Skip to content

Commit cb557b4

Browse files
Djip007Djip007
committed
clang-format
1 parent 10df7d0 commit cb557b4

File tree

6 files changed

+291
-288
lines changed

6 files changed

+291
-288
lines changed

‎ggml/src/ggml-cpu/amx/amx.cpp‎

Lines changed: 64 additions & 69 deletions
Original file line numberDiff line numberDiff line change
@@ -20,54 +20,55 @@
2020

2121
// AMX type_trais
2222
namespace ggml::cpu::amx {
23-
class tensor_traits : public ggml::cpu::tensor_traits {
23+
class tensor_traits : public ggml::cpu::tensor_traits {
24+
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
25+
size = ggml_backend_amx_desired_wsize(op);
26+
return true;
27+
}
2428

25-
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
26-
size = ggml_backend_amx_desired_wsize(op);
29+
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
30+
if (op->op == GGML_OP_MUL_MAT) {
31+
ggml_backend_amx_mul_mat(params, op);
2732
return true;
2833
}
29-
30-
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
31-
if (op->op == GGML_OP_MUL_MAT) {
32-
ggml_backend_amx_mul_mat(params, op);
33-
return true;
34-
}
35-
return false;
36-
}
37-
};
38-
39-
static ggml::cpu::tensor_traits* get_tensor_traits(ggml_backend_buffer_t , struct ggml_tensor *) {
40-
static tensor_traits traits;
41-
return &traits;
34+
return false;
4235
}
36+
};
37+
38+
static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
39+
static tensor_traits traits;
40+
return &traits;
4341
}
42+
} // namespace ggml::cpu::amx
4443

4544
// AMX buffer interface
4645
static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
4746
free(buffer->context);
4847
}
4948

5049
static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
51-
return (void *)(buffer->context);
50+
return (void *) (buffer->context);
5251
}
5352

5453
static void ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
55-
tensor->extra = (void *)ggml::cpu::amx::get_tensor_traits(buffer, tensor);
54+
tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor);
5655

5756
GGML_UNUSED(buffer);
5857
}
5958

60-
static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
61-
memset((char *)tensor->data + offset, value, size);
59+
static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
60+
uint8_t value, size_t offset, size_t size) {
61+
memset((char *) tensor->data + offset, value, size);
6262

6363
GGML_UNUSED(buffer);
6464
}
6565

66-
static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
66+
static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
67+
const void * data, size_t offset, size_t size) {
6768
if (qtype_has_amx_kernels(tensor->type)) {
6869
ggml_backend_amx_convert_weight(tensor, data, offset, size);
6970
} else {
70-
memcpy((char *)tensor->data + offset, data, size);
71+
memcpy((char *) tensor->data + offset, data, size);
7172
}
7273

7374
GGML_UNUSED(buffer);
@@ -136,49 +137,42 @@ static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_typ
136137
}
137138

138139
namespace ggml::cpu::amx {
139-
class extra_buffer_type : ggml::cpu::extra_buffer_type {
140-
bool supports_op(ggml_backend_dev_t , const struct ggml_tensor * op) override {
141-
// handle only 2d gemm for now
142-
auto is_contiguous_2d = [](const struct ggml_tensor * t) {
143-
return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
144-
};
145-
146-
if ( op->op == GGML_OP_MUL_MAT &&
147-
is_contiguous_2d(op->src[0]) && // src0 must be contiguous
148-
is_contiguous_2d(op->src[1]) && // src1 must be contiguous
149-
op->src[0]->buffer &&
150-
op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
151-
op->ne[0] % (TILE_N * 2) == 0 && // out_features is 32x
152-
(qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))
153-
)
154-
{
155-
// src1 must be host buffer
156-
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
157-
return false;
158-
}
159-
// src1 must be float32
160-
if (op->src[1]->type == GGML_TYPE_F32) {
161-
return true;
162-
}
140+
class extra_buffer_type : ggml::cpu::extra_buffer_type {
141+
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
142+
// handle only 2d gemm for now
143+
auto is_contiguous_2d = [](const struct ggml_tensor * t) {
144+
return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
145+
};
146+
147+
if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) && // src0 must be contiguous
148+
is_contiguous_2d(op->src[1]) && // src1 must be contiguous
149+
op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
150+
op->ne[0] % (TILE_N * 2) == 0 && // out_features is 32x
151+
(qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
152+
// src1 must be host buffer
153+
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
154+
return false;
163155
}
164-
return false;
165-
}
166-
167-
ggml::cpu::tensor_traits* get_tensor_traits(const struct ggml_tensor * op) override {
168-
if ( op->op == GGML_OP_MUL_MAT &&
169-
op->src[0]->buffer &&
170-
op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()
171-
)
172-
{
173-
return (ggml::cpu::tensor_traits*) op->src[0]->extra;
156+
// src1 must be float32
157+
if (op->src[1]->type == GGML_TYPE_F32) {
158+
return true;
174159
}
160+
}
161+
return false;
162+
}
175163

176-
return nullptr;
164+
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
165+
if (op->op == GGML_OP_MUL_MAT && op->src[0]->buffer &&
166+
op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()) {
167+
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
177168
}
178-
};
179-
}
180169

181-
static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
170+
return nullptr;
171+
}
172+
};
173+
} // namespace ggml::cpu::amx
174+
175+
static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
182176
return ggml_backend_amx_get_alloc_size(tensor);
183177

184178
GGML_UNUSED(buft);
@@ -200,25 +194,26 @@ static bool ggml_amx_init() {
200194
return true;
201195
#endif
202196
}
197+
203198
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
204199
static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
205200
/* .iface = */ {
206-
/* .get_name = */ ggml_backend_amx_buffer_type_get_name,
207-
/* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer,
208-
/* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment,
209-
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
210-
/* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size,
211-
/* .is_host = */ nullptr,
212-
},
201+
/* .get_name = */ ggml_backend_amx_buffer_type_get_name,
202+
/* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer,
203+
/* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment,
204+
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
205+
/* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size,
206+
/* .is_host = */ nullptr,
207+
},
213208
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
214209
/* .context = */ new ggml::cpu::amx::extra_buffer_type(),
215210
};
216211

217212
if (!ggml_amx_init()) {
218-
return NULL;
213+
return nullptr;
219214
}
220215

221216
return &ggml_backend_buffer_type_amx;
222217
}
223218

224-
#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)
219+
#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)

0 commit comments

Comments
 (0)