freedreno: enable draw/batch reordering by default
[mesa.git] / src / gallium / drivers / freedreno / freedreno_util.h
1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
2
3 /*
4 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 *
25 * Authors:
26 * Rob Clark <robclark@freedesktop.org>
27 */
28
29 #ifndef FREEDRENO_UTIL_H_
30 #define FREEDRENO_UTIL_H_
31
32 #include <freedreno_drmif.h>
33 #include <freedreno_ringbuffer.h>
34
35 #include "pipe/p_format.h"
36 #include "pipe/p_state.h"
37 #include "util/u_debug.h"
38 #include "util/u_math.h"
39 #include "util/u_half.h"
40 #include "util/u_dynarray.h"
41 #include "util/u_pack_color.h"
42
43 #include "disasm.h"
44 #include "adreno_common.xml.h"
45 #include "adreno_pm4.xml.h"
46
47 enum adreno_rb_depth_format fd_pipe2depth(enum pipe_format format);
48 enum pc_di_index_size fd_pipe2index(enum pipe_format format);
49 enum pipe_format fd_gmem_restore_format(enum pipe_format format);
50 enum adreno_rb_blend_factor fd_blend_factor(unsigned factor);
51 enum adreno_pa_su_sc_draw fd_polygon_mode(unsigned mode);
52 enum adreno_stencil_op fd_stencil_op(unsigned op);
53
54 #define A3XX_MAX_MIP_LEVELS 14
55 /* TBD if it is same on a2xx, but for now: */
56 #define MAX_MIP_LEVELS A3XX_MAX_MIP_LEVELS
57
58 #define A2XX_MAX_RENDER_TARGETS 1
59 #define A3XX_MAX_RENDER_TARGETS 4
60 #define A4XX_MAX_RENDER_TARGETS 8
61 #define A5XX_MAX_RENDER_TARGETS 8
62
63 #define MAX_RENDER_TARGETS A5XX_MAX_RENDER_TARGETS
64
65 #define FD_DBG_MSGS 0x0001
66 #define FD_DBG_DISASM 0x0002
67 #define FD_DBG_DCLEAR 0x0004
68 #define FD_DBG_DDRAW 0x0008
69 #define FD_DBG_NOSCIS 0x0010
70 #define FD_DBG_DIRECT 0x0020
71 #define FD_DBG_NOBYPASS 0x0040
72 #define FD_DBG_FRAGHALF 0x0080
73 #define FD_DBG_NOBIN 0x0100
74 #define FD_DBG_OPTMSGS 0x0200
75 #define FD_DBG_GLSL120 0x0400
76 #define FD_DBG_SHADERDB 0x0800
77 #define FD_DBG_FLUSH 0x1000
78 #define FD_DBG_DEQP 0x2000
79 #define FD_DBG_NIR 0x4000
80 #define FD_DBG_INORDER 0x8000
81 #define FD_DBG_BSTAT 0x10000
82 #define FD_DBG_NOGROW 0x20000
83
84 extern int fd_mesa_debug;
85 extern bool fd_binning_enabled;
86
87 #define DBG(fmt, ...) \
88 do { if (fd_mesa_debug & FD_DBG_MSGS) \
89 debug_printf("%s:%d: "fmt "\n", \
90 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
91
92 /* for conditionally setting boolean flag(s): */
93 #define COND(bool, val) ((bool) ? (val) : 0)
94
95 #define CP_REG(reg) ((0x4 << 16) | ((unsigned int)((reg) - (0x2000))))
96
97 static inline uint32_t DRAW(enum pc_di_primtype prim_type,
98 enum pc_di_src_sel source_select, enum pc_di_index_size index_size,
99 enum pc_di_vis_cull_mode vis_cull_mode,
100 uint8_t instances)
101 {
102 return (prim_type << 0) |
103 (source_select << 6) |
104 ((index_size & 1) << 11) |
105 ((index_size >> 1) << 13) |
106 (vis_cull_mode << 9) |
107 (1 << 14) |
108 (instances << 24);
109 }
110
111 /* for tracking cmdstream positions that need to be patched: */
112 struct fd_cs_patch {
113 uint32_t *cs;
114 uint32_t val;
115 };
116 #define fd_patch_num_elements(buf) ((buf)->size / sizeof(struct fd_cs_patch))
117 #define fd_patch_element(buf, i) util_dynarray_element(buf, struct fd_cs_patch, i)
