mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-12-25 02:44:36 +00:00
metal : add option to disable debug logs (close #2764)
This commit is contained in:
parent
611363ac79
commit
3a007648f2
@ -301,7 +301,7 @@ if (LLAMA_METAL)
|
||||
set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
|
||||
|
||||
add_compile_definitions(GGML_USE_METAL)
|
||||
add_compile_definitions(GGML_METAL_NDEBUG)
|
||||
#add_compile_definitions(GGML_METAL_NDEBUG)
|
||||
|
||||
# get full path to the file
|
||||
#add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
|
||||
|
2
Makefile
2
Makefile
@ -305,7 +305,7 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
|
||||
endif # LLAMA_HIPBLAS
|
||||
|
||||
ifdef LLAMA_METAL
|
||||
CFLAGS += -DGGML_USE_METAL -DGGML_METAL_NDEBUG
|
||||
CFLAGS += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG
|
||||
CXXFLAGS += -DGGML_USE_METAL
|
||||
LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
|
||||
OBJS += ggml-metal.o
|
||||
|
71
ggml-metal.m
71
ggml-metal.m
@ -11,6 +11,7 @@
|
||||
#define MIN(a, b) ((a) < (b) ? (a) : (b))
|
||||
#define MAX(a, b) ((a) > (b) ? (a) : (b))
|
||||
|
||||
// TODO: temporary - reuse llama.cpp logging
|
||||
#ifdef GGML_METAL_NDEBUG
|
||||
#define metal_printf(...)
|
||||
#else
|
||||
@ -113,7 +114,7 @@ static NSString * const msl_library_source = @"see metal.metal";
|
||||
@end
|
||||
|
||||
struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
||||
fprintf(stderr, "%s: allocating\n", __func__);
|
||||
metal_printf("%s: allocating\n", __func__);
|
||||
|
||||
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
|
||||
|
||||
@ -132,7 +133,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
||||
|
||||
ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
|
||||
if (error) {
|
||||
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
|
||||
metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
|
||||
return NULL;
|
||||
}
|
||||
}
|
||||
@ -146,11 +147,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
||||
//NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
|
||||
NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
|
||||
NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
|
||||
fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
|
||||
metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]);
|
||||
|
||||
NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
|
||||
if (error) {
|
||||
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
|
||||
metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@ -162,7 +163,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
||||
ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
|
||||
#endif
|
||||
if (error) {
|
||||
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
|
||||
metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
|
||||
return NULL;
|
||||
}
|
||||
}
|
||||
@ -174,11 +175,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
||||
#define GGML_METAL_ADD_KERNEL(name) \
|
||||
ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
|
||||
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
|
||||
fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
|
||||
metal_printf("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
|
||||
(int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
|
||||
(int) ctx->pipeline_##name.threadExecutionWidth); \
|
||||
if (error) { \
|
||||
fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
|
||||
metal_printf("%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
|
||||
return NULL; \
|
||||
}
|
||||
|
||||
@ -230,19 +231,19 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
||||
#undef GGML_METAL_ADD_KERNEL
|
||||
}
|
||||
|
||||
fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
|
||||
fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
|
||||
metal_printf("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
|
||||
metal_printf("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
|
||||
if (ctx->device.maxTransferRate != 0) {
|
||||
fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
|
||||
metal_printf("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
|
||||
} else {
|
||||
fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__);
|
||||
metal_printf("%s: maxTransferRate = built-in GPU\n", __func__);
|
||||
}
|
||||
|
||||
return ctx;
|
||||
}
|
||||
|
||||
void ggml_metal_free(struct ggml_metal_context * ctx) {
|
||||
fprintf(stderr, "%s: deallocating\n", __func__);
|
||||
metal_printf("%s: deallocating\n", __func__);
|
||||
#define GGML_METAL_DEL_KERNEL(name) \
|
||||
[ctx->function_##name release]; \
|
||||
[ctx->pipeline_##name release];
|
||||
@ -311,7 +312,7 @@ void * ggml_metal_host_malloc(size_t n) {
|
||||
void * data = NULL;
|
||||
const int result = posix_memalign((void **) &data, getpagesize(), n);
|
||||
if (result != 0) {
|
||||
fprintf(stderr, "%s: error: posix_memalign failed\n", __func__);
|
||||
metal_printf("%s: error: posix_memalign failed\n", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@ -339,7 +340,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
|
||||
// Metal buffer based on the host memory pointer
|
||||
//
|
||||
static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
|
||||
//fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
|
||||
//metal_printf("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
|
||||
|
||||
const int64_t tsize = ggml_nbytes(t);
|
||||
|
||||
@ -350,13 +351,13 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
|
||||
