xref: /aosp_15_r20/external/mesa3d/src/amd/vulkan/radv_cp_dma.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2016 Red Hat.
3  * Copyright © 2016 Bas Nieuwenhuizen
4  *
5  * based on si_state.c
6  * Copyright © 2015 Advanced Micro Devices, Inc.
7  *
8  * SPDX-License-Identifier: MIT
9  */
10 
11 #include "radv_cp_dma.h"
12 #include "radv_buffer.h"
13 #include "radv_cs.h"
14 #include "radv_debug.h"
15 #include "radv_shader.h"
16 #include "radv_sqtt.h"
17 #include "sid.h"
18 
19 /* Set this if you want the 3D engine to wait until CP DMA is done.
20  * It should be set on the last CP DMA packet. */
21 #define CP_DMA_SYNC (1 << 0)
22 
23 /* Set this if the source data was used as a destination in a previous CP DMA
24  * packet. It's for preventing a read-after-write (RAW) hazard between two
25  * CP DMA packets. */
26 #define CP_DMA_RAW_WAIT (1 << 1)
27 #define CP_DMA_USE_L2   (1 << 2)
28 #define CP_DMA_CLEAR    (1 << 3)
29 
30 /* Alignment for optimal performance. */
31 #define SI_CPDMA_ALIGNMENT 32
32 
33 /* The max number of bytes that can be copied per packet. */
34 static inline unsigned
cp_dma_max_byte_count(enum amd_gfx_level gfx_level)35 cp_dma_max_byte_count(enum amd_gfx_level gfx_level)
36 {
37    unsigned max = gfx_level >= GFX11  ? 32767
38                   : gfx_level >= GFX9 ? S_415_BYTE_COUNT_GFX9(~0u)
39                                       : S_415_BYTE_COUNT_GFX6(~0u);
40 
41    /* make it aligned for optimal performance */
42    return max & ~(SI_CPDMA_ALIGNMENT - 1);
43 }
44 
45 /* Emit a CP DMA packet to do a copy from one buffer to another, or to clear
46  * a buffer. The size must fit in bits [20:0]. If CP_DMA_CLEAR is set, src_va is a 32-bit
47  * clear value.
48  */
49 static void
radv_cs_emit_cp_dma(struct radv_device * device,struct radeon_cmdbuf * cs,bool predicating,uint64_t dst_va,uint64_t src_va,unsigned size,unsigned flags)50 radv_cs_emit_cp_dma(struct radv_device *device, struct radeon_cmdbuf *cs, bool predicating, uint64_t dst_va,
51                     uint64_t src_va, unsigned size, unsigned flags)
52 {
53    const struct radv_physical_device *pdev = radv_device_physical(device);
54    uint32_t header = 0, command = 0;
55 
56    assert(size <= cp_dma_max_byte_count(pdev->info.gfx_level));
57 
58    radeon_check_space(device->ws, cs, 9);
59    if (pdev->info.gfx_level >= GFX9)
60       command |= S_415_BYTE_COUNT_GFX9(size);
61    else
62       command |= S_415_BYTE_COUNT_GFX6(size);
63 
64    /* Sync flags. */
65    if (flags & CP_DMA_SYNC)
66       header |= S_411_CP_SYNC(1);
67 
68    if (flags & CP_DMA_RAW_WAIT)
69       command |= S_415_RAW_WAIT(1);
70 
71    /* Src and dst flags. */
72    if (pdev->info.gfx_level >= GFX9 && !(flags & CP_DMA_CLEAR) && src_va == dst_va)
73       header |= S_411_DST_SEL(V_411_NOWHERE); /* prefetch only */
74    else if (flags & CP_DMA_USE_L2)
75       header |= S_411_DST_SEL(V_411_DST_ADDR_TC_L2);
76 
77    if (flags & CP_DMA_CLEAR)
78       header |= S_411_SRC_SEL(V_411_DATA);
79    else if (flags & CP_DMA_USE_L2)
80       header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
81 
82    if (pdev->info.gfx_level >= GFX7) {
83       radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, predicating));
84       radeon_emit(cs, header);
85       radeon_emit(cs, src_va);       /* SRC_ADDR_LO [31:0] */
86       radeon_emit(cs, src_va >> 32); /* SRC_ADDR_HI [31:0] */
87       radeon_emit(cs, dst_va);       /* DST_ADDR_LO [31:0] */
88       radeon_emit(cs, dst_va >> 32); /* DST_ADDR_HI [31:0] */
89       radeon_emit(cs, command);
90    } else {
91       assert(!(flags & CP_DMA_USE_L2));
92       header |= S_411_SRC_ADDR_HI(src_va >> 32);
93       radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, predicating));
94       radeon_emit(cs, src_va);                  /* SRC_ADDR_LO [31:0] */
95       radeon_emit(cs, header);                  /* SRC_ADDR_HI [15:0] + flags. */
96       radeon_emit(cs, dst_va);                  /* DST_ADDR_LO [31:0] */