118
119 static inline enum pipe_format
120 pipe_surface_format(struct pipe_surface *psurf)
121 {
122 if (!psurf)
123 return PIPE_FORMAT_NONE;
124 return psurf->format;
125 }
126
127 static inline bool
128 fd_surface_half_precision(const struct pipe_surface *psurf)
129 {
130 enum pipe_format format;
131
132 if (!psurf)
133 return true;
134
135 format = psurf->format;
136
137 /* colors are provided in consts, which go through cov.f32f16, which will
138 * break these values
139 */
140 if (util_format_is_pure_integer(format))
141 return false;
142
143 /* avoid losing precision on 32-bit float formats */
144 if (util_format_is_float(format) &&
145 util_format_get_component_bits(format, UTIL_FORMAT_COLORSPACE_RGB, 0) == 32)
146 return false;
147
148 return true;
149 }
150
151 static inline unsigned
152 fd_sampler_first_level(const struct pipe_sampler_view *view)
153 {
154 if (view->target == PIPE_BUFFER)
155 return 0;
156 return view->u.tex.first_level;
157 }
158
159 static inline unsigned
160 fd_sampler_last_level(const struct pipe_sampler_view *view)
161 {
162 if (view->target == PIPE_BUFFER)
163 return 0;
164 return view->u.tex.last_level;
165 }
166
167 static inline bool
168 fd_half_precision(struct pipe_framebuffer_state *pfb)
169 {
170 unsigned i;
171
172 for (i = 0; i < pfb->nr_cbufs; i++)
173 if (!fd_surface_half_precision(pfb->cbufs[i]))
174 return false;
175
176 return true;
177 }
178
179 #define LOG_DWORDS 0
180
181 static inline void emit_marker(struct fd_ringbuffer *ring, int scratch_idx);
182 static inline void emit_marker5(struct fd_ringbuffer *ring, int scratch_idx);
183
184 static inline void
185 OUT_RING(struct fd_ringbuffer *ring, uint32_t data)
186 {
187 if (LOG_DWORDS) {
188 DBG("ring[%p]: OUT_RING %04x: %08x", ring,
189 (uint32_t)(ring->cur - ring->last_start), data);
190 }
191 fd_ringbuffer_emit(ring, data);
192 }
193
194 /* like OUT_RING() but appends a cmdstream patch point to 'buf' */
195 static inline void
196 OUT_RINGP(struct fd_ringbuffer *ring, uint32_t data,
197 struct util_dynarray *buf)
198 {
199 if (LOG_DWORDS) {
200 DBG("ring[%p]: OUT_RINGP %04x: %08x", ring,
201 (uint32_t)(ring->cur - ring->last_start), data);
202 }
203 util_dynarray_append(buf, struct fd_cs_patch, ((struct fd_cs_patch){
204 .cs = ring->cur++,
205 .val = data,
206 }));
207 }
208
209 /*
210 * NOTE: OUT_RELOC*() is 2 dwords (64b) on a5xx+
211 */
212
213 static inline void
214 OUT_RELOC(struct fd_ringbuffer *ring, struct fd_bo *bo,
215 uint32_t offset, uint64_t or, int32_t shift)
216 {
217 if (LOG_DWORDS) {
218 DBG("ring[%p]: OUT_RELOC %04x: %p+%u << %d", ring,
219 (uint32_t)(ring->cur - ring->last_start), bo, offset, shift);
220 }
221 debug_assert(offset < fd_bo_size(bo));
222 fd_ringbuffer_reloc2(ring, &(struct fd_reloc){
223 .bo = bo,
224 .flags = FD_RELOC_READ,
225 .offset = offset,
226 .or = or,
227 .shift = shift,
228 .orhi = or >> 32,
229 });
230 }
231
232 static inline void
233 OUT_RELOCW(struct fd_ringbuffer *ring, struct fd_bo *bo,
234 uint32_t offset, uint64_t or, int32_t shift)
235 {
236 if (LOG_DWORDS) {
237 DBG("ring[%p]: OUT_RELOCW %04x: %p+%u << %d", ring,
