turnip: make cond_exec helper easier to use
[mesa.git] / src / freedreno / vulkan / tu_cs.h
1 /*
2 * Copyright © 2019 Google LLC
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
22 */
23 #ifndef TU_CS_H
24 #define TU_CS_H
25
26 #include "tu_private.h"
27
28 #include "registers/adreno_pm4.xml.h"
29
30 void
31 tu_cs_init(struct tu_cs *cs,
32 struct tu_device *device,
33 enum tu_cs_mode mode,
34 uint32_t initial_size);
35
36 void
37 tu_cs_init_external(struct tu_cs *cs, uint32_t *start, uint32_t *end);
38
39 void
40 tu_cs_finish(struct tu_cs *cs);
41
42 void
43 tu_cs_begin(struct tu_cs *cs);
44
45 void
46 tu_cs_end(struct tu_cs *cs);
47
48 VkResult
49 tu_cs_begin_sub_stream(struct tu_cs *cs, uint32_t size, struct tu_cs *sub_cs);
50
51 VkResult
52 tu_cs_alloc(struct tu_cs *cs,
53 uint32_t count,
54 uint32_t size,
55 struct ts_cs_memory *memory);
56
57 struct tu_cs_entry
58 tu_cs_end_sub_stream(struct tu_cs *cs, struct tu_cs *sub_cs);
59
60 VkResult
61 tu_cs_reserve_space(struct tu_cs *cs, uint32_t reserved_size);
62
63 void
64 tu_cs_reset(struct tu_cs *cs);
65
66 VkResult
67 tu_cs_add_entries(struct tu_cs *cs, struct tu_cs *target);
68
69 /**
70 * Discard all entries. This allows \a cs to be reused while keeping the
71 * existing BOs and command packets intact.
72 */
73 static inline void
74 tu_cs_discard_entries(struct tu_cs *cs)
75 {
76 assert(cs->mode == TU_CS_MODE_GROW);
77 cs->entry_count = 0;
78 }
79
80 /**
81 * Get the size needed for tu_cs_emit_call.
82 */
83 static inline uint32_t
84 tu_cs_get_call_size(const struct tu_cs *cs)
85 {
86 assert(cs->mode == TU_CS_MODE_GROW);
87 /* each CP_INDIRECT_BUFFER needs 4 dwords */
88 return cs->entry_count * 4;
89 }
90
91 /**
92 * Assert that we did not exceed the reserved space.
93 */
94 static inline void
95 tu_cs_sanity_check(const struct tu_cs *cs)
96 {
97 assert(cs->start <= cs->cur);
98 assert(cs->cur <= cs->reserved_end);
99 assert(cs->reserved_end <= cs->end);
100 }
101
102 /**
103 * Emit a uint32_t value into a command stream, without boundary checking.
104 */
105 static inline void
106 tu_cs_emit(struct tu_cs *cs, uint32_t value)
107 {
108 assert(cs->cur < cs->reserved_end);
109 *cs->cur = value;
110 ++cs->cur;
111 }
112
113 /**
114 * Emit an array of uint32_t into a command stream, without boundary checking.
115 */
116 static inline void
117 tu_cs_emit_array(struct tu_cs *cs, const uint32_t *values, uint32_t length)
118 {
119 assert(cs->cur + length <= cs->reserved_end);
120 memcpy(cs->cur, values, sizeof(uint32_t) * length);
121 cs->cur += length;
122 }
123
124 static inline unsigned
125 tu_odd_parity_bit(unsigned val)
126 {
127 /* See: http://graphics.stanford.edu/~seander/bithacks.html#ParityParallel
128 * note that we want odd parity so 0x6996 is inverted.
129 */
130 val ^= val >> 16;
131 val ^= val >> 8;
132 val ^= val >> 4;
133 val &= 0xf;
134 return (~0x6996 >> val) & 1;
135 }
136
137 /**
138 * Get the size of the remaining space in the current BO.
