userspace-generated polygon list, for clear-only jobs without
1495 * hierarhical tiling. */
1496 #define MALI_TILER_USER 0xFFF
1498 /* Absent any geometry, the minimum size of the polygon list header */
1499 #define MALI_TILER_MINIMUM_HEADER_SIZE 0x200
1501 struct midgard_tiler_descriptor
{
1502 /* Size of the entire polygon list; see pan_tiler.c for the
1503 * computation. It's based on hierarchical tiling */
1505 u32 polygon_list_size
;
1507 /* Name known from the replay workaround in the kernel. What exactly is
1508 * flagged here is less known. We do that (tiler_hierarchy_mask & 0x1ff)
1509 * specifies a mask of hierarchy weights, which explains some of the
1510 * performance mysteries around setting it. We also see the bottom bit
1511 * of tiler_flags set in the kernel, but no comment why.
1513 * hierarchy_mask can have the TILER_DISABLED flag */
1518 /* See mali_tiler.c for an explanation */
1519 mali_ptr polygon_list
;
1520 mali_ptr polygon_list_body
;
1522 /* Names based on we see symmetry with replay jobs which name these
1525 mali_ptr heap_start
; /* tiler heap_free_address */
1528 /* Hierarchy weights. We know these are weights based on the kernel,
1529 * but I've never seen them be anything other than zero */
1533 enum mali_block_format
{
1534 MALI_BLOCK_TILED
= 0x0,
1535 MALI_BLOCK_UNKNOWN
= 0x1,
1536 MALI_BLOCK_LINEAR
= 0x2,
1537 MALI_BLOCK_AFBC
= 0x3,
1540 struct mali_sfbd_format
{
1544 /* mali_channel_swizzle */
1545 unsigned swizzle
: 12;
1548 unsigned nr_channels
: 2;
1553 enum mali_block_format block
: 2;
1559 /* Shared structure at the start of framebuffer descriptors, or used bare for
1560 * compute jobs, configuring stack and shared memory */
1562 struct mali_shared_memory
{
1563 u32 stack_shift
: 4;
1566 /* Configuration for shared memory for compute shaders.
1567 * shared_workgroup_count is logarithmic and may be computed for a
1568 * compute shader using shared memory as:
1570 * shared_workgroup_count = MAX2(ceil(log2(count_x)) + ... + ceil(log2(count_z), 10)
1572 * For compute shaders that don't use shared memory, or non-compute
1573 * shaders, this is set to ~0
1576 u32 shared_workgroup_count
: 5;
1577 u32 shared_unk1
: 3;
1578 u32 shared_shift
: 4;
1579 u32 shared_zero
: 20;
1581 mali_ptr scratchpad
;
1583 /* For compute shaders, the RAM backing of workgroup-shared memory. For
1584 * fragment shaders on Bifrost, apparently multisampling locations */
1586 mali_ptr shared_memory
;
1588 } __attribute__((packed
));
1590 /* Configures multisampling on Bifrost fragment jobs */
1592 struct bifrost_multisampling
{
1595 mali_ptr sample_locations
;
1597 } __attribute__((packed
));
1599 struct mali_single_framebuffer
{
1600 struct mali_shared_memory shared_memory
;
1601 struct mali_sfbd_format format
;
1606 /* Purposeful off-by-one in these fields should be accounted for by the
1607 * MALI_DIMENSION macro */
1614 u32 checksum_stride
;
1617 /* By default, the framebuffer is upside down from OpenGL's
1618 * perspective. Set framebuffer to the end and negate the stride to
1619 * flip in the Y direction */
1621 mali_ptr framebuffer
;
1626 /* Depth and stencil buffers are interleaved, it appears, as they are
1627 * set to the same address in captures. Both fields set to zero if the
1628 * buffer is not being cleared. Depending on GL_ENABLE magic, you might
1629 * get a zero enable despite the buffer being present; that still is
1632 mali_ptr depth_buffer
; // not SAME_VA
1633 u32 depth_stride_zero
: 4;
1634 u32 depth_stride
: 28;
1637 mali_ptr stencil_buffer
; // not SAME_VA
1638 u32 stencil_stride_zero
: 4;
1639 u32 stencil_stride
: 28;
1642 u32 clear_color_1
; // RGBA8888 from glClear, actually used by hardware
1643 u32 clear_color_2
; // always equal, but unclear function?
1644 u32 clear_color_3
; // always equal, but unclear function?
1645 u32 clear_color_4
; // always equal, but unclear function?
