Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-02-07 00:43:14 +01:00
Code Issues Packages Projects Releases Wiki Activity

try make CL run w/o tunning, but -ngl stucks no output. had to add task runer and profile id, many changes, see the f codes

Browse Source
This commit is contained in:
mqy 2023-06-16 20:32:12 +08:00
parent 5342dc075f
commit 6b83a3e16f
10 changed files with 435 additions and 281 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
examples/mulmat-tune/mulmat-tune.cpp
View File

@ -11,6 +11,10 @@
#define UNUSED(x) (void)(x) #define UNUSED(x) (void)(x)
// F16 has an pending Illegal Instruction error on macos-latest-cmake.
// So the workaround is to disable non-quantized ftypes.
// #define SUPPORT_NONE_Q_TYPE 1
static void print_build_tips(void) { static void print_build_tips(void) {
const char *a = "LLAMA_NO_ACCELERATE"; const char *a = "LLAMA_NO_ACCELERATE";
fprintf(stderr, "Tips on how to build with various backend vendors:\n\n"); fprintf(stderr, "Tips on how to build with various backend vendors:\n\n");
@ -62,11 +66,12 @@ static void usage(char *prog) {
"--model MODEL 3B | 7B | 13B | 30B | 65B", "--model MODEL 3B | 7B | 13B | 30B | 65B",
" default 7B", " default 7B",
"--ftype FTYPE ggml ftype:", "--ftype FTYPE ggml ftype:",
#ifdef SUPPORT_NONE_Q_TYPE
" 0: all F32", " 0: all F32",
" 1: mostly F16", " 1: mostly F16",
#endif
" 2: mostly Q4_0", " 2: mostly Q4_0",
" 3: mostly Q4_1", " 3: mostly Q4_1",
" 4: mostly Q4_1, some F16",
" 7: mostly Q8_0", " 7: mostly Q8_0",
" 8: mostly Q5_0", " 8: mostly Q5_0",
" 9: mostly Q5_1", " 9: mostly Q5_1",
@ -84,7 +89,7 @@ static void usage(char *prog) {
" requires: between [1, 3]", " requires: between [1, 3]",
"--n_threads NTH bench with this number of threads", "--n_threads NTH bench with this number of threads",
" requires: between [1, 16]", " requires: between [1, 16]",
" default 1", " default 4",
"--file FILE data file to write", "--file FILE data file to write",
" default stdout", " default stdout",
"-y always answer \"yes\" to all prompts", "-y always answer \"yes\" to all prompts",
@ -170,8 +175,22 @@ int main(int argc, char **argv) {
ftype = (enum ggml_ftype)v; ftype = (enum ggml_ftype)v;
} }
#ifndef SUPPORT_NONE_Q_TYPE
if (ftype == GGML_FTYPE_ALL_F32 || ftype == GGML_FTYPE_MOSTLY_F16) { if (ftype == GGML_FTYPE_ALL_F32 || ftype == GGML_FTYPE_MOSTLY_F16) {
fprintf(stderr, "none quantized type %d is not supported\n", ftype); fprintf(stderr, "error: none quantized type %d is not supported\n",
ftype);
return 1;
}
#endif
bool cond_1 = ftype >= GGML_FTYPE_MOSTLY_Q4_0 &&
ftype <= GGML_FTYPE_MOSTLY_Q4_1;
bool cond_2 =
ftype >= GGML_FTYPE_MOSTLY_Q8_0 && ftype <= GGML_FTYPE_MOSTLY_Q6_K;
if (!(cond_1 || cond_2)) {
fprintf(stderr, "error: type %d is not a known ggml ftype.\n",
ftype);
return 1; return 1;
} }
} }
@ -223,7 +242,7 @@ int main(int argc, char **argv) {
} }
} }
int n_threads = 1; int n_threads = 4;
{ {
if (arg_n_threads != NULL) { if (arg_n_threads != NULL) {
int v = atoi(arg_n_threads); int v = atoi(arg_n_threads);

2
ggml-opencl.cpp
View File

@ -1628,7 +1628,7 @@ bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_
} }
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) { void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
// GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst)); GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));
if (src0->type == GGML_TYPE_F32) { if (src0->type == GGML_TYPE_F32) {
ggml_cl_mul_mat_f32(src0, src1, dst); ggml_cl_mul_mat_f32(src0, src1, dst);

37
ggml-threading.c
View File

@ -170,7 +170,8 @@ struct ggml_compute_state_shared {
atomic_bool wait_on_done; atomic_bool wait_on_done;
atomic_bool stop; atomic_bool stop;
ggml_threading_task_runner *task_runner; // Default task runner, can be overriden by node.task_profile.runner.
ggml_task_runner *task_runner;
struct ggml_threading_context *ctx; struct ggml_threading_context *ctx;
}; };
@ -391,8 +392,10 @@ ggml_thread_ret_t ggml_threading_graph_compute_thread(void *data) {
} }
if (shared->n_tasks > 0 && state->has_work) { if (shared->n_tasks > 0 && state->has_work) {
enum ggml_compute_error err = ggml_task_runner *runner = state->node->task_profile.runner
shared->task_runner(&state->params, state->node); ? state->node->task_profile.runner
: shared->task_runner;
enum ggml_compute_error err = runner(&state->params, state->node);
GGML_ASSERT(err == GGML_COMPUTE_OK); GGML_ASSERT(err == GGML_COMPUTE_OK);
@ -427,8 +430,13 @@ ggml_threading_compute_tensor(struct ggml_threading_context *ctx,
size_t wsize) { size_t wsize) {
GGML_ASSERT(ctx); GGML_ASSERT(ctx);
GGML_ASSERT(node); GGML_ASSERT(node);
GGML_ASSERT(ctx->shared.task_runner); GGML_ASSERT(ctx->shared.task_runner);
ggml_task_runner *runner = ctx->shared.task_runner;
if (node->task_profile.runner) {
runner = node->task_profile.runner;
}
struct ggml_compute_state_shared *state_shared = &ctx->shared; struct ggml_compute_state_shared *state_shared = &ctx->shared;
// This is the params for main thread. // This is the params for main thread.
@ -491,7 +499,7 @@ START:
params.wsize = wsize; params.wsize = wsize;
params.wdata = wdata; params.wdata = wdata;
err = state_shared->task_runner(&params, node); err = runner(&params, node);
} }
// wait for tasks done. // wait for tasks done.
@ -509,11 +517,21 @@ START:
if (err != GGML_COMPUTE_OK) { if (err != GGML_COMPUTE_OK) {
if (err == GGML_COMPUTE_FALLBACK) { if (err == GGML_COMPUTE_FALLBACK) {
PRINT_DEBUG("[main] fallback from profile, id=%d\n",
node->task_profile.id);
GGML_ASSERT(node->task_profile.stages[1].backend >
GGML_TASK_BACKEND_CPU);
struct ggml_task_profile profiles[GGML_MAX_TASK_PROFILES]; struct ggml_task_profile profiles[GGML_MAX_TASK_PROFILES];
int n = ggml_get_task_profiles(node, profiles); int n = ggml_get_task_profiles(node, profiles);
GGML_ASSERT(n > 0); GGML_ASSERT(n > 0);
GGML_ASSERT(profiles[0].stages[1].backend ==
GGML_TASK_BACKEND_CPU);
memcpy(&node->task_profile, &profiles[0], memcpy(&node->task_profile, &profiles[0],
sizeof(struct ggml_task_profile)); sizeof(struct ggml_task_profile));
runner = ctx->shared.task_runner;
goto START; goto START;
} }
return err; return err;
@ -525,12 +543,13 @@ START:
struct ggml_threading_context * struct ggml_threading_context *
ggml_threading_start(int n_threads, ggml_threading_thread_runner *thread_runner, ggml_threading_start(int n_threads, ggml_threading_thread_runner *thread_runner,
ggml_threading_task_runner *task_stage_runner, ggml_task_runner *task_runner,
enum ggml_threading_features features, enum ggml_threading_features features,
int64_t stages_time[3]) { int64_t stages_time[3]) {
GGML_ASSERT(n_threads > 0); GGML_ASSERT(n_threads > 0);
GGML_ASSERT(thread_runner); if (thread_runner == NULL) {
GGML_ASSERT(task_stage_runner); thread_runner = ggml_threading_graph_compute_thread;
}
size_t ctx_sz = sizeof(struct ggml_threading_context); size_t ctx_sz = sizeof(struct ggml_threading_context);
struct ggml_threading_context *ctx = malloc(ctx_sz); struct ggml_threading_context *ctx = malloc(ctx_sz);
@ -545,7 +564,7 @@ ggml_threading_start(int n_threads, ggml_threading_thread_runner *thread_runner,
.wait_now = false, .wait_now = false,
.wait_on_done = false, .wait_on_done = false,
.stop = false, .stop = false,
.task_runner = task_stage_runner, .task_runner = task_runner,
.ctx = ctx, .ctx = ctx,
}; };