97       radeon_emit(cs, (dst_va >> 32) & 0xffff); /* DST_ADDR_HI [15:0] */
98       radeon_emit(cs, command);
99    }
100 }
101 
102 static void
radv_emit_cp_dma(struct radv_cmd_buffer * cmd_buffer,uint64_t dst_va,uint64_t src_va,unsigned size,unsigned flags)103 radv_emit_cp_dma(struct radv_cmd_buffer *cmd_buffer, uint64_t dst_va, uint64_t src_va, unsigned size, unsigned flags)
104 {
105    struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
106    struct radeon_cmdbuf *cs = cmd_buffer->cs;
107    bool predicating = cmd_buffer->state.predicating;
108 
109    radv_cs_emit_cp_dma(device, cs, predicating, dst_va, src_va, size, flags);
110 
111    /* CP DMA is executed in ME, but index buffers are read by PFP.
112     * This ensures that ME (CP DMA) is idle before PFP starts fetching
113     * indices. If we wanted to execute CP DMA in PFP, this packet
114     * should precede it.
115     */
116    if (flags & CP_DMA_SYNC) {
117       if (cmd_buffer->qf == RADV_QUEUE_GENERAL) {
118          radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, cmd_buffer->state.predicating));
119          radeon_emit(cs, 0);
120       }
121 
122       /* CP will see the sync flag and wait for all DMAs to complete. */
123       cmd_buffer->state.dma_is_busy = false;
124    }
125 
126    if (radv_device_fault_detection_enabled(device))
127       radv_cmd_buffer_trace_emit(cmd_buffer);
128 }
129 
130 void
radv_cs_cp_dma_prefetch(const struct radv_device * device,struct radeon_cmdbuf * cs,uint64_t va,unsigned size,bool predicating)131 radv_cs_cp_dma_prefetch(const struct radv_device *device, struct radeon_cmdbuf *cs, uint64_t va, unsigned size,
132                         bool predicating)
133 {
134    const struct radv_physical_device *pdev = radv_device_physical(device);
135    struct radeon_winsys *ws = device->ws;
136    enum amd_gfx_level gfx_level = pdev->info.gfx_level;
137    uint32_t header = 0, command = 0;
138 
139    if (gfx_level >= GFX11)
140       size = MIN2(size, 32768 - SI_CPDMA_ALIGNMENT);
141 
142    assert(size <= cp_dma_max_byte_count(gfx_level));
143 
144    radeon_check_space(ws, cs, 9);
145 
146    uint64_t aligned_va = va & ~(SI_CPDMA_ALIGNMENT - 1);
147    uint64_t aligned_size = ((va + size + SI_CPDMA_ALIGNMENT - 1) & ~(SI_CPDMA_ALIGNMENT - 1)) - aligned_va;
148 
149    if (gfx_level >= GFX9) {
150       command |= S_415_BYTE_COUNT_GFX9(aligned_size) | S_415_DISABLE_WR_CONFIRM_GFX9(1);
151       header |= S_411_DST_SEL(V_411_NOWHERE);
152    } else {
153       command |= S_415_BYTE_COUNT_GFX6(aligned_size) | S_415_DISABLE_WR_CONFIRM_GFX6(1);
154       header |= S_411_DST_SEL(V_411_DST_ADDR_TC_L2);
155    }
156 
157    header |= S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2);
158 
159    radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, predicating));
160    radeon_emit(cs, header);
161    radeon_emit(cs, aligned_va);       /* SRC_ADDR_LO [31:0] */
162    radeon_emit(cs, aligned_va >> 32); /* SRC_ADDR_HI [31:0] */
163    radeon_emit(cs, aligned_va);       /* DST_ADDR_LO [31:0] */
164    radeon_emit(cs, aligned_va >> 32); /* DST_ADDR_HI [31:0] */
165    radeon_emit(cs, command);
166 }
167 
168 void
radv_cp_dma_prefetch(struct radv_cmd_buffer * cmd_buffer,uint64_t va,unsigned size)169 radv_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va, unsigned size)
170 {
171    struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
172 
173    radv_cs_cp_dma_prefetch(device, cmd_buffer->cs, va, size, cmd_buffer->state.predicating);
174 
175    if (radv_device_fault_detection_enabled(device))
176       radv_cmd_buffer_trace_emit(cmd_buffer);
177 }
178 
179 static void
radv_cp_dma_prepare(struct radv_cmd_buffer * cmd_buffer,uint64_t byte_count,uint64_t remaining_size,unsigned * flags)180 radv_cp_dma_prepare(struct radv_cmd_buffer *cmd_buffer, uint64_t byte_count, uint64_t remaining_size, unsigned *flags)
181 {
182 
183    /* Flush the caches for the first copy only.
