• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2016 Red Hat.
3  * Copyright © 2016 Bas Nieuwenhuizen
4  *
5  * based in part on anv driver which is:
6  * Copyright © 2015 Intel Corporation
7  *
8  * Permission is hereby granted, free of charge, to any person obtaining a
9  * copy of this software and associated documentation files (the "Software"),
10  * to deal in the Software without restriction, including without limitation
11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12  * and/or sell copies of the Software, and to permit persons to whom the
13  * Software is furnished to do so, subject to the following conditions:
14  *
15  * The above copyright notice and this permission notice (including the next
16  * paragraph) shall be included in all copies or substantial portions of the
17  * Software.
18  *
19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
22  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25  * IN THE SOFTWARE.
26  */
27 
28 #include "radv_cs.h"
29 #include "radv_debug.h"
30 #include "radv_private.h"
31 #include "vk_semaphore.h"
32 #include "vk_sync.h"
33 
34 #include "ac_debug.h"
35 
36 enum radeon_ctx_priority
radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoKHR * pObj)37 radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoKHR *pObj)
38 {
39    /* Default to MEDIUM when a specific global priority isn't requested */
40    if (!pObj)
41       return RADEON_CTX_PRIORITY_MEDIUM;
42 
43    switch (pObj->globalPriority) {
44    case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_KHR:
45       return RADEON_CTX_PRIORITY_REALTIME;
46    case VK_QUEUE_GLOBAL_PRIORITY_HIGH_KHR:
47       return RADEON_CTX_PRIORITY_HIGH;
48    case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR:
49       return RADEON_CTX_PRIORITY_MEDIUM;
50    case VK_QUEUE_GLOBAL_PRIORITY_LOW_KHR:
51       return RADEON_CTX_PRIORITY_LOW;
52    default:
53       unreachable("Illegal global priority value");
54       return RADEON_CTX_PRIORITY_INVALID;
55    }
56 }
57 
58 static VkResult
radv_sparse_buffer_bind_memory(struct radv_device * device,const VkSparseBufferMemoryBindInfo * bind)59 radv_sparse_buffer_bind_memory(struct radv_device *device, const VkSparseBufferMemoryBindInfo *bind)
60 {
61    RADV_FROM_HANDLE(radv_buffer, buffer, bind->buffer);
62    VkResult result = VK_SUCCESS;
63 
64    struct radv_device_memory *mem = NULL;
65    VkDeviceSize resourceOffset = 0;
66    VkDeviceSize size = 0;
67    VkDeviceSize memoryOffset = 0;
68    for (uint32_t i = 0; i < bind->bindCount; ++i) {
69       struct radv_device_memory *cur_mem = NULL;
70 
71       if (bind->pBinds[i].memory != VK_NULL_HANDLE)
72          cur_mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
73       if (i && mem == cur_mem) {
74          if (mem) {
75             if (bind->pBinds[i].resourceOffset == resourceOffset + size &&
76                 bind->pBinds[i].memoryOffset == memoryOffset + size) {
77                size += bind->pBinds[i].size;
78                continue;
79             }
80          } else {
81             if (bind->pBinds[i].resourceOffset == resourceOffset + size) {
82                size += bind->pBinds[i].size;
83                continue;
84             }
85          }
86       }
87       if (size) {
88          result = device->ws->buffer_virtual_bind(device->ws, buffer->bo, resourceOffset, size, mem ? mem->bo : NULL,
89                                                   memoryOffset);
90          if (result != VK_SUCCESS)
91             return result;
92 
93          if (bind->pBinds[i].memory)
94             radv_rmv_log_sparse_add_residency(device, buffer->bo, memoryOffset);
95          else
96             radv_rmv_log_sparse_remove_residency(device, buffer->bo, memoryOffset);
97       }
98       mem = cur_mem;
99       resourceOffset = bind->pBinds[i].resourceOffset;
100       size = bind->pBinds[i].size;
101       memoryOffset = bind->pBinds[i].memoryOffset;
102    }
103    if (size) {
104       result = device->ws->buffer_virtual_bind(device->ws, buffer->bo, resourceOffset, size, mem ? mem->bo : NULL,
105                                                memoryOffset);
106 
107       if (mem)
108          radv_rmv_log_sparse_add_residency(device, buffer->bo, memoryOffset);
109       else
110          radv_rmv_log_sparse_remove_residency(device, buffer->bo, memoryOffset);
111    }
112 
113    return result;
114 }
115 
116 static VkResult
radv_sparse_image_opaque_bind_memory(struct radv_device * device,const VkSparseImageOpaqueMemoryBindInfo * bind)117 radv_sparse_image_opaque_bind_memory(struct radv_device *device, const VkSparseImageOpaqueMemoryBindInfo *bind)
118 {
119    RADV_FROM_HANDLE(radv_image, image, bind->image);
120    VkResult result;
121 
122    for (uint32_t i = 0; i < bind->bindCount; ++i) {
123       struct radv_device_memory *mem = NULL;
124 
125       if (bind->pBinds[i].memory != VK_NULL_HANDLE)
126          mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
127 
128       result =
129          device->ws->buffer_virtual_bind(device->ws, image->bindings[0].bo, bind->pBinds[i].resourceOffset,
130                                          bind->pBinds[i].size, mem ? mem->bo : NULL, bind->pBinds[i].memoryOffset);
131       if (result != VK_SUCCESS)
132          return result;
133 
134       if (bind->pBinds[i].memory)
135          radv_rmv_log_sparse_add_residency(device, image->bindings[0].bo, bind->pBinds[i].resourceOffset);
136       else
137          radv_rmv_log_sparse_remove_residency(device, image->bindings[0].bo, bind->pBinds[i].resourceOffset);
138    }
139 
140    return VK_SUCCESS;
141 }
142 
143 static VkResult
radv_sparse_image_bind_memory(struct radv_device * device,const VkSparseImageMemoryBindInfo * bind)144 radv_sparse_image_bind_memory(struct radv_device *device, const VkSparseImageMemoryBindInfo *bind)
145 {
146    RADV_FROM_HANDLE(radv_image, image, bind->image);
147    struct radeon_surf *surface = &image->planes[0].surface;
148    uint32_t bs = vk_format_get_blocksize(image->vk.format);
149    VkResult result;
150 
151    for (uint32_t i = 0; i < bind->bindCount; ++i) {
152       struct radv_device_memory *mem = NULL;
153       uint64_t offset, depth_pitch;
154       uint32_t pitch;
155       uint64_t mem_offset = bind->pBinds[i].memoryOffset;
156       const uint32_t layer = bind->pBinds[i].subresource.arrayLayer;
157       const uint32_t level = bind->pBinds[i].subresource.mipLevel;
158 
159       VkExtent3D bind_extent = bind->pBinds[i].extent;
160       bind_extent.width = DIV_ROUND_UP(bind_extent.width, vk_format_get_blockwidth(image->vk.format));
161       bind_extent.height = DIV_ROUND_UP(bind_extent.height, vk_format_get_blockheight(image->vk.format));
162 
163       VkOffset3D bind_offset = bind->pBinds[i].offset;
164       bind_offset.x /= vk_format_get_blockwidth(image->vk.format);
165       bind_offset.y /= vk_format_get_blockheight(image->vk.format);
166 
167       if (bind->pBinds[i].memory != VK_NULL_HANDLE)
168          mem = radv_device_memory_from_handle(bind->pBinds[i].memory);
169 
170       if (device->physical_device->rad_info.gfx_level >= GFX9) {
171          offset = surface->u.gfx9.surf_slice_size * layer + surface->u.gfx9.prt_level_offset[level];
172          pitch = surface->u.gfx9.prt_level_pitch[level];
173          depth_pitch = surface->u.gfx9.surf_slice_size;
174       } else {
175          depth_pitch = surface->u.legacy.level[level].slice_size_dw * 4;
176          offset = (uint64_t)surface->u.legacy.level[level].offset_256B * 256 + depth_pitch * layer;
177          pitch = surface->u.legacy.level[level].nblk_x;
178       }
179 
180       offset +=
181          bind_offset.z * depth_pitch + ((uint64_t)bind_offset.y * pitch * surface->prt_tile_depth +
182                                         (uint64_t)bind_offset.x * surface->prt_tile_height * surface->prt_tile_depth) *
183                                           bs;
184 
185       uint32_t aligned_extent_width = ALIGN(bind_extent.width, surface->prt_tile_width);
186       uint32_t aligned_extent_height = ALIGN(bind_extent.height, surface->prt_tile_height);
187       uint32_t aligned_extent_depth = ALIGN(bind_extent.depth, surface->prt_tile_depth);
188 
189       bool whole_subres = (bind_extent.height <= surface->prt_tile_height || aligned_extent_width == pitch) &&
190                           (bind_extent.depth <= surface->prt_tile_depth ||
191                            (uint64_t)aligned_extent_width * aligned_extent_height * bs == depth_pitch);
192 
193       if (whole_subres) {
194          uint64_t size = (uint64_t)aligned_extent_width * aligned_extent_height * aligned_extent_depth * bs;
195          result = device->ws->buffer_virtual_bind(device->ws, image->bindings[0].bo, offset, size, mem ? mem->bo : NULL,
196                                                   mem_offset);
197          if (result != VK_SUCCESS)
198             return result;
199 
200          if (bind->pBinds[i].memory)
201             radv_rmv_log_sparse_add_residency(device, image->bindings[0].bo, offset);
202          else
203             radv_rmv_log_sparse_remove_residency(device, image->bindings[0].bo, offset);
204 
205       } else {
206          uint32_t img_y_increment = pitch * bs * surface->prt_tile_depth;
207          uint32_t mem_y_increment = aligned_extent_width * bs * surface->prt_tile_depth;
208          uint64_t mem_z_increment = (uint64_t)aligned_extent_width * aligned_extent_height * bs;
209          uint64_t size = mem_y_increment * surface->prt_tile_height;
210          for (unsigned z = 0; z < bind_extent.depth;
211               z += surface->prt_tile_depth, offset += depth_pitch * surface->prt_tile_depth) {
212             for (unsigned y = 0; y < bind_extent.height; y += surface->prt_tile_height) {
213                result = device->ws->buffer_virtual_bind(
214                   device->ws, image->bindings[0].bo, offset + (uint64_t)img_y_increment * y, size, mem ? mem->bo : NULL,
215                   mem_offset + (uint64_t)mem_y_increment * y + mem_z_increment * z);
216                if (result != VK_SUCCESS)
217                   return result;
218 
219                if (bind->pBinds[i].memory)
220                   radv_rmv_log_sparse_add_residency(device, image->bindings[0].bo, offset);
221                else
222                   radv_rmv_log_sparse_remove_residency(device, image->bindings[0].bo, offset);
223             }
224          }
225       }
226    }
227 
228    return VK_SUCCESS;
229 }
230 
231 static VkResult
radv_queue_submit_bind_sparse_memory(struct radv_device * device,struct vk_queue_submit * submission)232 radv_queue_submit_bind_sparse_memory(struct radv_device *device, struct vk_queue_submit *submission)
233 {
234    for (uint32_t i = 0; i < submission->buffer_bind_count; ++i) {
235       VkResult result = radv_sparse_buffer_bind_memory(device, submission->buffer_binds + i);
236       if (result != VK_SUCCESS)
237          return result;
238    }
239 
240    for (uint32_t i = 0; i < submission->image_opaque_bind_count; ++i) {
241       VkResult result = radv_sparse_image_opaque_bind_memory(device, submission->image_opaque_binds + i);
242       if (result != VK_SUCCESS)
243          return result;
244    }
245 
246    for (uint32_t i = 0; i < submission->image_bind_count; ++i) {
247       VkResult result = radv_sparse_image_bind_memory(device, submission->image_binds + i);
248       if (result != VK_SUCCESS)
249          return result;
250    }
251 
252    return VK_SUCCESS;
253 }
254 
255 static VkResult
radv_queue_submit_empty(struct radv_queue * queue,struct vk_queue_submit * submission)256 radv_queue_submit_empty(struct radv_queue *queue, struct vk_queue_submit *submission)
257 {
258    struct radeon_winsys_ctx *ctx = queue->hw_ctx;
259    struct radv_winsys_submit_info submit = {
260       .ip_type = radv_queue_ring(queue),
261       .queue_index = queue->vk.index_in_family,
262    };
263 
264    return queue->device->ws->cs_submit(ctx, &submit, submission->wait_count, submission->waits,
265                                        submission->signal_count, submission->signals);
266 }
267 
268 static void
