Skip to content

Commit

Permalink
add ggml_cann prefix for acl funcs
Browse files Browse the repository at this point in the history
  • Loading branch information
@hipudding
hipudding committed Jul 17, 2024
1 parent 96e09b9 commit 57197b7
Show file tree
Hide file tree
Showing 4 changed files with 302 additions and 284 deletions.
44 changes: 23 additions & 21 deletions ggml/src/ggml-cann/acl_tensor.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@
#include <algorithm>
#include <cstring>

aclDataType type_mapping(ggml_type type) {
aclDataType ggml_cann_type_mapping(ggml_type type) {
switch (type) {
case GGML_TYPE_F32:
return ACL_FLOAT;
Expand All @@ -43,8 +43,9 @@ aclDataType type_mapping(ggml_type type) {
return ACL_DT_UNDEFINED;
}

aclTensor* create_acl_tensor(const ggml_tensor* tensor, int64_t* ne, size_t* nb,
int64_t dims, aclFormat format, size_t offset) {
aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
size_t* nb, int64_t dims, aclFormat format,
size_t offset) {
// If tensor is bcasted, Up to GGML_MAX_DIMS additional dimensions will be
// added.
int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
Expand All @@ -71,15 +72,15 @@ aclTensor* create_acl_tensor(const ggml_tensor* tensor, int64_t* ne, size_t* nb,
std::reverse(acl_ne, acl_ne + final_dims);
std::reverse(acl_stride, acl_stride + final_dims);

aclTensor* acl_tensor =
aclCreateTensor(acl_ne, final_dims, type_mapping(tensor->type),
acl_stride, offset / ggml_element_size(tensor), format,
&acl_storage_len, 1, tensor->data);
aclTensor* acl_tensor = aclCreateTensor(
acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
offset / ggml_element_size(tensor), format, &acl_storage_len, 1,
tensor->data);

return acl_tensor;
}

bool need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
for (int i = 0; i < GGML_MAX_DIMS; i++) {
if (t1->ne[i] != t0->ne[i] && t1->ne[i] != 1) {
return true;
Expand All @@ -88,9 +89,10 @@ bool need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
return false;
}

aclTensor* create_acl_tensor(void* data_ptr, aclDataType dtype,
size_t type_size, int64_t* ne, size_t* nb,
int64_t dims, aclFormat format, size_t offset) {
aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
size_t type_size, int64_t* ne, size_t* nb,
int64_t dims, aclFormat format,
size_t offset) {
int64_t tmp_ne[GGML_MAX_DIMS * 2];
int64_t tmp_stride[GGML_MAX_DIMS * 2];

Expand All @@ -114,9 +116,11 @@ aclTensor* create_acl_tensor(void* data_ptr, aclDataType dtype,
return acl_tensor;
}

int64_t get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
int64_t* bcast_src0_ne, int64_t* bcast_src1_ne,
size_t* bcast_src0_nb, size_t* bcast_src1_nb) {
int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
const ggml_tensor* src1,
int64_t* bcast_src0_ne,
int64_t* bcast_src1_ne, size_t* bcast_src0_nb,
size_t* bcast_src1_nb) {
GGML_ASSERT(ggml_can_repeat(src1, src0));
int bcast_dim_cnt = 0;
for (int i = 0; i < GGML_MAX_DIMS; i++) {
Expand All @@ -140,13 +144,11 @@ int64_t get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
return bcast_dim_cnt;
}

int64_t get_mul_mat_bcast_shape(const int64_t* input_ne,
const int64_t* weight_ne, const int64_t* dst_ne,
const size_t* input_nb, const size_t* weight_nb,
const size_t* dst_nb, int64_t* bcast_input_ne,
int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
size_t* bcast_input_nb, size_t* bcast_weight_nb,
size_t* bcast_dst_nb) {
int64_t ggml_cann_get_mulmat_bcast_shape(
const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb) {
// input and dst shoule in same shape, except first two dims.
GGML_ASSERT(input_ne[2] == dst_ne[2]);
GGML_ASSERT(input_ne[3] == dst_ne[3]);
Expand Down
46 changes: 22 additions & 24 deletions ggml/src/ggml-cann/acl_tensor.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,7 @@
* @return The corresponding aclDataType. If the input type is not recognized,
* ACL_DT_UNDEFINED is returned.
*/
aclDataType type_mapping(ggml_type type);
aclDataType ggml_cann_type_mapping(ggml_type type);

