Skip to content

Commit 2a22db5

Browse files
rfsalievrfsaliev
committed
Small code refactoring
1 parent fe450a3 commit 2a22db5

File tree

1 file changed

+76
-109
lines changed

1 file changed

+76
-109
lines changed

‎src/ggml-dnnl.cpp‎

Lines changed: 76 additions & 109 deletions
Original file line numberDiff line numberDiff line change
@@ -35,8 +35,25 @@ struct ggml_backend_dnnl_buffer_context
3535
std::vector<dnnl::memory> sub_mems;
3636
};
3737

38+
static void* get_memory_handle(const struct ggml_tensor * t) {
39+
auto buf_mem = dnnl::memory{(dnnl_memory_t)t->extra, true};
40+
auto buf_md = buf_mem.get_desc();
41+
auto buf_handle = buf_mem.get_data_handle();
42+
auto buf_offset = buf_md.get_submemory_offset();
43+
GGML_ASSERT((size_t)buf_offset == ((uintptr_t)t->data - DNNL_BUFFER_BASE));
44+
// auto buf_ctx = (ggml_backend_dnnl_buffer_context*)t->buffer->context;
45+
//auto parent_buf_handle = buf_ctx->mem.get_data_handle();
46+
47+
// FIXME: buf_handle + offset works for CPU only
48+
return (char*)buf_handle + buf_offset;
49+
}
50+
#else
51+
static void* get_memory_handle(const struct ggml_tensor * t) {
52+
return t->data;
53+
}
3854
#endif
3955

56+
4057
namespace {
4158
template <class T>
4259
struct dnnl_mem_ptr {
@@ -91,15 +108,6 @@ static bool ggml_dnnl_type_supported(enum ggml_type type) {
91108

92109
static bool ggml_dnnl_tensor_supported(const struct ggml_tensor * t) {
93110
auto type = t->type;
94-
GGML_TENSOR_LOCALS(int64_t, ne, t, ne)
95-
GGML_TENSOR_LOCALS(size_t, nb, t, nb)
96-
97-
98-
// cannot be transposed or permuted
99-
// GGML_ASSERT(nb0 == ggml_type_size(type));
100-
// GGML_ASSERT(nb0 <= nb1);
101-
// GGML_ASSERT(nb1 <= nb2);
102-
// GGML_ASSERT(nb2 <= nb3);
103111

104112
if (!ggml_dnnl_type_supported(type)) {
105113
return false;
@@ -166,25 +174,6 @@ dnnl::memory::desc ggml_tensor_to_dnnl_md(const struct ggml_tensor * t, bool tra
166174
return dnnl::memory::desc{adims, dt, strides};
167175
}
168176

169-
#if USE_DNNL_BUFFER
170-
static void* get_memory_handle(const struct ggml_tensor * t) {
171-
auto buf_mem = dnnl::memory{(dnnl_memory_t)t->extra, true};
172-
auto buf_md = buf_mem.get_desc();
173-
auto buf_handle = buf_mem.get_data_handle();
174-
auto buf_offset = buf_md.get_submemory_offset();
175-
GGML_ASSERT((size_t)buf_offset == ((uintptr_t)t->data - DNNL_BUFFER_BASE));
176-
// auto buf_ctx = (ggml_backend_dnnl_buffer_context*)t->buffer->context;
177-
//auto parent_buf_handle = buf_ctx->mem.get_data_handle();
178-
179-
// FIXME: buf_handle + offset works for CPU only
180-
return (char*)buf_handle + buf_offset;
181-
}
182-
#else
183-
static void* get_memory_handle(const struct ggml_tensor * t) {
184-
return t->data;
185-
}
186-
#endif
187-
188177
dnnl::memory ggml_tensor_to_dnnl_mem(ggml_backend_t backend, const struct ggml_tensor * t, bool transpose = false,
189178
dnnl::memory::data_type convert_to = dnnl::memory::data_type::undef,
190179
size_t ndims = GGML_MAX_DIMS) {
@@ -474,7 +463,7 @@ static ggml_status ggml_backend_dnnl_softmax(ggml_backend_t backend, struct ggml
474463
{DNNL_ARG_DST, src_mem},
475464
});
476465
}
477-
//float alpha = *reinterpret_cast<float*>(dst->op_params);
466+
478467
const int axis = src_mem.get_desc().get_dims().size() - 1;
479468
auto pd = dnnl::softmax_forward::primitive_desc{ctx->engine, dnnl::prop_kind::forward_inference, dnnl::algorithm::softmax_accurate, src_mem.get_desc(), dst_mem.get_desc(), axis};
480469
auto prim = dnnl::softmax_forward{pd};
@@ -493,7 +482,6 @@ static ggml_status ggml_backend_dnnl_norm(ggml_backend_t backend, struct ggml_te
493482
auto src_mem = ggml_tensor_to_dnnl_mem(backend, src);
494483
auto dst_mem = ggml_tensor_to_dnnl_mem(backend, dst);
495484

