CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-zdnn: attempt at manually changing the layout

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-07-30 15:33:13 +08:00
parent ad0cb30212
commit 7b50d057dd
1 changed files with 17 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
ggml/src/ggml-zdnn/ggml-zdnn-rewrite.cpp
View File

@@ -55,10 +55,6 @@ inline void ggml_zdnn_load_tensor(zdnn_ztensor & ztensor,
inline void ggml_zdnn_init_tensor(ggml_backend_zdnn_buffer * buffer, const ggml_tensor * tensor) {
switch (tensor->op) {
case GGML_OP_NONE:
// noop here because we will initialise it during the compute graph execution
return;
case GGML_OP_MUL_MAT:
{
zdnn_init_pre_transformed_desc(
@@ -133,12 +129,24 @@ static void ggml_zdnn_mul_mat_op(ggml_backend_zdnn_context * ctx, const ggml_ten
const int64_t bias_dim [GGML_MAX_DIMS] = { 1, 1, 1, output_cols };
const int64_t output_dim[GGML_MAX_DIMS] = { 1, 1, output_cols, output_rows };
ggml_zdnn_create_tensor(weights_extra->pre_tfm_desc, weights_extra->tfm_desc, weights_extra->ztensor, weights, weights_dim, ZDNN_2D);
ggml_zdnn_create_tensor(inputs_extra->pre_tfm_desc, inputs_extra->tfm_desc, inputs_extra->ztensor, inputs, inputs_dim, ZDNN_2D);
zdnn_init_pre_transformed_desc(ZDNN_2D,
ggml_zdnn_type_mapping(weights->type),
&weights_extra->pre_tfm_desc,
weights_dim[3], weights_dim[2],
weights_dim[1], weights_dim[0]);
zdnn_init_pre_transformed_desc(ZDNN_2D,
ggml_zdnn_type_mapping(inputs->type),
&inputs_extra->pre_tfm_desc,
inputs_dim[3], inputs_dim[2],
inputs_dim[1], inputs_dim[0]);
ZDNN_CHECK(zdnn_generate_transformed_desc(&weights_extra->pre_tfm_desc, &weights_extra->tfm_desc));
ZDNN_CHECK(zdnn_generate_transformed_desc(&inputs_extra->pre_tfm_desc, &inputs_extra->tfm_desc));
ggml_zdnn_create_tensor(ptd_bias, td_bias, zt_bias, output, bias_dim, ZDNN_1D);
// ggml_zdnn_create_tensor(ptd_output, td_output, zt_output, output, output_dim, ZDNN_2D);
std::raise(SIGINT);
void * bias_data = (void *)calloc(ne0, ggml_element_size(output));
ggml_zdnn_load_tensor(weights_extra->ztensor, weights->data);
ggml_zdnn_load_tensor(inputs_extra->ztensor, inputs->data);
@@ -161,13 +169,13 @@ static void ggml_zdnn_mul_mat_op(ggml_backend_zdnn_context * ctx, const ggml_ten
// inputs_extra->pre_tfm_desc.dim3,
// inputs_extra->pre_tfm_desc.dim4);
GGML_ASSERT(weights_extra->pre_tfm_desc.layout == ZDNN_2D && "weights_extra->pre_tfm_desc.layout must be ZDNN_2D");
GGML_ASSERT(inputs_extra->pre_tfm_desc.layout == ZDNN_2D && "inputs_extra->pre_tfm_desc.layout must be ZDNN_2D");
GGML_ASSERT(weights_extra->pre_tfm_desc.dim1 == weights->ne[0] && "weights_extra->pre_tfm_desc.dim1 must match weights->ne[0]");
GGML_ASSERT(weights_extra->pre_tfm_desc.dim2 == weights->ne[1] && "weights_extra->pre_tfm_desc.dim2 must match weights->ne[1]");
GGML_ASSERT(inputs_extra->pre_tfm_desc.dim1 == inputs->ne[0] && "inputs_extra->pre_tfm_desc.dim1 must match inputs->ne[0]");
GGML_ASSERT(inputs_extra->pre_tfm_desc.dim2 == inputs->ne[1] && "inputs_extra->pre_tfm_desc.dim2 must match inputs->ne[1]");
// std::raise(SIGINT);
ZDNN_CHECK(zdnn_matmul_transpose_op(&inputs_extra->ztensor, &weights_extra->ztensor, &zt_bias,
false, true, MATMUL_OP_ADDITION, &output_extra->ztensor));
ZDNN_CHECK(zdnn_transform_origtensor(&output_extra->ztensor, output->data));
@@ -413,7 +421,7 @@ static enum ggml_status ggml_backend_zdnn_buffer_init_tensor(ggml_backend_buffer
strncpy(zdnn_buffer->name, tensor->name, GGML_MAX_NAME - 1);
ggml_zdnn_init_tensor(zdnn_buffer.get(), tensor);
tensor->extra = zdnn_buffer.get(); // Stable pointer to heap-allocated object
tensor->extra = zdnn_buffer.get();
ctx->buffers.push_back(std::move(zdnn_buffer));
ctx->n_buffers++;
@@ -434,9 +442,6 @@ static void ggml_backend_zdnn_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
memcpy((char *)tensor->data + offset, data, size);
ggml_backend_zdnn_buffer * zdnn_buffer = (ggml_backend_zdnn_buffer *)tensor->extra;
if (tensor->op == GGML_OP_NONE) {
return;
}
ggml_zdnn_load_tensor(zdnn_buffer->ztensor, (void *)data);
GGML_UNUSED(buffer);