/**
* @brief Creates an ACL tensor from a ggml_tensor with optional shape.
Expand All @@ -59,7 +59,7 @@ aclDataType type_mapping(ggml_type type);
* @param offset Offset in bytes for the ACL tensor data. Defaults to 0.
* @return Pointer to the created ACL tensor.
*/
aclTensor* create_acl_tensor(const ggml_tensor* tensor, int64_t* ne = nullptr,
aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = nullptr,
size_t* nb = nullptr, int64_t dims = 0,
aclFormat format = ACL_FORMAT_ND,
size_t offset = 0);
Expand All @@ -83,7 +83,7 @@ aclTensor* create_acl_tensor(const ggml_tensor* tensor, int64_t* ne = nullptr,
* @param offset Offset in bytes for the ACL tensor data. Defaults to 0.
* @return Pointer to the created ACL tensor.
*/
aclTensor* create_acl_tensor(void* data_ptr, aclDataType dtype,
aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
size_t type_size, int64_t* ne, size_t* nb,
int64_t dims, aclFormat format = ACL_FORMAT_ND,
size_t offset = 0);
Expand All @@ -104,7 +104,7 @@ aclTensor* create_acl_tensor(void* data_ptr, aclDataType dtype,
* to 1. If such a dimension is found, broadcasting is required to align t1
* with t0 for element-wise operations.
*/
bool need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);

/**
* @brief Computes broadcast shapes and strides for two ggml_tensors.
Expand Down Expand Up @@ -159,19 +159,19 @@ bool need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
* dim1 in a inserted dim, should add nb for dim1,
* and all other nb moves to next in order.
*/
int64_t get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
int64_t* bcast_ne_src0, int64_t* bcast_ne_src1,
size_t* bcast_nb_src0, size_t* bcast_nb_src1);

// Bcast macro to avoid duplicate code.
#define BCAST_SHAPE(src0, src1) \
int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2]; \
int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2]; \
size_t bcast_##src0##_nb[GGML_MAX_DIMS * 2]; \
size_t bcast_##src1##_nb[GGML_MAX_DIMS * 2]; \
int64_t bcast_dims = \
get_bcast_shape(src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, \
bcast_##src0##_nb, bcast_##src1##_nb);
#define BCAST_SHAPE(src0, src1) \
int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2]; \
int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2]; \
size_t bcast_##src0##_nb[GGML_MAX_DIMS * 2]; \
size_t bcast_##src1##_nb[GGML_MAX_DIMS * 2]; \
int64_t bcast_dims = ggml_cann_get_bcast_shape( \
src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, bcast_##src0##_nb, \
bcast_##src1##_nb);

#define BCAST_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims

Expand Down Expand Up @@ -201,17 +201,15 @@ int64_t get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
* shapes needed for matrix multiplication. It ensures that dimensions where
* weight tensor requires expansion are appropriately handled to conform with
* broadcasting rules.
* @note compare with get_bcast_shape，mul_mat broadcast need add this new dim before
* cast dim.
* @sa get_bcast_shape
* @note compare with ggml_cann_get_bcast_shape，mul_mat broadcast need add this new dim
* before cast dim.
* @sa ggml_cann_get_bcast_shape
*/
int64_t get_mul_mat_bcast_shape(const int64_t* input_ne,
const int64_t* weight_ne, const int64_t* dst_ne,
const size_t* input_nb, const size_t* weight_nb,
const size_t* dst_nb, int64_t* bcast_input_ne,
int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
size_t* bcast_input_nb, size_t* bcast_weight_nb,
size_t* bcast_dst_nb);
int64_t ggml_cann_get_mulmat_bcast_shape(
const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb);

// Bcast macro to avoid duplicate code.
#define BCAST_MUL_MAT_SHAPE(input, weight, dst) \
Expand All @@ -221,7 +219,7 @@ int64_t get_mul_mat_bcast_shape(const int64_t* input_ne,
size_t bcast_##input##_nb[GGML_MAX_DIMS * 2]; \
size_t bcast_##weight##_nb[GGML_MAX_DIMS * 2]; \
size_t bcast_##dst##_nb[GGML_MAX_DIMS * 2]; \
int64_t bcast_dims = get_mul_mat_bcast_shape( \
int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape( \
input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, \
bcast_##input##_ne, bcast_##weight##_ne, bcast_##dst##_ne, \
bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
Expand Down
Loading

0 comments on commit 57197b7

Please sign in to comment.