radv_fill_shader_rings(struct radv_device * device,uint32_t * desc,struct radeon_winsys_bo * scratch_bo,uint32_t esgs_ring_size,struct radeon_winsys_bo * esgs_ring_bo,uint32_t gsvs_ring_size,struct radeon_winsys_bo * gsvs_ring_bo,struct radeon_winsys_bo * tess_rings_bo,struct radeon_winsys_bo * task_rings_bo,struct radeon_winsys_bo * mesh_scratch_ring_bo,uint32_t attr_ring_size,struct radeon_winsys_bo * attr_ring_bo)269 radv_fill_shader_rings(struct radv_device *device, uint32_t *desc, struct radeon_winsys_bo *scratch_bo,
270                        uint32_t esgs_ring_size, struct radeon_winsys_bo *esgs_ring_bo, uint32_t gsvs_ring_size,
271                        struct radeon_winsys_bo *gsvs_ring_bo, struct radeon_winsys_bo *tess_rings_bo,
272                        struct radeon_winsys_bo *task_rings_bo, struct radeon_winsys_bo *mesh_scratch_ring_bo,
273                        uint32_t attr_ring_size, struct radeon_winsys_bo *attr_ring_bo)
274 {
275    if (scratch_bo) {
276       uint64_t scratch_va = radv_buffer_get_va(scratch_bo);
277       uint32_t rsrc1 = S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);
278 
279       if (device->physical_device->rad_info.gfx_level >= GFX11)
280          rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
281       else
282          rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
283 
284       desc[0] = scratch_va;
285       desc[1] = rsrc1;
286    }
287 
288    desc += 4;
289 
290    if (esgs_ring_bo) {
291       uint64_t esgs_va = radv_buffer_get_va(esgs_ring_bo);
292 
293       /* stride 0, num records - size, add tid, swizzle, elsize4,
294          index stride 64 */
295       desc[0] = esgs_va;
296       desc[1] = S_008F04_BASE_ADDRESS_HI(esgs_va >> 32);
297       desc[2] = esgs_ring_size;
298       desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
299                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
300                 S_008F0C_INDEX_STRIDE(3) | S_008F0C_ADD_TID_ENABLE(1);
301 
302       if (device->physical_device->rad_info.gfx_level >= GFX11)
303          desc[1] |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
304       else
305          desc[1] |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
306 
307       if (device->physical_device->rad_info.gfx_level >= GFX11) {
308          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
309       } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
310          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
311                     S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
312       } else if (device->physical_device->rad_info.gfx_level >= GFX8) {
313          /* DATA_FORMAT is STRIDE[14:17] for MUBUF with ADD_TID_ENABLE=1 */
314          desc[3] |=
315             S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | S_008F0C_DATA_FORMAT(0) | S_008F0C_ELEMENT_SIZE(1);
316       } else {
317          desc[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
318                     S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) | S_008F0C_ELEMENT_SIZE(1);
319       }
320 
321       /* GS entry for ES->GS ring */
322       /* stride 0, num records - size, elsize0,
323          index stride 0 */
324       desc[4] = esgs_va;
325       desc[5] = S_008F04_BASE_ADDRESS_HI(esgs_va >> 32);
326       desc[6] = esgs_ring_size;
327       desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
328                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
329 
330       if (device->physical_device->rad_info.gfx_level >= GFX11) {
331          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
332       } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
333          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
334                     S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
335       } else {
336          desc[7] |=
337             S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
338       }
339    }
340 
341    desc += 8;
342 
343    if (gsvs_ring_bo) {
344       uint64_t gsvs_va = radv_buffer_get_va(gsvs_ring_bo);
345 
346       /* VS entry for GS->VS ring */
347       /* stride 0, num records - size, elsize0,
348          index stride 0 */
349       desc[0] = gsvs_va;
350       desc[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va >> 32);
351       desc[2] = gsvs_ring_size;
352       desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
353                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
354 
355       if (device->physical_device->rad_info.gfx_level >= GFX11) {
356          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
357       } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
358          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
359                     S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
360       } else {
361          desc[3] |=
362             S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
363       }
364 
365       /* stride gsvs_itemsize, num records 64
366          elsize 4, index stride 16 */
367       /* shader will patch stride and desc[2] */
368       desc[4] = gsvs_va;
369       desc[5] = S_008F04_BASE_ADDRESS_HI(gsvs_va >> 32);
370       desc[6] = 0;
371       desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
372                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
373                 S_008F0C_INDEX_STRIDE(1) | S_008F0C_ADD_TID_ENABLE(true);
374 
375       if (device->physical_device->rad_info.gfx_level >= GFX11)
376          desc[5] |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
377       else
378          desc[5] |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
379 
380       if (device->physical_device->rad_info.gfx_level >= GFX11) {
381          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
382       } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
383          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) |
384                     S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) | S_008F0C_RESOURCE_LEVEL(1);
385       } else if (device->physical_device->rad_info.gfx_level >= GFX8) {
386          /* DATA_FORMAT is STRIDE[14:17] for MUBUF with ADD_TID_ENABLE=1 */
387          desc[7] |=
388             S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | S_008F0C_DATA_FORMAT(0) | S_008F0C_ELEMENT_SIZE(1);
389       } else {
390          desc[7] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
391                     S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) | S_008F0C_ELEMENT_SIZE(1);
392       }
393    }
394 
395    desc += 8;
396 
397    if (tess_rings_bo) {
398       uint64_t tess_va = radv_buffer_get_va(tess_rings_bo);
399       uint64_t tess_offchip_va = tess_va + device->physical_device->hs.tess_offchip_ring_offset;
400 
401       desc[0] = tess_va;
402       desc[1] = S_008F04_BASE_ADDRESS_HI(tess_va >> 32);
403       desc[2] = device->physical_device->hs.tess_factor_ring_size;
404       desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
405                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
406 
407       if (device->physical_device->rad_info.gfx_level >= GFX11) {
408          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW);
409       } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
410          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
411                     S_008F0C_RESOURCE_LEVEL(1);
412       } else {
413          desc[3] |=
414             S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
415       }
416 
417       desc[4] = tess_offchip_va;
418       desc[5] = S_008F04_BASE_ADDRESS_HI(tess_offchip_va >> 32);
419       desc[6] = device->physical_device->hs.tess_offchip_ring_size;
420       desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
421                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
422 
423       if (device->physical_device->rad_info.gfx_level >= GFX11) {
424          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW);
425       } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
426          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_FLOAT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
427                     S_008F0C_RESOURCE_LEVEL(1);
428       } else {
429          desc[7] |=
430             S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) | S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
431       }
432    }
433 
434    desc += 8;
435 
436    if (task_rings_bo) {
437       uint64_t task_va = radv_buffer_get_va(task_rings_bo);
438       uint64_t task_draw_ring_va = task_va + device->physical_device->task_info.draw_ring_offset;
439       uint64_t task_payload_ring_va = task_va + device->physical_device->task_info.payload_ring_offset;
440 
441       desc[0] = task_draw_ring_va;
442       desc[1] = S_008F04_BASE_ADDRESS_HI(task_draw_ring_va >> 32);
443       desc[2] = device->physical_device->task_info.num_entries * AC_TASK_DRAW_ENTRY_BYTES;
444       desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
445                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
446 
447       if (device->physical_device->rad_info.gfx_level >= GFX11) {
448          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_UINT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
449       } else {
450          assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
451          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
452                     S_008F0C_RESOURCE_LEVEL(1);
453       }
454 
455       desc[4] = task_payload_ring_va;
456       desc[5] = S_008F04_BASE_ADDRESS_HI(task_payload_ring_va >> 32);
457       desc[6] = device->physical_device->task_info.num_entries * AC_TASK_PAYLOAD_ENTRY_BYTES;
458       desc[7] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
459                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
460 
461       if (device->physical_device->rad_info.gfx_level >= GFX11) {
462          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_UINT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
463       } else {
464          assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
465          desc[7] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
466                     S_008F0C_RESOURCE_LEVEL(1);
467       }
468    }
469 
470    desc += 8;
471 
472    if (mesh_scratch_ring_bo) {
473       uint64_t va = radv_buffer_get_va(mesh_scratch_ring_bo);
474 
475       desc[0] = va;
476       desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
477       desc[2] = RADV_MESH_SCRATCH_NUM_ENTRIES * RADV_MESH_SCRATCH_ENTRY_BYTES;
478       desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
479                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
480 
481       if (device->physical_device->rad_info.gfx_level >= GFX11) {
482          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_UINT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED);
483       } else {
484          assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
485          desc[3] |= S_008F0C_FORMAT(V_008F0C_GFX10_FORMAT_32_UINT) | S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_DISABLED) |
486                     S_008F0C_RESOURCE_LEVEL(1);
487       }
488    }
489 
490    desc += 4;
491 
492    if (attr_ring_bo) {
493       assert(device->physical_device->rad_info.gfx_level >= GFX11);
494 
495       uint64_t va = radv_buffer_get_va(attr_ring_bo);
496 
497       desc[0] = va;
498       desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) | S_008F04_SWIZZLE_ENABLE_GFX11(3) /* 16B */;
499       desc[2] = attr_ring_size;
500       desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) | S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
501                 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) | S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
502                 S_008F0C_FORMAT(V_008F0C_GFX11_FORMAT_32_32_32_32_FLOAT) | S_008F0C_INDEX_STRIDE(2) /* 32 elements */;
503    }
504 
505    desc += 4;
506 
507    /* add sample positions after all rings */
508    memcpy(desc, device->sample_locations_1x, 8);
509    desc += 2;