if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
|
||||
*offs = (size_t) ioffs;
|
||||
|
||||
//fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
|
||||
//metal_printf("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
|
||||
|
||||
return ctx->buffers[i].metal;
|
||||
}
|
||||
}
|
||||
|
||||
fprintf(stderr, "%s: error: buffer is nil\n", __func__);
|
||||
metal_printf("%s: error: buffer is nil\n", __func__);
|
||||
|
||||
return nil;
|
||||
}
|
||||
@ -368,7 +369,7 @@ bool ggml_metal_add_buffer(
|
||||
size_t size,
|
||||
size_t max_size) {
|
||||
if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
|
||||
fprintf(stderr, "%s: too many buffers\n", __func__);
|
||||
metal_printf("%s: too many buffers\n", __func__);
|
||||
return false;
|
||||
}
|
||||
|
||||
@ -378,7 +379,7 @@ bool ggml_metal_add_buffer(
|
||||
const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
|
||||
|
||||
if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
|
||||
fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
|
||||
metal_printf("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
@ -399,11 +400,11 @@ bool ggml_metal_add_buffer(
|
||||
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
|
||||
|
||||
if (ctx->buffers[ctx->n_buffers].metal == nil) {
|
||||
fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
|
||||
metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
|
||||
return false;
|
||||
}
|
||||
|
||||
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
|
||||
metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
|
||||
|
||||
++ctx->n_buffers;
|
||||
} else {
|
||||
@ -423,27 +424,27 @@ bool ggml_metal_add_buffer(
|
||||
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
|
||||
|
||||
if (ctx->buffers[ctx->n_buffers].metal == nil) {
|
||||
fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
|
||||
metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
|
||||
return false;
|
||||
}
|
||||
|
||||
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
|
||||
metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
|
||||
if (i + size_step < size) {
|
||||
fprintf(stderr, "\n");
|
||||
metal_printf("\n");
|
||||
}
|
||||
|
||||
++ctx->n_buffers;
|
||||
}
|
||||
}
|
||||
|
||||
fprintf(stderr, ", (%8.2f / %8.2f)",
|
||||
metal_printf(", (%8.2f / %8.2f)",
|
||||
ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
|
||||
ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
|
||||
|
||||
if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
|
||||
fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
|
||||
metal_printf(", warning: current allocated size is greater than the recommended max working set size\n");
|
||||
} else {
|
||||
fprintf(stderr, "\n");
|
||||
metal_printf("\n");
|
||||
}
|
||||
}
|
||||
|
||||
@ -453,8 +454,6 @@ bool ggml_metal_add_buffer(
|
||||
void ggml_metal_set_tensor(
|
||||
struct ggml_metal_context * ctx,
|
||||
struct ggml_tensor * t) {
|
||||
metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
|
||||
|
||||
size_t offs;
|
||||
id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
|
||||
|
||||
@ -464,8 +463,6 @@ void ggml_metal_set_tensor(
|
||||
void ggml_metal_get_tensor(
|
||||
struct ggml_metal_context * ctx,
|
||||
struct ggml_tensor * t) {
|
||||
metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
|
||||
|
||||
size_t offs;
|
||||
id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
|
||||
|
||||
@ -560,15 +557,13 @@ void ggml_metal_graph_find_concurrency(
|
||||
}
|
||||
|
||||
if (ctx->concur_list_len > GGML_MAX_CONCUR) {
|
||||
fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
|
||||
metal_printf("%s: too many elements for metal ctx->concur_list!\n", __func__);
|
||||
}
|
||||
}
|
||||
|
||||
void ggml_metal_graph_compute(
|
||||
struct ggml_metal_context * ctx,
|
||||
struct ggml_cgraph * gf) {
|
||||
metal_printf("%s: evaluating graph\n", __func__);
|
||||
|
||||
@autoreleasepool {
|
||||
|
||||
// if there is ctx->concur_list, dispatch concurrently
|
||||
@ -616,7 +611,7 @@ void ggml_metal_graph_compute(
|
||||
continue;
|
||||
}
|
||||
|
||||
metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
|
||||
//metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
|
||||
|
||||
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
|
||||
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
|
||||
@ -764,7 +759,7 @@ void ggml_metal_graph_compute(
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
|
||||
metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
|
||||
GGML_ASSERT(false);
|
||||
}
|
||||
} break;
|
||||
@ -923,7 +918,7 @@ void ggml_metal_graph_compute(
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "Asserting on type %d\n",(int)src0t);
|
||||
metal_printf("Asserting on type %d\n",(int)src0t);
|
||||
GGML_ASSERT(false && "not implemented");
|
||||
}
|
||||
};
|
||||
@ -1161,7 +1156,7 @@ void ggml_metal_graph_compute(
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
|
||||
metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
|
||||
GGML_ASSERT(false);
|
||||
}
|
||||
}
|
||||
@ -1186,7 +1181,7 @@ void ggml_metal_graph_compute(
|
||||
|
||||
MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
|
||||
if (status != MTLCommandBufferStatusCompleted) {
|
||||
fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
|
||||
metal_printf("%s: command buffer %d failed with status %lu\n", __func__, i, status);
|
||||
GGML_ASSERT(false);
|
||||
}
|
||||
}
|
||||
|
Loading…
Reference in New Issue
Block a user