Skip to content

metal : use virtual GPU address for private buffers #15985

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 2 commits into from
Closed

metal : use virtual GPU address for private buffers #15985

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
626fa1d
metal : use virtual GPU address for private buffers
ggerganov Sep 14, 2025
6045c5a
cont : put all buffers in the same virtual address space
ggerganov Sep 14, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
60 changes: 44 additions & 16 deletions ggml/src/ggml-metal/ggml-metal.m
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,8 @@

#import <Metal/Metal.h>

#include <stdatomic.h>

#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
Expand Down Expand Up @@ -41,6 +43,9 @@
static struct ggml_backend_reg g_ggml_backend_metal_reg;
static struct ggml_backend_device g_ggml_backend_metal_device;

// virtual address for GPU memory allocations
static atomic_uintptr_t g_addr_device = 0x000000400ULL;

// information about a Metal device
// note: assumes single GPU device - the default one
// TODO: support multiple GPU devices
Expand Down Expand Up @@ -1785,9 +1790,11 @@ static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
};

struct ggml_backend_metal_buffer_context {
void * all_data;
void * all_data; // for shared buffers
size_t all_size;

void * base_addr;

// if false, the Metal buffer data is allocated in private GPU memory and is not shared with the host
bool is_shared;

Expand Down Expand Up @@ -6033,33 +6040,42 @@ static void ggml_backend_metal_buffer_shared_free_buffer(ggml_backend_buffer_t b
}

static void * ggml_backend_metal_buffer_shared_get_base(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;

return ctx->all_data;
return ctx->base_addr;
}

static void ggml_backend_metal_buffer_shared_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *) tensor->buffer->context;

GGML_ASSERT(ctx->is_shared);

memset((char *)tensor->data + offset, value, size);
const ptrdiff_t base_offset = (char *)tensor->data - (char *)ctx->base_addr;
memset((char *) ctx->all_data + base_offset + offset, value, size);

GGML_UNUSED(buffer);
}

static void ggml_backend_metal_buffer_shared_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *) tensor->buffer->context;

GGML_ASSERT(ctx->is_shared);

memcpy((char *)tensor->data + offset, data, size);
const ptrdiff_t base_offset = (char *)tensor->data - (char *)ctx->base_addr;
memcpy((char *) ctx->all_data + base_offset + offset, data, size);

GGML_UNUSED(buffer);
}

static void ggml_backend_metal_buffer_shared_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *) tensor->buffer->context;

GGML_ASSERT(ctx->is_shared);

memcpy(data, (const char *)tensor->data + offset, size);
const ptrdiff_t base_offset = (char *)tensor->data - (char *)ctx->base_addr;
memcpy(data, (const char *) ctx->all_data + base_offset + offset, size);

GGML_UNUSED(buffer);
}

static bool ggml_backend_metal_buffer_shared_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
Expand Down Expand Up @@ -6109,7 +6125,7 @@ static void ggml_backend_metal_buffer_private_free_buffer(ggml_backend_buffer_t
static void * ggml_backend_metal_buffer_private_get_base(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;

return ctx->all_data;
return ctx->base_addr;
}

static void ggml_backend_metal_buffer_private_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
Expand Down Expand Up @@ -6327,12 +6343,14 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
ctx->all_data = ggml_metal_host_malloc(size_aligned);
ctx->is_shared = true;
} else {
// dummy, non-NULL value - we'll populate this after creating the Metal buffer below
// dummy, non-NULL value - not used
ctx->all_data = (void *) 0x000000400ULL;
ctx->is_shared = false;
}
ctx->all_size = size_aligned;

ctx->base_addr = (void *) atomic_fetch_add_explicit(&g_addr_device, size_aligned, memory_order_relaxed);

ctx->device = device;
ctx->queue = ctx_dev->mtl_queue;

Expand All @@ -6343,19 +6361,22 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
ctx->buffers[0].metal = nil;

if (size_aligned > 0) {
if (ctx_dev->use_shared_buffers) {
if (ctx_dev->use_shared_buffers && shared) {
ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
length:size_aligned
options:MTLResourceStorageModeShared
deallocator:nil];
} else {
ctx->buffers[0].metal = [device newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate];

ctx->all_data = (void *) (ctx->buffers[0].metal.gpuAddress);
// MTLBuffer.gpuAddress is not available on early OSes, so we use a virtual address instead
//if (@available(macOS 10.13, iOS 16.0, *)) {
// ctx->all_data = (void *) (ctx->buffers[0].metal.gpuAddress);
//}
}
}

ctx->buffers[0].data = ctx->all_data;
ctx->buffers[0].data = ctx->base_addr;
}

if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) {
Expand Down Expand Up @@ -6956,6 +6977,13 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_mapped(ggml_backen
size_aligned += (size_page - (size_aligned % size_page));
}

// note: I think this is preferred because we want to have both the mapped and non-mapped buffers in the same
// address space. not sure if there are any side-effects from this though.
//ctx->base_addr = (void *) atomic_fetch_add_explicit(&g_addr_device, size_aligned, memory_order_relaxed);
//
// note2: the above does not actually work
ctx->base_addr = ptr;

struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context;

GGML_ASSERT(ctx_dev->mtl_device != nil);
Expand All @@ -6967,7 +6995,7 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_mapped(ggml_backen

// the buffer fits into the max buffer size allowed by the device
if (size_aligned <= device.maxBufferLength) {
ctx->buffers[ctx->n_buffers].data = ptr;
ctx->buffers[ctx->n_buffers].data = ctx->base_addr;
ctx->buffers[ctx->n_buffers].size = size;
ctx->buffers[ctx->n_buffers].metal = nil;

Expand All @@ -6993,7 +7021,7 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_mapped(ggml_backen
for (size_t i = 0; i < size; i += size_step) {
const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);

ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) ptr + i);
ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) ctx->base_addr + i);
ctx->buffers[ctx->n_buffers].size = size_step_aligned;
ctx->buffers[ctx->n_buffers].metal = nil;

Expand Down
Loading