184     * Also wait for the previous CP DMA operations.
185     */
186    if (cmd_buffer->state.flush_bits) {
187       radv_emit_cache_flush(cmd_buffer);
188       *flags |= CP_DMA_RAW_WAIT;
189    }
190 
191    /* Do the synchronization after the last dma, so that all data
192     * is written to memory.
193     */
194    if (byte_count == remaining_size)
195       *flags |= CP_DMA_SYNC;
196 }
197 
198 static void
radv_cp_dma_realign_engine(struct radv_cmd_buffer * cmd_buffer,unsigned size)199 radv_cp_dma_realign_engine(struct radv_cmd_buffer *cmd_buffer, unsigned size)
200 {
201    uint64_t va;
202    uint32_t offset;
203    unsigned dma_flags = 0;
204    unsigned buf_size = SI_CPDMA_ALIGNMENT * 2;
205    void *ptr;
206 
207    assert(size < SI_CPDMA_ALIGNMENT);
208 
209    radv_cmd_buffer_upload_alloc(cmd_buffer, buf_size, &offset, &ptr);
210 
211    va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
212    va += offset;
213 
214    radv_cp_dma_prepare(cmd_buffer, size, size, &dma_flags);
215 
216    radv_emit_cp_dma(cmd_buffer, va, va + SI_CPDMA_ALIGNMENT, size, dma_flags);
217 }
218 
219 void
radv_cp_dma_buffer_copy(struct radv_cmd_buffer * cmd_buffer,uint64_t src_va,uint64_t dest_va,uint64_t size)220 radv_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer, uint64_t src_va, uint64_t dest_va, uint64_t size)
221 {
222    struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
223    const struct radv_physical_device *pdev = radv_device_physical(device);
224    enum amd_gfx_level gfx_level = pdev->info.gfx_level;
225    uint64_t main_src_va, main_dest_va;
226    uint64_t skipped_size = 0, realign_size = 0;
227 
228    /* Assume that we are not going to sync after the last DMA operation. */
229    cmd_buffer->state.dma_is_busy = true;
230 
231    if (pdev->info.family <= CHIP_CARRIZO || pdev->info.family == CHIP_STONEY) {
232       /* If the size is not aligned, we must add a dummy copy at the end
233        * just to align the internal counter. Otherwise, the DMA engine
234        * would slow down by an order of magnitude for following copies.
235        */
236       if (size % SI_CPDMA_ALIGNMENT)
237          realign_size = SI_CPDMA_ALIGNMENT - (size % SI_CPDMA_ALIGNMENT);
238 
239       /* If the copy begins unaligned, we must start copying from the next
240        * aligned block and the skipped part should be copied after everything
241        * else has been copied. Only the src alignment matters, not dst.