238 (uint32_t)(ring->cur - ring->last_start), bo, offset, shift);
239 }
240 debug_assert(offset < fd_bo_size(bo));
241 fd_ringbuffer_reloc2(ring, &(struct fd_reloc){
242 .bo = bo,
243 .flags = FD_RELOC_READ | FD_RELOC_WRITE,
244 .offset = offset,
245 .or = or,
246 .shift = shift,
247 .orhi = or >> 32,
248 });
249 }
250
251 static inline void BEGIN_RING(struct fd_ringbuffer *ring, uint32_t ndwords)
252 {
253 if (ring->cur + ndwords >= ring->end)
254 fd_ringbuffer_grow(ring, ndwords);
255 }
256
257 static inline uint32_t
258 __gpu_id(struct fd_ringbuffer *ring)
259 {
260 uint64_t val;
261 fd_pipe_get_param(ring->pipe, FD_GPU_ID, &val);
262 return val;
263 }
264
265 static inline void
266 OUT_PKT0(struct fd_ringbuffer *ring, uint16_t regindx, uint16_t cnt)
267 {
268 debug_assert(__gpu_id(ring) < 500);
269 BEGIN_RING(ring, cnt+1);
270 OUT_RING(ring, CP_TYPE0_PKT | ((cnt-1) << 16) | (regindx & 0x7FFF));
271 }
272
273 static inline void
274 OUT_PKT2(struct fd_ringbuffer *ring)
275 {
276 debug_assert(__gpu_id(ring) < 500);
277 BEGIN_RING(ring, 1);
278 OUT_RING(ring, CP_TYPE2_PKT);
279 }
280
281 static inline void
282 OUT_PKT3(struct fd_ringbuffer *ring, uint8_t opcode, uint16_t cnt)
283 {
284 debug_assert(__gpu_id(ring) < 500);
285 BEGIN_RING(ring, cnt+1);
286 OUT_RING(ring, CP_TYPE3_PKT | ((cnt-1) << 16) | ((opcode & 0xFF) << 8));
287 }
288
289 /*
290 * Starting with a5xx, pkt4/pkt7 are used instead of pkt0/pkt3
291 */
292
293 static inline unsigned
294 _odd_parity_bit(unsigned val)
295 {
296 /* See: http://graphics.stanford.edu/~seander/bithacks.html#ParityParallel
297 * note that we want odd parity so 0x6996 is inverted.
298 */
299 val ^= val >> 16;
300 val ^= val >> 8;
301 val ^= val >> 4;
302 val &= 0xf;
303 return (~0x6996 >> val) & 1;
304 }
305
306 static inline void
307 OUT_PKT4(struct fd_ringbuffer *ring, uint16_t regindx, uint16_t cnt)
308 {
309 BEGIN_RING(ring, cnt+1);
310 OUT_RING(ring, CP_TYPE4_PKT | cnt |
311 (_odd_parity_bit(cnt) << 7) |
312 ((regindx & 0x3ffff) << 8) |
313 ((_odd_parity_bit(regindx) << 27)));
314 }
315
316 static inline void
317 OUT_PKT7(struct fd_ringbuffer *ring, uint8_t opcode, uint16_t cnt)
318 {
319 BEGIN_RING(ring, cnt+1);
320 OUT_RING(ring, CP_TYPE7_PKT | cnt |
321 (_odd_parity_bit(cnt) << 15) |
322 ((opcode & 0x7f) << 16) |
323 ((_odd_parity_bit(opcode) << 23)));
324 }
325
326 static inline void
327 OUT_WFI(struct fd_ringbuffer *ring)
328 {
329 OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
330 OUT_RING(ring, 0x00000000);
331 }
332
333 static inline void
334 OUT_WFI5(struct fd_ringbuffer *ring)
335 {
336 OUT_PKT7(ring, CP_WAIT_FOR_IDLE, 0);
337 }
338
339 static inline void
340 __OUT_IB(struct fd_ringbuffer *ring, bool prefetch, struct fd_ringbuffer *target)
341 {
342 unsigned count = fd_ringbuffer_cmd_count(target);
343
344 debug_assert(__gpu_id(ring) < 500);
345
346 /* for debug after a lock up, write a unique counter value
347 * to scratch6 for each IB, to make it easier to match up
348 * register dumps to cmdstream. The combination of IB and
349 * DRAW (scratch7) is enough to "triangulate" the particular
350 * draw that caused lockup.