22
ggml-threading.h
View File

@ -21,27 +21,21 @@ enum ggml_threading_features {
GGML_THREADING_FEATURE_PERF = 1 << 1, GGML_THREADING_FEATURE_PERF = 1 << 1,
}; };
// Compute errors.
enum ggml_compute_error {
GGML_COMPUTE_OK = 0,
GGML_COMPUTE_FALLBACK = 1,
};
// The task runner to be called by main thread and workers.
typedef enum ggml_compute_error(ggml_threading_task_runner)(
struct ggml_compute_params *params, struct ggml_tensor *node);
// The thread runner to feed into OS threads. // The thread runner to feed into OS threads.
typedef ggml_thread_ret_t(ggml_threading_thread_runner)(void *data); typedef ggml_thread_ret_t(ggml_threading_thread_runner)(void *data);
// Init and start underlying workers if n_threads > 1. // Init and start underlying workers if n_threads > 1.
// //
// features: optional for configure threading additional features. // thread: optional OS thread runner, default value:
// see `ggml_threading_feature`, default 0. // `ggml_threading_graph_compute_thread`.
//
// features: optional for configure
// threading additional features. see `ggml_threading_feature`, default 0.
//
// stages_time: optional for collecting per-stage wall clock time. // stages_time: optional for collecting per-stage wall clock time.
struct ggml_threading_context * struct ggml_threading_context *
ggml_threading_start(int n_threads, ggml_threading_thread_runner *thread, ggml_threading_start(int n_threads, ggml_threading_thread_runner *thread,
ggml_threading_task_runner *task_stage_runner, ggml_task_runner *task_runner,
enum ggml_threading_features features, enum ggml_threading_features features,
int64_t stages_time[3]); int64_t stages_time[3]);
@ -60,7 +54,7 @@ ggml_threading_compute_tensor(struct ggml_threading_context *ctx,
// This is an experimental functionality for mulmat tune, as a thin wrapper. // This is an experimental functionality for mulmat tune, as a thin wrapper.
enum ggml_compute_error enum ggml_compute_error
ggml_compute_forward_wrapper(struct ggml_compute_params *params, ggml_compute_forward_wrapper(const struct ggml_compute_params *params,
struct ggml_tensor *tensor); struct ggml_tensor *tensor);
#ifdef __cplusplus #ifdef __cplusplus

