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
, int fd
)
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;
103 drmDevicePtr devinfo
;
106 r
= drmGetDevice(fd
, &devinfo
);
108 fprintf(stderr
, "amdgpu: drmGetDevice failed.\n");
111 ws
->info
.pci_domain
= devinfo
->businfo
.pci
->domain
;
112 ws
->info
.pci_bus
= devinfo
->businfo
.pci
->bus
;
113 ws
->info
.pci_dev
= devinfo
->businfo
.pci
->dev
;
114 ws
->info
.pci_func
= devinfo
->businfo
.pci
->func
;
115 drmFreeDevice(&devinfo
);
117 /* Query hardware and driver information. */
118 r
= amdgpu_query_gpu_info(ws
->dev
, &ws
->amdinfo
);
120 fprintf(stderr
, "amdgpu: amdgpu_query_gpu_info failed.\n");
124 r
= amdgpu_query_buffer_size_alignment(ws
->dev
, &alignment_info
);
126 fprintf(stderr
, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
130 r
= amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_VRAM
, 0, &vram
);
132 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
136 r
= amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_GTT
, 0, >t
);
138 fprintf(stderr
, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
142 r
= amdgpu_query_hw_ip_info(ws
->dev
, AMDGPU_HW_IP_DMA
, 0, &dma
);
144 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
148 r
= amdgpu_query_hw_ip_info(ws
->dev
, AMDGPU_HW_IP_UVD
, 0, &uvd
);
150 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(uvd) failed.\n");
154 r
= amdgpu_query_hw_ip_info(ws
->dev
, AMDGPU_HW_IP_VCE
, 0, &vce
);
156 fprintf(stderr
, "amdgpu: amdgpu_query_hw_ip_info(vce) failed.\n");
160 r
= amdgpu_query_firmware_version(ws
->dev
, AMDGPU_INFO_FW_VCE
, 0, 0,
161 &vce_version
, &vce_feature
);
163 fprintf(stderr
, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
167 /* Set chip identification. */
168 ws
->info
.pci_id
= ws
->amdinfo
.asic_id
; /* TODO: is this correct? */
169 ws
->info
.vce_harvest_config
= ws
->amdinfo
.vce_harvest_config
;
171 switch (ws
->info
.pci_id
) {
172 #define CHIPSET(pci_id, name, cfamily) case pci_id: ws->info.family = CHIP_##cfamily; break;
173 #include "pci_ids/radeonsi_pci_ids.h"
177 fprintf(stderr
, "amdgpu: Invalid PCI ID.\n");
181 if (ws
->info
.family
>= CHIP_TONGA
)
182 ws
->info
.chip_class
= VI
;
183 else if (ws
->info
.family
>= CHIP_BONAIRE
)
184 ws
->info
.chip_class
= CIK
;
186 fprintf(stderr
, "amdgpu: Unknown family.\n");
190 /* LLVM 3.6.1 is required for VI. */
191 if (ws
->info
.chip_class
>= VI
&&
192 HAVE_LLVM
== 0x0306 && MESA_LLVM_VERSION_PATCH
< 1) {
193 fprintf(stderr
, "amdgpu: LLVM 3.6.1 is required, got LLVM %i.%i.%i\n",
194 HAVE_LLVM
>> 8, HAVE_LLVM
& 255, MESA_LLVM_VERSION_PATCH
);
198 /* family and rev_id are for addrlib */
199 switch (ws
->info
.family
) {
201 ws
->family
= FAMILY_CI
;
202 ws
->rev_id
= CI_BONAIRE_M_A0
;
205 ws
->family
= FAMILY_KV
;
206 ws
->rev_id
= KV_SPECTRE_A0
;
209 ws
->family
= FAMILY_KV
;
210 ws
->rev_id
= KB_KALINDI_A0
;
213 ws
->family
= FAMILY_CI
;
214 ws
->rev_id
= CI_HAWAII_P_A0
;
217 ws
->family
= FAMILY_KV
;
218 ws
->rev_id
= ML_GODAVARI_A0
;
221 ws
->family
= FAMILY_VI
;
222 ws
->rev_id
= VI_TONGA_P_A0
;
225 ws
->family
= FAMILY_VI
;
226 ws
->rev_id
= VI_ICELAND_M_A0
;
229 ws
->family
= FAMILY_CZ
;
230 ws
->rev_id
= CARRIZO_A0
;
233 ws
->family
= FAMILY_CZ
;
234 ws
->rev_id
= STONEY_A0
;
237 ws
->family
= FAMILY_VI
;
238 ws
->rev_id
= VI_FIJI_P_A0
;
241 fprintf(stderr
, "amdgpu: Unknown family.\n");
245 ws
->addrlib
= amdgpu_addr_create(ws
);
247 fprintf(stderr
, "amdgpu: Cannot create addrlib.\n");
251 /* Set hardware information. */
252 ws
->info
.gart_size
= gtt
.heap_size
;
253 ws
->info
.vram_size
= vram
.heap_size
;