1647 /* Set to zero if not cleared */
1649 float clear_depth_1
; // float32, ditto
1650 float clear_depth_2
; // float32, ditto
1651 float clear_depth_3
; // float32, ditto
1652 float clear_depth_4
; // float32, ditto
1654 u32 clear_stencil
; // Exactly as it appears in OpenGL
1658 struct midgard_tiler_descriptor tiler
;
1660 /* More below this, maybe */
1661 } __attribute__((packed
));
1663 /* Format bits for the render target flags */
1665 #define MALI_MFBD_FORMAT_MSAA (1 << 1)
1666 #define MALI_MFBD_FORMAT_SRGB (1 << 2)
1668 struct mali_rt_format
{
1672 unsigned nr_channels
: 2; /* MALI_POSITIVE */
1676 enum mali_block_format block
: 2;
1679 unsigned swizzle
: 12;
1683 /* Disables MFBD preload. When this bit is set, the render target will
1684 * be cleared every frame. When this bit is clear, the hardware will
1685 * automatically wallpaper the render target back from main memory.
1686 * Unfortunately, MFBD preload is very broken on Midgard, so in
1687 * practice, this is a chicken bit that should always be set.
1688 * Discovered by accident, as all good chicken bits are. */
1690 unsigned no_preload
: 1;
1691 } __attribute__((packed
));
1693 struct mali_render_target
{
1694 struct mali_rt_format format
;
1699 /* Stuff related to ARM Framebuffer Compression. When AFBC is enabled,
1700 * there is an extra metadata buffer that contains 16 bytes per tile.
1701 * The framebuffer needs to be the same size as before, since we don't
1702 * know ahead of time how much space it will take up. The
1703 * framebuffer_stride is set to 0, since the data isn't stored linearly
1706 * When AFBC is disabled, these fields are zero.
1710 u32 stride
; // stride in units of tiles
1711 u32 unk
; // = 0x20000
1714 mali_ptr framebuffer
;
1717 u32 framebuffer_stride
: 28; // in units of bytes
1720 u32 clear_color_1
; // RGBA8888 from glClear, actually used by hardware
1721 u32 clear_color_2
; // always equal, but unclear function?
1722 u32 clear_color_3
; // always equal, but unclear function?
1723 u32 clear_color_4
; // always equal, but unclear function?
1724 } __attribute__((packed
));
1726 /* An optional part of mali_framebuffer. It comes between the main structure
1727 * and the array of render targets. It must be included if any of these are
1730 * - Transaction Elimination
1732 * - TODO: Anything else?
1736 #define MALI_EXTRA_PRESENT (0x10)
1739 #define MALI_EXTRA_ZS (0x4)
1741 struct mali_framebuffer_extra
{
1743 /* Each tile has an 8 byte checksum, so the stride is "width in tiles * 8" */
1744 u32 checksum_stride
;
1746 unsigned flags_lo
: 4;
1747 enum mali_block_format zs_block
: 2;
1748 unsigned flags_hi
: 26;
1751 /* Note: AFBC is only allowed for 24/8 combined depth/stencil. */
1753 mali_ptr depth_stencil_afbc_metadata
;
1754 u32 depth_stencil_afbc_stride
; // in units of tiles
1757 mali_ptr depth_stencil
;
1763 /* Depth becomes depth/stencil in case of combined D/S */
1765 u32 depth_stride_zero
: 4;
1766 u32 depth_stride
: 28;
1770 u32 stencil_stride_zero
: 4;
1771 u32 stencil_stride
: 28;
1780 } __attribute__((packed
));
1782 /* Flags for mfbd_flags */
1784 /* Enables writing depth results back to main memory (rather than keeping them
1785 * on-chip in the tile buffer and then discarding) */
1787 #define MALI_MFBD_DEPTH_WRITE (1 << 10)
1789 /* The MFBD contains the extra mali_framebuffer_extra section */
1791 #define MALI_MFBD_EXTRA (1 << 13)
1793 struct mali_framebuffer
{
1795 struct mali_shared_memory shared_memory
;
1796 struct bifrost_multisampling msaa
;
1800 u16 width1
, height1
;
1802 u16 width2
, height2
;
1803 u32 unk1
: 19; // = 0x01000
1804 u32 rt_count_1
: 2; // off-by-one (use MALI_POSITIVE)
1805 u32 unk2
: 3; // = 0
1806 u32 rt_count_2
: 3; // no off-by-one
1809 u32 clear_stencil
: 8;
1810 u32 mfbd_flags
: 24; // = 0x100
1814 struct midgard_tiler_descriptor tiler
;
1816 mali_ptr tiler_meta
;
1821 /* optional: struct mali_framebuffer_extra extra */
1822 /* struct mali_render_target rts[] */
1823 } __attribute__((packed
));
1825 #endif /* __PANFROST_JOB_H__ */