70
ggml-tune.c
View File

@ -44,9 +44,12 @@ ggml_mulmat_tune_task_backend_name(enum ggml_task_backend backend) {
} }
} }
const struct ggml_task_profile *ggml_mulmat_tune_select_task_profile( // NOTE: we can not use the profile from tune because the profiles do not
struct ggml_mulmat_tune *tune, int M, int N, int K, enum ggml_type src0_t, // contain fields such as runner, get_size.
enum ggml_type src1_t, int stages_time[3]) { int ggml_mulmat_tune_select_task_profile(struct ggml_mulmat_tune *tune, int M,
int N, int K, enum ggml_type src0_t,
enum ggml_type src1_t,
int stages_time[3]) {
GGML_ASSERT(tune); GGML_ASSERT(tune);
// TODO: default_mm_cache is thread-unsafe. // TODO: default_mm_cache is thread-unsafe.
@ -103,15 +106,15 @@ const struct ggml_task_profile *ggml_mulmat_tune_select_task_profile(
names[i] = ggml_mulmat_tune_task_backend_name( names[i] = ggml_mulmat_tune_task_backend_name(
prof->stages[i].backend); prof->stages[i].backend);
} }
printf("\n[tune] M: %3d, N: %5d, K: %5d, backends of the " printf("\n[tune] M: %3d, N: %5d, K: %5d, profile id: %d, "
"fastest profile: %s %s %s\n", "backends: %s %s %s\n",
M, N, K, names[0], names[1], names[2]); M, N, K, prof->id, names[0], names[1], names[2]);
#endif #endif
} }
} }
} }
return prof; return prof->id;
} }
void ggml_mulmat_tune_model_init(struct ggml_mulmat_tune_model *model, void ggml_mulmat_tune_model_init(struct ggml_mulmat_tune_model *model,
@ -264,10 +267,13 @@ void ggml_mulmat_tune_free(struct ggml_mulmat_tune *tune) {
if (shape->m_num > 0) { if (shape->m_num > 0) {
if (shape->arr_m) { if (shape->arr_m) {
free(shape->arr_m); free(shape->arr_m);
shape->arr_m = NULL;
} }
if (shape->items) { if (shape->items) {
free(shape->items); free(shape->items);
shape->items = NULL;
} }
shape->m_num = 0;
} }
} }
} }
@ -277,6 +283,11 @@ static bool ggml_mulmat_tune_write_profiles(
int rc; int rc;
for (int i = 0; i < n_profiles; i++) { for (int i = 0; i < n_profiles; i++) {
const struct ggml_task_profile *profile = &profiles[i]; const struct ggml_task_profile *profile = &profiles[i];
rc = fprintf(fp, "%d ", profile->id);
if (rc <= 0) {
return false;
}
for (int j = 0; j < 3; j++) { for (int j = 0; j < 3; j++) {
const struct ggml_task_stage *ts = &profile->stages[j]; const struct ggml_task_stage *ts = &profile->stages[j];
rc = fprintf(fp, "%2d %d %d", ts->backend, ts->parallel ? 1 : 0, rc = fprintf(fp, "%2d %d %d", ts->backend, ts->parallel ? 1 : 0,
@ -304,7 +315,6 @@ static bool
ggml_mulmat_tune_validate_internal(const struct ggml_mulmat_tune *tune, ggml_mulmat_tune_validate_internal(const struct ggml_mulmat_tune *tune,
const char *model, int ftype, int n_threads, const char *model, int ftype, int n_threads,
char *errbuf, int errbuf_len) { char *errbuf, int errbuf_len) {
if (tune->version != GGML_MULMAT_TUNE_VERSION) { if (tune->version != GGML_MULMAT_TUNE_VERSION) {
snprintf(errbuf, errbuf_len - 1, snprintf(errbuf, errbuf_len - 1,
"version mismatch, built-in: %d, " "version mismatch, built-in: %d, "
@ -348,14 +358,28 @@ ggml_mulmat_tune_validate_internal(const struct ggml_mulmat_tune *tune,
int n_profiles = ggml_get_task_profiles(&node, builtin_profiles); int n_profiles = ggml_get_task_profiles(&node, builtin_profiles);
if (n_profiles != shape->n_profiles) { if (n_profiles != shape->n_profiles) {
snprintf(errbuf, errbuf_len - 1, "task profiles mismatch(n_profiles)"); snprintf(errbuf, errbuf_len - 1,
"task profiles mismatch (n_profiles)");
return false; return false;
} }
// TODO: profiles order is relevant, too strict. // TODO: profiles order is relevant, too strict.
size_t sz = sizeof(struct ggml_task_profile) * n_profiles; // Only validate stages!
if (memcmp(builtin_profiles, shape->profiles, sz) != 0) { size_t sz = sizeof(struct ggml_task_stage) * 3;
snprintf(errbuf, errbuf_len - 1, "task profiles mismatch(profiles)"); bool matched = true;
for (int j = 0; j < n_profiles; j++) {
if (builtin_profiles[j].id != shape->profiles[j].id) {
return false;
}
if (memcmp(builtin_profiles[j].stages, shape->profiles[j].stages,
sz) != 0) {
matched = false;
break;
}
}
if (!matched) {
snprintf(errbuf, errbuf_len - 1,
"task profiles mismatch (profiles)");
printf("=== built-in profiles:\n"); printf("=== built-in profiles:\n");
ggml_mulmat_tune_write_profiles(stderr, builtin_profiles, ggml_mulmat_tune_write_profiles(stderr, builtin_profiles,
@ -426,6 +450,12 @@ bool ggml_mulmat_tune_read_data(struct ggml_mulmat_tune *tune, FILE *fp) {
for (int ip = 0; ip < shape->n_profiles; ip++) { for (int ip = 0; ip < shape->n_profiles; ip++) {
struct ggml_task_profile *profile = &shape->profiles[ip]; struct ggml_task_profile *profile = &shape->profiles[ip];
rc = fscanf(fp, "%d ", &profile->id);
if (rc <= 0) {
return false;
}
for (int j = 0; j < 3; j++) { for (int j = 0; j < 3; j++) {
struct ggml_task_stage *ts = &profile->stages[j]; struct ggml_task_stage *ts = &profile->stages[j];
int backend; int backend;
@ -777,6 +807,8 @@ bool ggml_mulmat_tune_bench(struct ggml_mulmat_tune *tune,
GGML_ASSERT(params); GGML_ASSERT(params);
GGML_ASSERT(params->model.name); GGML_ASSERT(params->model.name);
memset(tune, 0, sizeof(struct ggml_mulmat_tune));
enum ggml_task_backend backends[16]; enum ggml_task_backend backends[16];
int n_backends = ggml_mulmat_tune_get_builtin_task_backends(backends); int n_backends = ggml_mulmat_tune_get_builtin_task_backends(backends);
if (n_backends < 2) { if (n_backends < 2) {
@ -785,6 +817,15 @@ bool ggml_mulmat_tune_bench(struct ggml_mulmat_tune *tune,
return false; return false;
} }
if (params->model.ftype >= GGML_FTYPE_MOSTLY_Q2_K &&
params->model.ftype <= GGML_FTYPE_MOSTLY_Q6_K) {
#if defined(GGML_USE_CLBLAST)
printf("[tune] error: cl implementation does not support k_quants at "
"the time of writing this code, skip.\n");
return false;
#endif
}
bool ok = ggml_mulmat_tune_init(tune, params, ggml_get_task_profiles); bool ok = ggml_mulmat_tune_init(tune, params, ggml_get_task_profiles);
if (!ok) { if (!ok) {
return false; return false;
@ -816,9 +857,8 @@ bool ggml_mulmat_tune_bench(struct ggml_mulmat_tune *tune,
int64_t t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
struct ggml_threading_context *thrd_ctx = ggml_threading_start( struct ggml_threading_context *thrd_ctx = ggml_threading_start(
tune->n_threads, ggml_threading_graph_compute_thread, tune->n_threads, NULL, ggml_compute_forward_wrapper,
ggml_compute_forward_wrapper, GGML_THREADING_FEATURE_WAIT_ON_DONE, GGML_THREADING_FEATURE_WAIT_ON_DONE, stages_time);
stages_time);
for (int i_shape = 0; i_shape < tune->n_shapes; i_shape++) { for (int i_shape = 0; i_shape < tune->n_shapes; i_shape++) {
const struct ggml_mulmat_tune_shape *shape = &tune->shapes[i_shape]; const struct ggml_mulmat_tune_shape *shape = &tune->shapes[i_shape];

