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

ggml: apply the unpad operator patch

Browse source 
Signed-off-by: YiYing He <yiying@secondstate.io>
This commit is contained in:
YiYing He 2025-01-15 17:08:54 +08:00
parent 45a89e0cec
commit 8bb33d3285
8 changed files with 220 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

58
ggml/src/ggml-cpu/ggml-cpu.c
View file

@ -10660,6 +10660,59 @@ static void ggml_compute_forward_pad_reflect_1d(
}
}
static void ggml_compute_forward_unpad_f32(
const struct ggml_compute_params *params,
struct ggml_tensor *dst) {
const struct ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT( dst->nb[0] == sizeof(float));
const int ith = params->ith;
const int nth = params->nth;
GGML_TENSOR_UNARY_OP_LOCALS
float * dst_ptr = (float *) dst->data;
// TODO: optimize
for (int64_t i2 = 0; i2 < ne2; ++i2) {
for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
for (int64_t i0 = 0; i0 < ne0; ++i0) {
for (int64_t i3 = 0; i3 < ne3; ++i3) {
const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
dst_ptr[dst_idx] = *src_ptr;
}
}
}
}
}
}
static void ggml_compute_forward_unpad(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const struct ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_unpad_f32(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_arange
static void ggml_compute_forward_arange_f32(
@ -12953,6 +13006,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_pad_reflect_1d(params, tensor);
} break;
case GGML_OP_UNPAD:
{
ggml_compute_forward_unpad(params, tensor);
} break;
case GGML_OP_ARANGE:
{
ggml_compute_forward_arange(params, tensor);
@ -13300,6 +13357,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_PAD_REFLECT_1D:
case GGML_OP_UNPAD:
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_ARGSORT: