CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-21 12:16:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-zdnn: clean up matmul codepath

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-09-06 22:06:06 +08:00
parent 9ed5947a12
commit 99311bc98f
1 changed files with 3 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
ggml/src/ggml-zdnn/ggml-zdnn.cpp
View File

@@ -160,19 +160,6 @@ static void ggml_zdnn_mul_mat_op(ggml_backend_zdnn_context * ctx, const ggml_ten
} }
static void ggml_zdnn_mul_mat_dispatch(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { static void ggml_zdnn_mul_mat_dispatch(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
bool use_mul_mat_vec =
(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F16)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q =
ggml_is_quantized(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
bool use_mul_mat_q =
ggml_is_quantized(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
// debug helpers // debug helpers
// GGML_LOG_INFO("%s: use_mul_mat_vec = %d\n", __func__, use_mul_mat_vec); // GGML_LOG_INFO("%s: use_mul_mat_vec = %d\n", __func__, use_mul_mat_vec);
// GGML_LOG_INFO("%s: use_mul_mat_vec_q = %d\n", __func__, use_mul_mat_vec_q); // GGML_LOG_INFO("%s: use_mul_mat_vec_q = %d\n", __func__, use_mul_mat_vec_q);
@@ -184,25 +171,7 @@ static void ggml_zdnn_mul_mat_dispatch(ggml_backend_zdnn_context * ctx, const gg
// GGML_LOG_INFO("%s: src0 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name); // GGML_LOG_INFO("%s: src0 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
// GGML_LOG_INFO("%s: src1 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name); // GGML_LOG_INFO("%s: src1 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16
&& !ggml_is_transposed(src0) && !ggml_is_transposed(src1)
&& src1->ne[2] * src1->ne[3] > 1) {
// general KQ + KQV multi-batch
GGML_LOG_INFO("%s: using zdnn_mul_mat_batched for KQ + KQV multi-batch\n", __func__);
// ggml_zdnn_mul_mat_batched(ctx, src0, src1, dst);
} else if (use_mul_mat_vec) {
GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec for vector multiplication\n", __func__);
// ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec, nullptr);
} else if (use_mul_mat_vec_q) {
GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec_q for quantized vector multiplication\n", __func__);
// ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec_q, ggml_zdnn_quantize_row_q8_1);
} else if (use_mul_mat_q) {
GGML_LOG_INFO("%s: using zdnn_op_mul_mat_q for quantized matrix multiplication\n", __func__);
// ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_q, ggml_zdnn_quantize_mmq_q8_1);
} else {
// GGML_LOG_INFO("%s: using zdnn_op_mul_mat for general matrix multiplication\n", __func__);
ggml_zdnn_mul_mat_op(ctx, src0, src1, dst); ggml_zdnn_mul_mat_op(ctx, src0, src1, dst);
}
} }
static bool ggml_zdnn_compute_forward(ggml_backend_zdnn_context * ctx, ggml_tensor * dst) { static bool ggml_zdnn_compute_forward(ggml_backend_zdnn_context * ctx, ggml_tensor * dst) {