* next paragraph) shall be included in all copies or substantial portions
* of the Software.
*/
-/*
- * Authors:
- * Marek Olšák <maraeo@gmail.com>
- */
#include "amdgpu_cs.h"
#include "amdgpu_public.h"
+#include "util/u_cpu_detect.h"
#include "util/u_hash_table.h"
+#include "util/hash_table.h"
+#include "util/xmlconfig.h"
#include <amdgpu_drm.h>
#include <xf86drm.h>
#include <stdio.h>
#include <sys/stat.h>
-#include "amd/common/amdgpu_id.h"
+#include <fcntl.h>
+#include "amd/common/ac_llvm_util.h"
#include "amd/common/sid.h"
-#include "amd/common/gfx9d.h"
#ifndef AMDGPU_INFO_NUM_VRAM_CPU_PAGE_FAULTS
#define AMDGPU_INFO_NUM_VRAM_CPU_PAGE_FAULTS 0x1E
#endif
static struct util_hash_table *dev_tab = NULL;
-static mtx_t dev_tab_mutex = _MTX_INITIALIZER_NP;
+static simple_mtx_t dev_tab_mutex = _SIMPLE_MTX_INITIALIZER_NP;
+
+DEBUG_GET_ONCE_BOOL_OPTION(all_bos, "RADEON_ALL_BOS", false)
+
+static void handle_env_var_force_family(struct amdgpu_winsys *ws)
+{
+ const char *family = debug_get_option("SI_FORCE_FAMILY", NULL);
+ unsigned i;
+
+ if (!family)
+ return;
+
+ for (i = CHIP_TAHITI; i < CHIP_LAST; i++) {
+ if (!strcmp(family, ac_get_llvm_processor_name(i))) {
+ /* Override family and chip_class. */
+ ws->info.family = i;
+ ws->info.name = "GCN-NOOP";
+
+ if (i >= CHIP_NAVI10)
+ ws->info.chip_class = GFX10;
+ else if (i >= CHIP_VEGA10)
+ ws->info.chip_class = GFX9;
+ else if (i >= CHIP_TONGA)
+ ws->info.chip_class = GFX8;
+ else if (i >= CHIP_BONAIRE)
+ ws->info.chip_class = GFX7;
+ else
+ ws->info.chip_class = GFX6;
+
+ /* Don't submit any IBs. */
+ setenv("RADEON_NOOP", "1", 1);
+ return;
+ }
+ }
+
+ fprintf(stderr, "radeonsi: Unknown family: %s\n", family);
+ exit(1);
+}
/* Helper function to do the ioctls needed for setup and init. */
-static bool do_winsys_init(struct amdgpu_winsys *ws, int fd)
+static bool do_winsys_init(struct amdgpu_winsys *ws,
+ const struct pipe_screen_config *config,
+ int fd)
{
if (!ac_query_gpu_info(fd, ws->dev, &ws->info, &ws->amdinfo))
goto fail;
- /* LLVM 5.0 is required for GFX9. */
- if (ws->info.chip_class >= GFX9 && HAVE_LLVM < 0x0500) {
- fprintf(stderr, "amdgpu: LLVM 5.0 is required, got LLVM %i.%i\n",
- HAVE_LLVM >> 8, HAVE_LLVM & 255);
- goto fail;
- }
+ /* TODO: Enable this once the kernel handles it efficiently. */
+ if (ws->info.has_dedicated_vram)
+ ws->info.has_local_buffers = false;
+
+ handle_env_var_force_family(ws);
- ws->addrlib = amdgpu_addr_create(&ws->info, &ws->amdinfo);
+ ws->addrlib = amdgpu_addr_create(&ws->info, &ws->amdinfo, &ws->info.max_alignment);
if (!ws->addrlib) {
fprintf(stderr, "amdgpu: Cannot create addrlib.\n");
goto fail;
}
- ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL;
+ ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL ||
+ strstr(debug_get_option("AMD_DEBUG", ""), "check_vm") != NULL;
+ ws->debug_all_bos = debug_get_option_all_bos();
+ ws->reserve_vmid = strstr(debug_get_option("R600_DEBUG", ""), "reserve_vmid") != NULL ||
+ strstr(debug_get_option("AMD_DEBUG", ""), "reserve_vmid") != NULL;
+ ws->zero_all_vram_allocs = strstr(debug_get_option("R600_DEBUG", ""), "zerovram") != NULL ||
+ strstr(debug_get_option("AMD_DEBUG", ""), "zerovram") != NULL ||
