Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 06:39:25 +01:00
Code Issues Packages Projects Releases Wiki Activity

CANN: ROPE operator optimization (#10540)

Browse Source 
* [cann] ROPE operator optimization

Co-authored-by: noemotiovon <noemotiovon@gmail.com>
This commit is contained in:
Chenguang Li 2024-11-28 14:24:46 +08:00 committed by GitHub
parent 9f912511bc
commit b7420131bf
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 211 additions and 106 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

272
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -21,22 +21,23 @@
*/ */
#include "aclnn_ops.h" #include "aclnn_ops.h"
#include "ggml-impl.h"
#include <aclnnop/aclnn_addcdiv.h>
#include <aclnnop/aclnn_avgpool2d.h> #include <aclnnop/aclnn_avgpool2d.h>
#include <aclnnop/aclnn_batch_matmul.h>
#include <aclnnop/aclnn_cast.h> #include <aclnnop/aclnn_cast.h>
#include <aclnnop/aclnn_constant_pad_nd.h> #include <aclnnop/aclnn_constant_pad_nd.h>
#include <aclnnop/aclnn_copy.h> #include <aclnnop/aclnn_copy.h>
#include <aclnnop/aclnn_cos.h> #include <aclnnop/aclnn_cos.h>
#include <aclnnop/aclnn_div.h>
#include <aclnnop/aclnn_exp.h> #include <aclnnop/aclnn_exp.h>
#include <aclnnop/aclnn_fill_scalar.h> #include <aclnnop/aclnn_fill_scalar.h>
#include <aclnnop/aclnn_group_norm.h> #include <aclnnop/aclnn_group_norm.h>
#include <aclnnop/aclnn_index_fill_tensor.h> #include <aclnnop/aclnn_index_fill_tensor.h>
#include <aclnnop/aclnn_layer_norm.h> #include <aclnnop/aclnn_layer_norm.h>
#include <aclnnop/aclnn_mm.h>
#include <aclnnop/aclnn_batch_matmul.h>
#include <aclnnop/aclnn_matmul.h> #include <aclnnop/aclnn_matmul.h>
#include <aclnnop/aclnn_max_pool.h> #include <aclnnop/aclnn_max_pool.h>
#include <aclnnop/aclnn_mm.h>
#include <aclnnop/aclnn_permute.h> #include <aclnnop/aclnn_permute.h>
#include <aclnnop/aclnn_pow_tensor_tensor.h> #include <aclnnop/aclnn_pow_tensor_tensor.h>
#include <aclnnop/aclnn_reduce_sum.h> #include <aclnnop/aclnn_reduce_sum.h>
@ -56,6 +57,7 @@
#include <exception> #include <exception>
#include <vector> #include <vector>
#include "ggml-impl.h"
#include "kernels/ascendc_kernels.h" #include "kernels/ascendc_kernels.h"
#define GGML_COMMON_DECL_C #define GGML_COMMON_DECL_C
@ -1103,9 +1105,9 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
} }
/** /**
* @brief Creates an ACL tensor initialized with ones using a provided buffer. * @brief Creates an ACL tensor initialized with value using a provided buffer.
* *
* This function initializes a tensor with ones using the specified buffer and * This function initializes a tensor with value using the specified buffer and
* tensor parameters. * tensor parameters.
* *
* @param ctx The context for the CANN backend operations. * @param ctx The context for the CANN backend operations.
@ -1118,9 +1120,9 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
* @param type_size The size of each element in the tensor data type. * @param type_size The size of each element in the tensor data type.
* @param value The value to be used for initializing the tensor (default * @param value The value to be used for initializing the tensor (default
* is 1.0). * is 1.0).
* @return An ACL tensor initialized with ones. * @return An ACL tensor initialized with value.
