8000 metal : use residency sets · ggml-org/llama.cpp@febb813 · GitHub
[go: up one dir, main page]
Skip to content

Commit febb813

Browse files
ggerganovggerganov
committed
metal : use residency sets
1 parent 6da9021 commit febb813

File tree

1 file changed

+158
-5
lines changed

1 file changed

+158
-5
lines changed

‎ggml/src/ggml-metal/ggml-metal.m

Lines changed: 158 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -19,6 +19,8 @@
1919
// max number of MTLCommandBuffer used to submit a graph for processing
2020
#define GGML_METAL_MAX_COMMAND_BUFFERS 8
2121

22+
#define GGML_METAL_MAX_RESIDENCY_SETS 128
23+
2224
#define UNUSED(x) (void)(x)
2325

2426
// globals
@@ -37,6 +39,9 @@
3739
id<MTLDevice> mtl_device;
3840
int mtl_device_ref_count;
3941

42+
id<MTLResidencySet> mtl_residency_set[GGML_METAL_MAX_RESIDENCY_SETS];
43+
int mtl_residency_set_n;
44+
4045
bool has_simdgroup_reduction;
4146
bool has_simdgroup_mm;
4247
bool has_bfloat;
@@ -46,6 +51,8 @@
4651
} g_ggml_ctx_dev_main = {
4752
/*.mtl_device =*/ nil,
4853
/*.mtl_device_ref_count =*/ 0,
54+
/*.mtl_residency_set =*/ { nil },
55+
/*.mtl_residency_set_n =*/ 0,
4956
/*.has_simdgroup_reduction =*/ false,
5057
/*.has_simdgroup_mm =*/ false,
5158
/*.has_bfloat =*/ false,
@@ -95,6 +102,42 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
95102
}
96103
}
97104

105+
// add command queue
106+
static bool ggml_backend_metal_device_add_residency_set(struct ggml_backend_metal_device_context * ctx, id<MTLResidencySet> residency_set) {
107+
assert(ctx != NULL);
108+
assert(queue != nil);
109+
110+
if (ctx->mtl_residency_set_n >= GGML_METAL_MAX_RESIDENCY_SETS) {
111+
GGML_LOG_ERROR("%s: warning: maximum number of residency sets reached\n", __func__);
112+
return false;
113+
}
114+
115+
ctx->mtl_residency_set[ctx->mtl_residency_set_n++] = residency_set;
116+
117+
return true;
118+
}
119+
120+
// remove residency set
121+
// search for the residency set in the list, remove it and shift the remaining residency sets
122+
static bool ggml_backend_metal_device_remove_residency_set(struct ggml_backend_metal_device_context * ctx, id<MTLResidencySet> residency_set) {
123+
assert(ctx != NULL);
124+
assert(residency_set != nil);
125+
126+
for (int i = 0; i < ctx->mtl_residency_set_n; ++i) {
127+
if (ctx->mtl_residency_set[i] == residency_set) {
128+
for (int j = i; j < ctx->mtl_residency_set_n - 1; ++j) {
129+
ctx->mtl_residency_set[j] = ctx->mtl_residency_set[j + 1];
130+
}
131+
132+
ctx->mtl_residency_set_n--;
133+
134+
return true;
135+
}
136+
}
137+
138+
return false;
139+
}
140+
98141
// kernels
99142

100143
struct ggml_metal_kernel {
@@ -483,6 +526,11 @@ @implementation GGMLMetalClass
483526
GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
484527

485528
ctx->queue = [device newCommandQueue];
529+
if (ctx->queue == nil) {
530+
GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
531+
return NULL;
532+
}
533+
486534
ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
487535

488536
id<MTLLibrary> metal_library;
@@ -1035,6 +1083,8 @@ static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
10351083
// multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
10361084
int n_buffers;
10371085
struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
1086+
1087+
id<MTLResidencySet> residency_set;
10381088
};
10391089

