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

backup

Browse source
This commit is contained in:
Meng, Hengyu 2024-08-26 08:58:54 +00:00
parent b180cb352b
commit 6494509801
6 changed files with 81 additions and 33 deletions
Download patch file Download diff file Expand all files Collapse all files

11
ggml/include/ggml.h
View file

@ -1659,6 +1659,17 @@ extern "C" {
struct ggml_tensor * b,
int stride);
GGML_API struct ggml_tensor * ggml_conv_transpose_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s0,
int s1,
int p0,
int p1,
int d0,
int d1);
enum ggml_op_pool {
GGML_OP_POOL_MAX,
GGML_OP_POOL_AVG,

2
ggml/src/ggml-sycl.cpp
View file

@ -3895,7 +3895,7 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens
switch (tensor->op) {
case GGML_OP_CONV_TRANSPOSE_2D:
func = ggml_sycl_op_conv_2d;
func = ggml_sycl_op_conv_transpose_2d;
break;
case GGML_OP_CONV_TRANSPOSE_1D:
func = ggml_sycl_op_conv_transpose_1d;

2
ggml/src/ggml-sycl/conv.cpp
View file

@ -99,7 +99,7 @@ void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_
}
void ggml_sycl_op_conv_2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
void ggml_sycl_op_conv_transpose_2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst) {
const void * src0_d = (const void *)src0->data;
const void * src1_d = (const void *)src1->data;

2
ggml/src/ggml-sycl/conv.hpp
View file

@ -18,7 +18,7 @@
void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst);
void ggml_sycl_op_conv_2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
void ggml_sycl_op_conv_transpose_2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst);
#endif // GGML_SYCL_CONV_HPP

66
ggml/src/ggml.c
View file

@ -6770,35 +6770,7 @@ struct ggml_tensor * ggml_conv_2d(
int p1,
int d0,
int d1) {
#ifdef GGML_SYCL_DNNL
bool is_node = false;
if (a->grad || b->grad) {
GGML_ABORT("fatal error"); // TODO: implement backward
is_node = true;
}
const int64_t OH = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
const int64_t ne[4] = {
OW,
OH,
a->ne[3], // OC
b->ne[3], // N
};
struct ggml_tensor * result = ggml_new_tensor(ctx, b->type, 4, ne);
int32_t params[] = { s0, s1, p0, p1, d0, d1};
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_CONV_TRANSPOSE_2D;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = b;
return result;
#else
struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, GGML_TYPE_F16); // [N, OH, OW, IC * KH * KW]
struct ggml_tensor * result =
@ -6811,7 +6783,6 @@ struct ggml_tensor * ggml_conv_2d(
return result;
#endif
}
// ggml_conv_2d_sk_p0
@ -6837,6 +6808,43 @@ static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int
return (ins - 1) * s - 2 * p + ks;
}
struct ggml_tensor * ggml_conv_transpose_2d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int s0,
int s1,
int p0,
int p1,
int d0,
int d1) {
GGML_ASSERT(a->ne[3] == b->ne[2]);
bool is_node = false;
if (a->grad || b->grad) {
GGML_ABORT("fatal error"); // TODO: implement backward
is_node = true;
}
const int64_t ne[4] = {
ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1),
ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0),
a->ne[2], b->ne[3],
};
struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
int32_t params[] = { s0, s1, p0, p1, d0, d1};
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_CONV_TRANSPOSE_2D;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = b;
return result;
}
struct ggml_tensor * ggml_conv_transpose_2d_p0(
struct ggml_context * ctx,
struct ggml_tensor * a,