+ driQueryOptionb(config->options, "radeonsi_zerovram");
return true;
static void do_winsys_deinit(struct amdgpu_winsys *ws)
{
+ if (ws->reserve_vmid)
+ amdgpu_vm_unreserve_vmid(ws->dev, 0);
+
+ if (util_queue_is_initialized(&ws->cs_queue))
+ util_queue_destroy(&ws->cs_queue);
+
+ simple_mtx_destroy(&ws->bo_fence_lock);
+ for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
+ if (ws->bo_slabs[i].groups)
+ pb_slabs_deinit(&ws->bo_slabs[i]);
+ }
+ pb_cache_deinit(&ws->bo_cache);
+ util_hash_table_destroy(ws->bo_export_table);
+ simple_mtx_destroy(&ws->global_bo_list_lock);
+ simple_mtx_destroy(&ws->bo_export_table_lock);
+
AddrDestroy(ws->addrlib);
amdgpu_device_deinitialize(ws->dev);
+ FREE(ws);
}
static void amdgpu_winsys_destroy(struct radeon_winsys *rws)
{
- struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
+ struct amdgpu_screen_winsys *sws = amdgpu_screen_winsys(rws);
+ struct amdgpu_winsys *ws = sws->aws;
+ bool destroy;
- if (util_queue_is_initialized(&ws->cs_queue))
- util_queue_destroy(&ws->cs_queue);
+ /* When the reference counter drops to zero, remove the device pointer
+ * from the table.
+ * This must happen while the mutex is locked, so that
+ * amdgpu_winsys_create in another thread doesn't get the winsys
+ * from the table when the counter drops to 0.
+ */
+ simple_mtx_lock(&dev_tab_mutex);
- mtx_destroy(&ws->bo_fence_lock);
- pb_slabs_deinit(&ws->bo_slabs);
- pb_cache_deinit(&ws->bo_cache);
- mtx_destroy(&ws->global_bo_list_lock);
- do_winsys_deinit(ws);
+ destroy = pipe_reference(&ws->reference, NULL);
+ if (destroy && dev_tab) {
+ util_hash_table_remove(dev_tab, ws->dev);
+ if (util_hash_table_count(dev_tab) == 0) {
+ util_hash_table_destroy(dev_tab);
+ dev_tab = NULL;
+ }
+ }
+
+ simple_mtx_unlock(&dev_tab_mutex);
+
+ if (destroy)
+ do_winsys_deinit(ws);
+
+ close(sws->fd);
FREE(rws);
}
static void amdgpu_winsys_query_info(struct radeon_winsys *rws,
struct radeon_info *info)
{
- *info = ((struct amdgpu_winsys *)rws)->info;
+ *info = amdgpu_winsys(rws)->info;
}
-static bool amdgpu_cs_request_feature(struct radeon_winsys_cs *rcs,
+static bool amdgpu_cs_request_feature(struct radeon_cmdbuf *rcs,
enum radeon_feature_id fid,
bool enable)
{
static uint64_t amdgpu_query_value(struct radeon_winsys *rws,
enum radeon_value_id value)
{
- struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
+ struct amdgpu_winsys *ws = amdgpu_winsys(rws);
struct amdgpu_heap_info heap;
uint64_t retval = 0;
return ws->num_gfx_IBs;
case RADEON_NUM_SDMA_IBS:
return ws->num_sdma_IBs;
+ case RADEON_GFX_BO_LIST_COUNTER:
+ return ws->gfx_bo_list_counter;
+ case RADEON_GFX_IB_SIZE_COUNTER:
+ return ws->gfx_ib_size_counter;
case RADEON_NUM_BYTES_MOVED:
amdgpu_query_info(ws->dev, AMDGPU_INFO_NUM_BYTES_MOVED, 8, &retval);
return retval;
case RADEON_CURRENT_MCLK:
amdgpu_query_sensor_info(ws->dev, AMDGPU_INFO_SENSOR_GFX_MCLK, 4, &retval);
return retval;
- case RADEON_GPU_RESET_COUNTER:
- assert(0);
- return 0;
case RADEON_CS_THREAD_TIME:
return util_queue_get_thread_time_nano(&ws->cs_queue, 0);
}
unsigned reg_offset,
unsigned num_registers, uint32_t *out)
{
- struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
+ struct amdgpu_winsys *ws = amdgpu_winsys(rws);