510    memcpy(desc, device->sample_locations_2x, 16);
511    desc += 4;
512    memcpy(desc, device->sample_locations_4x, 32);
513    desc += 8;
514    memcpy(desc, device->sample_locations_8x, 64);
515 }
516 
517 static void
radv_emit_gs_ring_sizes(struct radv_device * device,struct radeon_cmdbuf * cs,struct radeon_winsys_bo * esgs_ring_bo,uint32_t esgs_ring_size,struct radeon_winsys_bo * gsvs_ring_bo,uint32_t gsvs_ring_size)518 radv_emit_gs_ring_sizes(struct radv_device *device, struct radeon_cmdbuf *cs, struct radeon_winsys_bo *esgs_ring_bo,
519                         uint32_t esgs_ring_size, struct radeon_winsys_bo *gsvs_ring_bo, uint32_t gsvs_ring_size)
520 {
521    if (!esgs_ring_bo && !gsvs_ring_bo)
522       return;
523 
524    if (esgs_ring_bo)
525       radv_cs_add_buffer(device->ws, cs, esgs_ring_bo);
526 
527    if (gsvs_ring_bo)
528       radv_cs_add_buffer(device->ws, cs, gsvs_ring_bo);
529 
530    if (device->physical_device->rad_info.gfx_level >= GFX7) {
531       radeon_set_uconfig_reg_seq(cs, R_030900_VGT_ESGS_RING_SIZE, 2);
532       radeon_emit(cs, esgs_ring_size >> 8);
533       radeon_emit(cs, gsvs_ring_size >> 8);
534    } else {
535       radeon_set_config_reg_seq(cs, R_0088C8_VGT_ESGS_RING_SIZE, 2);
536       radeon_emit(cs, esgs_ring_size >> 8);
537       radeon_emit(cs, gsvs_ring_size >> 8);
538    }
539 }
540 
541 static void
radv_emit_tess_factor_ring(struct radv_device * device,struct radeon_cmdbuf * cs,struct radeon_winsys_bo * tess_rings_bo)542 radv_emit_tess_factor_ring(struct radv_device *device, struct radeon_cmdbuf *cs, struct radeon_winsys_bo *tess_rings_bo)
543 {
544    uint64_t tf_va;
545    uint32_t tf_ring_size;
546    if (!tess_rings_bo)
547       return;
548 
549    tf_ring_size = device->physical_device->hs.tess_factor_ring_size / 4;
550    tf_va = radv_buffer_get_va(tess_rings_bo);
551 
552    radv_cs_add_buffer(device->ws, cs, tess_rings_bo);
553 
554    if (device->physical_device->rad_info.gfx_level >= GFX7) {
555       if (device->physical_device->rad_info.gfx_level >= GFX11) {
556          /* TF_RING_SIZE is per SE on GFX11. */
557          tf_ring_size /= device->physical_device->rad_info.max_se;
558       }
559 
560       radeon_set_uconfig_reg(cs, R_030938_VGT_TF_RING_SIZE, S_030938_SIZE(tf_ring_size));
561       radeon_set_uconfig_reg(cs, R_030940_VGT_TF_MEMORY_BASE, tf_va >> 8);
562 
563       if (device->physical_device->rad_info.gfx_level >= GFX10) {
564          radeon_set_uconfig_reg(cs, R_030984_VGT_TF_MEMORY_BASE_HI, S_030984_BASE_HI(tf_va >> 40));
565       } else if (device->physical_device->rad_info.gfx_level == GFX9) {
566          radeon_set_uconfig_reg(cs, R_030944_VGT_TF_MEMORY_BASE_HI, S_030944_BASE_HI(tf_va >> 40));
567       }
568 
569       radeon_set_uconfig_reg(cs, R_03093C_VGT_HS_OFFCHIP_PARAM, device->physical_device->hs.hs_offchip_param);
570    } else {
571       radeon_set_config_reg(cs, R_008988_VGT_TF_RING_SIZE, S_008988_SIZE(tf_ring_size));
572       radeon_set_config_reg(cs, R_0089B8_VGT_TF_MEMORY_BASE, tf_va >> 8);
573       radeon_set_config_reg(cs, R_0089B0_VGT_HS_OFFCHIP_PARAM, device->physical_device->hs.hs_offchip_param);
574    }
575 }
576 
577 static VkResult
radv_initialise_task_control_buffer(struct radv_device * device,struct radeon_winsys_bo * task_rings_bo)578 radv_initialise_task_control_buffer(struct radv_device *device, struct radeon_winsys_bo *task_rings_bo)
579 {
580    uint32_t *ptr = (uint32_t *)device->ws->buffer_map(task_rings_bo);
581    if (!ptr)
582       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
583 
584    const uint32_t num_entries = device->physical_device->task_info.num_entries;
585    const uint64_t task_va = radv_buffer_get_va(task_rings_bo);
586    const uint64_t task_draw_ring_va = task_va + device->physical_device->task_info.draw_ring_offset;
587    assert((task_draw_ring_va & 0xFFFFFF00) == (task_draw_ring_va & 0xFFFFFFFF));
588 
589    /* 64-bit write_ptr */
590    ptr[0] = num_entries;
591    ptr[1] = 0;
592    /* 64-bit read_ptr */
593    ptr[2] = num_entries;
594    ptr[3] = 0;
595    /* 64-bit dealloc_ptr */
596    ptr[4] = num_entries;
597    ptr[5] = 0;
598    /* num_entries */
599    ptr[6] = num_entries;
600    /* 64-bit draw ring address */
601    ptr[7] = task_draw_ring_va;
602    ptr[8] = task_draw_ring_va >> 32;
603 
604    device->ws->buffer_unmap(task_rings_bo);
605    return VK_SUCCESS;
606 }
607 
608 static void
radv_emit_task_rings(struct radv_device * device,struct radeon_cmdbuf * cs,struct radeon_winsys_bo * task_rings_bo,bool compute)609 radv_emit_task_rings(struct radv_device *device, struct radeon_cmdbuf *cs, struct radeon_winsys_bo *task_rings_bo,
610                      bool compute)
611 {
612    if (!task_rings_bo)
613       return;
614 
615    const uint64_t task_ctrlbuf_va = radv_buffer_get_va(task_rings_bo);
616    assert(radv_is_aligned(task_ctrlbuf_va, 256));
617    radv_cs_add_buffer(device->ws, cs, task_rings_bo);
618 
619    /* Tell the GPU where the task control buffer is. */
620    radeon_emit(cs, PKT3(PKT3_DISPATCH_TASK_STATE_INIT, 1, 0) | PKT3_SHADER_TYPE_S(!!compute));
621    /* bits [31:8]: control buffer address lo, bits[7:0]: reserved (set to zero) */
622    radeon_emit(cs, task_ctrlbuf_va & 0xFFFFFF00);
623    /* bits [31:0]: control buffer address hi */
624    radeon_emit(cs, task_ctrlbuf_va >> 32);
625 }
626 
627 static void
radv_emit_graphics_scratch(struct radv_device * device,struct radeon_cmdbuf * cs,uint32_t size_per_wave,uint32_t waves,struct radeon_winsys_bo * scratch_bo)628 radv_emit_graphics_scratch(struct radv_device *device, struct radeon_cmdbuf *cs, uint32_t size_per_wave, uint32_t waves,
629                            struct radeon_winsys_bo *scratch_bo)
630 {
631    const struct radeon_info *info = &device->physical_device->rad_info;
632 
633    if (!scratch_bo)
634       return;
635 
636    radv_cs_add_buffer(device->ws, cs, scratch_bo);
637 
638    if (info->gfx_level >= GFX11) {
639       uint64_t va = radv_buffer_get_va(scratch_bo);
640 
641       /* WAVES is per SE for SPI_TMPRING_SIZE. */
642       waves /= info->num_se;
643 
644       radeon_set_context_reg_seq(cs, R_0286E8_SPI_TMPRING_SIZE, 3);
645       radeon_emit(cs, S_0286E8_WAVES(waves) | S_0286E8_WAVESIZE(DIV_ROUND_UP(size_per_wave, 256)));
646       radeon_emit(cs, va >> 8);  /* SPI_GFX_SCRATCH_BASE_LO */
647       radeon_emit(cs, va >> 40); /* SPI_GFX_SCRATCH_BASE_HI */
648    } else {
649       radeon_set_context_reg(cs, R_0286E8_SPI_TMPRING_SIZE,
650                              S_0286E8_WAVES(waves) | S_0286E8_WAVESIZE(DIV_ROUND_UP(size_per_wave, 1024)));
651    }
652 }
653 
654 static void
radv_emit_compute_scratch(struct radv_device * device,struct radeon_cmdbuf * cs,uint32_t size_per_wave,uint32_t waves,struct radeon_winsys_bo * compute_scratch_bo)655 radv_emit_compute_scratch(struct radv_device *device, struct radeon_cmdbuf *cs, uint32_t size_per_wave, uint32_t waves,
656                           struct radeon_winsys_bo *compute_scratch_bo)
657 {
658    const struct radeon_info *info = &device->physical_device->rad_info;
659    uint64_t scratch_va;
660    uint32_t rsrc1;
661 
662    if (!compute_scratch_bo)
663       return;
664 
665    scratch_va = radv_buffer_get_va(compute_scratch_bo);
666    rsrc1 = S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);
667 
668    if (info->gfx_level >= GFX11)
669       rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX11(1);
670    else
671       rsrc1 |= S_008F04_SWIZZLE_ENABLE_GFX6(1);
672 
673    radv_cs_add_buffer(device->ws, cs, compute_scratch_bo);
674 
675    if (info->gfx_level >= GFX11) {
676       radeon_set_sh_reg_seq(cs, R_00B840_COMPUTE_DISPATCH_SCRATCH_BASE_LO, 2);
677       radeon_emit(cs, scratch_va >> 8);
678       radeon_emit(cs, scratch_va >> 40);
679 
680       waves /= info->num_se;
681    }
682 
683    radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0, 2);
684    radeon_emit(cs, scratch_va);
685    radeon_emit(cs, rsrc1);
686 
687    radeon_set_sh_reg(
688       cs, R_00B860_COMPUTE_TMPRING_SIZE,
689       S_00B860_WAVES(waves) | S_00B860_WAVESIZE(DIV_ROUND_UP(size_per_wave, info->gfx_level >= GFX11 ? 256 : 1024)));
690 }
691 
692 static void
radv_emit_compute_shader_pointers(struct radv_device * device,struct radeon_cmdbuf * cs,struct radeon_winsys_bo * descriptor_bo)693 radv_emit_compute_shader_pointers(struct radv_device *device, struct radeon_cmdbuf *cs,
694                                   struct radeon_winsys_bo *descriptor_bo)
695 {
696    if (!descriptor_bo)
697       return;
698 
699    uint64_t va = radv_buffer_get_va(descriptor_bo);
700    radv_cs_add_buffer(device->ws, cs, descriptor_bo);
701 
702    /* Compute shader user data 0-1 have the scratch pointer (unlike GFX shaders),
703     * so emit the descriptor pointer to user data 2-3 instead (task_ring_offsets arg).
704     */
705    radv_emit_shader_pointer(device, cs, R_00B908_COMPUTE_USER_DATA_2, va, true);
706 }
707 
708 static void
radv_emit_graphics_shader_pointers(struct radv_device * device,struct radeon_cmdbuf * cs,struct radeon_winsys_bo * descriptor_bo)709 radv_emit_graphics_shader_pointers(struct radv_device *device, struct radeon_cmdbuf *cs,
710                                    struct radeon_winsys_bo *descriptor_bo)
711 {
712    uint64_t va;
713 
714    if (!descriptor_bo)
715       return;
716 
717    va = radv_buffer_get_va(descriptor_bo);
718 
719    radv_cs_add_buffer(device->ws, cs, descriptor_bo);
720 
721    if (device->physical_device->rad_info.gfx_level >= GFX11) {
722       uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0, R_00B420_SPI_SHADER_PGM_LO_HS,
723                          R_00B220_SPI_SHADER_PGM_LO_GS};
724 
725       for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
726          radv_emit_shader_pointer(device, cs, regs[i], va, true);
727       }
728    } else if (device->physical_device->rad_info.gfx_level >= GFX10) {
729       uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0, R_00B130_SPI_SHADER_USER_DATA_VS_0,
730                          R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS, R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS};
731 
732       for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
733          radv_emit_shader_pointer(device, cs, regs[i], va, true);
734       }
735    } else if (device->physical_device->rad_info.gfx_level == GFX9) {
736       uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0, R_00B130_SPI_SHADER_USER_DATA_VS_0,
737                          R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS, R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS};
738 
739       for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
740          radv_emit_shader_pointer(device, cs, regs[i], va, true);
741       }
742    } else {
743       uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0, R_00B130_SPI_SHADER_USER_DATA_VS_0,
744                          R_00B230_SPI_SHADER_USER_DATA_GS_0, R_00B330_SPI_SHADER_USER_DATA_ES_0,
745                          R_00B430_SPI_SHADER_USER_DATA_HS_0, R_00B530_SPI_SHADER_USER_DATA_LS_0};
746 
747       for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
748          radv_emit_shader_pointer(device, cs, regs[i], va, true);
749       }
750    }
751 }
752 
753 static void
radv_emit_attribute_ring(struct radv_device * device,struct radeon_cmdbuf * cs,struct radeon_winsys_bo * attr_ring_bo,uint32_t attr_ring_size)754 radv_emit_attribute_ring(struct radv_device *device, struct radeon_cmdbuf *cs, struct radeon_winsys_bo *attr_ring_bo,
755                          uint32_t attr_ring_size)
756 {
757    const struct radv_physical_device *pdevice = device->physical_device;
758    uint64_t va;
759 
760    if (!attr_ring_bo)
761       return;
762 
763    assert(pdevice->rad_info.gfx_level >= GFX11);
764 
765    va = radv_buffer_get_va(attr_ring_bo);
766    assert((va >> 32) == pdevice->rad_info.address32_hi);
767 
768    radv_cs_add_buffer(device->ws, cs, attr_ring_bo);
769 
770    /* We must wait for idle using an EOP event before changing the attribute ring registers. Use the
771     * bottom-of-pipe EOP event, but increment the PWS counter instead of writing memory.