351 */
352 emit_marker(ring, 6);
353
354 for (unsigned i = 0; i < count; i++) {
355 uint32_t dwords;
356 OUT_PKT3(ring, prefetch ? CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
357 dwords = fd_ringbuffer_emit_reloc_ring_full(ring, target, i) / 4;
358 assert(dwords > 0);
359 OUT_RING(ring, dwords);
360 OUT_PKT2(ring);
361 }
362
363 emit_marker(ring, 6);
364 }
365
366 static inline void
367 __OUT_IB5(struct fd_ringbuffer *ring, struct fd_ringbuffer *target)
368 {
369 unsigned count = fd_ringbuffer_cmd_count(target);
370
371 /* for debug after a lock up, write a unique counter value
372 * to scratch6 for each IB, to make it easier to match up
373 * register dumps to cmdstream. The combination of IB and
374 * DRAW (scratch7) is enough to "triangulate" the particular
375 * draw that caused lockup.
376 */
377 emit_marker5(ring, 6);
378
379 for (unsigned i = 0; i < count; i++) {
380 uint32_t dwords;
381 OUT_PKT7(ring, CP_INDIRECT_BUFFER, 3);
382 dwords = fd_ringbuffer_emit_reloc_ring_full(ring, target, i) / 4;
383 assert(dwords > 0);
384 OUT_RING(ring, dwords);
385 }
386
387 emit_marker5(ring, 6);
388 }
389
390 /* CP_SCRATCH_REG4 is used to hold base address for query results: */
391 // XXX annoyingly scratch regs move on a5xx.. and additionally different
392 // packet types.. so freedreno_query_hw is going to need a bit of
393 // rework..
394 #define HW_QUERY_BASE_REG REG_AXXX_CP_SCRATCH_REG4
395
396 static inline void
397 emit_marker(struct fd_ringbuffer *ring, int scratch_idx)
398 {
399 extern unsigned marker_cnt;
400 unsigned reg = REG_AXXX_CP_SCRATCH_REG0 + scratch_idx;
401 assert(reg != HW_QUERY_BASE_REG);
402 if (reg == HW_QUERY_BASE_REG)
403 return;
404 OUT_PKT0(ring, reg, 1);
405 OUT_RING(ring, ++marker_cnt);
406 }
407
408 static inline void
409 emit_marker5(struct fd_ringbuffer *ring, int scratch_idx)
410 {
411 extern unsigned marker_cnt;
412 //XXX unsigned reg = REG_A5XX_CP_SCRATCH_REG(scratch_idx);
413 unsigned reg = 0x00000b78 + scratch_idx;
414 assert(reg != HW_QUERY_BASE_REG);
415 if (reg == HW_QUERY_BASE_REG)
416 return;
417 OUT_WFI5(ring);
418 OUT_PKT4(ring, reg, 1);
419 OUT_RING(ring, ++marker_cnt);
420 OUT_WFI5(ring);
421 }
422
423 /* helper to get numeric value from environment variable.. mostly
424 * just leaving this here because it is helpful to brute-force figure
425 * out unknown formats, etc, which blob driver does not support:
426 */
427 static inline uint32_t env2u(const char *envvar)
428 {
429 char *str = getenv(envvar);
430 if (str)
431 return strtoul(str, NULL, 0);
432 return 0;
433 }
434
435 static inline uint32_t
436 pack_rgba(enum pipe_format format, const float *rgba)
437 {
438 union util_color uc;
439 util_pack_color(rgba, format, &uc);
440 return uc.ui[0];
441 }
442
443 /*
444 * swap - swap value of @a and @b
445 */
446 #define swap(a, b) \
447 do { __typeof(a) __tmp = (a); (a) = (b); (b) = __tmp; } while (0)
448
449 #define foreach_bit(b, mask) \
450 for (uint32_t _m = (mask); _m && ({(b) = u_bit_scan(&_m); 1;});)
451
452 #endif /* FREEDRENO_UTIL_H_ */