10401090
// finds the Metal buffer that contains the tensor data on the GPU device
@@ -4039,6 +4089,20 @@ static enum ggml_status ggml_metal_graph_compute(
40394089
struct ggml_backend_metal_context * ctx = backend->context;
40404090
struct ggml_backend_metal_device_context * ctx_dev = backend->device->context;
40414091

4092+
static bool is_first = true;
4093+
if (is_first) {
4094+
is_first = false;
4095+
GGML_LOG_INFO("%s: adding %d residency sets\n", __func__, ctx_dev->mtl_residency_set_n);
4096+
[ctx->queue addResidencySets:ctx_dev->mtl_residency_set count:ctx_dev->mtl_residency_set_n];
4097+
}
4098+
4099+
//for (int i = 0; i < ctx_dev->mtl_residency_set_n; ++i) {
4100+
// GGML_LOG_INFO("%s: residency set %d allocations size = %zu\n", __func__, i, [ctx_dev->mtl_residency_set[i] allocatedSize]);
4101+
// [ctx_dev->mtl_residency_set[i] requestResidency];
4102+
//}
4103+
4104+
int64_t t_start_us = ggml_time_us();
4105+
40424106
// number of nodes encoded by the main thread (empirically determined)
40434107
const int n_main = 128;
40444108

@@ -4086,19 +4150,23 @@ static enum ggml_status ggml_metal_graph_compute(
40864150
// the main thread commits the first few commands immediately
40874151
// command_buffer[n_cb]
40884152
{
4089-
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
4153+
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBuffer];
40904154
ctx->command_buffers[n_cb] = command_buffer;
40914155

4156+
[command_buffer useResidencySets:ctx_dev->mtl_residency_set count:ctx_dev->mtl_residency_set_n];
4157+
40924158
[command_buffer enqueue];
40934159
ctx->enc F438 ode_async(n_cb);
40944160
}
40954161

40964162
// prepare the rest of the command buffers asynchronously
40974163
// command_buffer[0.. n_cb)
40984164
for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
4099-
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
4165+
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBuffer];
41004166
ctx->command_buffers[cb_idx] = command_buffer;
41014167

4168+
[command_buffer useResidencySets:ctx_dev->mtl_residency_set count:ctx_dev->mtl_residency_set_n];
4169+
41024170
// always enqueue the first two command buffers
41034171
// enqueue all of the command buffers if we don't need to abort
41044172
if (cb_idx < 2 || ctx->abort_callback == NULL) {
@@ -4163,6 +4231,10 @@ static enum ggml_status ggml_metal_graph_compute(
41634231
}
41644232
}
41654233

4234+
int64_t t_end_us = ggml_time_us();
4235+
4236+
GGML_LOG_DEBUG("%s: compute graph took %8.2f ms\n", __func__, (t_end_us - t_start_us) / 1000.0);
4237+
41664238
return GGML_STATUS_SUCCESS;
41674239
}
41684240

@@ -4176,6 +4248,13 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
41764248
for (int i = 0; i < ctx->n_buffers; i++) {
41774249
[ctx->buffers[i].metal release];
41784250
}
4251+
4252+
ggml_backend_metal_device_remove_residency_set(buffer->buft->device->context, ctx->residency_set);
4253+
4254+
[ctx->residency_set endResidency];
4255+
[ctx->residency_set removeAllAllocations];
4256+
[ctx->residency_set release];
4257+
41794258
ggml_backend_metal_device_rel(buffer->buft->device->context);
41804259

41814260
if (ctx->owned) {
@@ -4284,7 +4363,8 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
42844363
size_aligned += (size_page - (size_aligned % size_page));
42854364
}
42864365

4287-
id<MTLDevice> device = ggml_backend_metal_device_acq(buft->device->context);
4366+
struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)buft->device->context;
4367+
id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
42884368

42894369
ctx->all_data = ggml_metal_host_malloc(size_aligned);
42904370
ctx->all_size = size_aligned;
@@ -4307,10 +4387,34 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
43074387
if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) {
43084388
GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
43094389
free(ctx);
4310-
ggml_backend_metal_device_rel(buft->device->context);
4390+
ggml_backend_metal_device_rel(ctx_dev);
43114391
return NULL;
43124392
}
43134393