772     */
773    radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, 6, 0));
774    radeon_emit(cs, S_490_EVENT_TYPE(V_028A90_BOTTOM_OF_PIPE_TS) | S_490_EVENT_INDEX(5) | S_490_PWS_ENABLE(1));
775    radeon_emit(cs, 0); /* DST_SEL, INT_SEL, DATA_SEL */
776    radeon_emit(cs, 0); /* ADDRESS_LO */
777    radeon_emit(cs, 0); /* ADDRESS_HI */
778    radeon_emit(cs, 0); /* DATA_LO */
779    radeon_emit(cs, 0); /* DATA_HI */
780    radeon_emit(cs, 0); /* INT_CTXID */
781 
782    /* Wait for the PWS counter. */
783    radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 6, 0));
784    radeon_emit(cs, S_580_PWS_STAGE_SEL(V_580_CP_ME) | S_580_PWS_COUNTER_SEL(V_580_TS_SELECT) | S_580_PWS_ENA2(1) |
785                       S_580_PWS_COUNT(0));
786    radeon_emit(cs, 0xffffffff); /* GCR_SIZE */
787    radeon_emit(cs, 0x01ffffff); /* GCR_SIZE_HI */
788    radeon_emit(cs, 0);          /* GCR_BASE_LO */
789    radeon_emit(cs, 0);          /* GCR_BASE_HI */
790    radeon_emit(cs, S_585_PWS_ENA(1));
791    radeon_emit(cs, 0); /* GCR_CNTL */
792 
793    /* The PS will read inputs from this address. */
794    radeon_set_uconfig_reg(cs, R_031118_SPI_ATTRIBUTE_RING_BASE, va >> 16);
795    radeon_set_uconfig_reg(cs, R_03111C_SPI_ATTRIBUTE_RING_SIZE,
796                           S_03111C_MEM_SIZE(((attr_ring_size / pdevice->rad_info.max_se) >> 16) - 1) |
797                              S_03111C_BIG_PAGE(pdevice->rad_info.discardable_allows_big_page) | S_03111C_L1_POLICY(1));
798 }
799 
800 static void
radv_init_graphics_state(struct radeon_cmdbuf * cs,struct radv_device * device)801 radv_init_graphics_state(struct radeon_cmdbuf *cs, struct radv_device *device)
802 {
803    if (device->gfx_init) {
804       struct radeon_winsys *ws = device->ws;
805 
806       ws->cs_execute_ib(cs, device->gfx_init, 0, device->gfx_init_size_dw & 0xffff, false);
807 
808       radv_cs_add_buffer(device->ws, cs, device->gfx_init);
809    } else {
810       radv_emit_graphics(device, cs);
811    }
812 }
813 
814 static void
radv_init_compute_state(struct radeon_cmdbuf * cs,struct radv_device * device)815 radv_init_compute_state(struct radeon_cmdbuf *cs, struct radv_device *device)
816 {
817    radv_emit_compute(device, cs);
818 }
819 
820 static VkResult
radv_update_preamble_cs(struct radv_queue_state * queue,struct radv_device * device,const struct radv_queue_ring_info * needs)821 radv_update_preamble_cs(struct radv_queue_state *queue, struct radv_device *device,
822                         const struct radv_queue_ring_info *needs)
823 {
824    struct radeon_winsys *ws = device->ws;
825    struct radeon_winsys_bo *scratch_bo = queue->scratch_bo;
826    struct radeon_winsys_bo *descriptor_bo = queue->descriptor_bo;
827    struct radeon_winsys_bo *compute_scratch_bo = queue->compute_scratch_bo;
828    struct radeon_winsys_bo *esgs_ring_bo = queue->esgs_ring_bo;
829    struct radeon_winsys_bo *gsvs_ring_bo = queue->gsvs_ring_bo;
830    struct radeon_winsys_bo *tess_rings_bo = queue->tess_rings_bo;
831    struct radeon_winsys_bo *task_rings_bo = queue->task_rings_bo;
832    struct radeon_winsys_bo *mesh_scratch_ring_bo = queue->mesh_scratch_ring_bo;
833    struct radeon_winsys_bo *attr_ring_bo = queue->attr_ring_bo;
834    struct radeon_winsys_bo *gds_bo = queue->gds_bo;
835    struct radeon_winsys_bo *gds_oa_bo = queue->gds_oa_bo;
836    struct radeon_cmdbuf *dest_cs[3] = {0};
837    const uint32_t ring_bo_flags = RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING;
838    VkResult result = VK_SUCCESS;
839 
840    const bool add_sample_positions = !queue->ring_info.sample_positions && needs->sample_positions;
841    const uint32_t scratch_size = needs->scratch_size_per_wave * needs->scratch_waves;
842    const uint32_t queue_scratch_size = queue->ring_info.scratch_size_per_wave * queue->ring_info.scratch_waves;
843 
844    if (scratch_size > queue_scratch_size) {
845       result = ws->buffer_create(ws, scratch_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0,
846                                  &scratch_bo);
847       if (result != VK_SUCCESS)
848          goto fail;
849       radv_rmv_log_command_buffer_bo_create(device, scratch_bo, 0, 0, scratch_size);
850    }
851 
852    const uint32_t compute_scratch_size = needs->compute_scratch_size_per_wave * needs->compute_scratch_waves;
853    const uint32_t compute_queue_scratch_size =
854       queue->ring_info.compute_scratch_size_per_wave * queue->ring_info.compute_scratch_waves;
855    if (compute_scratch_size > compute_queue_scratch_size) {
856       result = ws->buffer_create(ws, compute_scratch_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags,
857                                  RADV_BO_PRIORITY_SCRATCH, 0, &compute_scratch_bo);
858       if (result != VK_SUCCESS)
859          goto fail;
860       radv_rmv_log_command_buffer_bo_create(device, compute_scratch_bo, 0, 0, compute_scratch_size);
861    }
862 
863    if (needs->esgs_ring_size > queue->ring_info.esgs_ring_size) {
864       result = ws->buffer_create(ws, needs->esgs_ring_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags,
865                                  RADV_BO_PRIORITY_SCRATCH, 0, &esgs_ring_bo);
866       if (result != VK_SUCCESS)
867          goto fail;
868       radv_rmv_log_command_buffer_bo_create(device, esgs_ring_bo, 0, 0, needs->esgs_ring_size);
869    }
870 
871    if (needs->gsvs_ring_size > queue->ring_info.gsvs_ring_size) {
872       result = ws->buffer_create(ws, needs->gsvs_ring_size, 4096, RADEON_DOMAIN_VRAM, ring_bo_flags,
873                                  RADV_BO_PRIORITY_SCRATCH, 0, &gsvs_ring_bo);
874       if (result != VK_SUCCESS)
875          goto fail;
876       radv_rmv_log_command_buffer_bo_create(device, gsvs_ring_bo, 0, 0, needs->gsvs_ring_size);
877    }
878 
879    if (!queue->ring_info.tess_rings && needs->tess_rings) {
880       uint64_t tess_rings_size =
881          device->physical_device->hs.tess_offchip_ring_offset + device->physical_device->hs.tess_offchip_ring_size;
882       result = ws->buffer_create(ws, tess_rings_size, 256, RADEON_DOMAIN_VRAM, ring_bo_flags, RADV_BO_PRIORITY_SCRATCH,
883                                  0, &tess_rings_bo);
884       if (result != VK_SUCCESS)
885          goto fail;
886       radv_rmv_log_command_buffer_bo_create(device, tess_rings_bo, 0, 0, tess_rings_size);
887    }
888 
889    if (!queue->ring_info.task_rings && needs->task_rings) {
890       assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
891 
892       /* We write the control buffer from the CPU, so need to grant CPU access to the BO.
893        * The draw ring needs to be zero-initialized otherwise the ready bits will be incorrect.
894        */
895       uint32_t task_rings_bo_flags =
896          RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_ZERO_VRAM;
897 
898       result = ws->buffer_create(ws, device->physical_device->task_info.bo_size_bytes, 256, RADEON_DOMAIN_VRAM,
899                                  task_rings_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0, &task_rings_bo);
900       if (result != VK_SUCCESS)
901          goto fail;
902       radv_rmv_log_command_buffer_bo_create(device, task_rings_bo, 0, 0,
903                                             device->physical_device->task_info.bo_size_bytes);
904 
905       result = radv_initialise_task_control_buffer(device, task_rings_bo);
906       if (result != VK_SUCCESS)
907          goto fail;
908    }
909 
910    if (!queue->ring_info.mesh_scratch_ring && needs->mesh_scratch_ring) {
911       assert(device->physical_device->rad_info.gfx_level >= GFX10_3);
912       result = ws->buffer_create(ws, RADV_MESH_SCRATCH_NUM_ENTRIES * RADV_MESH_SCRATCH_ENTRY_BYTES, 256,
913                                  RADEON_DOMAIN_VRAM, ring_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0, &mesh_scratch_ring_bo);
914 
915       if (result != VK_SUCCESS)
916          goto fail;
917       radv_rmv_log_command_buffer_bo_create(device, mesh_scratch_ring_bo, 0, 0,
918                                             RADV_MESH_SCRATCH_NUM_ENTRIES * RADV_MESH_SCRATCH_ENTRY_BYTES);
919    }
920 
921    if (needs->attr_ring_size > queue->ring_info.attr_ring_size) {
922       assert(device->physical_device->rad_info.gfx_level >= GFX11);
923       result = ws->buffer_create(ws, needs->attr_ring_size, 2 * 1024 * 1024 /* 2MiB */, RADEON_DOMAIN_VRAM,
924                                  RADEON_FLAG_32BIT | RADEON_FLAG_DISCARDABLE | ring_bo_flags, RADV_BO_PRIORITY_SCRATCH,
925                                  0, &attr_ring_bo);
926       if (result != VK_SUCCESS)
927          goto fail;
928       radv_rmv_log_command_buffer_bo_create(device, attr_ring_bo, 0, 0, needs->attr_ring_size);
929    }
930 
931    if (!queue->ring_info.gds && needs->gds) {
932       assert(device->physical_device->rad_info.gfx_level >= GFX10);
933 
934       /* 4 streamout GDS counters.
935        * We need 256B (64 dw) of GDS, otherwise streamout hangs.
936        */
937       result = ws->buffer_create(ws, 256, 4, RADEON_DOMAIN_GDS, ring_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0, &gds_bo);
938       if (result != VK_SUCCESS)
939          goto fail;
940 
941       /* Add the GDS BO to our global BO list to prevent the kernel to emit a GDS switch and reset
942        * the state when a compute queue is used.
943        */
944       result = device->ws->buffer_make_resident(ws, gds_bo, true);
945       if (result != VK_SUCCESS)
946          goto fail;
947    }
948 
949    if (!queue->ring_info.gds_oa && needs->gds_oa) {
950       assert(device->physical_device->rad_info.gfx_level >= GFX10);
951 
952       result = ws->buffer_create(ws, 1, 1, RADEON_DOMAIN_OA, ring_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0, &gds_oa_bo);
953       if (result != VK_SUCCESS)
954          goto fail;
955 
956       /* Add the GDS OA BO to our global BO list to prevent the kernel to emit a GDS switch and
957        * reset the state when a compute queue is used.
958        */
959       result = device->ws->buffer_make_resident(ws, gds_oa_bo, true);
960       if (result != VK_SUCCESS)
961          goto fail;
962    }
963 
964    /* Re-initialize the descriptor BO when any ring BOs changed.
965     *
966     * Additionally, make sure to create the descriptor BO for the compute queue
967     * when it uses the task shader rings. The task rings BO is shared between the
968     * GFX and compute queues and already initialized here.