9
ggml-tune.h
View File

@ -10,7 +10,7 @@
extern "C" { extern "C" {
#endif #endif
#define GGML_MULMAT_TUNE_VERSION 8 #define GGML_MULMAT_TUNE_VERSION 9
#define GGML_MULMAT_N_SHAPES 4 #define GGML_MULMAT_N_SHAPES 4
#define GGML_MULMAT_CACHE_LEN 16 #define GGML_MULMAT_CACHE_LEN 16
@ -98,10 +98,11 @@ struct ggml_mulmat_tune_params {
}; };
// NOTE: stages_time is filled if not null. // NOTE: stages_time is filled if not null.
const struct ggml_task_profile * // Return profile id.
ggml_mulmat_tune_select_task_profile(struct ggml_mulmat_tune *tune, int M, int ggml_mulmat_tune_select_task_profile(struct ggml_mulmat_tune *tune, int M,
int N, int K, enum ggml_type src0_t, int N, int K, enum ggml_type src0_t,
enum ggml_type src1_t, int stages_time[3]); enum ggml_type src1_t,
int stages_time[3]);
bool ggml_mulmat_tune_validate(const struct ggml_mulmat_tune *tune, bool ggml_mulmat_tune_validate(const struct ggml_mulmat_tune *tune,
const char *model_name, int ftype, const char *model_name, int ftype,

