vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : add GGML_PAD_REFLECT_1D operation (ggml/1034)

Browse source 
* ggml_pad_reflect_1d defined in header

* implemented on CPU

* called the forward pass

* impl Metal kernel

* added Metal kernel

* added OP_PAD_REFLECT_1D in test-backend-ops.cpp

* add test-pad-reflect-1d test case

* test case support multiple backend
This commit is contained in:
PAB 2024-12-03 20:20:04 +01:00 • committed by Georgi Gerganov
parent d405804be8
commit c2082d93a8
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
6 changed files with 192 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

37
ggml/src/ggml.c
View file

@ -950,6 +950,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"POOL_2D_BACK",
"UPSCALE",
"PAD",
"PAD_REFLECT_1D",
"ARANGE",
"TIMESTEP_EMBEDDING",
"ARGSORT",
@ -983,7 +984,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"OPT_STEP_ADAMW",
};
static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@ -1045,6 +1046,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"pool_2d_back(x)",
"upscale(x)",
"pad(x)",
"pad_reflect_1d(x)",
"arange(start, stop, step)",
"timestep_embedding(timesteps, dim, max_period)",
"argsort(x)",
@ -1078,7 +1080,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"adamw(x)",
};
static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@ -4097,6 +4099,37 @@ struct ggml_tensor * ggml_pad(
return result;
}
// ggml_pad_reflect_1d
struct ggml_tensor * ggml_pad_reflect_1d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int p0,
int p1) {
GGML_ASSERT(p0 >= 0);
GGML_ASSERT(p1 >= 0);
GGML_ASSERT(p0 < a->ne[0]); // padding length on each size must be less than the
GGML_ASSERT(p1 < a->ne[0]); // existing length of the dimension being padded
GGML_ASSERT(ggml_is_contiguous(a));
GGML_ASSERT(a->type == GGML_TYPE_F32);
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
a->ne[0] + p0 + p1,
a->ne[1],
a->ne[2],
a->ne[3]);
int32_t params[] = { p0, p1 };
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_PAD_REFLECT_1D;
result->src[0] = a;
return result;
}
// ggml_arange
struct ggml_tensor * ggml_arange(