*/ */
static aclTensor* aclnn_ones(ggml_backend_cann_context& ctx, void* buffer, static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer,
size_t n_bytes, int64_t* ne, int64_t dims, size_t n_bytes, int64_t* ne, int64_t dims,
aclDataType type, size_t type_size, aclDataType type, size_t type_size,
float value = 1.0f) { float value = 1.0f) {
@ -1165,7 +1167,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src); size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes); ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
aclTensor* acl_gamma = aclnn_ones( aclTensor* acl_gamma = aclnn_values(
ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1, ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
ggml_cann_type_mapping(src->type), ggml_element_size(src)); ggml_cann_type_mapping(src->type), ggml_element_size(src));
@ -1209,8 +1211,8 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes); ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
aclTensor* mask_tensor = aclTensor* mask_tensor =
aclnn_ones(ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes,
GGML_MAX_DIMS, ggml_cann_type_mapping(src->type), src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
ggml_element_size(src), value); ggml_element_size(src), value);
uint64_t workspaceSize = 0; uint64_t workspaceSize = 0;
@ -1768,6 +1770,92 @@ static void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream())); ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream()));
} }
/**
* @brief Performs element-wise division of tensor1 by tensor2 , multiplies the
result by the scalar value and adds it to self .
*
* Performs element-wise division of tensor1 by tensor2,
* multiplies the result by the scalar value and adds it to self .
* The operation is defined as:
* \f[
* \text{out}_i = \text{selft}_i + \text{value} \times
\frac{\text{tensor1}_i}{\text{tensor2}_i}
* \f]
* @param ctx The context for the CANN backend operations.
* @param acl_self The source tensor on which the addcdiv function will be
applied.
* @param tensor1 Numerator tensor.
* @param tensor2 Denominator tensor.
* @param value The value to be used for coefficient.
*/
static void aclnn_inplace_addcdiv(ggml_backend_cann_context& ctx,
aclTensor* acl_self, aclTensor* tensor1,
aclTensor* tensor2, float value) {
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
ACL_CHECK(aclnnInplaceAddcdivGetWorkspaceSize(
acl_self, tensor1, tensor2, acl_value, &workspaceSize, &executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get();
}
ACL_CHECK(aclnnInplaceAddcdiv(workspaceAddr, workspaceSize, executor,
ctx.stream()));
}
/**
* @brief Matrix division, optionally in-place.
*
* This function division each element of the source tensor `acl_src` by the
* tensor `acl_other` and stores the result in the destination tensor `acl_dst`.
* If `inplace` is true, `acl_dst` will not be used and the operation is
* performed in-place on `acl_src`. The operation is defined as: \f[
* \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i}
* \f]
*
* @param ctx The context for the CANN backend operations.
* @param acl_src Numerator tensor..
* @param acl_other Denominator tensor.
* @param acl_dst The destination tensor where the result will be stored if
* `inplace` is false.
* @param inplace Flag indicating whether to perform the operation in-place on
* `acl_src`.
*/
static void aclnn_div_tensor(ggml_backend_cann_context& ctx, aclTensor* acl_src,
aclTensor* acl_other, aclTensor* acl_dst,
bool inplace) {
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
if (inplace) {
ACL_CHECK(aclnnInplaceDivGetWorkspaceSize(acl_src, acl_other,
&workspaceSize, &executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get();
}
ACL_CHECK(aclnnInplaceDiv(workspaceAddr, workspaceSize, executor,
ctx.stream()));
} else {
ACL_CHECK(aclnnDivGetWorkspaceSize(acl_src, acl_other, acl_dst,
&workspaceSize, &executor));
if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get();
}
ACL_CHECK(
aclnnDiv(workspaceAddr, workspaceSize, executor, ctx.stream()));
}
}
void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
ggml_tensor* dst) { ggml_tensor* dst) {
const ggml_tensor* src = dst->src[0]; const ggml_tensor* src = dst->src[0];
@ -2311,12 +2399,13 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ctx.stream())); ctx.stream()));
switch (src0->type) { switch (src0->type) {
case GGML_TYPE_F32: case GGML_TYPE_F32: {
{
#ifdef ASCEND_310P #ifdef ASCEND_310P
// Special operation for get_row_f32 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes // Special operation for get_row_f32 kernel of 310P: clear the
// content of dest data buffer when row is not aligned to 32 bytes
if ((src0->ne[0] % 8) != 0) { if ((src0->ne[0] % 8) != 0) {
size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32); size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] *
src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len)); ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
} }
#endif #endif
@ -2329,12 +2418,15 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
((ggml_tensor*)dst->extra)->nb); ((ggml_tensor*)dst->extra)->nb);
break; break;
} }
case GGML_TYPE_F16: case GGML_TYPE_F16: {
{
#ifdef ASCEND_310P #ifdef ASCEND_310P
// Special operation for get_row_f16 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes // Special operation for get_row_f16 kernel of 310P: clear the
// content of dest data buffer when row is not aligned to 32 bytes
if ((src0->ne[0] % 16) != 0) { if ((src0->ne[0] % 16) != 0) {
size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32); // out is also f32, even input is f16 size_t dst_len =
src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] *
ggml_type_size(
GGML_TYPE_F32); // out is also f32, even input is f16
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len)); ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
} }
#endif #endif
@ -2459,8 +2551,9 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
* @param acl_dst The destination tensor where the result of the matrix * @param acl_dst The destination tensor where the result of the matrix
* multiplication will be stored. * multiplication will be stored.