355
ggml.c
View File

@ -8500,8 +8500,6 @@ static void ggml_compute_forward_mul_f32(
const int ith = params->ith; const int ith = params->ith;
const int nth = params->nth; const int nth = params->nth;
enum ggml_task_backend comp_backend = dst->task_profile.stages[GGML_TASK_COMPUTE].backend;
if (comp_backend == GGML_TASK_BACKEND_GPU_CL) {
#ifdef GGML_USE_CLBLAST #ifdef GGML_USE_CLBLAST
if (src1->backend == GGML_BACKEND_GPU) { if (src1->backend == GGML_BACKEND_GPU) {
if (ith == 0) { if (ith == 0) {
@ -8509,10 +8507,7 @@ static void ggml_compute_forward_mul_f32(
} }
return; return;
} }
#else
GGML_ASSERT(false);
#endif #endif
};
const int64_t nr = ggml_nrows(src0); const int64_t nr = ggml_nrows(src0);
@ -9938,7 +9933,7 @@ static void ggml_compute_forward_rms_norm_back(
} }
} }
// CPU only
static void ggml_compute_forward_mul_mat_f32( static void ggml_compute_forward_mul_mat_f32(
const struct ggml_compute_params * params, const struct ggml_compute_params * params,
const struct ggml_tensor * src0, const struct ggml_tensor * src0,
@ -10010,18 +10005,6 @@ static void ggml_compute_forward_mul_mat_f32(
// compute by src0 rows // compute by src0 rows
enum ggml_task_backend comp_backend = dst->task_profile.stages[GGML_TASK_COMPUTE].backend; enum ggml_task_backend comp_backend = dst->task_profile.stages[GGML_TASK_COMPUTE].backend;
if (comp_backend == GGML_TASK_BACKEND_GPU_CL) {
#if defined(GGML_USE_CLBLAST)
GGML_ASSERT(params->nth == 1);
GGML_ASSERT(params->type == GGML_TASK_COMPUTE);
ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
return;
#else
GGML_ASSERT(false);
#endif
}
GGML_ASSERT(comp_backend & GGML_TASK_BACKEND_CPU); GGML_ASSERT(comp_backend & GGML_TASK_BACKEND_CPU);
if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) { if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) {
@ -10104,6 +10087,7 @@ static void ggml_compute_forward_mul_mat_f32(
//} //}
} }
// CPU only.
static void ggml_compute_forward_mul_mat_f16_f32( static void ggml_compute_forward_mul_mat_f16_f32(
const struct ggml_compute_params * params, const struct ggml_compute_params * params,
const struct ggml_tensor * src0, const struct ggml_tensor * src0,
@ -10168,19 +10152,9 @@ static void ggml_compute_forward_mul_mat_f16_f32(
// nb01 >= nb00 - src0 is not transposed // nb01 >= nb00 - src0 is not transposed
// compute by src0 rows // compute by src0 rows
enum ggml_task_backend init_backend = dst->task_profile.stages[GGML_TASK_INIT].backend;
enum ggml_task_backend comp_backend = dst->task_profile.stages[GGML_TASK_COMPUTE].backend; enum ggml_task_backend comp_backend = dst->task_profile.stages[GGML_TASK_COMPUTE].backend;
if (comp_backend == GGML_TASK_BACKEND_GPU_CL) {
#if defined(GGML_USE_CLBLAST)
GGML_ASSERT(params->type == GGML_TASK_COMPUTE);
ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
return;
#else
GGML_ASSERT(false);
#endif
}
enum ggml_task_backend init_backend = dst->task_profile.stages[GGML_TASK_INIT].backend;
GGML_ASSERT(comp_backend & GGML_TASK_BACKEND_CPU); GGML_ASSERT(comp_backend & GGML_TASK_BACKEND_CPU);
if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) { if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) {
@ -10304,6 +10278,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
//} //}
} }
// CPU only
static void ggml_compute_forward_mul_mat_q_f32( static void ggml_compute_forward_mul_mat_q_f32(
const struct ggml_compute_params * params, const struct ggml_compute_params * params,
const struct ggml_tensor * src0, const struct ggml_tensor * src0,
@ -10373,20 +10348,8 @@ static void ggml_compute_forward_mul_mat_q_f32(
// nb01 >= nb00 - src0 is not transposed // nb01 >= nb00 - src0 is not transposed
// compute by src0 rows // compute by src0 rows
enum ggml_task_backend comp_backend = dst->task_profile.stages[GGML_TASK_COMPUTE].backend;
if (comp_backend == GGML_TASK_BACKEND_GPU_CL) {
#if defined(GGML_USE_CLBLAST)
GGML_ASSERT(params->nth == 1);
GGML_ASSERT(params->type == GGML_TASK_COMPUTE);
ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
return;
#else
GGML_ASSERT(false);
#endif
}
enum ggml_task_backend init_backend = dst->task_profile.stages[GGML_TASK_INIT].backend; enum ggml_task_backend init_backend = dst->task_profile.stages[GGML_TASK_INIT].backend;
enum ggml_task_backend comp_backend = dst->task_profile.stages[GGML_TASK_COMPUTE].backend;
GGML_ASSERT(comp_backend & GGML_TASK_BACKEND_CPU); GGML_ASSERT(comp_backend & GGML_TASK_BACKEND_CPU);
if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) { if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) {
@ -14294,30 +14257,9 @@ static void ggml_compute_forward_cross_entropy_loss_back(
///////////////////////////////// /////////////////////////////////
static enum ggml_compute_error ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) { static enum ggml_compute_error ggml_compute_forward(const struct ggml_compute_params * params, struct ggml_tensor * tensor) {
GGML_ASSERT(params); GGML_ASSERT(params);
enum ggml_task_backend comp_backend = tensor->task_profile.stages[GGML_TASK_COMPUTE].backend;
if (comp_backend == GGML_TASK_BACKEND_GPU_CUDA) {
#if defined(GGML_USE_CUBLAS)
bool skip_cpu = ggml_cuda_compute_forward(params, tensor);
if (skip_cpu) {
return GGML_COMPUTE_OK;
}
GGML_ASSERT(tensor->src0->backend == GGML_BACKEND_CPU);
GGML_ASSERT(tensor->src1 == NULL || tensor->src1->backend == GGML_BACKEND_CPU);
return GGML_COMPUTE_FALLBACK;
#else
GGML_ASSERT(false);
#endif
}
// if (tensor->task_profile.stages[params->type].backend > GGML_TASK_BACKEND_CPU) {
// printf("mulmat: test fallback\n");
// return GGML_COMPUTE_FALLBACK;
// }
switch (tensor->op) { switch (tensor->op) {
case GGML_OP_DUP: case GGML_OP_DUP:
{ {
@ -14568,13 +14510,6 @@ static enum ggml_compute_error ggml_compute_forward(struct ggml_compute_params *
return GGML_COMPUTE_OK; return GGML_COMPUTE_OK;
} }
enum ggml_compute_error ggml_compute_forward_wrapper(struct ggml_compute_params *params,
struct ggml_tensor *tensor) {
// We call ggml_compute_forward because the CUDA mul_mat entry point
// was moved out of `ggml_compute_forward_mul_mat`.
return ggml_compute_forward(params, tensor);
}
//////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////
static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, bool inplace) { static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, bool inplace) {
@ -15524,12 +15459,67 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg
// ---- task profiles ---- // ---- task profiles ----
// TODO: replace with ggml_compute_forward_cuda
// DO NOT check matrix size further.
#if defined(GGML_USE_CUBLAS)
static enum ggml_compute_error ggml_compute_forward_cuda(
const struct ggml_compute_params * params,
struct ggml_tensor * tensor) {
GGML_ASSERT (ggml_cuda_can_mul_mat(tensor->src0, tensor->src1, tensor));
if (ggml_cuda_compute_forward(params, tensor)) {
return GGML_COMPUTE_OK;
}
GGML_ASSERT(tensor->src0->backend == GGML_BACKEND_CPU);
GGML_ASSERT(tensor->src1 == NULL || tensor->src1->backend == GGML_BACKEND_CPU);
return GGML_COMPUTE_FALLBACK;
}
#endif
// TODO: replace with ggml_cl_mul_mat.
// DO NOT check matrix size further.
#if defined(GGML_USE_CLBLAST)
static enum ggml_compute_error ggml_compute_forward_cl(
const struct ggml_compute_params * params,
struct ggml_tensor * tensor) {
switch (tensor->op) {
case GGML_OP_MUL_MAT:
GGML_ASSERT(ggml_cl_can_mul_mat(tensor->src0, tensor->src1, tensor));
ggml_cl_mul_mat(tensor->src0, tensor->src1, tensor, params->wdata, params->wsize);
return GGML_COMPUTE_OK;
default:
break;
}
GGML_ASSERT(false);
}
static int ggml_compute_forward_get_wsize_cl (struct ggml_tensor *tensor) {
switch (tensor->op) {
case GGML_OP_MUL_MAT:
return ggml_cl_mul_mat_get_wsize(tensor->src0, tensor->src1, tensor);
default:
break;
}
return -1;
}
#endif
// The wrapper for external mulmat tune tool.
enum ggml_compute_error ggml_compute_forward_wrapper(const struct ggml_compute_params *params,
struct ggml_tensor *tensor) {
// We call ggml_compute_forward because the CUDA mul_mat entry point
// was moved out of `ggml_compute_forward_mul_mat`.
return ggml_compute_forward(params, tensor);
}
// Implement `ggml_task_profiles_provider`. // Implement `ggml_task_profiles_provider`.
// Fill `profiles` for the `node` and return number of profiles. // Fill `profiles` for the `node` and return number of profiles.
// //
// NOTE: the node may be incompleted from testing or tunning, so please assert // NOTE: the node may be incompleted from testing or tunning, so please assert
// everything used here. // everything used here.
inline int ggml_get_task_profiles( //
// TODO: configure cuda for none mul_mat nodes.
int ggml_get_task_profiles(
struct ggml_tensor *node, struct ggml_tensor *node,
struct ggml_task_profile profiles[GGML_MAX_TASK_PROFILES]) { struct ggml_task_profile profiles[GGML_MAX_TASK_PROFILES]) {
GGML_ASSERT(node); GGML_ASSERT(node);
@ -15595,6 +15585,8 @@ inline int ggml_get_task_profiles(
} break; } break;
case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT:
case GGML_OP_OUT_PROD: { case GGML_OP_OUT_PROD: {
// CPU only profiles.
// CUDA/CL: see end of function.
GGML_ASSERT(node->src0); GGML_ASSERT(node->src0);
GGML_ASSERT(node->src1); GGML_ASSERT(node->src1);
@ -15614,16 +15606,6 @@ inline int ggml_get_task_profiles(
p[i].stages[1].wait = true; p[i].stages[1].wait = true;
i++; i++;
#endif #endif
#if defined(GGML_USE_CUBLAS)
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CUDA;
p[i].stages[1].wait = true;
i++;
#elif defined(GGML_USE_CLBLAST)
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CL;
p[i].stages[1].wait = true;
i++;
#endif
} else if (src0_t == GGML_TYPE_F16) { } else if (src0_t == GGML_TYPE_F16) {
p[i].stages[0].backend = GGML_TASK_BACKEND_CPU; p[i].stages[0].backend = GGML_TASK_BACKEND_CPU;
p[i].stages[1].backend = GGML_TASK_BACKEND_CPU; p[i].stages[1].backend = GGML_TASK_BACKEND_CPU;
@ -15635,16 +15617,6 @@ inline int ggml_get_task_profiles(
p[i].stages[1].wait = true; p[i].stages[1].wait = true;
i++; i++;
#endif #endif
#if defined(GGML_USE_CUBLAS)
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CUDA;
p[i].stages[1].wait = true;
i++;
#elif defined(GGML_USE_CLBLAST)
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CL;
p[i].stages[1].wait = true;
i++;
#endif
} else if (ggml_is_quantized(src0_t)) { } else if (ggml_is_quantized(src0_t)) {
p[i].stages[0].backend = GGML_TASK_BACKEND_CPU; p[i].stages[0].backend = GGML_TASK_BACKEND_CPU;
p[i].stages[1].backend = GGML_TASK_BACKEND_CPU; p[i].stages[1].backend = GGML_TASK_BACKEND_CPU;
