Skip to content

Commit a9b39f4

Browse files
slarenslaren
committed
vulkan : add backend registry / device interfaces
1 parent d5ed2b9 commit a9b39f4

File tree

4 files changed

+216
-116
lines changed

4 files changed

+216
-116
lines changed

‎ggml/include/ggml-vulkan.h

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -24,6 +24,8 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
2424
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
2525
GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
2626

27+
GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
28+
2729
#ifdef __cplusplus
2830
}
2931
#endif

‎ggml/src/ggml-backend.cpp

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -525,6 +525,10 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
525525
#include "ggml-cuda.h"
526526
#endif
527527

528+
#ifdef GGML_USE_VULKAN
529+
#include "ggml-vulkan.h"
530+
#endif
531+
528532
struct ggml_backend_registry {
529533
std::vector<ggml_backend_reg_t> backends;
530534
std::vector<ggml_backend_dev_t> devices;
@@ -534,6 +538,10 @@ struct ggml_backend_registry {
534538
register_backend(ggml_backend_cuda_reg());
535539
#endif
536540

541+
#ifdef GGML_USE_VULKAN
542+
register_backend(ggml_backend_vk_reg());
543+
#endif
544+
537545
register_backend(ggml_backend_cpu_reg());
538546

539547
// TODO: sycl, metal, vulkan, kompute, cann

‎ggml/src/ggml-vulkan.cpp

Lines changed: 204 additions & 74 deletions
Original file line numberDiff line numberDiff line change
@@ -1926,7 +1926,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
19261926
if (device->fp16) {
19271927
device_extensions.push_back("VK_KHR_shader_float16_int8");
19281928
}
1929-
device->name = device->properties.deviceName.data();
1929+
device->name = GGML_VK_NAME + std::to_string(idx);
19301930

19311931
device_create_info = {
19321932
vk::DeviceCreateFlags(),
@@ -1953,7 +1953,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
19531953

19541954
device->buffer_type = {
19551955
/* .iface = */ ggml_backend_vk_buffer_type_interface,
1956-
/* .device = */ nullptr,
1956+
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), idx),
19571957
/* .context = */ new ggml_backend_vk_buffer_type_context{ device->name, device },
19581958
};
19591959

@@ -6363,7 +6363,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
63636363
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
63646364
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
63656365
},
6366-
/* .device = */ nullptr,
6366+
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), 0),
63676367
/* .context = */ nullptr,
63686368
};
63696369

@@ -6566,9 +6566,135 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
65666566
UNUSED(backend);
65676567
}
65686568

6569-
static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
6570-
// ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context;
6569+
// TODO: enable async and synchronize
6570+
static ggml_backend_i ggml_backend_vk_interface = {
6571+
/* .get_name = */ ggml_backend_vk_name,
6572+
/* .free = */ ggml_backend_vk_free,
6573+
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
6574+
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
6575+
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
6576+
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
6577+
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
6578+
/* .graph_plan_create = */ NULL,
6579+
/* .graph_plan_free = */ NULL,
6580+
/* .graph_plan_update = */ NULL,
6581+
/* .graph_plan_compute = */ NULL,
6582+
/* .graph_compute = */ ggml_backend_vk_graph_compute,
6583+
/* .supports_op = */ NULL,
6584+
/* .supports_buft = */ NULL,
6585+
/* .offload_op = */ NULL,
6586+
/* .event_record = */ NULL,
6587+
/* .event_wait = */ NULL,
6588+
};
6589+
6590+
static ggml_guid_t ggml_backend_vk_guid() {
6591+
static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
6592+
return &guid;
6593+
}
6594+
6595+
ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
6596+
VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
6597+
6598+
ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
6599+
ggml_vk_init(ctx, dev_num);
6600+
6601+
ggml_backend_t vk_backend = new ggml_backend {
6602+
/* .guid = */ ggml_backend_vk_guid(),
6603+
/* .interface = */ ggml_backend_vk_interface,
6604+
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), dev_num),
6605+
/* .context = */ ctx,
6606+
};
6607+
6608+
return vk_backend;
6609+
}
6610+
6611+
bool ggml_backend_is_vk(ggml_backend_t backend) {
6612+
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
6613+
}
6614+
6615+
int ggml_backend_vk_get_device_count() {
6616+
return ggml_vk_get_device_count();
6617+
}
6618+
6619+
void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
6620+
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
6621+
int dev_idx = vk_instance.device_indices[device];
6622+
ggml_vk_get_device_description(dev_idx, description, description_size);
6623+
}
6624+
6625+
void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
6626+
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
6627+
6628+
vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
6629+
6630+
vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
6631+
6632+
for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
6633+
if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
6634+
*total = heap.size;
6635+
*free = heap.size;
6636+
break;
6637+
}
6638+
}
6639+
}
6640+
6641+
//////////////////////////
6642+
6643+
struct ggml_backend_vk_device_context {
6644+
int device;
6645+
std::string name;
6646+
std::string description;
6647+
};
6648+
6649+
static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
6650+
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
6651+
return ctx->name.c_str();
6652+
}
6653+
6654+
static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t dev) {
6655+
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
6656+
return ctx->description.c_str();
6657+
}
6658+
6659+
static void ggml_backend_vk_device_get_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
6660+
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)device->context;
6661+
ggml_backend_vk_get_device_memory(ctx->device, free, total);
6662+
}
6663+
6664+
static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
6665+
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
6666+
return ggml_backend_vk_buffer_type(ctx->device);
6667+
}
6668+
6669+
static ggml_backend_buffer_type_t ggml_backend_vk_device_get_host_buffer_type(ggml_backend_dev_t dev) {
6670+
UNUSED(dev);
6671+
return ggml_backend_vk_host_buffer_type();
6672+
}
65716673