254 /* convert the shader clock from KHz to MHz */
255 ws
->info
.max_shader_clock
= ws
->amdinfo
.max_engine_clk
/ 1000;
256 ws
->info
.max_se
= ws
->amdinfo
.num_shader_engines
;
257 ws
->info
.max_sh_per_se
= ws
->amdinfo
.num_shader_arrays_per_engine
;
258 ws
->info
.has_uvd
= uvd
.available_rings
!= 0;
259 ws
->info
.vce_fw_version
=
260 vce
.available_rings
? vce_version
: 0;
261 ws
->info
.has_userptr
= TRUE
;
262 ws
->info
.num_render_backends
= ws
->amdinfo
.rb_pipes
;
263 ws
->info
.clock_crystal_freq
= ws
->amdinfo
.gpu_counter_freq
;
264 ws
->info
.num_tile_pipes
= cik_get_num_tile_pipes(&ws
->amdinfo
);
265 ws
->info
.pipe_interleave_bytes
= 256 << ((ws
->amdinfo
.gb_addr_cfg
>> 4) & 0x7);
266 ws
->info
.has_virtual_memory
= TRUE
;
267 ws
->info
.has_sdma
= dma
.available_rings
!= 0;
269 /* Get the number of good compute units. */
270 ws
->info
.num_good_compute_units
= 0;
271 for (i
= 0; i
< ws
->info
.max_se
; i
++)
272 for (j
= 0; j
< ws
->info
.max_sh_per_se
; j
++)
273 ws
->info
.num_good_compute_units
+=
274 util_bitcount(ws
->amdinfo
.cu_bitmap
[i
][j
]);
276 memcpy(ws
->info
.si_tile_mode_array
, ws
->amdinfo
.gb_tile_mode
,
277 sizeof(ws
->amdinfo
.gb_tile_mode
));
278 ws
->info
.si_tile_mode_array_valid
= TRUE
;
279 ws
->info
.enabled_rb_mask
= ws
->amdinfo
.enabled_rb_pipes_mask
;
281 memcpy(ws
->info
.cik_macrotile_mode_array
, ws
->amdinfo
.gb_macro_tile_mode
,
282 sizeof(ws
->amdinfo
.gb_macro_tile_mode
));
283 ws
->info
.cik_macrotile_mode_array_valid
= TRUE
;
285 ws
->gart_page_size
= alignment_info
.size_remote
;
291 AddrDestroy(ws
->addrlib
);
292 amdgpu_device_deinitialize(ws
->dev
);
297 static void amdgpu_winsys_destroy(struct radeon_winsys
*rws
)
299 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
301 pipe_mutex_destroy(ws
->bo_fence_lock
);
302 pb_cache_deinit(&ws
->bo_cache
);
303 pipe_mutex_destroy(ws
->global_bo_list_lock
);
304 AddrDestroy(ws
->addrlib
);
305 amdgpu_device_deinitialize(ws
->dev
);
309 static void amdgpu_winsys_query_info(struct radeon_winsys
*rws
,
310 struct radeon_info
*info
)
312 *info
= ((struct amdgpu_winsys
*)rws
)->info
;
315 static boolean
amdgpu_cs_request_feature(struct radeon_winsys_cs
*rcs
,
316 enum radeon_feature_id fid
,
322 static uint64_t amdgpu_query_value(struct radeon_winsys
*rws
,
323 enum radeon_value_id value
)
325 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
326 struct amdgpu_heap_info heap
;
330 case RADEON_REQUESTED_VRAM_MEMORY
:
331 return ws
->allocated_vram
;
332 case RADEON_REQUESTED_GTT_MEMORY
:
333 return ws
->allocated_gtt
;
334 case RADEON_BUFFER_WAIT_TIME_NS
:
335 return ws
->buffer_wait_time
;
336 case RADEON_TIMESTAMP
:
337 amdgpu_query_info(ws
->dev
, AMDGPU_INFO_TIMESTAMP
, 8, &retval
);
339 case RADEON_NUM_CS_FLUSHES
:
340 return ws
->num_cs_flushes
;
341 case RADEON_NUM_BYTES_MOVED
:
342 amdgpu_query_info(ws
->dev
, AMDGPU_INFO_NUM_BYTES_MOVED
, 8, &retval
);
344 case RADEON_VRAM_USAGE
:
345 amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_VRAM
, 0, &heap
);
346 return heap
.heap_usage
;
347 case RADEON_GTT_USAGE
:
348 amdgpu_query_heap_info(ws
->dev
, AMDGPU_GEM_DOMAIN_GTT
, 0, &heap
);
349 return heap
.heap_usage
;
350 case RADEON_GPU_TEMPERATURE
:
351 case RADEON_CURRENT_SCLK
:
352 case RADEON_CURRENT_MCLK
:
354 case RADEON_GPU_RESET_COUNTER
:
361 static bool amdgpu_read_registers(struct radeon_winsys
*rws
,
363 unsigned num_registers
, uint32_t *out
)
365 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
367 return amdgpu_read_mm_registers(ws
->dev
, reg_offset
/ 4, num_registers
,
368 0xffffffff, 0, out
) == 0;
