AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

radv/winsys: use an array for the global BO list instead of a list

Browse Source 
This allows to remove one 64-bit pointer from radv_amdgpu_winsys_bo.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8859>
This commit is contained in:
Samuel Pitoiset
2021-02-03 10:47:08 +01:00
parent 0b29a8a206
commit eb625b7a5f
5 changed files with 87 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
+58 -30
View File
@@ -317,6 +317,50 @@ static void radv_amdgpu_log_bo(struct radv_amdgpu_winsys_bo *bo,
u_rwlock_wrunlock(&ws->log_bo_list_lock);
}
static int radv_amdgpu_global_bo_list_add(struct radv_amdgpu_winsys_bo *bo)
{
struct radv_amdgpu_winsys *ws = bo->ws;
if (!ws->debug_all_bos)
return VK_SUCCESS;
u_rwlock_wrlock(&ws->global_bo_list.lock);
if (ws->global_bo_list.count == ws->global_bo_list.capacity) {
unsigned capacity = MAX2(4, ws->global_bo_list.capacity * 2);
void *data = realloc(ws->global_bo_list.bos,
capacity * sizeof(struct radv_amdgpu_winsys_bo *));
if (!data) {
u_rwlock_wrunlock(&ws->global_bo_list.lock);
return VK_ERROR_OUT_OF_HOST_MEMORY;
}
ws->global_bo_list.bos = (struct radv_amdgpu_winsys_bo **)data;
ws->global_bo_list.capacity = capacity;
}
ws->global_bo_list.bos[ws->global_bo_list.count++] = bo;
u_rwlock_wrunlock(&ws->global_bo_list.lock);
return VK_SUCCESS;
}
static void radv_amdgpu_global_bo_list_del(struct radv_amdgpu_winsys_bo *bo)
{
struct radv_amdgpu_winsys *ws = bo->ws;
if (!ws->debug_all_bos)
return;
u_rwlock_wrlock(&ws->global_bo_list.lock);
for(unsigned i = ws->global_bo_list.count; i-- > 0;) {
if (ws->global_bo_list.bos[i] == bo) {
ws->global_bo_list.bos[i] = ws->global_bo_list.bos[ws->global_bo_list.count - 1];
--ws->global_bo_list.count;
break;
}
}
u_rwlock_wrunlock(&ws->global_bo_list.lock);
}
static void radv_amdgpu_winsys_bo_destroy(struct radeon_winsys_bo *_bo)
{
struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
@@ -334,12 +378,7 @@ static void radv_amdgpu_winsys_bo_destroy(struct radeon_winsys_bo *_bo)
free(bo->bos);
free(bo->ranges);
} else {
if (bo->ws->debug_all_bos) {
u_rwlock_wrlock(&bo->ws->global_bo_list_lock);
list_del(&bo->global_list_item);
bo->ws->num_buffers--;
u_rwlock_wrunlock(&bo->ws->global_bo_list_lock);
}
radv_amdgpu_global_bo_list_del(bo);
radv_amdgpu_bo_va_op(bo->ws, bo->bo, 0, bo->size, bo->base.va,
0, 0, AMDGPU_VA_OP_UNMAP);
amdgpu_bo_free(bo->bo);
@@ -363,18 +402,6 @@ static void radv_amdgpu_winsys_bo_destroy(struct radeon_winsys_bo *_bo)
FREE(bo);
}
static void radv_amdgpu_add_buffer_to_global_list(struct radv_amdgpu_winsys_bo *bo)
{
struct radv_amdgpu_winsys *ws = bo->ws;
if (bo->ws->debug_all_bos) {
u_rwlock_wrlock(&ws->global_bo_list_lock);
list_addtail(&bo->global_list_item, &ws->global_bo_list);
ws->num_buffers++;
u_rwlock_wrunlock(&ws->global_bo_list_lock);
}
}
static struct radeon_winsys_bo *
radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws,
uint64_t size,
@@ -532,7 +559,7 @@ radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws,
p_atomic_add(&ws->allocated_gtt,
align64(bo->size, ws->info.gart_page_size));
radv_amdgpu_add_buffer_to_global_list(bo);
radv_amdgpu_global_bo_list_add(bo);
radv_amdgpu_log_bo(bo, false);
return (struct radeon_winsys_bo *)bo;
@@ -641,7 +668,7 @@ radv_amdgpu_winsys_bo_from_ptr(struct radeon_winsys *_ws,
p_atomic_add(&ws->allocated_gtt,
align64(bo->size, ws->info.gart_page_size));
radv_amdgpu_add_buffer_to_global_list(bo);
radv_amdgpu_global_bo_list_add(bo);
radv_amdgpu_log_bo(bo, false);
return (struct radeon_winsys_bo *)bo;