969     */
970    if ((queue->qf == RADV_QUEUE_COMPUTE && !descriptor_bo && task_rings_bo) || scratch_bo != queue->scratch_bo ||
971        esgs_ring_bo != queue->esgs_ring_bo || gsvs_ring_bo != queue->gsvs_ring_bo ||
972        tess_rings_bo != queue->tess_rings_bo || task_rings_bo != queue->task_rings_bo ||
973        mesh_scratch_ring_bo != queue->mesh_scratch_ring_bo || attr_ring_bo != queue->attr_ring_bo ||
974        add_sample_positions) {
975       const uint32_t size = 304;
976 
977       result = ws->buffer_create(ws, size, 4096, RADEON_DOMAIN_VRAM,
978                                  RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_READ_ONLY,
979                                  RADV_BO_PRIORITY_DESCRIPTOR, 0, &descriptor_bo);
980       if (result != VK_SUCCESS)
981          goto fail;
982    }
983 
984    if (descriptor_bo != queue->descriptor_bo) {
985       uint32_t *map = (uint32_t *)ws->buffer_map(descriptor_bo);
986       if (!map) {
987          result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
988          goto fail;
989       }
990 
991       radv_fill_shader_rings(device, map, scratch_bo, needs->esgs_ring_size, esgs_ring_bo, needs->gsvs_ring_size,
992                              gsvs_ring_bo, tess_rings_bo, task_rings_bo, mesh_scratch_ring_bo, needs->attr_ring_size,
993                              attr_ring_bo);
994 
995       ws->buffer_unmap(descriptor_bo);
996    }
997 
998    for (int i = 0; i < 3; ++i) {
999       enum rgp_flush_bits sqtt_flush_bits = 0;
1000       struct radeon_cmdbuf *cs = NULL;
1001       cs = ws->cs_create(ws, radv_queue_family_to_ring(device->physical_device, queue->qf), false);
1002       if (!cs) {
1003          result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
1004          goto fail;
1005       }
1006 
1007       radeon_check_space(ws, cs, 512);
1008       dest_cs[i] = cs;
1009 
1010       if (scratch_bo)
1011          radv_cs_add_buffer(ws, cs, scratch_bo);
1012 
1013       /* Emit initial configuration. */
1014       switch (queue->qf) {
1015       case RADV_QUEUE_GENERAL:
1016          if (queue->uses_shadow_regs)
1017             radv_emit_shadow_regs_preamble(cs, device, queue);
1018          radv_init_graphics_state(cs, device);
1019 
1020          if (esgs_ring_bo || gsvs_ring_bo || tess_rings_bo || task_rings_bo) {
1021             radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1022             radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1023 
1024             radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1025             radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1026          }
1027 
1028          radv_emit_gs_ring_sizes(device, cs, esgs_ring_bo, needs->esgs_ring_size, gsvs_ring_bo, needs->gsvs_ring_size);
1029          radv_emit_tess_factor_ring(device, cs, tess_rings_bo);
1030          radv_emit_task_rings(device, cs, task_rings_bo, false);
1031          radv_emit_attribute_ring(device, cs, attr_ring_bo, needs->attr_ring_size);
1032          radv_emit_graphics_shader_pointers(device, cs, descriptor_bo);
1033          radv_emit_compute_scratch(device, cs, needs->compute_scratch_size_per_wave, needs->compute_scratch_waves,
1034                                    compute_scratch_bo);
1035          radv_emit_graphics_scratch(device, cs, needs->scratch_size_per_wave, needs->scratch_waves, scratch_bo);
1036          break;
1037       case RADV_QUEUE_COMPUTE:
1038          radv_init_compute_state(cs, device);
1039 
1040          if (task_rings_bo) {
1041             radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1042             radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
1043          }
1044 
1045          radv_emit_task_rings(device, cs, task_rings_bo, true);
1046          radv_emit_compute_shader_pointers(device, cs, descriptor_bo);
1047          radv_emit_compute_scratch(device, cs, needs->compute_scratch_size_per_wave, needs->compute_scratch_waves,
1048                                    compute_scratch_bo);
1049          break;
1050       default:
1051          break;
1052       }
1053 
1054       if (i < 2) {
1055          /* The two initial preambles have a cache flush at the beginning. */
1056          const enum amd_gfx_level gfx_level = device->physical_device->rad_info.gfx_level;
1057          enum radv_cmd_flush_bits flush_bits = RADV_CMD_FLAG_INV_ICACHE | RADV_CMD_FLAG_INV_SCACHE |
1058                                                RADV_CMD_FLAG_INV_VCACHE | RADV_CMD_FLAG_INV_L2 |
1059                                                RADV_CMD_FLAG_START_PIPELINE_STATS;
1060 
1061          if (i == 0) {
1062             /* The full flush preamble should also wait for previous shader work to finish. */
1063             flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH;
1064             if (queue->qf == RADV_QUEUE_GENERAL)
1065                flush_bits |= RADV_CMD_FLAG_PS_PARTIAL_FLUSH;
1066          }
1067 
1068          radv_cs_emit_cache_flush(ws, cs, gfx_level, NULL, 0, queue->qf, flush_bits, &sqtt_flush_bits, 0);
1069       }
1070 
1071       result = ws->cs_finalize(cs);
1072       if (result != VK_SUCCESS)
1073          goto fail;
1074    }
1075 
1076    if (queue->initial_full_flush_preamble_cs)
1077       ws->cs_destroy(queue->initial_full_flush_preamble_cs);
1078 
1079    if (queue->initial_preamble_cs)
1080       ws->cs_destroy(queue->initial_preamble_cs);
1081 
1082    if (queue->continue_preamble_cs)
1083       ws->cs_destroy(queue->continue_preamble_cs);
1084 
1085    queue->initial_full_flush_preamble_cs = dest_cs[0];
1086    queue->initial_preamble_cs = dest_cs[1];
1087    queue->continue_preamble_cs = dest_cs[2];
1088 
1089    if (scratch_bo != queue->scratch_bo) {
1090       if (queue->scratch_bo) {
1091          ws->buffer_destroy(ws, queue->scratch_bo);
1092          radv_rmv_log_command_buffer_bo_destroy(device, queue->scratch_bo);
1093       }
1094       queue->scratch_bo = scratch_bo;
1095    }
1096 
1097    if (compute_scratch_bo != queue->compute_scratch_bo) {
1098       if (queue->compute_scratch_bo) {
1099          ws->buffer_destroy(ws, queue->compute_scratch_bo);
1100          radv_rmv_log_command_buffer_bo_destroy(device, queue->compute_scratch_bo);
1101       }
1102       queue->compute_scratch_bo = compute_scratch_bo;
1103    }
1104 
1105    if (esgs_ring_bo != queue->esgs_ring_bo) {
1106       if (queue->esgs_ring_bo) {
1107          ws->buffer_destroy(ws, queue->esgs_ring_bo);
1108          radv_rmv_log_command_buffer_bo_destroy(device, queue->esgs_ring_bo);
1109       }
1110       queue->esgs_ring_bo = esgs_ring_bo;
1111    }
1112 
1113    if (gsvs_ring_bo != queue->gsvs_ring_bo) {
1114       if (queue->gsvs_ring_bo) {
1115          ws->buffer_destroy(ws, queue->gsvs_ring_bo);
1116          radv_rmv_log_command_buffer_bo_destroy(device, queue->gsvs_ring_bo);
1117       }
1118       queue->gsvs_ring_bo = gsvs_ring_bo;
1119    }
1120 
1121    if (descriptor_bo != queue->descriptor_bo) {
1122       if (queue->descriptor_bo)
1123          ws->buffer_destroy(ws, queue->descriptor_bo);
1124       queue->descriptor_bo = descriptor_bo;
1125    }
1126 
1127    queue->tess_rings_bo = tess_rings_bo;
1128    queue->task_rings_bo = task_rings_bo;
1129    queue->mesh_scratch_ring_bo = mesh_scratch_ring_bo;
1130    queue->attr_ring_bo = attr_ring_bo;
1131    queue->gds_bo = gds_bo;
1132    queue->gds_oa_bo = gds_oa_bo;
1133    queue->ring_info = *needs;
1134    return VK_SUCCESS;
1135 fail:
1136    for (int i = 0; i < ARRAY_SIZE(dest_cs); ++i)
1137       if (dest_cs[i])
1138          ws->cs_destroy(dest_cs[i]);
1139    if (descriptor_bo && descriptor_bo != queue->descriptor_bo)
1140       ws->buffer_destroy(ws, descriptor_bo);
1141    if (scratch_bo && scratch_bo != queue->scratch_bo)
1142       ws->buffer_destroy(ws, scratch_bo);
1143    if (compute_scratch_bo && compute_scratch_bo != queue->compute_scratch_bo)
1144       ws->buffer_destroy(ws, compute_scratch_bo);
1145    if (esgs_ring_bo && esgs_ring_bo != queue->esgs_ring_bo)
1146       ws->buffer_destroy(ws, esgs_ring_bo);
1147    if (gsvs_ring_bo && gsvs_ring_bo != queue->gsvs_ring_bo)
1148       ws->buffer_destroy(ws, gsvs_ring_bo);
1149    if (tess_rings_bo && tess_rings_bo != queue->tess_rings_bo)
1150       ws->buffer_destroy(ws, tess_rings_bo);
1151    if (task_rings_bo && task_rings_bo != queue->task_rings_bo)
1152       ws->buffer_destroy(ws, task_rings_bo);
1153    if (attr_ring_bo && attr_ring_bo != queue->attr_ring_bo)
1154       ws->buffer_destroy(ws, attr_ring_bo);
1155    if (gds_bo && gds_bo != queue->gds_bo) {
1156       ws->buffer_make_resident(ws, queue->gds_bo, false);
1157       ws->buffer_destroy(ws, gds_bo);
1158    }
1159    if (gds_oa_bo && gds_oa_bo != queue->gds_oa_bo) {
1160       ws->buffer_make_resident(ws, queue->gds_oa_bo, false);
1161       ws->buffer_destroy(ws, gds_oa_bo);
1162    }
1163 
1164    return vk_error(queue, result);
1165 }
1166 
1167 static VkResult
radv_update_preambles(struct radv_queue_state * queue,struct radv_device * device,struct vk_command_buffer * const * cmd_buffers,uint32_t cmd_buffer_count,bool * use_perf_counters,bool * has_follower)1168 radv_update_preambles(struct radv_queue_state *queue, struct radv_device *device,
1169                       struct vk_command_buffer *const *cmd_buffers, uint32_t cmd_buffer_count, bool *use_perf_counters,
1170                       bool *has_follower)
1171 {
1172    if (queue->qf != RADV_QUEUE_GENERAL && queue->qf != RADV_QUEUE_COMPUTE) {
1173       for (uint32_t j = 0; j < cmd_buffer_count; j++) {
1174          struct radv_cmd_buffer *cmd_buffer = container_of(cmd_buffers[j], struct radv_cmd_buffer, vk);
1175 
1176          *has_follower |= !!cmd_buffer->gang.cs;
1177       }
1178 
1179       return VK_SUCCESS;
1180    }
1181 
1182    /* Figure out the needs of the current submission.
1183     * Start by copying the queue's current info.
1184     * This is done because we only allow two possible behaviours for these buffers:
1185     * - Grow when the newly needed amount is larger than what we had
1186     * - Allocate the max size and reuse it, but don't free it until the queue is destroyed
1187     */
1188    struct radv_queue_ring_info needs = queue->ring_info;
1189    *use_perf_counters = false;
1190    *has_follower = false;
1191 
1192    for (uint32_t j = 0; j < cmd_buffer_count; j++) {
1193       struct radv_cmd_buffer *cmd_buffer = container_of(cmd_buffers[j], struct radv_cmd_buffer, vk);
1194 
1195       needs.scratch_size_per_wave = MAX2(needs.scratch_size_per_wave, cmd_buffer->scratch_size_per_wave_needed);
1196       needs.scratch_waves = MAX2(needs.scratch_waves, cmd_buffer->scratch_waves_wanted);
1197       needs.compute_scratch_size_per_wave =
1198          MAX2(needs.compute_scratch_size_per_wave, cmd_buffer->compute_scratch_size_per_wave_needed);
1199       needs.compute_scratch_waves = MAX2(needs.compute_scratch_waves, cmd_buffer->compute_scratch_waves_wanted);
1200       needs.esgs_ring_size = MAX2(needs.esgs_ring_size, cmd_buffer->esgs_ring_size_needed);
1201       needs.gsvs_ring_size = MAX2(needs.gsvs_ring_size, cmd_buffer->gsvs_ring_size_needed);
1202       needs.tess_rings |= cmd_buffer->tess_rings_needed;
1203       needs.task_rings |= cmd_buffer->task_rings_needed;
1204       needs.mesh_scratch_ring |= cmd_buffer->mesh_scratch_ring_needed;
1205       needs.gds |= cmd_buffer->gds_needed;
1206       needs.gds_oa |= cmd_buffer->gds_oa_needed;
1207       needs.sample_positions |= cmd_buffer->sample_positions_needed;
1208       *use_perf_counters |= cmd_buffer->state.uses_perf_counters;
1209       *has_follower |= !!cmd_buffer->gang.cs;
1210    }
1211 
1212    /* Sanitize scratch size information. */
1213    needs.scratch_waves =
1214       needs.scratch_size_per_wave ? MIN2(needs.scratch_waves, UINT32_MAX / needs.scratch_size_per_wave) : 0;
1215    needs.compute_scratch_waves =
1216       needs.compute_scratch_size_per_wave
1217          ? MIN2(needs.compute_scratch_waves, UINT32_MAX / needs.compute_scratch_size_per_wave)
1218          : 0;
1219 
1220    if (device->physical_device->rad_info.gfx_level >= GFX11 && queue->qf == RADV_QUEUE_GENERAL) {
1221       needs.attr_ring_size =
1222          device->physical_device->rad_info.attribute_ring_size_per_se * device->physical_device->rad_info.max_se;
1223    }
1224 
1225    /* Return early if we already match these needs.