@ -15658,16 +15630,6 @@ inline int ggml_get_task_profiles(
p[i].stages[1].wait = true; p[i].stages[1].wait = true;
i++; i++;
#endif #endif
#if defined(GGML_USE_CUBLAS)
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CUDA;
p[i].stages[1].wait = true;
i++;
#elif defined(GGML_USE_CLBLAST)
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CL;
p[i].stages[1].wait = true;
i++;
#endif
} }
n_profiles = i; n_profiles = i;
} break; } break;
@ -15757,7 +15719,43 @@ inline int ggml_get_task_profiles(
GGML_ASSERT(false); GGML_ASSERT(false);
} }
#if defined(GGML_USE_CUBLAS)
switch (node->op) {
case GGML_OP_ADD:
case GGML_OP_MUL:
case GGML_OP_SILU:
case GGML_OP_RMS_NORM:
case GGML_OP_MUL_MAT:
case GGML_OP_RESHAPE:
case GGML_OP_ROPE: {
int i = n_profiles;
p[i].runner = ggml_compute_forward_cuda;
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CUDA;
p[i].stages[1].wait = true;
++n_profiles;
} break;
default: {
} break;
}
#elif defined(GGML_USE_CLBLAST)
switch (node->op) {
case GGML_OP_MUL_MAT: {
int i = n_profiles;
p[i].runner = ggml_compute_forward_cl;
p[i].get_wsize = ggml_compute_forward_get_wsize_cl;
p[i].stages[1].backend = GGML_TASK_BACKEND_GPU_CL;
p[i].stages[1].wait = true;
++n_profiles;
} break;
default: {
} break;
}
#endif
GGML_ASSERT(n_profiles > 0 && n_profiles <= GGML_MAX_TASK_PROFILES); GGML_ASSERT(n_profiles > 0 && n_profiles <= GGML_MAX_TASK_PROFILES);
for (int i = 0; i < n_profiles; i++) {
profiles[i].id = i + 1;
}
return n_profiles; return n_profiles;
} }
@ -15769,7 +15767,7 @@ static const struct ggml_task_profile *ggml_mulmat_get_task_profile(
GGML_ASSERT(node); GGML_ASSERT(node);
GGML_ASSERT(node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_OUT_PROD); GGML_ASSERT(node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_OUT_PROD);
GGML_ASSERT(profiles); GGML_ASSERT(profiles);
GGML_ASSERT(n_profiles >= 2); GGML_ASSERT(n_profiles > 0);
enum ggml_type src0_t = node->src0->type; enum ggml_type src0_t = node->src0->type;
enum ggml_type src1_t = node->src1->type; enum ggml_type src1_t = node->src1->type;
@ -15790,12 +15788,16 @@ static const struct ggml_task_profile *ggml_mulmat_get_task_profile(
if (cond_match) { if (cond_match) {
#if defined(GGML_USE_TUNE) #if defined(GGML_USE_TUNE)
if (tune != NULL) { if (tune != NULL) {
prof = ggml_mulmat_tune_select_task_profile(tune, M, N, K, src0_t, GGML_ASSERT(n_profiles >= 2);
int id = ggml_mulmat_tune_select_task_profile(tune, M, N, K, src0_t,
src1_t, stages_time_us); src1_t, stages_time_us);
if (prof != NULL) { for (int i = 0; i < n_profiles; i++) {
if (profiles[i].id == id) {
prof = &profiles[i];
return prof; return prof;
} }
} }
}
#else #else
UNUSED(tune); UNUSED(tune);
UNUSED(stages_time_us); UNUSED(stages_time_us);
@ -15841,35 +15843,23 @@ static const struct ggml_task_profile *ggml_mulmat_get_task_profile(
return prof; return prof;
} }
void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { void ggml_graph_compute_set_tensor_task_proile(struct ggml_tensor *node,
int n_threads = cgraph->n_threads; struct ggml_cgraph *cgraph) {
// Pre-specified.
struct ggml_threading_context *thrd_ctx = ggml_threading_start( for (int i = 0; i < 3; i++) {
n_threads, ggml_threading_graph_compute_thread, ggml_compute_forward, if (node->task_profile.stages[i].backend > 0) {
GGML_THREADING_FEATURE_WAIT_ON_DONE, NULL); return;
}
// initialize tasks + work buffer
{
// int64_t t0 = ggml_time_us();
size_t work_size = 0;
struct ggml_task_profile profiles[GGML_MAX_TASK_PROFILES];
// thread scheduling for the different operations
for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * node = cgraph->nodes[i];
if (node->op == GGML_OP_NONE || node->op == GGML_OP_CONT) {
continue;
} }
struct ggml_task_profile profiles[GGML_MAX_TASK_PROFILES];
int n_profiles = ggml_get_task_profiles(node, profiles); int n_profiles = ggml_get_task_profiles(node, profiles);
const struct ggml_task_profile *profile = NULL; const struct ggml_task_profile *profile = NULL;
// Adapt node->backend: assume GPU at COMPUTE stage. // GPU offloading. A special case of pre-specified task_profile.
if (node->backend == GGML_BACKEND_GPU || if (node->backend == GGML_BACKEND_GPU || node->backend == GGML_BACKEND_GPU_SPLIT) {
node->backend == GGML_BACKEND_GPU_SPLIT) { if (node->op != GGML_OP_MUL_MAT && node->op != GGML_OP_OUT_PROD) {
enum ggml_task_backend be; enum ggml_task_backend be;
if (ggml_cpu_has_cublas()) { if (ggml_cpu_has_cublas()) {
be = GGML_TASK_BACKEND_GPU_CUDA; be = GGML_TASK_BACKEND_GPU_CUDA;
@ -15886,23 +15876,28 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
} }
} }
GGML_ASSERT(profile); GGML_ASSERT(profile);
} else { GGML_ASSERT(!cgraph->tune);
GGML_ASSERT(node->backend == GGML_BACKEND_CPU);
memcpy(&node->task_profile, profile, sizeof(struct ggml_task_profile));
return;
}
} }
bool profile_copied = false; // mul_mat: GGML_OP_MUL_MAT and GGML_OP_OUT_PROD.
if (node->op == GGML_OP_MUL_MAT) { if (node->op == GGML_OP_MUL_MAT) {
#if defined(GGML_USE_TUNE) #if defined(GGML_USE_TUNE)
GGML_ASSERT(node->backend == GGML_BACKEND_CPU);
int stages_time_us[3]; int stages_time_us[3];
profile = ggml_mulmat_get_task_profile( profile = ggml_mulmat_get_task_profile(node, profiles, n_profiles,
node, profiles, n_profiles, cgraph->tune, stages_time_us); cgraph->tune, stages_time_us);
GGML_ASSERT(profile); GGML_ASSERT(profile);
memcpy(&node->task_profile, profile, sizeof(struct ggml_task_profile));
if (cgraph->tune) { if (cgraph->tune) {
memcpy(&node->task_profile, profile, memcpy(&node->task_profile, profile,
sizeof(struct ggml_task_profile)); sizeof(struct ggml_task_profile));
profile_copied = true;
// Do not wait if the estimated execution time is too small // Do not wait if the estimated execution time is too small
// (e.g. less than 0.1 ms) // (e.g. less than 0.1 ms)
@ -15916,45 +15911,69 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
} }
} }
} }
return;
#else #else
profile = ggml_mulmat_get_task_profile(node, profiles, profile = ggml_mulmat_get_task_profile(node, profiles, n_profiles, NULL,
n_profiles, NULL, NULL); NULL);
GGML_ASSERT(profile); GGML_ASSERT(profile);
memcpy(&node->task_profile, profile, sizeof(struct ggml_task_profile));
return;
#endif #endif
} else if (node->op == GGML_OP_OUT_PROD) { // FIXME: is is right? } else if (node->op == GGML_OP_OUT_PROD) { // FIXME: is this correct?
profile = ggml_mulmat_get_task_profile(node, profiles, profile = ggml_mulmat_get_task_profile(node, profiles, n_profiles, NULL,
n_profiles, NULL, NULL); NULL);
GGML_ASSERT(profile); GGML_ASSERT(profile);
} else { memcpy(&node->task_profile, profile, sizeof(struct ggml_task_profile));
profile = &profiles[0]; return;
GGML_ASSERT(profile->stages[1].backend ==
GGML_TASK_BACKEND_CPU);
} }
if (!profile_copied) { // default.
memcpy(&node->task_profile, profile, profile = &profiles[0];
sizeof(struct ggml_task_profile)); GGML_ASSERT(profile->stages[1].backend == GGML_TASK_BACKEND_CPU);
memcpy(&node->task_profile, profile, sizeof(struct ggml_task_profile));
} }
void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) {
int n_threads = cgraph->n_threads;
struct ggml_threading_context *thrd_ctx = ggml_threading_start(
n_threads, NULL, ggml_compute_forward,
GGML_THREADING_FEATURE_WAIT_ON_DONE, NULL);
// initialize tasks + work buffer
{
// int64_t t0 = ggml_time_us();
size_t work_size = 0;
// thread scheduling for the different operations
for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * node = cgraph->nodes[i];
GGML_ASSERT (node->op != GGML_OP_NONE);
struct ggml_task_stage *stages = node->task_profile.stages; struct ggml_task_stage *stages = node->task_profile.stages;
// Workrounnd to set node->backend. ggml_graph_compute_set_tensor_task_proile(node, cgraph);
for (int j = 0; j < 3; j++) {
if (node->backend == GGML_BACKEND_CPU && //
(stages[j].backend & GGML_TASK_BACKEND_GPU)) { // Allocate temp buffer `wdata` for CPU.
if (ggml_cpu_has_cublas() || ggml_cpu_has_clblast()) { // NOTE: GPU MAY fallback to CPU, so we have to cover all possible cases.
node->backend = GGML_BACKEND_GPU; //
} else {
GGML_ASSERT(false); if (node->task_profile.get_wsize) {
} int sz = node->task_profile.get_wsize(node);
if (sz >= 0) {
work_size = MAX(work_size, (size_t)sz);
continue;
} }
} }
//printf("op: %d, comp backend: %d\n", node->op, node->task_profile.stages[1].backend);
// compute stage n_tasks. // compute stage n_tasks.
int n_tasks = stages[1].parallel ? n_threads : 1; int n_tasks = stages[1].parallel ? n_threads : 1;
// Allocate temp buffer `wdata` for CPU.
// NOTE: GPU MAY fallback to CPU, so we have to cover all possible cases.
switch (node->op) { switch (node->op) {
case GGML_OP_CPY: case GGML_OP_CPY:
case GGML_OP_DUP: case GGML_OP_DUP:
@ -16012,20 +16031,12 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
{ {
} break; } break;
case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT:
case GGML_OP_OUT_PROD: // FIXME: is is right? case GGML_OP_OUT_PROD: // FIXME: is this correct?
{ {
size_t cur = 0; size_t cur = 0;
enum ggml_task_backend comp_backend = stages[GGML_TASK_COMPUTE].backend; enum ggml_task_backend comp_backend = stages[GGML_TASK_COMPUTE].backend;
GGML_ASSERT(comp_backend != GGML_TASK_BACKEND_NONE); GGML_ASSERT(comp_backend != GGML_TASK_BACKEND_NONE);
if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) {
if (comp_backend == GGML_TASK_BACKEND_GPU_CL) {
#if defined(GGML_USE_CLBLAST)
GGML_ASSERT(ggml_cl_can_mul_mat(node->src0, node->src1, node));
cur = ggml_cl_mul_mat_get_wsize(node->src0, node->src1, node);
#else
GGML_ASSERT(false);
#endif
} else if (comp_backend == GGML_TASK_BACKEND_CPU_BLAS) {
GGML_ASSERT(ggml_cpu_has_cpublas()); GGML_ASSERT(ggml_cpu_has_cpublas());
GGML_ASSERT(node->src1->type == GGML_TYPE_F32); GGML_ASSERT(node->src1->type == GGML_TYPE_F32);
@ -16039,11 +16050,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
} else { } else {
GGML_ASSERT(false); GGML_ASSERT(false);
} }
} else if (comp_backend == GGML_TASK_BACKEND_CPU || comp_backend == GGML_TASK_BACKEND_GPU_CUDA) { } else { // CPU or GPU fallback
if (comp_backend == GGML_TASK_BACKEND_GPU_CUDA) {
GGML_ASSERT(ggml_cpu_has_cublas());
}
GGML_ASSERT(node->src1->type == GGML_TYPE_F32); GGML_ASSERT(node->src1->type == GGML_TYPE_F32);
if (node->src0->type == GGML_TYPE_F32) { if (node->src0->type == GGML_TYPE_F32) {
@ -16056,8 +16063,6 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
} else { } else {
GGML_ASSERT(false); GGML_ASSERT(false);
} }
} else {
GGML_ASSERT(false);
} }
work_size = MAX(work_size, cur); work_size = MAX(work_size, cur);