139 */
140 static inline uint32_t
141 tu_cs_get_space(const struct tu_cs *cs)
142 {
143 return cs->end - cs->cur;
144 }
145
146 static inline void
147 tu_cs_reserve(struct tu_cs *cs, uint32_t reserved_size)
148 {
149 if (cs->mode != TU_CS_MODE_GROW) {
150 assert(tu_cs_get_space(cs) >= reserved_size);
151 assert(cs->reserved_end == cs->end);
152 return;
153 }
154
155 if (tu_cs_get_space(cs) >= reserved_size &&
156 cs->entry_count < cs->entry_capacity) {
157 cs->reserved_end = cs->cur + reserved_size;
158 return;
159 }
160
161 VkResult result = tu_cs_reserve_space(cs, reserved_size);
162 /* TODO: set this error in tu_cs and use it */
163 assert(result == VK_SUCCESS);
164 }
165
166 /**
167 * Emit a type-4 command packet header into a command stream.
168 */
169 static inline void
170 tu_cs_emit_pkt4(struct tu_cs *cs, uint16_t regindx, uint16_t cnt)
171 {
172 tu_cs_reserve(cs, cnt + 1);
173 tu_cs_emit(cs, CP_TYPE4_PKT | cnt | (tu_odd_parity_bit(cnt) << 7) |
174 ((regindx & 0x3ffff) << 8) |
175 ((tu_odd_parity_bit(regindx) << 27)));
176 }
177
178 /**
179 * Emit a type-7 command packet header into a command stream.
180 */
181 static inline void
182 tu_cs_emit_pkt7(struct tu_cs *cs, uint8_t opcode, uint16_t cnt)
183 {
184 tu_cs_reserve(cs, cnt + 1);
185 tu_cs_emit(cs, CP_TYPE7_PKT | cnt | (tu_odd_parity_bit(cnt) << 15) |
186 ((opcode & 0x7f) << 16) |
187 ((tu_odd_parity_bit(opcode) << 23)));
188 }
189
190 static inline void
191 tu_cs_emit_wfi(struct tu_cs *cs)
192 {
193 tu_cs_emit_pkt7(cs, CP_WAIT_FOR_IDLE, 0);
194 }
195
196 static inline void
197 tu_cs_emit_qw(struct tu_cs *cs, uint64_t value)
198 {
199 tu_cs_emit(cs, (uint32_t) value);
200 tu_cs_emit(cs, (uint32_t) (value >> 32));
201 }
202
203 static inline void
204 tu_cs_emit_write_reg(struct tu_cs *cs, uint16_t reg, uint32_t value)
205 {
206 tu_cs_emit_pkt4(cs, reg, 1);
207 tu_cs_emit(cs, value);
208 }
209
210 /**
211 * Emit a CP_INDIRECT_BUFFER command packet.
212 */
213 static inline void
214 tu_cs_emit_ib(struct tu_cs *cs, const struct tu_cs_entry *entry)
215 {
216 assert(entry->bo);
217 assert(entry->size && entry->offset + entry->size <= entry->bo->size);
218 assert(entry->size % sizeof(uint32_t) == 0);
219 assert(entry->offset % sizeof(uint32_t) == 0);
220
221 tu_cs_emit_pkt7(cs, CP_INDIRECT_BUFFER, 3);
222 tu_cs_emit_qw(cs, entry->bo->iova + entry->offset);
223 tu_cs_emit(cs, entry->size / sizeof(uint32_t));
224 }
225
226 /**
227 * Emit a CP_INDIRECT_BUFFER command packet for each entry in the target
228 * command stream.
229 */
230 static inline void
231 tu_cs_emit_call(struct tu_cs *cs, const struct tu_cs *target)
232 {
233 assert(target->mode == TU_CS_MODE_GROW);
234 for (uint32_t i = 0; i < target->entry_count; i++)
235 tu_cs_emit_ib(cs, target->entries + i);
236 }
237
238 /* Helpers for bracketing a large sequence of commands of unknown size inside
239 * a CP_COND_REG_EXEC packet.