return amdgpu_read_mm_registers(ws->dev, reg_offset / 4, num_registers,
0xffffffff, 0, out) == 0;
}
-static unsigned hash_dev(void *key)
+static unsigned hash_pointer(void *key)
{
-#if defined(PIPE_ARCH_X86_64)
- return pointer_to_intptr(key) ^ (pointer_to_intptr(key) >> 32);
-#else
- return pointer_to_intptr(key);
-#endif
+ return _mesa_hash_pointer(key);
}
-static int compare_dev(void *key1, void *key2)
+static int compare_pointers(void *key1, void *key2)
{
return key1 != key2;
}
static bool amdgpu_winsys_unref(struct radeon_winsys *rws)
{
- struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
- bool destroy;
-
- /* When the reference counter drops to zero, remove the device pointer
- * from the table.
- * This must happen while the mutex is locked, so that
- * amdgpu_winsys_create in another thread doesn't get the winsys
- * from the table when the counter drops to 0. */
- mtx_lock(&dev_tab_mutex);
+ /* radeon_winsys corresponds to amdgpu_screen_winsys, which is never
+ * referenced multiple times, so amdgpu_winsys_destroy always needs to be
+ * called. It handles reference counting for amdgpu_winsys.
+ */
+ return true;
+}
- destroy = pipe_reference(&ws->reference, NULL);
- if (destroy && dev_tab)
- util_hash_table_remove(dev_tab, ws->dev);
+static void amdgpu_pin_threads_to_L3_cache(struct radeon_winsys *rws,
+ unsigned cache)
+{
+ struct amdgpu_winsys *ws = amdgpu_winsys(rws);
- mtx_unlock(&dev_tab_mutex);
- return destroy;
+ util_pin_thread_to_L3(ws->cs_queue.threads[0], cache,
+ util_cpu_caps.cores_per_L3);
}
PUBLIC struct radeon_winsys *
-amdgpu_winsys_create(int fd, radeon_screen_create_t screen_create)
+amdgpu_winsys_create(int fd, const struct pipe_screen_config *config,
+ radeon_screen_create_t screen_create)
{
- struct amdgpu_winsys *ws;
- drmVersionPtr version = drmGetVersion(fd);
+ struct amdgpu_screen_winsys *ws;
+ struct amdgpu_winsys *aws;
amdgpu_device_handle dev;
uint32_t drm_major, drm_minor, r;
- /* The DRM driver version of amdgpu is 3.x.x. */
- if (version->version_major != 3) {
- drmFreeVersion(version);
+ ws = CALLOC_STRUCT(amdgpu_screen_winsys);
+ if (!ws)
return NULL;
- }
- drmFreeVersion(version);
+
+ ws->fd = fcntl(fd, F_DUPFD_CLOEXEC, 0);
/* Look up the winsys from the dev table. */
- mtx_lock(&dev_tab_mutex);
+ simple_mtx_lock(&dev_tab_mutex);
if (!dev_tab)
- dev_tab = util_hash_table_create(hash_dev, compare_dev);
+ dev_tab = util_hash_table_create(hash_pointer, compare_pointers);
/* Initialize the amdgpu device. This should always return the same pointer
* for the same fd. */
r = amdgpu_device_initialize(fd, &drm_major, &drm_minor, &dev);
if (r) {
- mtx_unlock(&dev_tab_mutex);
fprintf(stderr, "amdgpu: amdgpu_device_initialize failed.\n");
- return NULL;
+ goto fail;
}
/* Lookup a winsys if we have already created one for this device. */
- ws = util_hash_table_get(dev_tab, dev);
- if (ws) {
- pipe_reference(NULL, &ws->reference);
- mtx_unlock(&dev_tab_mutex);
- return &ws->base;
+ aws = util_hash_table_get(dev_tab, dev);
+ if (aws) {
+ pipe_reference(NULL, &aws->reference);
+ simple_mtx_unlock(&dev_tab_mutex);
+
+ /* Release the device handle, because we don't need it anymore.