1226     * Note that it's not possible for any of the needed values to be less
1227     * than what the queue already had, because we only ever increase the allocated size.
1228     */
1229    if (queue->initial_full_flush_preamble_cs && queue->ring_info.scratch_size_per_wave == needs.scratch_size_per_wave &&
1230        queue->ring_info.scratch_waves == needs.scratch_waves &&
1231        queue->ring_info.compute_scratch_size_per_wave == needs.compute_scratch_size_per_wave &&
1232        queue->ring_info.compute_scratch_waves == needs.compute_scratch_waves &&
1233        queue->ring_info.esgs_ring_size == needs.esgs_ring_size &&
1234        queue->ring_info.gsvs_ring_size == needs.gsvs_ring_size && queue->ring_info.tess_rings == needs.tess_rings &&
1235        queue->ring_info.task_rings == needs.task_rings &&
1236        queue->ring_info.mesh_scratch_ring == needs.mesh_scratch_ring &&
1237        queue->ring_info.attr_ring_size == needs.attr_ring_size && queue->ring_info.gds == needs.gds &&
1238        queue->ring_info.gds_oa == needs.gds_oa && queue->ring_info.sample_positions == needs.sample_positions)
1239       return VK_SUCCESS;
1240 
1241    return radv_update_preamble_cs(queue, device, &needs);
1242 }
1243 
1244 static VkResult
radv_create_gang_wait_preambles_postambles(struct radv_queue * queue)1245 radv_create_gang_wait_preambles_postambles(struct radv_queue *queue)
1246 {
1247    if (queue->gang_sem_bo)
1248       return VK_SUCCESS;
1249 
1250    VkResult r = VK_SUCCESS;
1251    struct radv_device *device = queue->device;
1252    struct radeon_winsys *ws = device->ws;
1253    const enum amd_ip_type leader_ip = radv_queue_family_to_ring(device->physical_device, queue->state.qf);
1254    struct radeon_winsys_bo *gang_sem_bo = NULL;
1255 
1256    /* Gang semaphores BO.
1257     * DWORD 0: used in preambles, gang leader writes, gang members wait.
1258     * DWORD 1: used in postambles, gang leader waits, gang members write.
1259     */
1260    r = ws->buffer_create(ws, 8, 4, RADEON_DOMAIN_VRAM, RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_ZERO_VRAM,
1261                          RADV_BO_PRIORITY_SCRATCH, 0, &gang_sem_bo);
1262    if (r != VK_SUCCESS)
1263       return r;
1264 
1265    struct radeon_cmdbuf *leader_pre_cs = ws->cs_create(ws, leader_ip, false);
1266    struct radeon_cmdbuf *leader_post_cs = ws->cs_create(ws, leader_ip, false);
1267    struct radeon_cmdbuf *ace_pre_cs = ws->cs_create(ws, AMD_IP_COMPUTE, false);
1268    struct radeon_cmdbuf *ace_post_cs = ws->cs_create(ws, AMD_IP_COMPUTE, false);
1269 
1270    if (!leader_pre_cs || !leader_post_cs || !ace_pre_cs || !ace_post_cs) {
1271       r = VK_ERROR_OUT_OF_DEVICE_MEMORY;
1272       goto fail;
1273    }
1274 
1275    radeon_check_space(ws, leader_pre_cs, 256);
1276    radeon_check_space(ws, leader_post_cs, 256);
1277    radeon_check_space(ws, ace_pre_cs, 256);
1278    radeon_check_space(ws, ace_post_cs, 256);
1279 
1280    radv_cs_add_buffer(ws, leader_pre_cs, gang_sem_bo);
1281    radv_cs_add_buffer(ws, leader_post_cs, gang_sem_bo);
1282    radv_cs_add_buffer(ws, ace_pre_cs, gang_sem_bo);
1283    radv_cs_add_buffer(ws, ace_post_cs, gang_sem_bo);
1284 
1285    const uint64_t ace_wait_va = radv_buffer_get_va(gang_sem_bo);
1286    const uint64_t leader_wait_va = ace_wait_va + 4;
1287    const uint32_t zero = 0;
1288    const uint32_t one = 1;
1289 
1290    /* Preambles for gang submission.
1291     * Make gang members wait until the gang leader starts.
1292     * Userspace is required to emit this wait to make sure it behaves correctly
1293     * in a multi-process environment, because task shader dispatches are not
1294     * meant to be executed on multiple compute engines at the same time.
1295     */
1296    radv_cp_wait_mem(ace_pre_cs, RADV_QUEUE_COMPUTE, WAIT_REG_MEM_GREATER_OR_EQUAL, ace_wait_va, 1, 0xffffffff);
1297    radv_cs_write_data(device, ace_pre_cs, RADV_QUEUE_COMPUTE, V_370_ME, ace_wait_va, 1, &zero, false);
1298    radv_cs_write_data(device, leader_pre_cs, queue->state.qf, V_370_ME, ace_wait_va, 1, &one, false);
1299 
1300    /* Create postambles for gang submission.
1301     * This ensures that the gang leader waits for the whole gang,
1302     * which is necessary because the kernel signals the userspace fence
1303     * as soon as the gang leader is done, which may lead to bugs because the
1304     * same command buffers could be submitted again while still being executed.
1305     */
1306    radv_cp_wait_mem(leader_post_cs, queue->state.qf, WAIT_REG_MEM_GREATER_OR_EQUAL, leader_wait_va, 1, 0xffffffff);
1307    radv_cs_write_data(device, leader_post_cs, queue->state.qf, V_370_ME, leader_wait_va, 1, &zero, false);
1308    radv_cs_emit_write_event_eop(ace_post_cs, device->physical_device->rad_info.gfx_level, RADV_QUEUE_COMPUTE,
1309                                 V_028A90_BOTTOM_OF_PIPE_TS, 0, EOP_DST_SEL_MEM, EOP_DATA_SEL_VALUE_32BIT,
1310                                 leader_wait_va, 1, 0);
1311 
1312    r = ws->cs_finalize(leader_pre_cs);
1313    if (r != VK_SUCCESS)
1314       goto fail;
1315    r = ws->cs_finalize(leader_post_cs);
1316    if (r != VK_SUCCESS)
1317       goto fail;
1318    r = ws->cs_finalize(ace_pre_cs);
1319    if (r != VK_SUCCESS)
1320       goto fail;
1321    r = ws->cs_finalize(ace_post_cs);
1322    if (r != VK_SUCCESS)
1323       goto fail;
1324 
1325    queue->gang_sem_bo = gang_sem_bo;
1326    queue->state.gang_wait_preamble_cs = leader_pre_cs;
1327    queue->state.gang_wait_postamble_cs = leader_post_cs;
1328    queue->follower_state->gang_wait_preamble_cs = ace_pre_cs;
1329    queue->follower_state->gang_wait_postamble_cs = ace_post_cs;
1330 
1331    return VK_SUCCESS;
1332 
1333 fail:
1334    if (leader_pre_cs)
1335       ws->cs_destroy(leader_pre_cs);
1336    if (leader_post_cs)
1337       ws->cs_destroy(leader_post_cs);
1338    if (ace_pre_cs)
1339       ws->cs_destroy(ace_pre_cs);
1340    if (ace_post_cs)
1341       ws->cs_destroy(ace_post_cs);
1342    if (gang_sem_bo)
1343       ws->buffer_destroy(ws, gang_sem_bo);
1344 
1345    return r;
1346 }
1347 
1348 static bool
radv_queue_init_follower_state(struct radv_queue * queue)1349 radv_queue_init_follower_state(struct radv_queue *queue)
1350 {
1351    if (queue->follower_state)
1352       return true;
1353 
1354    queue->follower_state = calloc(1, sizeof(struct radv_queue_state));
1355    if (!queue->follower_state)
1356       return false;
1357 
1358    queue->follower_state->qf = RADV_QUEUE_COMPUTE;
1359    return true;
1360 }
1361 
1362 static VkResult
radv_update_gang_preambles(struct radv_queue * queue)1363 radv_update_gang_preambles(struct radv_queue *queue)
1364 {
1365    if (!radv_queue_init_follower_state(queue))
1366       return VK_ERROR_OUT_OF_HOST_MEMORY;
1367 
1368    VkResult r = VK_SUCCESS;
1369 
1370    /* Copy task rings state.
1371     * Task shaders that are submitted on the ACE queue need to share
1372     * their ring buffers with the mesh shaders on the GFX queue.
1373     */
1374    queue->follower_state->ring_info.task_rings = queue->state.ring_info.task_rings;
1375    queue->follower_state->task_rings_bo = queue->state.task_rings_bo;
1376 
1377    /* Copy some needed states from the parent queue state.
1378     * These can only increase so it's okay to copy them as-is without checking.
1379     * Note, task shaders use the scratch size from their graphics pipeline.
1380     */
1381    struct radv_queue_ring_info needs = queue->follower_state->ring_info;
1382    needs.compute_scratch_size_per_wave = queue->state.ring_info.scratch_size_per_wave;
1383    needs.compute_scratch_waves = queue->state.ring_info.scratch_waves;
1384    needs.task_rings = queue->state.ring_info.task_rings;
1385 
1386    r = radv_update_preamble_cs(queue->follower_state, queue->device, &needs);
1387    if (r != VK_SUCCESS)
1388       return r;
1389 
1390    r = radv_create_gang_wait_preambles_postambles(queue);
1391    if (r != VK_SUCCESS)
1392       return r;
1393 
1394    return VK_SUCCESS;
1395 }
1396 
1397 static struct radeon_cmdbuf *
radv_create_perf_counter_lock_cs(struct radv_device * device,unsigned pass,bool unlock)1398 radv_create_perf_counter_lock_cs(struct radv_device *device, unsigned pass, bool unlock)
1399 {
1400    struct radeon_cmdbuf **cs_ref = &device->perf_counter_lock_cs[pass * 2 + (unlock ? 1 : 0)];
1401    struct radeon_cmdbuf *cs;
1402 
1403    if (*cs_ref)
1404       return *cs_ref;
1405 
1406    cs = device->ws->cs_create(device->ws, AMD_IP_GFX, false);
1407    if (!cs)
1408       return NULL;
1409 
1410    ASSERTED unsigned cdw = radeon_check_space(device->ws, cs, 21);
1411 
1412    radv_cs_add_buffer(device->ws, cs, device->perf_counter_bo);
1413 
1414    if (!unlock) {
1415       uint64_t mutex_va = radv_buffer_get_va(device->perf_counter_bo) + PERF_CTR_BO_LOCK_OFFSET;
1416       radeon_emit(cs, PKT3(PKT3_ATOMIC_MEM, 7, 0));
1417       radeon_emit(cs, ATOMIC_OP(TC_OP_ATOMIC_CMPSWAP_32) | ATOMIC_COMMAND(ATOMIC_COMMAND_LOOP));
1418       radeon_emit(cs, mutex_va);       /* addr lo */
1419       radeon_emit(cs, mutex_va >> 32); /* addr hi */
1420       radeon_emit(cs, 1);              /* data lo */
1421       radeon_emit(cs, 0);              /* data hi */
1422       radeon_emit(cs, 0);              /* compare data lo */
1423       radeon_emit(cs, 0);              /* compare data hi */
1424       radeon_emit(cs, 10);             /* loop interval */
1425    }
1426 
1427    uint64_t va = radv_buffer_get_va(device->perf_counter_bo) + PERF_CTR_BO_PASS_OFFSET;
1428    uint64_t unset_va = va + (unlock ? 8 * pass : 0);
1429    uint64_t set_va = va + (unlock ? 0 : 8 * pass);
1430 
1431    radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
1432    radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_IMM) | COPY_DATA_DST_SEL(COPY_DATA_DST_MEM) | COPY_DATA_COUNT_SEL |
1433                       COPY_DATA_WR_CONFIRM);
1434    radeon_emit(cs, 0); /* immediate */
1435    radeon_emit(cs, 0);
1436    radeon_emit(cs, unset_va);
1437    radeon_emit(cs, unset_va >> 32);
1438 
1439    radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
1440    radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_IMM) | COPY_DATA_DST_SEL(COPY_DATA_DST_MEM) | COPY_DATA_COUNT_SEL |
1441                       COPY_DATA_WR_CONFIRM);
1442    radeon_emit(cs, 1); /* immediate */
1443    radeon_emit(cs, 0);
1444    radeon_emit(cs, set_va);
1445    radeon_emit(cs, set_va >> 32);
1446 
1447    if (unlock) {
1448       uint64_t mutex_va = radv_buffer_get_va(device->perf_counter_bo) + PERF_CTR_BO_LOCK_OFFSET;
1449 
1450       radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
1451       radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_IMM) | COPY_DATA_DST_SEL(COPY_DATA_DST_MEM) | COPY_DATA_COUNT_SEL |
1452                          COPY_DATA_WR_CONFIRM);
1453       radeon_emit(cs, 0); /* immediate */
1454       radeon_emit(cs, 0);
1455       radeon_emit(cs, mutex_va);
1456       radeon_emit(cs, mutex_va >> 32);
1457    }
1458 
1459    assert(cs->cdw <= cdw);
1460 
1461    VkResult result = device->ws->cs_finalize(cs);
1462    if (result != VK_SUCCESS) {
1463       device->ws->cs_destroy(cs);
1464       return NULL;
1465    }
1466 
1467    /* All the casts are to avoid MSVC errors around pointer truncation in a non-taken
1468     * alternative.
