Skip to content
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

Jump to bottom

Improve ADD_REL_POS perf in SAM by doing it inplace #466

Merged
merged 7 commits into from
Aug 21, 2023
Merged

Improve ADD_REL_POS perf in SAM by doing it inplace #466

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
58d9081
Improve ADD_REL_POS perf in SAM by doing it inplace
Aug 20, 2023
9d6eaa8
Add non-inplace version for the GGML_OP_ADD_REL_POS
Aug 21, 2023
ca3e0fc
Fix map_unary warnings and refactor LayerNorm2d + remove ggml_cont in it
Aug 21, 2023
d65d74d
Fix Mac printf format warnings
Aug 21, 2023
6e13e28
sam : add ggml_graph_print() comment
ggerganov Aug 21, 2023
d4b3b46
ggml : add broadcast support for BLAS ggml_mul_mat() (#460)
ggerganov Aug 21, 2023
415dbf1
Remove not needed build_forward_expand from add-rel-pos unit test
Aug 21, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
Prev Previous commit
Next Next commit
Add non-inplace version for the GGML_OP_ADD_REL_POS
  • Loading branch information
YavorGIvanov committed Aug 21, 2023
commit 9d6eaa8388ff7f88298e5e18e8eced47e2831451
6 changes: 6 additions & 0 deletions include/ggml/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1385,6 +1385,12 @@ extern "C" {

// used in sam

GGML_API struct ggml_tensor * ggml_add_rel_pos(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph);

GGML_API struct ggml_tensor * ggml_add_rel_pos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
Expand Down
55 changes: 42 additions & 13 deletions src/ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3930,6 +3930,7 @@ static void ggml_setup_op_has_task_pass(void) {
p[GGML_OP_CONV_TRANSPOSE_2D ] = true;
p[GGML_OP_FLASH_ATTN_BACK ] = true;
p[GGML_OP_CROSS_ENTROPY_LOSS ] = true;
p[GGML_OP_ADD_REL_POS ] = true;
}

{ // FINALIZE
Expand Down Expand Up @@ -7328,34 +7329,57 @@ struct ggml_tensor * ggml_get_rel_pos(

// ggml_add_rel_pos

struct ggml_tensor * ggml_add_rel_pos_inplace(
static struct ggml_tensor * ggml_add_rel_pos_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph) {
GGML_ASSERT(pw->ne[0] == ph->ne[0]);
GGML_ASSERT(pw->ne[1] == ph->ne[1]);
GGML_ASSERT(pw->ne[2] == ph->ne[2]);
GGML_ASSERT(pw->ne[3] == ph->ne[3]);
GGML_ASSERT(pw->ne[3] == a->ne[2]);
GGML_ASSERT(pw->ne[0]*ph->ne[0] == a->ne[0]);
GGML_ASSERT(pw->ne[1]*pw->ne[2] == a->ne[1]);
struct ggml_tensor * ph,
bool inplace) {
GGML_ASSERT(ggml_are_same_shape(pw, ph));
GGML_ASSERT(ggml_is_contiguous(a));
GGML_ASSERT(ggml_is_contiguous(pw));
GGML_ASSERT(ggml_is_contiguous(ph));
GGML_ASSERT(pw->type == GGML_TYPE_F32);
GGML_ASSERT(ph->type == GGML_TYPE_F32);
GGML_ASSERT(pw->type == GGML_TYPE_F32);
GGML_ASSERT(pw->ne[3] == a->ne[2]);
GGML_ASSERT(pw->ne[0]*pw->ne[0] == a->ne[0]);
GGML_ASSERT(pw->ne[1]*pw->ne[2] == a->ne[1]);

bool is_node = false;

if (!inplace && (a->grad || pw->grad || ph->grad)) {
is_node = true;
}

struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params_i32(result, 0, inplace ? 1 : 0);

struct ggml_tensor * result = ggml_view_tensor(ctx, a);
result->op = GGML_OP_ADD_REL_POS;
result->grad = NULL;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = pw;
result->src[2] = ph;

return result;
}


struct ggml_tensor * ggml_add_rel_pos(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph) {
return ggml_add_rel_pos_impl(ctx, a, pw, ph, false);
}

struct ggml_tensor * ggml_add_rel_pos_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * pw,
struct ggml_tensor * ph) {
return ggml_add_rel_pos_impl(ctx, a, pw, ph, true);
}

// gmml_unary

static struct ggml_tensor * ggml_unary_impl(
Expand Down Expand Up @@ -14607,7 +14631,12 @@ static void ggml_compute_forward_add_rel_pos_f32(
const struct ggml_tensor * src1,
const struct ggml_tensor * src2,
struct ggml_tensor * dst) {
UNUSED(src0);

const bool inplace = (bool) ((int32_t *) dst->op_params)[0];
if (!inplace && params->type == GGML_TASK_INIT) {
memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst));
return;
}
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
return;
}
Expand Down
22 changes: 15 additions & 7 deletions tests/test-rel-pos.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,21 +59,29 @@ int main(int argc, const char** argv) {
struct ggml_tensor * rw_f32 = ggml_cpy(ctx, rw, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 3, 2, 2));
struct ggml_tensor * rh_f32 = ggml_cpy(ctx, rh, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 3, 2, 2));

struct ggml_tensor * out = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 9, 4);
float* out_d = (float*)out->data;
for (int i = 0; i < ggml_nelements(out); ++i) {
out_d[i] = 1.f;
struct ggml_tensor * in = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 9, 4);
struct ggml_tensor * out_inplace = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 9, 4);
float * in_d = (float*)in->data;
float * out_inplace_d = (float*)out_inplace->data;
for (int i = 0; i < ggml_nelements(in); ++i) {
in_d[i] = 1.f;
out_inplace_d[i] = 1.f;
}

out = ggml_add_rel_pos_inplace(ctx, out, rw_f32, rh_f32);

struct ggml_tensor * out = ggml_add_rel_pos(ctx, in, rw_f32, rh_f32);
struct ggml_cgraph gf = ggml_build_forward(out);
ggml_build_forward_expand(&gf, rw_f32);
ggml_build_forward_expand(&gf, rh_f32);

ggml_graph_compute_with_ctx(ctx, &gf, 1);

out_inplace = ggml_add_rel_pos_inplace(ctx, out_inplace, rw_f32, rh_f32);
struct ggml_cgraph gf_2 = ggml_build_forward(out_inplace);
ggml_build_forward_expand(&gf_2, rw_f32);
ggml_build_forward_expand(&gf_2, rh_f32);
ggml_graph_compute_with_ctx(ctx, &gf_2, 1);

check_tensor(out, (float*)expected_out, 9, 4, 1);
check_tensor(out_inplace, (float*)expected_out, 9, 4, 1);
}

return 0;
Expand Down