496-
//float alpha = *reinterpret_cast<float*>(dst->op_params);
497485
float eps = ((const float *)(dst->op_params))[0];
498486

499487
GGML_ASSERT(eps > 0.0f);
@@ -805,6 +793,62 @@ static ggml_status ggml_backend_dnnl_node_compute(ggml_backend_t backend, struct
805793
*/
806794
}
807795

796+
static bool ggml_backend_dnnl_node_supported(ggml_backend_t backend, const struct ggml_tensor * node) {
797+
GGML_UNUSED(backend);
798+
// return false;
799+
switch (node->op) {
800+
case GGML_OP_NONE:
801+
return true;
802+
case GGML_OP_RESHAPE:
803+
case GGML_OP_VIEW:
804+
case GGML_OP_PERMUTE:
805+
case GGML_OP_TRANSPOSE:
806+
case GGML_OP_ADD:
807+
case GGML_OP_ADD1:
808+
case GGML_OP_SUB:
809+
case GGML_OP_MUL:
810+
case GGML_OP_DIV:
811+
case GGML_OP_SQR:
812+
case GGML_OP_SQRT:
813+
case GGML_OP_LOG:
814+
case GGML_OP_CONT:
815+
case GGML_OP_CPY:
816+
case GGML_OP_DUP:
817+
case GGML_OP_SCALE:
818+
case GGML_OP_DIAG_MASK_ZERO:
819+
case GGML_OP_DIAG_MASK_INF:
820+
case GGML_OP_SOFT_MAX:
821+
case GGML_OP_GET_ROWS:
822+
return ggml_dnnl_tensor_supported(node) && ggml_dnnl_tensor_supported(node->src[0]);
823+
case GGML_OP_MUL_MAT:
824+
return ggml_compute_forward_mul_mat_use_dnnl(node);
825+
case GGML_OP_UNARY:
826+
{
827+
enum ggml_unary_op uop = ggml_get_unary_op(node);
828+
switch(uop) {
829+
case GGML_UNARY_OP_ABS:
830+
case GGML_UNARY_OP_TANH:
831+
case GGML_UNARY_OP_ELU:
832+
case GGML_UNARY_OP_RELU:
833+
case GGML_UNARY_OP_GELU:
834+
case GGML_UNARY_OP_GELU_QUICK:
835+
case GGML_UNARY_OP_HARDSWISH:
836+
case GGML_UNARY_OP_HARDSIGMOID:
837+
return ggml_dnnl_tensor_supported(node) && ggml_dnnl_tensor_supported(node->src[0]);
838+
default:
839+
// GGML_UNARY_OP_SGN,
840+
// GGML_UNARY_OP_NEG,
841+
// GGML_UNARY_OP_STEP,
842+
// GGML_UNARY_OP_SILU,
843+
return false;
844+
}
845+
}
846+
847+
default:
848+
return false;
849+
}
850+
}
851+
808852
// buffer interface
809853

810854
#if USE_DNNL_BUFFER
@@ -949,23 +993,6 @@ GGML_CALL static const char * ggml_backend_dnnl_buffer_type_get_name(ggml_backen
949993
GGML_UNUSED(buft);
950994
}
951995