6674+
static enum ggml_backend_dev_type ggml_backend_vk_device_get_type(ggml_backend_dev_t dev) {
6675+
UNUSED(dev);
6676+
return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
6677+
}
6678+
6679+
static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
6680+
props->name = ggml_backend_vk_device_get_name(dev);
6681+
props->description = ggml_backend_vk_device_get_description(dev);
6682+
props->type = ggml_backend_vk_device_get_type(dev);
6683+
ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
6684+
props->caps = {
6685+
/* async */ false,
6686+
/* host_buffer */ true,
6687+
/* events */ false,
6688+
};
6689+
}
6690+
6691+
static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
6692+
UNUSED(params);
6693+
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
6694+
return ggml_backend_vk_init(ctx->device);
6695+
}
6696+
6697+
static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
65726698
switch (op->op) {
65736699
case GGML_OP_UNARY:
65746700
switch (ggml_get_unary_op(op)) {
@@ -6686,97 +6812,101 @@ static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tenso
66866812
return false;
66876813
}
66886814

6689-
UNUSED(backend);
6690-
}
6691-
6692-
static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
6693-
const int min_batch_size = 32;
6694-
6695-
return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
6696-
(op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
6697-
6698-
UNUSED(backend);
6815+
UNUSED(dev);
66996816
}
67006817

6701-
static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
6818+
static bool ggml_backend_vk_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
67026819
if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
67036820
return false;
67046821
}
67056822

6823+
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
67066824
ggml_backend_vk_buffer_type_context * buft_ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
6707-
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
6708-
6709-
return buft_ctx->device == ctx->device;
6710-
}
6711-
6712-
// TODO: enable async and synchronize
6713-
static ggml_backend_i ggml_backend_vk_interface = {
6714-
/* .get_name = */ ggml_backend_vk_name,
6715-
/* .free = */ ggml_backend_vk_free,
6716-
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
6717-
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
6718-
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
6719-
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
6720-
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
6721-
/* .graph_plan_create = */ NULL,
6722-
/* .graph_plan_free = */ NULL,
6723-
/* .graph_plan_update = */ NULL,
6724-
/* .graph_plan_compute = */ NULL,
6725-
/* .graph_compute = */ ggml_backend_vk_graph_compute,
6726-
/* .supports_op = */ ggml_backend_vk_supports_op,
6727-
/* .supports_buft = */ ggml_backend_vk_supports_buft,
6728-
/* .offload_op = */ ggml_backend_vk_offload_op,
6729-
/* .event_record = */ NULL,
6730-
/* .event_wait = */ NULL,
6731-
};
67326825

6733-
static ggml_guid_t ggml_backend_vk_guid() {
6734-
static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
6735-
return &guid;
6826+
return buft_ctx->device->idx == ctx->device;
67366827
}
67376828

6738-
ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
6739-
VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
6829+
static bool ggml_backend_vk_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
6830+
const int min_batch_size = 32;
67406831

6741-
ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
6742-
ggml_vk_init(ctx, dev_num);
6832+
return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
6833+
(op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
67436834