240 */
241 static inline void
242 tu_cond_exec_start(struct tu_cs *cs, uint32_t cond_flags)
243 {
244 assert(cs->mode == TU_CS_MODE_GROW);
245 assert(!cs->cond_flags && cond_flags);
246
247 tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
248 tu_cs_emit(cs, cond_flags);
249
250 cs->cond_flags = cond_flags;
251 cs->cond_dwords = cs->cur;
252
253 /* Emit dummy DWORD field here */
254 tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(0));
255 }
256 #define CP_COND_EXEC_0_RENDER_MODE_GMEM \
257 (CP_COND_REG_EXEC_0_MODE(RENDER_MODE) | CP_COND_REG_EXEC_0_GMEM)
258 #define CP_COND_EXEC_0_RENDER_MODE_SYSMEM \
259 (CP_COND_REG_EXEC_0_MODE(RENDER_MODE) | CP_COND_REG_EXEC_0_SYSMEM)
260
261 static inline void
262 tu_cond_exec_end(struct tu_cs *cs)
263 {
264 assert(cs->cond_flags);
265
266 cs->cond_flags = 0;
267 /* Subtract one here to account for the DWORD field itself. */
268 *cs->cond_dwords = cs->cur - cs->cond_dwords - 1;
269 }
270
271 #define fd_reg_pair tu_reg_value
272 #define __bo_type struct tu_bo *
273
274 #include "a6xx.xml.h"
275 #include "a6xx-pack.xml.h"
276
277 #define __assert_eq(a, b) \
278 do { \
279 if ((a) != (b)) { \
280 fprintf(stderr, "assert failed: " #a " (0x%x) != " #b " (0x%x)\n", a, b); \
281 assert((a) == (b)); \
282 } \
283 } while (0)
284
285 #define __ONE_REG(i, regs) \
286 do { \
287 if (i < ARRAY_SIZE(regs) && regs[i].reg > 0) { \
288 __assert_eq(regs[0].reg + i, regs[i].reg); \
289 if (regs[i].bo) { \
290 uint64_t v = regs[i].bo->iova + regs[i].bo_offset; \
291 v >>= regs[i].bo_shift; \
292 v |= regs[i].value; \
293 \
294 *p++ = v; \
295 *p++ = v >> 32; \
296 } else { \
297 *p++ = regs[i].value; \
298 if (regs[i].is_address) \
299 *p++ = regs[i].value >> 32; \
300 } \
301 } \
302 } while (0)
303
304 /* Emits a sequence of register writes in order using a pkt4. This will check
305 * (at runtime on a !NDEBUG build) that the registers were actually set up in
306 * order in the code.
307 *
308 * Note that references to buffers aren't automatically added to the CS,
309 * unlike in freedreno. We are clever in various places to avoid duplicating
310 * the reference add work.
311 *
312 * Also, 64-bit address registers don't have a way (currently) to set a 64-bit
313 * address without having a reference to a BO, since the .dword field in the
314 * register's struct is only 32-bit wide. We should fix this in the pack
315 * codegen later.
316 */
317 #define tu_cs_emit_regs(cs, ...) do { \
318 const struct fd_reg_pair regs[] = { __VA_ARGS__ }; \
319 unsigned count = ARRAY_SIZE(regs); \
320 \
321 STATIC_ASSERT(count > 0); \
322 STATIC_ASSERT(count <= 16); \
323 \
324 tu_cs_emit_pkt4(cs, regs[0].reg, count); \
325 uint32_t *p = cs->cur; \
326 __ONE_REG( 0, regs); \
327 __ONE_REG( 1, regs); \
328 __ONE_REG( 2, regs); \
329 __ONE_REG( 3, regs); \
330 __ONE_REG( 4, regs); \
331 __ONE_REG( 5, regs); \
332 __ONE_REG( 6, regs); \
333 __ONE_REG( 7, regs); \
334 __ONE_REG( 8, regs); \
335 __ONE_REG( 9, regs); \
336 __ONE_REG(10, regs); \
337 __ONE_REG(11, regs); \
338 __ONE_REG(12, regs); \
339 __ONE_REG(13, regs); \
340 __ONE_REG(14, regs); \
341 __ONE_REG(15, regs); \
342 cs->cur = p; \
343 } while (0)
344
345 #endif /* TU_CS_H */