*/ */
static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_input, static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx,
aclTensor* acl_weight, aclTensor* acl_dst) { aclTensor* acl_input, aclTensor* acl_weight,
aclTensor* acl_dst) {
int8_t cube_math_type = 2; int8_t cube_math_type = 2;
uint64_t workspaceSize = 0; uint64_t workspaceSize = 0;
aclOpExecutor* executor; aclOpExecutor* executor;
@ -2475,8 +2568,7 @@ static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_inpu
workspaceAddr = workspace_allocator.get(); workspaceAddr = workspace_allocator.get();
} }
ACL_CHECK( ACL_CHECK(aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
} }
/** /**
@ -2496,8 +2588,9 @@ static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_inpu
* @param acl_dst The destination tensor where the result of the matrix * @param acl_dst The destination tensor where the result of the matrix
* multiplication will be stored. * multiplication will be stored.
*/ */
static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx, aclTensor* acl_input, static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx,
aclTensor* acl_weight, aclTensor* acl_dst) { aclTensor* acl_input, aclTensor* acl_weight,
aclTensor* acl_dst) {
int8_t cube_math_type = 2; int8_t cube_math_type = 2;
uint64_t workspaceSize = 0; uint64_t workspaceSize = 0;
aclOpExecutor* executor; aclOpExecutor* executor;
@ -2548,16 +2641,12 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
aclTensor* acl_input_tensor = aclTensor* acl_input_tensor =
ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims); ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims);
int64_t transpose_ne[] = { int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0],
bcast_weight_ne[1], bcast_weight_ne[0],
bcast_weight_ne[2], bcast_weight_ne[3], bcast_weight_ne[2], bcast_weight_ne[3],
bcast_weight_ne[4], bcast_weight_ne[5] bcast_weight_ne[4], bcast_weight_ne[5]};
}; size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
size_t transpose_nb[] = {
bcast_weight_nb[1], bcast_weight_nb[0],
bcast_weight_nb[2], bcast_weight_nb[3], bcast_weight_nb[2], bcast_weight_nb[3],
bcast_weight_nb[4], bcast_weight_nb[5] bcast_weight_nb[4], bcast_weight_nb[5]};
};
aclTensor* acl_weight_tensor = aclTensor* acl_weight_tensor =
ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims); ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
aclTensor* acl_dst = aclTensor* acl_dst =
@ -2617,7 +2706,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
// scale stored at the end of weight. Also need transpose. // scale stored at the end of weight. Also need transpose.