29
ggml.h
View File

@ -390,11 +390,40 @@ extern "C" {
bool wait; bool wait;
}; };
struct ggml_tensor;
struct ggml_compute_params;
// Compute errors.
enum ggml_compute_error {
GGML_COMPUTE_OK = 0,
GGML_COMPUTE_FALLBACK = 1,
};
// The task runner to be called by main thread and workers.
typedef enum ggml_compute_error(ggml_task_runner)(
const struct ggml_compute_params *params,
struct ggml_tensor *node);
// Get wsize for node computing.
// When return -1: should be explained as `fallback to CPU`, caller MUST
// determine how much memory to reserve for this node.
typedef int (ggml_task_get_wsize)(struct ggml_tensor *tensor);
// config for computing a tensor. // config for computing a tensor.
struct ggml_task_profile { struct ggml_task_profile {
// profile id, start from 1.
int id;
// index 0: INIT, 1: COMPUTE, 2: FINALIZE // index 0: INIT, 1: COMPUTE, 2: FINALIZE
struct ggml_task_stage stages[3]; struct ggml_task_stage stages[3];
// Optional task runner, overrides threading's task runner.
ggml_task_runner *runner;
// Optional function to return required wsize for wdata.
ggml_task_get_wsize *get_wsize;
// Optional flag for development.
// MUST be used only in testing codes. // MUST be used only in testing codes.
uint8_t dev_flags[4]; uint8_t dev_flags[4];
}; };