952-
// GGML_CALL static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
953-
// size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned
954-
// void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
955-
// if (data == NULL) {
956-
// fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
957-
// return NULL;
958-
// }
959-
960-
// return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
961-
// }
962-
963-
// GGML_CALL static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
964-
// return TENSOR_ALIGNMENT;
965-
966-
// GGML_UNUSED(buft);
967-
// }
968-
969996
GGML_CALL static bool ggml_backend_dnnl_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
970997
return ggml_backend_is_dnnl(backend) || ggml_backend_is_cpu(backend);
971998

@@ -1037,67 +1064,7 @@ GGML_CALL static ggml_status ggml_backend_dnnl_graph_compute(ggml_backend_t back
10371064
}
10381065

10391066
GGML_CALL static bool ggml_backend_dnnl_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
1040-
GGML_UNUSED(backend);
1041-
// return false;
1042-
switch (op->op) {
1043-
case GGML_OP_NONE:
1044-
return true;
1045-
case GGML_OP_RESHAPE:
1046-
case GGML_OP_VIEW:
1047-
case GGML_OP_PERMUTE:
1048-
case GGML_OP_TRANSPOSE:
1049-
case GGML_OP_ADD:
1050-
case GGML_OP_ADD1:
1051-
case GGML_OP_SUB:
1052-
case GGML_OP_MUL:
1053-
case GGML_OP_DIV:
1054-
case GGML_OP_SQR:
1055-
case GGML_OP_SQRT:
1056-
case GGML_OP_LOG:
1057-
case GGML_OP_CONT:
1058-
case GGML_OP_CPY:
1059-
case GGML_OP_DUP:
1060-
case GGML_OP_SCALE:
1061-
case GGML_OP_DIAG_MASK_ZERO:
1062-
case GGML_OP_DIAG_MASK_INF:
1063-
case GGML_OP_SOFT_MAX:
1064-
case GGML_OP_GET_ROWS:
1065-
return ggml_dnnl_tensor_supported(op) && ggml_dnnl_tensor_supported(op->src[0]);
1066-
case GGML_OP_MUL_MAT:
1067-
return ggml_compute_forward_mul_mat_use_dnnl(op);
1068-
case GGML_OP_UNARY:
1069-
{
1070-
enum ggml_unary_op uop = ggml_get_unary_op(op);
1071-
switch(uop) {
1072-
case GGML_UNARY_OP_ABS:
1073-
case GGML_UNARY_OP_TANH:
1074-
case GGML_UNARY_OP_ELU:
1075-
case GGML_UNARY_OP_RELU:
1076-
case GGML_UNARY_OP_GELU:
1077-
case GGML_UNARY_OP_GELU_QUICK:
1078-
case GGML_UNARY_OP_HARDSWISH:
1079-
case GGML_UNARY_OP_HARDSIGMOID:
1080-
return ggml_dnnl_tensor_supported(op) && ggml_dnnl_tensor_supported(op->src[0]);
1081-
default:
1082-
// GGML_UNARY_OP_SGN,
1083-
// GGML_UNARY_OP_NEG,
1084-
// GGML_UNARY_OP_STEP,
1085-
// GGML_UNARY_OP_SILU,
1086-
return false;
1087-
}
1088-
}
1089-
1090-
default:
1091-
return false;
1092-
}
1093-
}
1094-
1095-
GGML_CALL static bool ggml_backend_dnnl_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
1096-
// const int min_batch_size = 32;
1097-
1098-
return ggml_backend_dnnl_supports_op(backend, op);
1099-
1100-
// GGML_UNUSED(backend);
1067+
return ggml_backend_dnnl_node_supported(backend, op);
11011068
}
11021069

11031070
static struct ggml_backend_i dnnl_backend_i = {
@@ -1113,7 +1080,7 @@ static struct ggml_backend_i dnnl_backend_i = {
11131080
/* .graph_plan_compute = */ NULL,
11141081
/* .graph_compute = */ ggml_backend_dnnl_graph_compute,
11151082
/* .supports_op = */ ggml_backend_dnnl_supports_op,
1116-
/* .offload_op = */ NULL, //ggml_backend_dnnl_offload_op,
1083+
/* .offload_op = */ NULL,
11171084
/* .event_new = */ NULL,
11181085
/* .event_free = */ NULL,
11191086
/* .event_record = */ NULL,

0 commit comments

Comments
 (0)