@@ -723,7 +750,7 @@ radv_amdgpu_winsys_bo_from_fd(struct radeon_winsys *_ws,
p_atomic_add(&ws->allocated_gtt,
align64(bo->size, ws->info.gart_page_size));
radv_amdgpu_add_buffer_to_global_list(bo);
radv_amdgpu_global_bo_list_add(bo);
radv_amdgpu_log_bo(bo, false);
return (struct radeon_winsys_bo *)bo;
@@ -938,28 +965,29 @@ static void radv_amdgpu_dump_bo_ranges(struct radeon_winsys *_ws, FILE *file)
{
struct radv_amdgpu_winsys *ws = radv_amdgpu_winsys(_ws);
if (ws->debug_all_bos) {
struct radv_amdgpu_winsys_bo *bo;
struct radv_amdgpu_winsys_bo **bos = NULL;
int i = 0;
u_rwlock_rdlock(&ws->global_bo_list_lock);
bos = malloc(sizeof(*bos) * ws->num_buffers);
u_rwlock_rdlock(&ws->global_bo_list.lock);
bos = malloc(sizeof(*bos) * ws->global_bo_list.count);
if (!bos) {
u_rwlock_rdunlock(&ws->global_bo_list_lock);
u_rwlock_rdunlock(&ws->global_bo_list.lock);
fprintf(file, " Failed to allocate memory to sort VA ranges for dumping\n");
return;
}
LIST_FOR_EACH_ENTRY(bo, &ws->global_bo_list, global_list_item) {
bos[i++] = bo;
for (i = 0; i < ws->global_bo_list.count; i++) {
bos[i] = ws->global_bo_list.bos[i];
}
qsort(bos, ws->num_buffers, sizeof(bos[0]), radv_amdgpu_bo_va_compare);
for (i = 0; i < ws->num_buffers; ++i) {
qsort(bos, ws->global_bo_list.count, sizeof(bos[0]), radv_amdgpu_bo_va_compare);
for (i = 0; i < ws->global_bo_list.count; ++i) {
fprintf(file, " VA=%.16llx-%.16llx, handle=%d%s\n",
(long long)bos[i]->base.va, (long long)(bos[i]->base.va + bos[i]->size),
bos[i]->bo_handle, bos[i]->is_virtual ? " sparse" : "");
}
free(bos);
u_rwlock_rdunlock(&ws->global_bo_list_lock);
u_rwlock_rdunlock(&ws->global_bo_list.lock);
} else
fprintf(file, " To get BO VA ranges, please specify RADV_DEBUG=allbos\n");
}
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h
-1
View File
@@ -54,7 +54,6 @@ struct radv_amdgpu_winsys_bo {
amdgpu_bo_handle bo;
bool is_shared;
uint32_t bo_handle;
struct list_head global_list_item;
};
/* virtual bo */
struct {
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+17 -19
View File
@@ -724,17 +724,14 @@ radv_amdgpu_get_bo_list(struct radv_amdgpu_winsys *ws,
unsigned num_handles = 0;
if (ws->debug_all_bos) {
struct radv_amdgpu_winsys_bo *bo;
handles = malloc(sizeof(handles[0]) * ws->num_buffers);
handles = malloc(sizeof(handles[0]) * ws->global_bo_list.count);
if (!handles) {
return VK_ERROR_OUT_OF_HOST_MEMORY;
}
LIST_FOR_EACH_ENTRY(bo, &ws->global_bo_list, global_list_item) {
assert(num_handles < ws->num_buffers);
handles[num_handles].bo_handle = bo->bo_handle;
handles[num_handles].bo_priority = bo->priority;
for (uint32_t i = 0; i < ws->global_bo_list.count; i++) {
handles[i].bo_handle = ws->global_bo_list.bos[i]->bo_handle;
handles[i].bo_priority = ws->global_bo_list.bos[i]->priority;
num_handles++;
}
} else if (count == 1 && !num_extra_bo && !extra_cs && !radv_bo_list &&
@@ -919,7 +916,7 @@ radv_amdgpu_winsys_cs_submit_chained(struct radeon_winsys_ctx *_ctx,
}
if (aws->debug_all_bos)
u_rwlock_rdlock(&aws->global_bo_list_lock);
u_rwlock_rdlock(&aws->global_bo_list.lock);
/* Get the BO list. */
result = radv_amdgpu_get_bo_list(cs0->ws, cs_array, cs_count, NULL, 0,
@@ -960,7 +957,7 @@ radv_amdgpu_winsys_cs_submit_chained(struct radeon_winsys_ctx *_ctx,
fail:
if (aws->debug_all_bos)
u_rwlock_rdunlock(&aws->global_bo_list_lock);
u_rwlock_rdunlock(&aws->global_bo_list.lock);
return result;
}
@@ -994,7 +991,7 @@ radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
number_of_ibs = cs_count + !!initial_preamble_cs;
if (aws->debug_all_bos)
u_rwlock_rdlock(&aws->global_bo_list_lock);