+ * This function is returning an existing winsys instance, which
+ * has its own device handle.
+ */
+ amdgpu_device_deinitialize(dev);
+ } else {
+ /* Create a new winsys. */
+ aws = CALLOC_STRUCT(amdgpu_winsys);
+ if (!aws)
+ goto fail;
+
+ aws->dev = dev;
+ aws->info.drm_major = drm_major;
+ aws->info.drm_minor = drm_minor;
+
+ if (!do_winsys_init(aws, config, fd))
+ goto fail_alloc;
+
+ /* Create managers. */
+ pb_cache_init(&aws->bo_cache, RADEON_MAX_CACHED_HEAPS,
+ 500000, aws->check_vm ? 1.0f : 2.0f, 0,
+ (aws->info.vram_size + aws->info.gart_size) / 8,
+ amdgpu_bo_destroy, amdgpu_bo_can_reclaim);
+
+ unsigned min_slab_order = 9; /* 512 bytes */
+ unsigned max_slab_order = 18; /* 256 KB - higher numbers increase memory usage */
+ unsigned num_slab_orders_per_allocator = (max_slab_order - min_slab_order) /
+ NUM_SLAB_ALLOCATORS;
+
+ /* Divide the size order range among slab managers. */
+ for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
+ unsigned min_order = min_slab_order;
+ unsigned max_order = MIN2(min_order + num_slab_orders_per_allocator,
+ max_slab_order);
+
+ if (!pb_slabs_init(&aws->bo_slabs[i],
+ min_order, max_order,
+ RADEON_MAX_SLAB_HEAPS,
+ aws,
+ amdgpu_bo_can_reclaim_slab,
+ amdgpu_bo_slab_alloc,
+ amdgpu_bo_slab_free)) {
+ amdgpu_winsys_destroy(&ws->base);
+ simple_mtx_unlock(&dev_tab_mutex);
+ return NULL;
+ }
+
+ min_slab_order = max_order + 1;
+ }
+
+ aws->info.min_alloc_size = 1 << aws->bo_slabs[0].min_order;
+
+ /* init reference */
+ pipe_reference_init(&aws->reference, 1);
+
+ LIST_INITHEAD(&aws->global_bo_list);
+ aws->bo_export_table = util_hash_table_create(hash_pointer, compare_pointers);
+
+ (void) simple_mtx_init(&aws->global_bo_list_lock, mtx_plain);
+ (void) simple_mtx_init(&aws->bo_fence_lock, mtx_plain);
+ (void) simple_mtx_init(&aws->bo_export_table_lock, mtx_plain);
+
+ if (!util_queue_init(&aws->cs_queue, "cs", 8, 1,
+ UTIL_QUEUE_INIT_RESIZE_IF_FULL)) {
+ amdgpu_winsys_destroy(&ws->base);
+ simple_mtx_unlock(&dev_tab_mutex);
+ return NULL;
+ }
+
+ util_hash_table_set(dev_tab, dev, aws);
+
+ if (aws->reserve_vmid) {
+ r = amdgpu_vm_reserve_vmid(dev, 0);
+ if (r) {
+ amdgpu_winsys_destroy(&ws->base);
+ simple_mtx_unlock(&dev_tab_mutex);
+ return NULL;
+ }
+ }
}
- /* Create a new winsys. */
- ws = CALLOC_STRUCT(amdgpu_winsys);
- if (!ws)
- goto fail;
-
- ws->dev = dev;