6744-
ggml_backend_t vk_backend = new ggml_backend {
6745-
/* .guid = */ ggml_backend_vk_guid(),
6746-
/* .interface = */ ggml_backend_vk_interface,
6747-
/* .device = */ nullptr,
6748-
/* .context = */ ctx,
6749-
};
6835+
UNUSED(dev);
6836+
}
6837+
6838+
static const struct ggml_backend_device_i ggml_backend_vk_device_i = {
6839+
/* .get_name = */ ggml_backend_vk_device_get_name,
6840+
/* .get_description = */ ggml_backend_vk_device_get_description,
6841+
/* .get_memory = */ ggml_backend_vk_device_get_memory,
6842+
/* .get_type = */ ggml_backend_vk_device_get_type,
6843+
/* .get_props = */ ggml_backend_vk_device_get_props,
6844+
/* .init_backend = */ ggml_backend_vk_device_init,
6845+
/* .get_buffer_type = */ ggml_backend_vk_device_get_buffer_type,
6846+
/* .get_host_buffer_type = */ ggml_backend_vk_device_get_host_buffer_type,
6847+
/* .buffer_from_host_ptr = */ NULL,
6848+
/* .supports_op = */ ggml_backend_vk_device_supports_op,
6849+
/* .supports_buft = */ ggml_backend_vk_device_supports_buft,
6850+
/* .offload_op = */ ggml_backend_vk_device_offload_op,
6851+
/* .event_new = */ NULL,
6852+
/* .event_free = */ NULL,
6853+
/* .event_synchronize = */ NULL,
6854+
};
67506855

6751-
return vk_backend;
6856+
static const char * ggml_backend_vk_reg_get_name(ggml_backend_reg_t reg) {
6857+
UNUSED(reg);
6858+
return GGML_VK_NAME;
67526859
}
67536860

6754-
bool ggml_backend_is_vk(ggml_backend_t backend) {
6755-
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
6861+
static size_t ggml_backend_vk_reg_get_device_count(ggml_backend_reg_t reg) {
6862+
UNUSED(reg);
6863+
return ggml_backend_vk_get_device_count();
67566864
}
67576865

6758-
int ggml_backend_vk_get_device_count() {
6759-
return ggml_vk_get_device_count();
6760-
}
6866+
static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg, size_t device) {
6867+
static std::vector<ggml_backend_dev_t> devices;
67616868

6762-
void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
6763-
ggml_vk_get_device_description(device, description, description_size);
6764-
}
6869+
static bool initialized = false;
67656870

6766-
void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
6767-
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
6871+
{
6872+
static std::mutex mutex;
6873+
std::lock_guard<std::mutex> lock(mutex);
6874+
if (!initialized) {
6875+
for (size_t i = 0; i < ggml_backend_vk_get_device_count(); i++) {
6876+
ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
6877+
char desc[256];
6878+
ggml_backend_vk_get_device_description(i, desc, sizeof(desc));
6879+
ctx->device = i;
6880+
ctx->name = GGML_VK_NAME + std::to_string(i);
6881+
ctx->description = desc;
6882+
devices.push_back(new ggml_backend_device {
6883+
/* .iface = */ ggml_backend_vk_device_i,
6884+
/* .reg = */ reg,
6885+
/* .context = */ ctx,
6886+
});
6887+
}
6888+
initialized = true;
6889+
}
6890+
}
67686891

6769-
vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
6892+
GGML_ASSERT(device < devices.size());
6893+
return devices[device];
6894+
}
67706895

6771-
vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
6896+
static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
6897+
/* .get_name = */ ggml_backend_vk_reg_get_name,
6898+
/* .get_device_count = */ ggml_backend_vk_reg_get_device_count,
6899+
/* .get_device = */ ggml_backend_vk_reg_get_device,
6900+
/* .get_proc_address = */ NULL,
6901+
};
67726902

6773-
for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
6774-
if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
6775-
*total = heap.size;
6776-
*free = heap.size;
6777-
break;
6778-
}
6779-
}
6903+
ggml_backend_reg_t ggml_backend_vk_reg() {
6904+
static ggml_backend_reg reg = {
6905+
/* .iface = */ ggml_backend_vk_reg_i,
6906+
/* .context = */ nullptr,
6907+
};
6908+
6909+
return &reg;
67806910
}
67816911

67826912
// Extension availability

0 commit comments

Comments
 (0)