2 * Copyright © 2009 Corbin Simpson <MostAwesomeDude@gmail.com>
3 * Copyright © 2009 Joakim Sindholt <opensource@zhasha.com>
4 * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
5 * Copyright © 2015 Advanced Micro Devices, Inc.
8 * Permission is hereby granted, free of charge, to any person obtaining
9 * a copy of this software and associated documentation files (the
10 * "Software"), to deal in the Software without restriction, including
11 * without limitation the rights to use, copy, modify, merge, publish,
12 * distribute, sub license, and/or sell copies of the Software, and to
13 * permit persons to whom the Software is furnished to do so, subject to
14 * the following conditions:
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
17 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
18 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
19 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
20 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
22 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 * The above copyright notice and this permission notice (including the
26 * next paragraph) shall be included in all copies or substantial portions
31 * Marek Olšák <maraeo@gmail.com>
34 #include "amdgpu_cs.h"
35 #include "amdgpu_public.h"
37 #include "util/u_hash_table.h"
38 #include <amdgpu_drm.h>
42 #include "amdgpu_id.h"
44 #define CIK_TILE_MODE_COLOR_2D 14
46 #define CIK__GB_TILE_MODE__PIPE_CONFIG(x) (((x) >> 6) & 0x1f)
47 #define CIK__PIPE_CONFIG__ADDR_SURF_P2 0
48 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16 4
49 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16 5
50 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32 6
51 #define CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32 7
52 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16 8
53 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16 9
54 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16 10
55 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16 11
56 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16 12
57 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32 13
58 #define CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32 14
59 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16 16
60 #define CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16 17
62 static struct util_hash_table
*dev_tab
= NULL
;
63 pipe_static_mutex(dev_tab_mutex
);
65 static unsigned cik_get_num_tile_pipes(struct amdgpu_gpu_info
*info
)
67 unsigned mode2d
= info
->gb_tile_mode
[CIK_TILE_MODE_COLOR_2D
];
69 switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d
)) {
70 case CIK__PIPE_CONFIG__ADDR_SURF_P2
:
72 case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16
:
73 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16
:
74 case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32
:
75 case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32
:
77 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16
:
78 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16
:
79 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16
:
80 case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16
:
81 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16
:
82 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32
:
83 case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32
:
85 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16
:
86 case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16
:
89 fprintf(stderr
, "Invalid CIK pipe configuration, assuming P2\n");