- ws->info.drm_major = drm_major;
- ws->info.drm_minor = drm_minor;
-
- if (!do_winsys_init(ws, fd))
- goto fail_alloc;
-
- /* Create managers. */
- pb_cache_init(&ws->bo_cache, 500000, ws->check_vm ? 1.0f : 2.0f, 0,
- (ws->info.vram_size + ws->info.gart_size) / 8,
- amdgpu_bo_destroy, amdgpu_bo_can_reclaim);
-
- if (!pb_slabs_init(&ws->bo_slabs,
- AMDGPU_SLAB_MIN_SIZE_LOG2, AMDGPU_SLAB_MAX_SIZE_LOG2,
- 12, /* number of heaps (domain/flags combinations) */
- ws,
- amdgpu_bo_can_reclaim_slab,
- amdgpu_bo_slab_alloc,
- amdgpu_bo_slab_free))
- goto fail_cache;
-
- ws->info.min_alloc_size = 1 << AMDGPU_SLAB_MIN_SIZE_LOG2;
-
- /* init reference */
- pipe_reference_init(&ws->reference, 1);
+ ws->aws = aws;
/* Set functions. */
ws->base.unref = amdgpu_winsys_unref;
ws->base.cs_request_feature = amdgpu_cs_request_feature;
ws->base.query_value = amdgpu_query_value;
ws->base.read_registers = amdgpu_read_registers;
+ ws->base.pin_threads_to_L3_cache = amdgpu_pin_threads_to_L3_cache;
amdgpu_bo_init_functions(ws);
amdgpu_cs_init_functions(ws);
amdgpu_surface_init_functions(ws);
- LIST_INITHEAD(&ws->global_bo_list);
- (void) mtx_init(&ws->global_bo_list_lock, mtx_plain);
- (void) mtx_init(&ws->bo_fence_lock, mtx_plain);
-
- if (!util_queue_init(&ws->cs_queue, "amdgpu_cs", 8, 1)) {
- amdgpu_winsys_destroy(&ws->base);
- mtx_unlock(&dev_tab_mutex);
- return NULL;
- }
-
/* Create the screen at the end. The winsys must be initialized
* completely.
*
* Alternatively, we could create the screen based on "ws->gen"
* and link all drivers into one binary blob. */
- ws->base.screen = screen_create(&ws->base);
+ ws->base.screen = screen_create(&ws->base, config);
if (!ws->base.screen) {
amdgpu_winsys_destroy(&ws->base);
- mtx_unlock(&dev_tab_mutex);
+ simple_mtx_unlock(&dev_tab_mutex);
return NULL;
}
- util_hash_table_set(dev_tab, dev, ws);
-
/* We must unlock the mutex once the winsys is fully initialized, so that
* other threads attempting to create the winsys from the same fd will
* get a fully initialized winsys and not just half-way initialized. */
- mtx_unlock(&dev_tab_mutex);
+ simple_mtx_unlock(&dev_tab_mutex);
return &ws->base;
-fail_cache:
- pb_cache_deinit(&ws->bo_cache);
- do_winsys_deinit(ws);
fail_alloc:
- FREE(ws);
+ FREE(aws);
fail:
- mtx_unlock(&dev_tab_mutex);
+ close(ws->fd);
+ FREE(ws);
+ simple_mtx_unlock(&dev_tab_mutex);
return NULL;
}