4394+
{
4395+
MTLResidencySetDescriptor * desc;
4396+
desc = [[MTLResidencySetDescriptor alloc] init];
4397+
desc.label = @"Primary residency set";
4398+
desc.initialCapacity = ctx->n_buffers;
4399+
4400+
NSError *error;
4401+
ctx->residency_set = [device newResidencySetWithDescriptor:desc error:&error];
4402+
if (error) {
4403+
GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
4404+
return NULL;
4405+
}
4406+
4407+
for (int i = 0; i < ctx->n_buffers; i++) {
4408+
[ctx->residency_set addAllocation:ctx->buffers[i].metal];
4409+
}
4410+
4411+
[ctx->residency_set commit];
4412+
[ctx->residency_set requestResidency];
4413+
4414+
// track the residency set in the device context
4415+
ggml_backend_metal_device_add_residency_set(ctx_dev, ctx->residency_set);
4416+
}
4417+
43144418
//ggml_backend_metal_log_allocated_size(device, size_aligned);
43154419

43164420
return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
@@ -4400,7 +4504,8 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
44004504
size_aligned += (size_page - (size_aligned % size_page));
44014505
}
44024506

4403-
id<MTLDevice> device = ggml_backend_metal_device_acq(&g_ggml_ctx_dev_main);
4507+
struct ggml_backend_metal_device_context * ctx_dev = &g_ggml_ctx_dev_main;
4508+
id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
44044509

44054510
// the buffer fits into the max buffer size allowed by the device
44064511
if (size_aligned <= device.maxBufferLength) {
@@ -4453,6 +4558,30 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
44534558
}
44544559
}
44554560

4561+
{
4562+
MTLResidencySetDescriptor * desc;
4563+
desc = [[MTLResidencySetDescriptor alloc] init];
4564+
desc.label = @"Primary residency set";
4565+
desc.initialCapacity = ctx->n_buffers;
4566+
4567+
NSError *error;
4568+
ctx->residency_set = [device newResidencySetWithDescriptor:desc error:&error];
4569+
if (error) {
4570+
GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
4571+
return NULL;
4572+
}
4573+
4574+
for (int i = 0; i < ctx->n_buffers; i++) {
4575+
[ctx->residency_set addAllocation:ctx->buffers[i].metal];
4576+
}
4577+
4578+
[ctx->residency_set commit];
4579+
[ctx->residency_set requestResidency];
4580+
4581+
// track the residency set in the device context
4582+
ggml_backend_metal_device_add_residency_set(ctx_dev, ctx->residency_set);
4583+
}
4584+
44564585
return ggml_backend_buffer_init(ggml_backend_metal_buffer_from_ptr_type(), ggml_backend_metal_buffer_i, ctx, size);
44574586
}
44584587

@@ -4766,6 +4895,30 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_back
47664895
}
47674896
}
47684897

4898+
{
4899+
MTLResidencySetDescriptor * desc;
4900+
desc = [[MTLResidencySetDescriptor alloc] init];
4901+
desc.label = @"Primary residency set";
4902+
desc.initialCapacity = ctx->n_buffers;
4903+
4904+
NSError *error;
4905+
ctx->residency_set = [device newResidencySetWithDescriptor:desc error:&error];
4906+
if (error) {
4907+
GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
4908+
return NULL;
4909+
}
4910+
4911+
for (int i = 0; i < ctx->n_buffers; i++) {
4912+
[ctx->residency_set addAllocation:ctx->buffers[i].metal];
4913+
}
4914+
4915+
[ctx->residency_set commit];
4916+
[ctx->residency_set requestResidency];
4917+
4918+
// track the residency set in the device context
4919+
ggml_backend_metal_device_add_residency_set(ctx_dev, ctx->residency_set);
4920+
}
4921+
47694922
return ggml_backend_buffer_init(ggml_backend_metal_buffer_from_ptr_type(), ggml_backend_metal_buffer_i, ctx, size);
47704923
}
47714924

0 commit comments

Comments
 (0)
0