61
tests/test-ggml-threading.c
View File

@ -42,7 +42,8 @@ static const int n_repeat = 10;
static int work_done_arr[MAX_N_THREADS]; static int work_done_arr[MAX_N_THREADS];
static enum ggml_compute_error static enum ggml_compute_error
mock_task_runner(struct ggml_compute_params *params, struct ggml_tensor *node) { mock_task_runner(const struct ggml_compute_params *params,
struct ggml_tensor *node) {
int64_t loops = node->task_profile.dev_flags[1] * 1000 * 1000; int64_t loops = node->task_profile.dev_flags[1] * 1000 * 1000;
if (node->task_profile.stages[params->type].parallel) { if (node->task_profile.stages[params->type].parallel) {
loops /= params->nth; loops /= params->nth;
@ -79,9 +80,8 @@ int test_driver(int id, struct ggml_tensor *node, int n_threads) {
int t0 = (int)ggml_time_us(); int t0 = (int)ggml_time_us();
struct ggml_threading_context *ctx = struct ggml_threading_context *ctx = ggml_threading_start(
ggml_threading_start(n_threads, ggml_threading_graph_compute_thread, n_threads, NULL, mock_task_runner, features, /*stages_time*/ NULL);
mock_task_runner, features, /*stages_time*/ NULL);
int t1 = (int)ggml_time_us(); int t1 = (int)ggml_time_us();
@ -141,7 +141,7 @@ int test_driver(int id, struct ggml_tensor *node, int n_threads) {
} }
static enum ggml_compute_error static enum ggml_compute_error
mock_task_runner_fallback(struct ggml_compute_params *params, mock_task_runner_fallback(const struct ggml_compute_params *params,
struct ggml_tensor *node) { struct ggml_tensor *node) {
UNUSED(params); UNUSED(params);
if (node->backend == GGML_BACKEND_GPU) { if (node->backend == GGML_BACKEND_GPU) {
@ -158,7 +158,7 @@ mock_task_runner_fallback(struct ggml_compute_params *params,
// thus it is not parallelled. // thus it is not parallelled.
int test_fallback(struct ggml_tensor *node) { int test_fallback(struct ggml_tensor *node) {
struct ggml_threading_context *ctx = ggml_threading_start( struct ggml_threading_context *ctx = ggml_threading_start(
1, ggml_threading_graph_compute_thread, mock_task_runner_fallback, 1, NULL, mock_task_runner_fallback,
/*features*/ GGML_THREADING_FEATURE_NONE, /*stages_time*/ NULL); /*features*/ GGML_THREADING_FEATURE_NONE, /*stages_time*/ NULL);
enum ggml_compute_error err = enum ggml_compute_error err =
@ -177,6 +177,38 @@ int test_fallback(struct ggml_tensor *node) {
return 0; return 0;
} }
static enum ggml_compute_error
customized_node_runner(const struct ggml_compute_params *params,
struct ggml_tensor *node) {
UNUSED(params);
// Reset runner thus caller will know it was called.
node->task_profile.runner = NULL;
return GGML_COMPUTE_OK;
}
// Test when node->task_profile.runner is not NULL.
int test_customized_node_runner(struct ggml_tensor *node) {
struct ggml_threading_context *ctx = ggml_threading_start(
1, NULL, mock_task_runner,
/*features*/ GGML_THREADING_FEATURE_NONE, /*stages_time*/ NULL);
node->task_profile.runner = customized_node_runner;
enum ggml_compute_error err =
ggml_threading_compute_tensor(ctx, node, /*wdata*/ NULL, /*wsize*/ 0);
ggml_threading_stop(ctx);
if (err != GGML_COMPUTE_OK) {
// should not happen.
abort();
}
if (node->task_profile.runner != NULL) {
return 2;
}
return 0;
}
int main(void) { int main(void) {
ggml_time_init(); ggml_time_init();
@ -367,7 +399,10 @@ int main(void) {
} }
} }
// fallback
{ {
printf("[test-ggml-threading] test fallback ...\n");
++n_tests; ++n_tests;
// required by getting task profiles. // required by getting task profiles.
@ -382,9 +417,21 @@ int main(void) {
node.src1 = &src1; node.src1 = &src1;
node.backend = GGML_BACKEND_GPU; node.backend = GGML_BACKEND_GPU;
stages[1].backend = GGML_TASK_BACKEND_GPU;
if (test_fallback(&node) == 0) { if (test_fallback(&node) == 0) {
++n_passed; ++n_passed;
printf("\n[test-ggml-threading] test fallback: ok\n\n"); printf("[test-ggml-threading] test fallback: ok\n\n");
}
}
// customized node runner
{
printf("[test-ggml-threading] test customized node runner ...\n");
++n_tests;
if (test_customized_node_runner(&node) == 0) {
++n_passed;
printf("[test-ggml-threading] test customized node runner: ok\n\n");
} }
} }

10
tests/test-ggml-tune.c
View File

@ -72,7 +72,9 @@ static int bench(void) {
// GGML_FTYPE_ALL_F32, // GGML_FTYPE_ALL_F32,
// GGML_FTYPE_MOSTLY_F16, // GGML_FTYPE_MOSTLY_F16,
GGML_FTYPE_MOSTLY_Q4_0, GGML_FTYPE_MOSTLY_Q4_0,
#if defined(GGML_USE_K_QUANTS)
GGML_FTYPE_MOSTLY_Q4_K, GGML_FTYPE_MOSTLY_Q4_K,
#endif
}; };
int n_ftypes = sizeof(ftypes) / sizeof(ftypes[0]); int n_ftypes = sizeof(ftypes) / sizeof(ftypes[0]);
@ -132,16 +134,14 @@ int estimate_time_non_zero_NK(void) {
int time[3]; // 3 profiles. int time[3]; // 3 profiles.
}; };
struct ggml_mulmat_tune tune = { struct ggml_mulmat_tune tune;
.version = 1,
.ftype = GGML_FTYPE_MOSTLY_Q4_0,
};
enum ggml_ftype ftype = GGML_FTYPE_MOSTLY_Q4_0;
const int m_num = 2; const int m_num = 2;
const int n_threads = 1; // useless. const int n_threads = 1; // useless.
struct ggml_mulmat_tune_params params; struct ggml_mulmat_tune_params params;
init_params(&params, tune.ftype, m_num, n_threads); init_params(&params, ftype, m_num, n_threads);
ggml_mulmat_tune_init(&tune, &params, ggml_task_profiles_mock_qxx_provider); ggml_mulmat_tune_init(&tune, &params, ggml_task_profiles_mock_qxx_provider);