242        */
243       if (src_va % SI_CPDMA_ALIGNMENT) {
244          skipped_size = SI_CPDMA_ALIGNMENT - (src_va % SI_CPDMA_ALIGNMENT);
245          /* The main part will be skipped if the size is too small. */
246          skipped_size = MIN2(skipped_size, size);
247          size -= skipped_size;
248       }
249    }
250    main_src_va = src_va + skipped_size;
251    main_dest_va = dest_va + skipped_size;
252 
253    while (size) {
254       unsigned dma_flags = 0;
255       unsigned byte_count = MIN2(size, cp_dma_max_byte_count(gfx_level));
256 
257       if (pdev->info.gfx_level >= GFX9) {
258          /* DMA operations via L2 are coherent and faster.
259           * TODO: GFX7-GFX8 should also support this but it
260           * requires tests/benchmarks.
261           *
262           * Also enable on GFX9 so we can use L2 at rest on GFX9+. On Raven
263           * this didn't seem to be worse.
264           *
265           * Note that we only use CP DMA for sizes < RADV_BUFFER_OPS_CS_THRESHOLD,
266           * which is 4k at the moment, so this is really unlikely to cause
267           * significant thrashing.
268           */
269          dma_flags |= CP_DMA_USE_L2;
270       }
271 
272       radv_cp_dma_prepare(cmd_buffer, byte_count, size + skipped_size + realign_size, &dma_flags);
273 
274       dma_flags &= ~CP_DMA_SYNC;
275 
276       radv_emit_cp_dma(cmd_buffer, main_dest_va, main_src_va, byte_count, dma_flags);
277 
278       size -= byte_count;
279       main_src_va += byte_count;
280       main_dest_va += byte_count;
281    }
282 
283    if (skipped_size) {
284       unsigned dma_flags = 0;
285 
286       radv_cp_dma_prepare(cmd_buffer, skipped_size, size + skipped_size + realign_size, &dma_flags);
287 
288       radv_emit_cp_dma(cmd_buffer, dest_va, src_va, skipped_size, dma_flags);
289    }
290    if (realign_size)
291       radv_cp_dma_realign_engine(cmd_buffer, realign_size);
292 }
293 
294 void
radv_cp_dma_clear_buffer(struct radv_cmd_buffer * cmd_buffer,uint64_t va,uint64_t size,unsigned value)295 radv_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va, uint64_t size, unsigned value)
296 {
297    struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
298    const struct radv_physical_device *pdev = radv_device_physical(device);
299 
300    if (!size)
301       return;
302 
303    assert(va % 4 == 0 && size % 4 == 0);
304 
305    enum amd_gfx_level gfx_level = pdev->info.gfx_level;
306 
307    /* Assume that we are not going to sync after the last DMA operation. */
308    cmd_buffer->state.dma_is_busy = true;
309 
310    while (size) {
311       unsigned byte_count = MIN2(size, cp_dma_max_byte_count(gfx_level));
312       unsigned dma_flags = CP_DMA_CLEAR;
313 
314       if (pdev->info.gfx_level >= GFX9) {
315          /* DMA operations via L2 are coherent and faster.
316           * TODO: GFX7-GFX8 should also support this but it
317           * requires tests/benchmarks.
318           *
319           * Also enable on GFX9 so we can use L2 at rest on GFX9+.
320           */
321          dma_flags |= CP_DMA_USE_L2;
322       }
323 
324       radv_cp_dma_prepare(cmd_buffer, byte_count, size, &dma_flags);
325 
326       /* Emit the clear packet. */
327       radv_emit_cp_dma(cmd_buffer, va, value, byte_count, dma_flags);
328 
329       size -= byte_count;
330       va += byte_count;
331    }
332 }
333 
334 void
radv_cp_dma_wait_for_idle(struct radv_cmd_buffer * cmd_buffer)335 radv_cp_dma_wait_for_idle(struct radv_cmd_buffer *cmd_buffer)
336 {
337    struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
338    const struct radv_physical_device *pdev = radv_device_physical(device);
339 
340    if (pdev->info.gfx_level < GFX7)
341       return;
342 
343    if (!cmd_buffer->state.dma_is_busy)
344       return;
345 
346    /* Issue a dummy DMA that copies zero bytes.
347     *
348     * The DMA engine will see that there's no work to do and skip this
349     * DMA request, however, the CP will see the sync flag and still wait
350     * for all DMAs to complete.
351     */
352    radv_emit_cp_dma(cmd_buffer, 0, 0, 0, CP_DMA_SYNC);
353 
354    cmd_buffer->state.dma_is_busy = false;
355 }
356