90 assert(!"this should never occur");
95 /* Helper function to do the ioctls needed for setup and init. */
96 static boolean
do_winsys_init(struct amdgpu_winsys
*ws
)
98 struct amdgpu_buffer_size_alignments alignment_info
= {};
99 struct amdgpu_heap_info vram
, gtt
;
100 struct drm_amdgpu_info_hw_ip dma
= {}, uvd
= {}, vce
= {};
101 uint32_t vce_version
= 0, vce_feature
= 0;
104 /* Query hardware and driver information. */
105 r
= amdgpu_query_gpu_info(ws
->dev
, &ws
->amdinfo
);
107 fprintf(stderr
, "amdgpu: amdgpu_query_gpu_info failed.\n");
111 r
= amdgpu_query_buffer_size_alignment(ws
->dev
, &alignment_info
);
113 fprintf(stderr
, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
117 r
= amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_VRAM
, 0, &vram
);
119 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
123 r
= amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_GTT
, 0, >t
);
125 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
129 r
= amdgpu_query_hw_ip_info(ws
->dev
, AMDGPU_HW_IP_DMA
, 0, &dma
);
131 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
135 r
= amdgpu_query_hw_ip_info(ws
->dev
, AMDGPU_HW_IP_UVD
, 0, &uvd
);
137 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd) failed.\n");
141 r
= amdgpu_query_hw_ip_info(ws
->dev
, AMDGPU_HW_IP_VCE
, 0, &vce
);
143 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vce) failed.\n");
147 r
= amdgpu_query_firmware_version(ws
->dev
, AMDGPU_INFO_FW_VCE
, 0, 0,
148 &vce_version
, &vce_feature
);
150 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
154 /* Set chip identification. */
155 ws
->info
.pci_id
= ws
->amdinfo
.asic_id
; /* TODO: is this correct? */
156 ws
->info
.vce_harvest_config
= ws
->amdinfo
.vce_harvest_config
;
158 switch (ws
->info
.pci_id
) {
159 #define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; break;
160 #include "pci_ids/radeonsi_pci_ids.h"
164 fprintf(stderr
, "amdgpu: Invalid PCI ID.\n");
168 if (ws
->info
.family
>= CHIP_TONGA
)
169 ws
->info
.chip_class
= VI
;
170 else if (ws
->info
.family
>= CHIP_BONAIRE
)
171 ws
->info
.chip_class
= CIK
;
173 fprintf(stderr
, "amdgpu: Unknown family.\n");
177 /* LLVM 3.6 is required for VI. */
178 if (ws
->info
.chip_class
>= VI
&&
179 (HAVE_LLVM
< 0x0306 ||
180 (HAVE_LLVM
== 0x0306 && MESA_LLVM_VERSION_PATCH
< 1))) {
181 fprintf(stderr
, "amdgpu: LLVM 3.6.1 is required, got LLVM %i.%i.%i\n",
182 HAVE_LLVM
>> 8, HAVE_LLVM
& 255, MESA_LLVM_VERSION_PATCH
);
186 /* family and rev_id are for addrlib */
187 switch (ws
->info
.family
) {
189 ws
->family
= FAMILY_CI
;
190 ws
->rev_id
= CI_BONAIRE_M_A0
;
193 ws
->family
= FAMILY_KV
;
194 ws
->rev_id
= KV_SPECTRE_A0
;
197 ws
->family
= FAMILY_KV
;
198 ws
->rev_id
= KB_KALINDI_A0
;
201 ws
->family
= FAMILY_CI
;
202 ws
->rev_id
= CI_HAWAII_P_A0
;
205 ws
->family
= FAMILY_KV
;
206 ws
->rev_id
= ML_GODAVARI_A0
;
209 ws
->family
= FAMILY_VI
;
210 ws
->rev_id
= VI_TONGA_P_A0
;
213 ws
->family
= FAMILY_VI
;
214 ws
->rev_id
= VI_ICELAND_M_A0
;
217 ws
->family
= FAMILY_CZ
;
218 ws
->rev_id
= CARRIZO_A0
;
221 ws
->family
= FAMILY_CZ
;
222 ws
->rev_id
= STONEY_A0
;
225 ws
->family
= FAMILY_VI
;
226 ws
->rev_id
= VI_FIJI_P_A0
;
229 fprintf(stderr
, "amdgpu: Unknown family.\n");
233 ws
->addrlib
= amdgpu_addr_create(ws
);
235 fprintf(stderr
, "amdgpu: Cannot create addrlib.\n");
239 /* Set hardware information. */
240 ws
->info
.gart_size
= gtt
.heap_size
;
241 ws
->info
.vram_size
= vram
.heap_size
;
242 /* convert the shader clock from KHz to MHz */
243 ws
->info
.max_shader_clock
= ws
->amdinfo
