nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

CUDA & CPU: support F32 kernel type for CONV_TRANSPOSE_2D (#17094)

Browse Source 
* Refactor CUDA 2D transpose implementation to support multiple kernel types and improve parameter handling

- Introduced a `conv2d_transpose_params` struct for better parameter management.
- Updated `conv2d_transpose_kernel` to be templated for different kernel types (float and half).
- Modified `ggml_cuda_conv_2d_transpose_p0` to handle both F16 and F32 kernel types.
- Enhanced test cases to validate functionality for both kernel types.

* Refactor test cases for 2D convolution transpose to support dynamic kernel types

- Updated `test_conv_transpose_2d` structure to improve parameter handling by reordering constructor arguments.
- Enhanced test case generation to iterate over kernel types, allowing for flexible testing of different configurations.
- Removed hardcoded kernel type instances in favor of a loop for better maintainability and scalability.

* Refactor ggml_compute_forward_conv_transpose_2d to support both F16 and F32 tensor types.

* Refactor conv2d transpose kernel to use a template for kernel type, enhancing flexibility for different data types.
Update test cases to include both F16 and F32 tensor types for comprehensive coverage.

* Update ggml/src/ggml-cuda/conv2d-transpose.cu

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

* Update ggml/src/ggml-cpu/ggml-cpu.c

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

* Refactor conv2d transpose implementation by removing the conv2d_transpose_params struct and dispatching with direct kernel launch.

* Enhance cpu conv2d transpose implementation by introducing a templated kernel type for improved flexibility with F16 and F32 data types.

---------

Co-authored-by: Aman Gupta <amangupta052@gmail.com>
This commit is contained in:
Yihao Wang
2026-03-26 10:19:14 +08:00
committed by GitHub
co-authored by Aman Gupta
parent c0159f9c1f
commit 0a524f2404
5 changed files with 122 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ggml-cpu.c
+6 -2
View File
@@ -2871,8 +2871,12 @@ struct ggml_cplan ggml_graph_plan(
const int64_t ne11 = node->src[1]->ne[1]; // H
const int64_t ne12 = node->src[1]->ne[2]; // Channels In
cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
GGML_ASSERT(node->src[0]->type == GGML_TYPE_F16 || node->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(node->src[1]->type == GGML_TYPE_F32);
cur += ggml_type_size(node->src[0]->type) * ne00 * ne01 * ne02 * ne03;
cur += ggml_type_size(node->src[0]->type) * ne10 * ne11 * ne12;
} break;
case GGML_OP_TOP_K:
{
ggml/src/ggml-cpu/ops.cpp
+50 -19
View File
@@ -6923,16 +6923,15 @@ void ggml_compute_forward_conv_3d(
ggml_compute_forward_conv_3d_impl(params, src0, src1, dst, src0->type);
}
// ggml_compute_forward_conv_transpose_2d
void ggml_compute_forward_conv_transpose_2d(
const ggml_compute_params * params,
ggml_tensor * dst) {
template <typename kernel_t>
static void ggml_compute_forward_conv_transpose_2d_impl(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -6943,7 +6942,7 @@ void ggml_compute_forward_conv_transpose_2d(
const int nk = ne00*ne01*ne02*ne03;
GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
GGML_ASSERT(nb00 == ggml_type_size(src0->type));
GGML_ASSERT(nb10 == sizeof(float));
if (ith == 0) {
@@ -6951,12 +6950,12 @@ void ggml_compute_forward_conv_transpose_2d(
// permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout)
{
ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
kernel_t * const wdata = (kernel_t *) params->wdata + 0;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i03*nb03 + i02*nb02);
ggml_fp16_t * dst_data = wdata + i02*ne01*ne00*ne03;
const kernel_t * const src = (kernel_t *)((char *) src0->data + i03*nb03 + i02*nb02);
kernel_t * dst_data = wdata + i02*ne01*ne00*ne03;
for (int64_t i01 = 0; i01 < ne01; i01++) {
for (int64_t i00 = 0; i00 < ne00; i00++) {
dst_data[i01*ne00*ne03 + i00*ne03 + i03] = src[i01 * ne00 + i00];
@@ -6968,13 +6967,17 @@ void ggml_compute_forward_conv_transpose_2d(
// permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh)
{
ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk;
kernel_t * const wdata = (kernel_t *) params->wdata + nk;
for (int i12 = 0; i12 < ne12; i12++) {
for (int i11 = 0; i11 < ne11; i11++) {
const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11);
ggml_fp16_t * dst_data = wdata + i11*ne10*ne12;
kernel_t * dst_data = wdata + i11*ne10*ne12;
for (int i10 = 0; i10 < ne10; i10++) {
dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
} else {
dst_data[i10*ne12 + i12] = src[i10];
}
}
}
}
@@ -6996,21 +6999,27 @@ void ggml_compute_forward_conv_transpose_2d(
const int ip0 = dp*ith;
const int ip1 = MIN(ip0 + dp, np);
ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
ggml_fp16_t * const wdata_src = wdata + nk;
kernel_t * const wdata = (kernel_t *) params->wdata + 0;
kernel_t * const wdata_src = wdata + nk;
for (int i2 = ip0; i2 < ip1; i2++) { // Cout
float * dst_data = (float *)((char *) dst->data + i2*nb2);
ggml_fp16_t * wdata_kernel = wdata + i2*ne01*ne00*ne03;
kernel_t * wdata_kernel = wdata + i2*ne01*ne00*ne03;
for (int i11 = 0; i11 < ne11; i11++) {
for (int i10 = 0; i10 < ne10; i10++) {
const int i1n = i11*ne10*ne12 + i10*ne12;
for (int i01 = 0; i01 < ne01; i01++) {
for (int i00 = 0; i00 < ne00; i00++) {
float v = 0;
ggml_vec_dot_f16(ne03, &v, 0,
wdata_src + i1n, 0,
wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
ggml_vec_dot_f16(ne03, &v, 0,
wdata_src + i1n, 0,
wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
} else {
ggml_vec_dot_f32(ne03, &v, 0,
wdata_src + i1n, 0,
wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
}
dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
}
}
@@ -7019,6 +7028,28 @@ void ggml_compute_forward_conv_transpose_2d(
}
}
void ggml_compute_forward_conv_transpose_2d(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F16:
{
ggml_compute_forward_conv_transpose_2d_impl<ggml_fp16_t>(params, dst);
} break;
case GGML_TYPE_F32:
{
ggml_compute_forward_conv_transpose_2d_impl<float>(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_conv_2d_dw
struct ggml_conv_2d_dw_params {