size_t scale_elem_size = sizeof(uint16_t); size_t scale_elem_size = sizeof(uint16_t);
size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size, scale_elem_size}; size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
scale_elem_size};
size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size; size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
char* scale_offset = (char*)src0->data + weight_size; char* scale_offset = (char*)src0->data + weight_size;
@ -2632,7 +2722,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
// case in // case in
if (src1->type != GGML_TYPE_F16) { if (src1->type != GGML_TYPE_F16) {
aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1); aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
input_buffer = input_alloctor.alloc(ggml_nelements(src1) * input_elem_size); input_buffer =
input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
int64_t* input_cast_ne = src1->ne; int64_t* input_cast_ne = src1->ne;
size_t input_cast_nb[GGML_MAX_DIMS]; size_t input_cast_nb[GGML_MAX_DIMS];
@ -2642,9 +2733,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
} }
aclTensor* acl_input_tensor = ggml_cann_create_tensor( aclTensor* acl_input_tensor = ggml_cann_create_tensor(
input_buffer, input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne,
ACL_FLOAT16, input_cast_nb, GGML_MAX_DIMS);
input_elem_size, input_cast_ne, input_cast_nb, GGML_MAX_DIMS);
aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16); aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16);
ACL_CHECK(aclDestroyTensor(acl_input_tensor)); ACL_CHECK(aclDestroyTensor(acl_input_tensor));
@ -2655,7 +2745,8 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
size_t output_elem_size = sizeof(uint16_t); size_t output_elem_size = sizeof(uint16_t);
size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size}; size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size};
ggml_cann_pool_alloc output_allocator(ctx.pool()); ggml_cann_pool_alloc output_allocator(ctx.pool());
void* output_buffer = output_allocator.alloc(ggml_nelements(dst) * output_elem_size); void* output_buffer =
output_allocator.alloc(ggml_nelements(dst) * output_elem_size);
size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size; size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size;
// aclnn // aclnn
@ -2679,7 +2770,9 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
// first split // first split
int64_t weight_ne_offset = 0; int64_t weight_ne_offset = 0;
int64_t weight_ne[2] = {max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size, src0->ne[0]}; int64_t weight_ne[2] = {
max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size,
src0->ne[0]};
int64_t scale_ne_offset = 0; int64_t scale_ne_offset = 0;
int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0}; int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0};
int64_t output_ne_offset = 0; int64_t output_ne_offset = 0;
@ -2687,24 +2780,21 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
aclTensor* acl_weight_tensor = ggml_cann_create_tensor( aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
(char*)src0->data + batch0 * weight_stride, (char*)src0->data + batch0 * weight_stride,
ggml_cann_type_mapping(type), ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
weight_elem_size, weight_ne, weight_nb, 2, weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
ACL_FORMAT_ND, weight_ne_offset);
aclTensor* acl_scale_tensor = ggml_cann_create_tensor( aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
scale_offset + batch0 * scale_stride, scale_offset + batch0 * scale_stride, ACL_FLOAT16,
ACL_FLOAT16, scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
scale_elem_size, scale_ne, scale_nb, 2, scale_ne_offset);
ACL_FORMAT_ND, scale_ne_offset);
aclTensor* acl_output_tensor = ggml_cann_create_tensor( aclTensor* acl_output_tensor = ggml_cann_create_tensor(
(char*)output_buffer + batch1 * output_stride, (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
ACL_FLOAT16, output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
output_elem_size, output_ne, output_nb, 2, output_ne_offset);
ACL_FORMAT_ND, output_ne_offset);
ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize( ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
acl_input_tensor, acl_weight_tensor, acl_scale_tensor, acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
nullptr, nullptr, nullptr, nullptr, QK8_0, nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
acl_output_tensor, &workspaceSize, &executor)); &workspaceSize, &executor));
if (workspaceAddr == nullptr) { if (workspaceAddr == nullptr) {
workspaceAddr = workspace_allocator.alloc(workspaceSize); workspaceAddr = workspace_allocator.alloc(workspaceSize);