1469     */
1470    if (p_atomic_cmpxchg((uintptr_t *)cs_ref, 0, (uintptr_t)cs) != 0) {
1471       device->ws->cs_destroy(cs);
1472    }
1473 
1474    return *cs_ref;
1475 }
1476 
1477 static void
radv_get_shader_upload_sync_wait(struct radv_device * device,uint64_t shader_upload_seq,struct vk_sync_wait * out_sync_wait)1478 radv_get_shader_upload_sync_wait(struct radv_device *device, uint64_t shader_upload_seq,
1479                                  struct vk_sync_wait *out_sync_wait)
1480 {
1481    struct vk_semaphore *semaphore = vk_semaphore_from_handle(device->shader_upload_sem);
1482    struct vk_sync *sync = vk_semaphore_get_active_sync(semaphore);
1483    *out_sync_wait = (struct vk_sync_wait){
1484       .sync = sync,
1485       .wait_value = shader_upload_seq,
1486       .stage_mask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
1487    };
1488 }
1489 
1490 static VkResult
radv_queue_submit_normal(struct radv_queue * queue,struct vk_queue_submit * submission)1491 radv_queue_submit_normal(struct radv_queue *queue, struct vk_queue_submit *submission)
1492 {
1493    struct radeon_winsys_ctx *ctx = queue->hw_ctx;
1494    bool use_ace = false;
1495    bool use_perf_counters = false;
1496    VkResult result;
1497    uint64_t shader_upload_seq = 0;
1498    uint32_t wait_count = submission->wait_count;
1499    struct vk_sync_wait *waits = submission->waits;
1500 
1501    result = radv_update_preambles(&queue->state, queue->device, submission->command_buffers,
1502                                   submission->command_buffer_count, &use_perf_counters, &use_ace);
1503    if (result != VK_SUCCESS)
1504       return result;
1505 
1506    if (use_ace) {
1507       result = radv_update_gang_preambles(queue);
1508       if (result != VK_SUCCESS)
1509          return result;
1510    }
1511 
1512    const unsigned cmd_buffer_count = submission->command_buffer_count;
1513    const unsigned max_cs_submission = radv_device_fault_detection_enabled(queue->device) ? 1 : cmd_buffer_count;
1514    const unsigned cs_array_size = (use_ace ? 2 : 1) * MIN2(max_cs_submission, cmd_buffer_count);
1515 
1516    struct radeon_cmdbuf **cs_array = malloc(sizeof(struct radeon_cmdbuf *) * cs_array_size);
1517    if (!cs_array)
1518       return VK_ERROR_OUT_OF_HOST_MEMORY;
1519 
1520    if (radv_device_fault_detection_enabled(queue->device))
1521       simple_mtx_lock(&queue->device->trace_mtx);
1522 
1523    for (uint32_t j = 0; j < submission->command_buffer_count; j++) {
1524       struct radv_cmd_buffer *cmd_buffer = (struct radv_cmd_buffer *)submission->command_buffers[j];
1525       shader_upload_seq = MAX2(shader_upload_seq, cmd_buffer->shader_upload_seq);
1526    }
1527 
1528    if (shader_upload_seq > queue->last_shader_upload_seq) {
1529       /* Patch the wait array to add waiting for referenced shaders to upload. */
1530       struct vk_sync_wait *new_waits = malloc(sizeof(struct vk_sync_wait) * (wait_count + 1));
1531       if (!new_waits) {
1532          result = VK_ERROR_OUT_OF_HOST_MEMORY;
1533          goto fail;
1534       }
1535 
1536       memcpy(new_waits, submission->waits, sizeof(struct vk_sync_wait) * submission->wait_count);
1537       radv_get_shader_upload_sync_wait(queue->device, shader_upload_seq, &new_waits[submission->wait_count]);
1538 
1539       waits = new_waits;
1540       wait_count += 1;
1541    }
1542 
1543    /* For fences on the same queue/vm amdgpu doesn't wait till all processing is finished
1544     * before starting the next cmdbuffer, so we need to do it here.
1545     */
1546    const bool need_wait = wait_count > 0;
1547    unsigned num_initial_preambles = 0;
1548    unsigned num_continue_preambles = 0;
1549    unsigned num_postambles = 0;
1550    struct radeon_cmdbuf *initial_preambles[5] = {0};
1551    struct radeon_cmdbuf *continue_preambles[5] = {0};
1552    struct radeon_cmdbuf *postambles[3] = {0};
1553 
1554    if (queue->state.qf == RADV_QUEUE_GENERAL || queue->state.qf == RADV_QUEUE_COMPUTE) {
1555       initial_preambles[num_initial_preambles++] =
1556          need_wait ? queue->state.initial_full_flush_preamble_cs : queue->state.initial_preamble_cs;
1557 
1558       continue_preambles[num_continue_preambles++] = queue->state.continue_preamble_cs;
1559 
1560       if (use_perf_counters) {
1561          /* RADV only supports perf counters on the GFX queue currently. */
1562          assert(queue->state.qf == RADV_QUEUE_GENERAL);
1563 
1564          /* Create the lock/unlock CS. */
1565          struct radeon_cmdbuf *perf_ctr_lock_cs =
1566             radv_create_perf_counter_lock_cs(queue->device, submission->perf_pass_index, false);
1567          struct radeon_cmdbuf *perf_ctr_unlock_cs =
1568             radv_create_perf_counter_lock_cs(queue->device, submission->perf_pass_index, true);
1569 
1570          if (!perf_ctr_lock_cs || !perf_ctr_unlock_cs) {
1571             result = VK_ERROR_OUT_OF_HOST_MEMORY;
1572             goto fail;
1573          }
1574 
1575          initial_preambles[num_initial_preambles++] = perf_ctr_lock_cs;
1576          continue_preambles[num_continue_preambles++] = perf_ctr_lock_cs;
1577          postambles[num_postambles++] = perf_ctr_unlock_cs;
1578       }
1579    }
1580 
1581    const unsigned num_1q_initial_preambles = num_initial_preambles;
1582    const unsigned num_1q_continue_preambles = num_continue_preambles;
1583    const unsigned num_1q_postambles = num_postambles;
1584 
1585    if (use_ace) {
1586       initial_preambles[num_initial_preambles++] = queue->state.gang_wait_preamble_cs;
1587       initial_preambles[num_initial_preambles++] = queue->follower_state->gang_wait_preamble_cs;
1588       initial_preambles[num_initial_preambles++] =
1589          need_wait ? queue->follower_state->initial_full_flush_preamble_cs : queue->follower_state->initial_preamble_cs;
1590 
1591       continue_preambles[num_continue_preambles++] = queue->state.gang_wait_preamble_cs;
1592       continue_preambles[num_continue_preambles++] = queue->follower_state->gang_wait_preamble_cs;
1593       continue_preambles[num_continue_preambles++] = queue->follower_state->continue_preamble_cs;
1594 
1595       postambles[num_postambles++] = queue->follower_state->gang_wait_postamble_cs;
1596       postambles[num_postambles++] = queue->state.gang_wait_postamble_cs;
1597    }
1598 
1599    struct radv_winsys_submit_info submit = {
1600       .ip_type = radv_queue_ring(queue),
1601       .queue_index = queue->vk.index_in_family,
1602       .cs_array = cs_array,
1603       .cs_count = 0,
1604       .initial_preamble_count = num_1q_initial_preambles,
1605       .continue_preamble_count = num_1q_continue_preambles,
1606       .postamble_count = num_1q_postambles,
1607       .initial_preamble_cs = initial_preambles,
1608       .continue_preamble_cs = continue_preambles,
1609       .postamble_cs = postambles,
1610       .uses_shadow_regs = queue->state.uses_shadow_regs,
1611    };
1612 
1613    for (uint32_t j = 0, advance; j < cmd_buffer_count; j += advance) {
1614       advance = MIN2(max_cs_submission, cmd_buffer_count - j);
1615       const bool last_submit = j + advance == cmd_buffer_count;
1616       bool submit_ace = false;
1617       unsigned num_submitted_cs = 0;
1618 
1619       if (radv_device_fault_detection_enabled(queue->device))
1620          *queue->device->trace_id_ptr = 0;
1621 
1622       struct radeon_cmdbuf *chainable = NULL;
1623       struct radeon_cmdbuf *chainable_ace = NULL;
1624 
1625       /* Add CS from submitted command buffers. */
1626       for (unsigned c = 0; c < advance; ++c) {
1627          struct radv_cmd_buffer *cmd_buffer = (struct radv_cmd_buffer *)submission->command_buffers[j + c];
1628          assert(cmd_buffer->vk.level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1629          const bool can_chain_next = !(cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT);
1630 
1631          /* Follower needs to be before the gang leader because the last CS must match the queue's IP type. */
1632          if (cmd_buffer->gang.cs) {
1633             queue->device->ws->cs_unchain(cmd_buffer->gang.cs);
1634             if (!chainable_ace || !queue->device->ws->cs_chain(chainable_ace, cmd_buffer->gang.cs, false)) {
1635                cs_array[num_submitted_cs++] = cmd_buffer->gang.cs;
1636 
1637                /* Prevent chaining the gang leader when the follower couldn't be chained.
1638                 * Otherwise, they would be in the wrong order.