371 static unsigned hash_dev(void *key
)
373 #if defined(PIPE_ARCH_X86_64)
374 return pointer_to_intptr(key
) ^ (pointer_to_intptr(key
) >> 32);
376 return pointer_to_intptr(key
);
380 static int compare_dev(void *key1
, void *key2
)
385 static bool amdgpu_winsys_unref(struct radeon_winsys
*rws
)
387 struct amdgpu_winsys
*ws
= (struct amdgpu_winsys
*)rws
;
390 /* When the reference counter drops to zero, remove the device pointer
392 * This must happen while the mutex is locked, so that
393 * amdgpu_winsys_create in another thread doesn't get the winsys
394 * from the table when the counter drops to 0. */
395 pipe_mutex_lock(dev_tab_mutex
);
397 destroy
= pipe_reference(&ws
->reference
, NULL
);
398 if (destroy
&& dev_tab
)
399 util_hash_table_remove(dev_tab
, ws
->dev
);
401 pipe_mutex_unlock(dev_tab_mutex
);
405 PUBLIC
struct radeon_winsys
*
406 amdgpu_winsys_create(int fd
, radeon_screen_create_t screen_create
)
408 struct amdgpu_winsys
*ws
;
409 drmVersionPtr version
= drmGetVersion(fd
);
410 amdgpu_device_handle dev
;
411 uint32_t drm_major
, drm_minor
, r
;
413 /* The DRM driver version of amdgpu is 3.x.x. */
414 if (version
->version_major
!= 3) {
415 drmFreeVersion(version
);
418 drmFreeVersion(version
);
420 /* Look up the winsys from the dev table. */
421 pipe_mutex_lock(dev_tab_mutex
);
423 dev_tab
= util_hash_table_create(hash_dev
, compare_dev
);
425 /* Initialize the amdgpu device. This should always return the same pointer
426 * for the same fd. */
427 r
= amdgpu_device_initialize(fd
, &drm_major
, &drm_minor
, &dev
);
429 pipe_mutex_unlock(dev_tab_mutex
);
430 fprintf(stderr
, "amdgpu: amdgpu_device_initialize failed.\n");
434 /* Lookup a winsys if we have already created one for this device. */
435 ws
= util_hash_table_get(dev_tab
, dev
);
437 pipe_reference(NULL
, &ws
->reference
);
438 pipe_mutex_unlock(dev_tab_mutex
);
442 /* Create a new winsys. */
443 ws
= CALLOC_STRUCT(amdgpu_winsys
);
445 pipe_mutex_unlock(dev_tab_mutex
);
450 ws
->info
.drm_major
= drm_major
;
451 ws
->info
.drm_minor
= drm_minor
;
453 if (!do_winsys_init(ws
, fd
))
456 /* Create managers. */
457 pb_cache_init(&ws
->bo_cache
, 500000, 2.0f
, 0,
458 (ws
->info
.vram_size
+ ws
->info
.gart_size
) / 8,
459 amdgpu_bo_destroy
, amdgpu_bo_can_reclaim
);
462 pipe_reference_init(&ws
->reference
, 1);
465 ws
->base
.unref
= amdgpu_winsys_unref
;
466 ws
->base
.destroy
= amdgpu_winsys_destroy
;
467 ws
->base
.query_info
= amdgpu_winsys_query_info
;
468 ws
->base
.cs_request_feature
= amdgpu_cs_request_feature
;
469 ws
->base
.query_value
= amdgpu_query_value
;
470 ws
->base
.read_registers
= amdgpu_read_registers
;
472 amdgpu_bo_init_functions(ws
);
473 amdgpu_cs_init_functions(ws
);
474 amdgpu_surface_init_functions(ws
);
476 LIST_INITHEAD(&ws
->global_bo_list
);
477 pipe_mutex_init(ws
->global_bo_list_lock
);
478 pipe_mutex_init(ws
->bo_fence_lock
);
480 /* Create the screen at the end. The winsys must be initialized
483 * Alternatively, we could create the screen based on "ws->gen"
484 * and link all drivers into one binary blob. */
485 ws
->base
.screen
= screen_create(&ws
->base
);
486 if (!ws
->base
.screen
) {
487 amdgpu_winsys_destroy(&ws
->base
);
488 pipe_mutex_unlock(dev_tab_mutex
);
492 util_hash_table_set(dev_tab
, dev
, ws
);
494 /* We must unlock the mutex once the winsys is fully initialized, so that
495 * other threads attempting to create the winsys from the same fd will
496 * get a fully initialized winsys and not just half-way initialized. */
497 pipe_mutex_unlock(dev_tab_mutex
);
502 pipe_mutex_unlock(dev_tab_mutex
);
503 pb_cache_deinit(&ws
->bo_cache
);