.max_engine_clk
/ 1000;
244 ws
->info
.max_se
= ws
->amdinfo
.num_shader_engines
;
245 ws
->info
.max_sh_per_se
= ws
->amdinfo
.num_shader_arrays_per_engine
;
246 ws
->info
.has_uvd
= uvd
.available_rings
!= 0;
247 ws
->info
.vce_fw_version
=
248 vce
.available_rings
? vce_version
: 0;
249 ws
->info
.has_userptr
= TRUE
;
250 ws
->info
.num_render_backends
= ws
->amdinfo
.rb_pipes
;
251 ws
->info
.clock_crystal_freq
= ws
->amdinfo
.gpu_counter_freq
;
252 ws
->info
.num_tile_pipes
= cik_get_num_tile_pipes(&ws
->amdinfo
);
253 ws
->info
.pipe_interleave_bytes
= 256 << ((ws
->amdinfo
.gb_addr_cfg
>> 4) & 0x7);
254 ws
->info
.has_virtual_memory
= TRUE
;
255 ws
->info
.has_sdma
= dma
.available_rings
!= 0;
257 /* Get the number of good compute units. */
258 ws
->info
.num_good_compute_units
= 0;
259 for (i
= 0; i
< ws
->info
.max_se
; i
++)
260 for (j
= 0; j
< ws
->info
.max_sh_per_se
; j
++)
261 ws
->info
.num_good_compute_units
+=
262 util_bitcount(ws
->amdinfo
.cu_bitmap
[i
][j
]);
264 memcpy(ws
->info
.si_tile_mode_array
, ws
->amdinfo
.gb_tile_mode
,
265 sizeof(ws
->amdinfo
.gb_tile_mode
));
266 ws
->info
.si_tile_mode_array_valid
= TRUE
;
267 ws
->info
.enabled_rb_mask
= ws
->amdinfo
.enabled_rb_pipes_mask
;
269 memcpy(ws
->info
.cik_macrotile_mode_array
, ws
->amdinfo
.gb_macro_tile_mode
,
270 sizeof(ws
->amdinfo
.gb_macro_tile_mode
));
271 ws
->info
.cik_macrotile_mode_array_valid
= TRUE
;
273 ws
->gart_page_size
= alignment_info
.size_remote
;
279 AddrDestroy(ws
->addrlib
);
280 amdgpu_device_deinitialize(ws
->dev
);
285 static void amdgpu_winsys_destroy(struct radeon_winsys
*rws
)
287 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
289 pipe_mutex_destroy(ws
->bo_fence_lock
);
290 pb_cache_deinit(&ws
->bo_cache
);
291 pipe_mutex_destroy(ws
->global_bo_list_lock
);
292 AddrDestroy(ws
->addrlib
);
293 amdgpu_device_deinitialize(ws
->dev
);
297 static void amdgpu_winsys_query_info(struct radeon_winsys
*rws
,
298 struct radeon_info
*info
)
300 *info
= ((struct amdgpu_winsys
*)rws
)->info
;
303 static boolean
amdgpu_cs_request_feature(struct radeon_winsys_cs
*rcs
,
304 enum radeon_feature_id fid
,
310 static uint64_t amdgpu_query_value(struct radeon_winsys
*rws
,
311 enum radeon_value_id value
)
313 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
314 struct amdgpu_heap_info heap
;
318 case RADEON_REQUESTED_VRAM_MEMORY
:
319 return ws
->allocated_vram
;
320 case RADEON_REQUESTED_GTT_MEMORY
:
321 return ws
->allocated_gtt
;
322 case RADEON_BUFFER_WAIT_TIME_NS
:
323 return ws
->buffer_wait_time
;
324 case RADEON_TIMESTAMP
:
325 amdgpu_query_info(ws
->dev
, AMDGPU_INFO_TIMESTAMP
, 8, &retval
);
327 case RADEON_NUM_CS_FLUSHES
:
328 return ws
->num_cs_flushes
;
329 case RADEON_NUM_BYTES_MOVED
:
330 amdgpu_query_info(ws
->dev
, AMDGPU_INFO_NUM_BYTES_MOVED
, 8, &retval
);
332 case RADEON_VRAM_USAGE
:
333 amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_VRAM
, 0, &heap
);
334 return heap
.heap_usage
;
335 case RADEON_GTT_USAGE
:
336 amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_GTT
, 0, &heap
);
337 return heap
.heap_usage
;
338 case RADEON_GPU_TEMPERATURE
:
339 case RADEON_CURRENT_SCLK
:
340 case RADEON_CURRENT_MCLK
:
342 case RADEON_GPU_RESET_COUNTER
:
349 static bool amdgpu_read_registers(struct radeon_winsys
*rws
,
351 unsigned num_registers
, uint32_t *out
)
353 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
355 return amdgpu_read_mm_registers(ws
->dev
, reg_offset
/ 4, num_registers
,
356 0xffffffff, 0, out
) == 0;
359 static unsigned hash_dev(void *key
)