u_rwlock_rdlock(&aws->global_bo_list.lock);
/* Get the BO list. */
result = radv_amdgpu_get_bo_list(cs0->ws, &cs_array[0], cs_count, NULL, 0,
@@ -1050,7 +1047,7 @@ radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
fail:
if (aws->debug_all_bos)
u_rwlock_rdunlock(&aws->global_bo_list_lock);
u_rwlock_rdunlock(&aws->global_bo_list.lock);
return result;
}
@@ -1209,7 +1206,7 @@ radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
}
if (aws->debug_all_bos)
u_rwlock_rdlock(&aws->global_bo_list_lock);
u_rwlock_rdlock(&aws->global_bo_list.lock);
result = radv_amdgpu_get_bo_list(cs0->ws, &cs_array[i], cnt,
(struct radv_amdgpu_winsys_bo **)bos,
@@ -1220,7 +1217,7 @@ radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
free(ibs);
free(bos);
if (aws->debug_all_bos)
u_rwlock_rdunlock(&aws->global_bo_list_lock);
u_rwlock_rdunlock(&aws->global_bo_list.lock);
return result;
}
@@ -1239,7 +1236,7 @@ radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
free(request.handles);
if (aws->debug_all_bos)
u_rwlock_rdunlock(&aws->global_bo_list_lock);
u_rwlock_rdunlock(&aws->global_bo_list.lock);
for (unsigned j = 0; j < number_of_ibs; j++) {
ws->buffer_destroy(bos[j]);
@@ -1310,17 +1307,18 @@ static void *radv_amdgpu_winsys_get_cpu_addr(void *_cs, uint64_t addr)
}
}
if(cs->ws->debug_all_bos) {
u_rwlock_rdlock(&cs->ws->global_bo_list_lock);
list_for_each_entry(struct radv_amdgpu_winsys_bo, bo,
&cs->ws->global_bo_list, global_list_item) {
u_rwlock_rdlock(&cs->ws->global_bo_list.lock);
for (uint32_t i = 0; i < cs->ws->global_bo_list.count; i++) {
struct radv_amdgpu_winsys_bo *bo = cs->ws->global_bo_list.bos[i];
if (addr >= bo->base.va && addr - bo->base.va < bo->size) {
if (amdgpu_bo_cpu_map(bo->bo, &ret) == 0) {
u_rwlock_rdunlock(&cs->ws->global_bo_list_lock);
u_rwlock_rdunlock(&cs->ws->global_bo_list.lock);
return (char *)ret + (addr - bo->base.va);
}
}
}
u_rwlock_rdunlock(&cs->ws->global_bo_list_lock);
u_rwlock_rdunlock(&cs->ws->global_bo_list.lock);
}
return ret;
}
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
+4 -3
View File
@@ -166,8 +166,10 @@ static void radv_amdgpu_winsys_destroy(struct radeon_winsys *rws)
amdgpu_cs_destroy_syncobj(ws->dev, ws->syncobj[i]);
free(ws->syncobj);
u_rwlock_destroy(&ws->global_bo_list.lock);
free(ws->global_bo_list.bos);
pthread_mutex_destroy(&ws->syncobj_lock);
u_rwlock_destroy(&ws->global_bo_list_lock);
u_rwlock_destroy(&ws->log_bo_list_lock);
ac_addrlib_destroy(ws->addrlib);
amdgpu_device_deinitialize(ws->dev);
@@ -204,8 +206,7 @@ radv_amdgpu_winsys_create(int fd, uint64_t debug_flags, uint64_t perftest_flags)
ws->zero_all_vram_allocs = debug_flags & RADV_DEBUG_ZERO_VRAM;
ws->use_llvm = debug_flags & RADV_DEBUG_LLVM;
ws->cs_bo_domain = radv_cmdbuffer_domain(&ws->info, perftest_flags);
list_inithead(&ws->global_bo_list);
u_rwlock_init(&ws->global_bo_list_lock);
u_rwlock_init(&ws->global_bo_list.lock);
list_inithead(&ws->log_bo_list);
u_rwlock_init(&ws->log_bo_list_lock);
pthread_mutex_init(&ws->syncobj_lock, NULL);
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.h
+8 -4
View File
@@ -50,15 +50,19 @@ struct radv_amdgpu_winsys {
bool zero_all_vram_allocs;
bool use_local_bos;
bool use_llvm;
unsigned num_buffers;
struct u_rwlock global_bo_list_lock;
struct list_head global_bo_list;
uint64_t allocated_vram;
uint64_t allocated_vram_vis;
uint64_t allocated_gtt;
/* Global BO list */
struct {
struct radv_amdgpu_winsys_bo **bos;
uint32_t count;
uint32_t capacity;
struct u_rwlock lock;
} global_bo_list;
/* syncobj cache */
pthread_mutex_t syncobj_lock;
uint32_t *syncobj;