} }
@ -2717,28 +2807,29 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
// other splits // other splits
for (int64_t split = 1; split < split_size; split++) { for (int64_t split = 1; split < split_size; split++) {
weight_ne_offset += weight_elem_size * weight_ne[0] * weight_ne[1]; weight_ne_offset +=
weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1] ? src0->ne[1] - (max_elem_size * split) : max_elem_size; weight_elem_size * weight_ne[0] * weight_ne[1];
weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1]
? src0->ne[1] - (max_elem_size * split)
: max_elem_size;
scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1]; scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1];
scale_ne[0] = weight_ne[0]; scale_ne[0] = weight_ne[0];
output_ne_offset += output_elem_size * output_ne[0] * output_ne[1]; output_ne_offset +=
output_elem_size * output_ne[0] * output_ne[1];
output_ne[0] = weight_ne[0]; output_ne[0] = weight_ne[0];
acl_weight_tensor = ggml_cann_create_tensor( acl_weight_tensor = ggml_cann_create_tensor(
(char*)src0->data + batch0 * weight_stride, (char*)src0->data + batch0 * weight_stride,
ggml_cann_type_mapping(type), ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
weight_elem_size, weight_ne, weight_nb, 2, weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
ACL_FORMAT_ND, weight_ne_offset);
acl_scale_tensor = ggml_cann_create_tensor( acl_scale_tensor = ggml_cann_create_tensor(
scale_offset + batch0 * scale_stride, scale_offset + batch0 * scale_stride, ACL_FLOAT16,
ACL_FLOAT16, scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
scale_elem_size, scale_ne, scale_nb, 2, scale_ne_offset);
ACL_FORMAT_ND, scale_ne_offset);
acl_output_tensor = ggml_cann_create_tensor( acl_output_tensor = ggml_cann_create_tensor(
(char*)output_buffer + batch1 * output_stride, (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
ACL_FLOAT16, output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
output_elem_size, output_ne, output_nb, 2, output_ne_offset);
ACL_FORMAT_ND, output_ne_offset);
ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize( ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
acl_input_tensor, acl_weight_tensor, acl_scale_tensor, acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
@ -2766,11 +2857,11 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
} }
aclTensor* acl_output_tensor = ggml_cann_create_tensor( aclTensor* acl_output_tensor = ggml_cann_create_tensor(
output_buffer, output_buffer, ACL_FLOAT16, output_elem_size, output_cast_ne,
ACL_FLOAT16, output_cast_nb, GGML_MAX_DIMS);
output_elem_size, output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst); aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type)); aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor,
ggml_cann_type_mapping(dst->type));
ACL_CHECK(aclDestroyTensor(acl_output_tensor)); ACL_CHECK(aclDestroyTensor(acl_output_tensor));
ACL_CHECK(aclDestroyTensor(acl_dst_tensor)); ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
@ -2873,12 +2964,14 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst, static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
aclTensor* acl_cos_repeat_tensor, aclTensor* acl_cos_repeat_tensor,
aclTensor* acl_sin_repeat_tensor, aclTensor* acl_sin_repeat_tensor,
float theta_scale, bool is_neox) { float theta_scale, float freq_scale,
bool is_neox) {
// int sin/cos cache, cache has different repeat method depond on // int sin/cos cache, cache has different repeat method depond on
// @param.is_neox // @param.is_neox
ggml_tensor* src0 = dst->src[0]; // input ggml_tensor* src0 = dst->src[0]; // input
ggml_tensor* src1 = dst->src[1]; // position ggml_tensor* src1 = dst->src[1]; // position
ggml_tensor* src2 = dst->src[2]; // freq_factors
// arange, [0,1,...,ne0/2] // arange, [0,1,...,ne0/2]
int64_t arange_length = src0->ne[0] / 2; int64_t arange_length = src0->ne[0] / 2;
@ -2907,11 +3000,25 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(), ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
arange_length * sizeof(float_t)); arange_length * sizeof(float_t));
void* theta_scale_buffer = theta_scale_allocator.get(); void* theta_scale_buffer = theta_scale_allocator.get();
aclTensor* acl_theta_scale_tensor = aclnn_ones( aclTensor* acl_theta_scale_tensor = aclnn_values(
ctx, theta_scale_buffer, arange_length * sizeof(float_t), arange_ne, ctx, theta_scale_buffer, arange_length * sizeof(float_t), arange_ne,
GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), theta_scale); GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), theta_scale);
