happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

cpu: skip NOPs to avoid barriers (#17133) 

* cpu: skip NOPs to avoid barriers

* cpu: use ggml_op_is_empty
This commit is contained in:
Max Krasnyansky 2025-11-10 12:44:49 -08:00 committed by GitHub
parent 13730c183b
commit 395e286bc9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 21 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -1807,22 +1807,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{ {
ggml_compute_forward_cont(params, tensor); ggml_compute_forward_cont(params, tensor);
} break; } break;
case GGML_OP_RESHAPE:
{
ggml_compute_forward_reshape(params, tensor);
} break;
case GGML_OP_VIEW:
{
ggml_compute_forward_view(params, tensor);
} break;
case GGML_OP_PERMUTE:
{
ggml_compute_forward_permute(params, tensor);
} break;
case GGML_OP_TRANSPOSE:
{
ggml_compute_forward_transpose(params, tensor);
} break;
case GGML_OP_GET_ROWS: case GGML_OP_GET_ROWS:
{ {
ggml_compute_forward_get_rows(params, tensor); ggml_compute_forward_get_rows(params, tensor);
@ -2042,6 +2026,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{ {
// nop // nop
} break; } break;
case GGML_OP_RESHAPE:
{
// nop
} break;
case GGML_OP_PERMUTE:
{
// nop
} break;
case GGML_OP_VIEW:
{
// nop
} break;
case GGML_OP_TRANSPOSE:
{
// nop
} break;
case GGML_OP_COUNT: case GGML_OP_COUNT:
{ {
GGML_ABORT("fatal error"); GGML_ABORT("fatal error");
@ -2884,6 +2884,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) { for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
struct ggml_tensor * node = cgraph->nodes[node_n]; struct ggml_tensor * node = cgraph->nodes[node_n];
if (ggml_op_is_empty(node->op)) {
// skip NOPs
continue;
}
ggml_compute_forward(&params, node); ggml_compute_forward(&params, node);
if (state->ith == 0 && cplan->abort_callback && if (state->ith == 0 && cplan->abort_callback &&

40
ggml/src/ggml-cpu/ops.cpp
View File

@ -4455,46 +4455,6 @@ void ggml_compute_forward_cont(
ggml_compute_forward_dup(params, dst); ggml_compute_forward_dup(params, dst);
} }
// ggml_compute_forward_reshape
void ggml_compute_forward_reshape(
const ggml_compute_params * params,
ggml_tensor * dst) {
// NOP
GGML_UNUSED(params);
GGML_UNUSED(dst);
}
// ggml_compute_forward_view
void ggml_compute_forward_view(
const ggml_compute_params * params,
ggml_tensor * dst) {
// NOP
GGML_UNUSED(params);
GGML_UNUSED(dst);
}
// ggml_compute_forward_permute
void ggml_compute_forward_permute(
const ggml_compute_params * params,
ggml_tensor * dst) {
// NOP
GGML_UNUSED(params);
GGML_UNUSED(dst);
}
// ggml_compute_forward_transpose
void ggml_compute_forward_transpose(
const ggml_compute_params * params,
ggml_tensor * dst) {
// NOP
GGML_UNUSED(params);
GGML_UNUSED(dst);
}
// ggml_compute_forward_get_rows // ggml_compute_forward_get_rows
static void ggml_compute_forward_get_rows_q( static void ggml_compute_forward_get_rows_q(

4
ggml/src/ggml-cpu/ops.h
View File

@ -51,10 +51,6 @@ void ggml_compute_forward_scale(const struct ggml_compute_params * params, struc
void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst); void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);