361 #if defined(PIPE_ARCH_X86_64)
362 return pointer_to_intptr(key
) ^ (pointer_to_intptr(key
) >> 32);
364 return pointer_to_intptr(key
);
368 static int compare_dev(void *key1
, void *key2
)
373 static bool amdgpu_winsys_unref(struct radeon_winsys
*rws
)
375 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
378 /* When the reference counter drops to zero, remove the device pointer
380 * This must happen while the mutex is locked, so that
381 * amdgpu_winsys_create in another thread doesn't get the winsys
382 * from the table when the counter drops to 0. */
383 pipe_mutex_lock(dev_tab_mutex
);
385 destroy
= pipe_reference(&ws
->reference
, NULL
);
386 if (destroy
&& dev_tab
)
387 util_hash_table_remove(dev_tab
, ws
->dev
);
389 pipe_mutex_unlock(dev_tab_mutex
);
393 PUBLIC
struct radeon_winsys
*
394 amdgpu_winsys_create(int fd
, radeon_screen_create_t screen_create
)
396 struct amdgpu_winsys
*ws
;
397 drmVersionPtr version
= drmGetVersion(fd
);
398 amdgpu_device_handle dev
;
399 uint32_t drm_major
, drm_minor
, r
;
401 /* The DRM driver version of amdgpu is 3.x.x. */
402 if (version
->version_major
!= 3) {
403 drmFreeVersion(version
);
406 drmFreeVersion(version
);
408 /* Look up the winsys from the dev table. */
409 pipe_mutex_lock(dev_tab_mutex
);
411 dev_tab
= util_hash_table_create(hash_dev
, compare_dev
);
413 /* Initialize the amdgpu device. This should always return the same pointer
414 * for the same fd. */
415 r
= amdgpu_device_initialize(fd
, &drm_major
, &drm_minor
, &dev
);
417 pipe_mutex_unlock(dev_tab_mutex
);
418 fprintf(stderr
, "amdgpu: amdgpu_device_initialize failed.\n");
422 /* Lookup a winsys if we have already created one for this device. */
423 ws
= util_hash_table_get(dev_tab
, dev
);
425 pipe_reference(NULL
, &ws
->reference
);
426 pipe_mutex_unlock(dev_tab_mutex
);
430 /* Create a new winsys. */
431 ws
= CALLOC_STRUCT(amdgpu_winsys
);
433 pipe_mutex_unlock(dev_tab_mutex
);
438 ws
->info
.drm_major
= drm_major
;
439 ws
->info
.drm_minor
= drm_minor
;
441 if (!do_winsys_init(ws
))
444 /* Create managers. */
445 pb_cache_init(&ws
->bo_cache
, 500000, 2.0f
, 0,
446 (ws
->info
.vram_size
+ ws
->info
.gart_size
) / 8,
447 amdgpu_bo_destroy
, amdgpu_bo_can_reclaim
);
450 pipe_reference_init(&ws
->reference
, 1);
453 ws
->base
.unref
= amdgpu_winsys_unref
;
454 ws
->base
.destroy
= amdgpu_winsys_destroy
;
455 ws
->base
.query_info
= amdgpu_winsys_query_info
;
456 ws
->base
.cs_request_feature
= amdgpu_cs_request_feature
;
457 ws
->base
.query_value
= amdgpu_query_value
;
458 ws
->base
.read_registers
= amdgpu_read_registers
;
460 amdgpu_bo_init_functions(ws
);
461 amdgpu_cs_init_functions(ws
);
462 amdgpu_surface_init_functions(ws
);
464 LIST_INITHEAD(&ws
->global_bo_list
);
465 pipe_mutex_init(ws
->global_bo_list_lock
);
466 pipe_mutex_init(ws
->bo_fence_lock
);
468 /* Create the screen at the end. The winsys must be initialized
471 * Alternatively, we could create the screen based on "ws->gen"
472 * and link all drivers into one binary blob. */
473 ws
->base
.screen
= screen_create(&ws
->base
);
474 if (!ws
->base
.screen
) {
475 amdgpu_winsys_destroy(&ws
->base
);
476 pipe_mutex_unlock(dev_tab_mutex
);
480 util_hash_table_set(dev_tab
, dev
, ws
);
482 /* We must unlock the mutex once the winsys is fully initialized, so that
483 * other threads attempting to create the winsys from the same fd will
484 * get a fully initialized winsys and not just half-way initialized. */
485 pipe_mutex_unlock(dev_tab_mutex
);
490 pipe_mutex_unlock(dev_tab_mutex
);
491 pb_cache_deinit(&ws
->bo_cache
);