aclnn_pow_tensor_tensor(ctx, acl_theta_scale_tensor, acl_arange_tensor); aclnn_pow_tensor_tensor(ctx, acl_theta_scale_tensor, acl_arange_tensor);
// freq_scale
if (freq_scale != 1) {
aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
}
// freq_factors
if (src2) {
aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
src2->data, ggml_cann_type_mapping(src2->type),
ggml_type_size(src2->type), arange_ne, arange_nb, GGML_MAX_DIMS);
aclnn_div_tensor(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor,
nullptr, true);
}
// position // position
GGML_ASSERT(src1->type == GGML_TYPE_I32); GGML_ASSERT(src1->type == GGML_TYPE_I32);
int64_t position_length = src1->ne[0]; int64_t position_length = src1->ne[0];
@ -2940,6 +3047,16 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor, aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
acl_theta_tensor); acl_theta_tensor);
// // power[] * position[] * freq_scale / freq_factors[]
// ggml_cann_pool_alloc theta_final_allocator(ctx.pool(),
// theta_length *
// sizeof(float_t));
// aclTensor* acl_theat_final_tensor = aclnn_zero(
// ctx, theta_final_allocator.get(), sizeof(float_t) * theta_length,
// theta_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t));
// aclnn_inplace_addcdiv(ctx, acl_theat_final_tensor, acl_theta_tensor,
// acl_freq_factors_tensor, freq_scale);
// permute: [0,1,2,3]->[0,2,1,3] // permute: [0,1,2,3]->[0,2,1,3]
int64_t permute_ne[] = {arange_length, 1, position_length, 1}; int64_t permute_ne[] = {arange_length, 1, position_length, 1};
size_t permute_nb[GGML_MAX_DIMS]; size_t permute_nb[GGML_MAX_DIMS];
@ -3038,8 +3155,6 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float)); memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float)); memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
// TODO: with freq_factors
GGML_ASSERT(src2 == NULL);
// TODO: attn_factor != 1 // TODO: attn_factor != 1
GGML_ASSERT(attn_factor == 1); GGML_ASSERT(attn_factor == 1);
// TODO: n_dims <= ne0 // TODO: n_dims <= ne0
@ -3047,8 +3162,6 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
GGML_ASSERT(n_dims % 2 == 0); GGML_ASSERT(n_dims % 2 == 0);
// TODO: ext_factor != 0 // TODO: ext_factor != 0
GGML_ASSERT(ext_factor == 0); GGML_ASSERT(ext_factor == 0);
// TODO: freq_scale != 1
GGML_ASSERT(freq_scale == 1);
// TODO: type == GGML_TYPE_F16 // TODO: type == GGML_TYPE_F16
GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src0->type == GGML_TYPE_F32);
@ -3081,7 +3194,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t), ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS); sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor, aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
theta_scale, is_neox); theta_scale, freq_scale, is_neox);
uint64_t workspaceSize = 0; uint64_t workspaceSize = 0;
aclOpExecutor* executor; aclOpExecutor* executor;
@ -3096,7 +3209,8 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
aclTensor* acl_x = ggml_cann_create_tensor(src0); aclTensor* acl_x = ggml_cann_create_tensor(src0);
aclTensor* acl_dst = ggml_cann_create_tensor(dst); aclTensor* acl_dst = ggml_cann_create_tensor(dst);
ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize( ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize(
acl_x, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode, acl_dst, &workspaceSize, &executor)); acl_x, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode,
acl_dst, &workspaceSize, &executor));
if (workspaceSize > 0) { if (workspaceSize > 0) {
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
workspaceAddr = workspace_allocator.get(); workspaceAddr = workspace_allocator.get();

9
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -1738,13 +1738,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
} }
case GGML_OP_ROPE: { case GGML_OP_ROPE: {
// TODO: with ops-test v == 1 // TODO: with ops-test v == 1
float * freq_scale = (float*)((int32_t*)op->op_params + 6);
float * ext_factor = (float*)((int32_t*)op->op_params + 7); float * ext_factor = (float*)((int32_t*)op->op_params + 7);
float * attn_factor = (float*)((int32_t*)op->op_params + 8); float * attn_factor = (float*)((int32_t*)op->op_params + 8);
// TODO: with freq_factors
if (op->src[2] != NULL) {
return false;
}
// TODO: n_dims <= ne0 // TODO: n_dims <= ne0
if (op->src[0]->ne[0] != op->op_params[1]) { if (op->src[0]->ne[0] != op->op_params[1]) {
return false; return false;
@ -1753,10 +1748,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
if (*ext_factor != 0) { if (*ext_factor != 0) {
return false; return false;
} }
// TODO: freq_scale != 1
if (*freq_scale != 1) {
return false;
}
// TODO: attn_factor != 1 // TODO: attn_factor != 1
if (*attn_factor != 1) { if (*attn_factor != 1) {
return false; return false;