1639                 */
1640                chainable = NULL;
1641             }
1642 
1643             chainable_ace = can_chain_next ? cmd_buffer->gang.cs : NULL;
1644             submit_ace = true;
1645          }
1646 
1647          queue->device->ws->cs_unchain(cmd_buffer->cs);
1648          if (!chainable || !queue->device->ws->cs_chain(chainable, cmd_buffer->cs, queue->state.uses_shadow_regs)) {
1649             /* don't submit empty command buffers to the kernel. */
1650             if ((radv_queue_ring(queue) != AMD_IP_VCN_ENC && radv_queue_ring(queue) != AMD_IP_UVD) ||
1651                 cmd_buffer->cs->cdw != 0)
1652                cs_array[num_submitted_cs++] = cmd_buffer->cs;
1653          }
1654 
1655          chainable = can_chain_next ? cmd_buffer->cs : NULL;
1656       }
1657 
1658       submit.cs_count = num_submitted_cs;
1659       submit.initial_preamble_count = submit_ace ? num_initial_preambles : num_1q_initial_preambles;
1660       submit.continue_preamble_count = submit_ace ? num_continue_preambles : num_1q_continue_preambles;
1661       submit.postamble_count = submit_ace ? num_postambles : num_1q_postambles;
1662 
1663       result = queue->device->ws->cs_submit(ctx, &submit, j == 0 ? wait_count : 0, waits,
1664                                             last_submit ? submission->signal_count : 0, submission->signals);
1665 
1666       if (result != VK_SUCCESS)
1667          goto fail;
1668 
1669       if (radv_device_fault_detection_enabled(queue->device)) {
1670          radv_check_gpu_hangs(queue, &submit);
1671       }
1672 
1673       if (queue->device->tma_bo) {
1674          radv_check_trap_handler(queue);
1675       }
1676 
1677       initial_preambles[0] = queue->state.initial_preamble_cs;
1678       initial_preambles[1] = !use_ace ? NULL : queue->follower_state->initial_preamble_cs;
1679    }
1680 
1681    queue->last_shader_upload_seq = MAX2(queue->last_shader_upload_seq, shader_upload_seq);
1682 
1683    radv_dump_printf_data(queue->device);
1684 
1685 fail:
1686    free(cs_array);
1687    if (waits != submission->waits)
1688       free(waits);
1689    if (radv_device_fault_detection_enabled(queue->device))
1690       simple_mtx_unlock(&queue->device->trace_mtx);
1691 
1692    return result;
1693 }
1694 
1695 static void
radv_report_gpuvm_fault(struct radv_device * device)1696 radv_report_gpuvm_fault(struct radv_device *device)
1697 {
1698    struct radv_winsys_gpuvm_fault_info fault_info = {0};
1699 
1700    if (!radv_vm_fault_occurred(device, &fault_info))
1701       return;
1702 
1703    fprintf(stderr, "radv: GPUVM fault detected at address 0x%08" PRIx64 ".\n", fault_info.addr);
1704    ac_print_gpuvm_fault_status(stderr, device->physical_device->rad_info.gfx_level, fault_info.status);
1705 }
1706 
1707 static VkResult
radv_queue_sparse_submit(struct vk_queue * vqueue,struct vk_queue_submit * submission)1708 radv_queue_sparse_submit(struct vk_queue *vqueue, struct vk_queue_submit *submission)
1709 {
1710    struct radv_queue *queue = (struct radv_queue *)vqueue;
1711    struct radv_device *device = queue->device;
1712    VkResult result;
1713 
1714    result = radv_queue_submit_bind_sparse_memory(device, submission);
1715    if (result != VK_SUCCESS)
1716       goto fail;
1717 
1718    /* We do a CPU wait here, in part to avoid more winsys mechanisms. In the likely kernel explicit
1719     * sync mechanism, we'd need to do a CPU wait anyway. Haven't seen this be a perf issue yet, but
1720     * we have to make sure the queue always has its submission thread enabled. */
1721    result = vk_sync_wait_many(&device->vk, submission->wait_count, submission->waits, 0, UINT64_MAX);
1722    if (result != VK_SUCCESS)
1723       goto fail;
1724 
1725    /* Ignore all the commandbuffers. They're necessarily empty anyway. */
1726 
1727    for (unsigned i = 0; i < submission->signal_count; ++i) {
1728       result = vk_sync_signal(&device->vk, submission->signals[i].sync, submission->signals[i].signal_value);
1729       if (result != VK_SUCCESS)
1730          goto fail;
1731    }
1732 
1733 fail:
1734    if (result != VK_SUCCESS) {
1735       /* When something bad happened during the submission, such as
1736        * an out of memory issue, it might be hard to recover from
1737        * this inconsistent state. To avoid this sort of problem, we
1738        * assume that we are in a really bad situation and return
1739        * VK_ERROR_DEVICE_LOST to ensure the clients do not attempt
1740        * to submit the same job again to this device.
1741        */
1742       radv_report_gpuvm_fault(queue->device);
1743       result = vk_device_set_lost(&queue->device->vk, "vkQueueSubmit() failed");
1744    }
1745    return result;
1746 }
1747 
1748 static VkResult
radv_queue_submit(struct vk_queue * vqueue,struct vk_queue_submit * submission)1749 radv_queue_submit(struct vk_queue *vqueue, struct vk_queue_submit *submission)
1750 {
1751    struct radv_queue *queue = (struct radv_queue *)vqueue;
1752    VkResult result;
1753 
1754    if (queue->device->instance->drirc.legacy_sparse_binding) {
1755       result = radv_queue_submit_bind_sparse_memory(queue->device, submission);
1756       if (result != VK_SUCCESS)
1757          goto fail;
1758    } else {
1759       assert(!submission->buffer_bind_count && !submission->image_bind_count && !submission->image_opaque_bind_count);
1760    }
1761 
1762    if (!submission->command_buffer_count && !submission->wait_count && !submission->signal_count)
1763       return VK_SUCCESS;
1764 
1765    if (!submission->command_buffer_count) {
1766       result = radv_queue_submit_empty(queue, submission);
1767    } else {
1768       result = radv_queue_submit_normal(queue, submission);
1769    }
1770 
1771 fail:
1772    if (result != VK_SUCCESS) {
1773       /* When something bad happened during the submission, such as
1774        * an out of memory issue, it might be hard to recover from
1775        * this inconsistent state. To avoid this sort of problem, we
1776        * assume that we are in a really bad situation and return
1777        * VK_ERROR_DEVICE_LOST to ensure the clients do not attempt
1778        * to submit the same job again to this device.
1779        */
1780       radv_report_gpuvm_fault(queue->device);
1781       result = vk_device_set_lost(&queue->device->vk, "vkQueueSubmit() failed");
1782    }
1783    return result;
1784 }
1785 
1786 bool
radv_queue_internal_submit(struct radv_queue * queue,struct radeon_cmdbuf * cs)1787 radv_queue_internal_submit(struct radv_queue *queue, struct radeon_cmdbuf *cs)
1788 {
1789    struct radeon_winsys_ctx *ctx = queue->hw_ctx;
1790    struct radv_winsys_submit_info submit = {
1791       .ip_type = radv_queue_ring(queue),
1792       .queue_index = queue->vk.index_in_family,
1793       .cs_array = &cs,
1794       .cs_count = 1,
1795    };
1796 
1797    VkResult result = queue->device->ws->cs_submit(ctx, &submit, 0, NULL, 0, NULL);
1798    if (result != VK_SUCCESS)
1799       return false;
1800 
1801    return true;
1802 }
1803 
1804 int
radv_queue_init(struct radv_device * device,struct radv_queue * queue,int idx,const VkDeviceQueueCreateInfo * create_info,const VkDeviceQueueGlobalPriorityCreateInfoKHR * global_priority)1805 radv_queue_init(struct radv_device *device, struct radv_queue *queue, int idx,
1806                 const VkDeviceQueueCreateInfo *create_info,
1807                 const VkDeviceQueueGlobalPriorityCreateInfoKHR *global_priority)
1808 {
1809    queue->device = device;
1810    queue->priority = radv_get_queue_global_priority(global_priority);
1811    queue->hw_ctx = device->hw_ctx[queue->priority];
1812    queue->state.qf = vk_queue_to_radv(device->physical_device, create_info->queueFamilyIndex);
1813    queue->gang_sem_bo = NULL;
1814 
1815    VkResult result = vk_queue_init(&queue->vk, &device->vk, create_info, idx);
1816    if (result != VK_SUCCESS)
1817       return result;
1818 
1819    queue->state.uses_shadow_regs = device->uses_shadow_regs && queue->state.qf == RADV_QUEUE_GENERAL;
1820    if (queue->state.uses_shadow_regs) {
1821       result = radv_create_shadow_regs_preamble(device, &queue->state);
1822       if (result != VK_SUCCESS)
1823          goto fail;
1824       result = radv_init_shadowed_regs_buffer_state(device, queue);
1825       if (result != VK_SUCCESS)
1826          goto fail;
1827    }
1828 
1829    if (queue->state.qf == RADV_QUEUE_SPARSE) {
1830       queue->vk.driver_submit = radv_queue_sparse_submit;
1831       vk_queue_enable_submit_thread(&queue->vk);
1832    } else {
1833       queue->vk.driver_submit = radv_queue_submit;
1834    }
1835    return VK_SUCCESS;
1836 fail:
1837    vk_queue_finish(&queue->vk);
1838    return result;
1839 }
1840 
1841 static void
radv_queue_state_finish(struct radv_queue_state * queue,struct radv_device * device)1842 radv_queue_state_finish(struct radv_queue_state *queue, struct radv_device *device)
1843 {
1844    radv_destroy_shadow_regs_preamble(queue, device->ws);
1845    if (queue->initial_full_flush_preamble_cs)
1846       device->ws->cs_destroy(queue->initial_full_flush_preamble_cs);
1847    if (queue->initial_preamble_cs)
1848       device->ws->cs_destroy(queue->initial_preamble_cs);
1849    if (queue->continue_preamble_cs)
1850       device->ws->cs_destroy(queue->continue_preamble_cs);
1851    if (queue->gang_wait_preamble_cs)
1852       device->ws->cs_destroy(queue->gang_wait_preamble_cs);
1853    if (queue->gang_wait_postamble_cs)
1854       device->ws->cs_destroy(queue->gang_wait_postamble_cs);
1855    if (queue->descriptor_bo)
1856       device->ws->buffer_destroy(device->ws, queue->descriptor_bo);
1857    if (queue->scratch_bo) {
1858       device->ws->buffer_destroy(device->ws, queue->scratch_bo);
1859       radv_rmv_log_command_buffer_bo_destroy(device, queue->scratch_bo);
1860    }
1861    if (queue->esgs_ring_bo) {
1862       radv_rmv_log_command_buffer_bo_destroy(device, queue->esgs_ring_bo);
1863       device->ws->buffer_destroy(device->ws, queue->esgs_ring_bo);
1864    }
1865    if (queue->gsvs_ring_bo) {
1866       radv_rmv_log_command_buffer_bo_destroy(device, queue->gsvs_ring_bo);
1867       device->ws->buffer_destroy(device->ws, queue->gsvs_ring_bo);
1868    }
1869    if (queue->tess_rings_bo) {
1870       radv_rmv_log_command_buffer_bo_destroy(device, queue->tess_rings_bo);
1871       device->ws->buffer_destroy(device->ws, queue->tess_rings_bo);
1872    }
1873    if (queue->task_rings_bo) {
1874       radv_rmv_log_command_buffer_bo_destroy(device, queue->task_rings_bo);
1875       device->ws->buffer_destroy(device->ws, queue->task_rings_bo);
1876    }
1877    if (queue->mesh_scratch_ring_bo) {
1878       radv_rmv_log_command_buffer_bo_destroy(device, queue->mesh_scratch_ring_bo);
1879       device->ws->buffer_destroy(device->ws, queue->mesh_scratch_ring_bo);
1880    }
1881    if (queue->attr_ring_bo) {
1882       radv_rmv_log_command_buffer_bo_destroy(device, queue->attr_ring_bo);
1883       device->ws->buffer_destroy(device->ws, queue->attr_ring_bo);
1884    }
1885    if (queue->gds_bo) {
1886       device->ws->buffer_make_resident(device->ws, queue->gds_bo, false);
1887       device->ws->buffer_destroy(device->ws, queue->gds_bo);
1888    }
1889    if (queue->gds_oa_bo) {
1890       device->ws->buffer_make_resident(device->ws, queue->gds_oa_bo, false);
1891       device->ws->buffer_destroy(device->ws, queue->gds_oa_bo);
1892    }
1893    if (queue->compute_scratch_bo) {
1894       radv_rmv_log_command_buffer_bo_destroy(device, queue->compute_scratch_bo);
1895       device->ws->buffer_destroy(device->ws, queue->compute_scratch_bo);
1896    }
1897 }
1898 
1899 void
radv_queue_finish(struct radv_queue * queue)1900 radv_queue_finish(struct radv_queue *queue)
1901 {
1902    if (queue->follower_state) {
1903       /* Prevent double free */
1904       queue->follower_state->task_rings_bo = NULL;
1905 
1906       /* Clean up the internal ACE queue state. */
1907       radv_queue_state_finish(queue->follower_state, queue->device);
1908       free(queue->follower_state);
1909    }
1910 
1911    if (queue->gang_sem_bo)
1912       queue->device->ws->buffer_destroy(queue->device->ws, queue->gang_sem_bo);
1913 
1914    radv_queue_state_finish(&queue->state, queue->device);
1915    vk_queue_finish(&queue->vk);
1916 }
1917