CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-14 11:07:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

CANN: Support more ops (#12841)

Browse Source 
* [CANN]Support Opt LOG && MEAN && PAD_REFLECT_1D

* [CANN]Support COUNT_EQUAL && STEP && SGN

* [CANN]codestyle adjustment

* [CANN]codestyle adjustment

---------

Signed-off-by: noemotiovon <noemotiovon@gmail.com>
This commit is contained in:
Chenguang Li
2025-04-10 08:51:52 +08:00
committed by GitHub
parent 11d07e1e69
commit fe5b78c896
4 changed files with 173 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
ggml/src/ggml-cann/aclnn_ops.h
View File

@@ -42,6 +42,8 @@
#include <aclnnop/aclnn_sqrt.h>
#include <aclnnop/aclnn_sin.h>
#include <aclnnop/aclnn_cos.h>
#include <aclnnop/aclnn_log.h>
#include <aclnnop/aclnn_sign.h>
#include "acl_tensor.h"
#include "common.h"
@@ -650,6 +652,67 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* ds
*/
void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Computes the mean of a ggml tensor element-wise using the CANN backend.
*
* @details This function calculates the element-wise mean of the input tensor.
* The result is written to the destination tensor `dst`.
* The mean is computed by averaging the values across the entire tensor.
*
* This operation is optimized using the CANN backend for high-performance inference or training.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the mean result will be stored.
* dst->op is expected to be `GGML_OP_MEAN`.
*/
void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies 1D reflect padding to a ggml tensor using the CANN backend.
*
* @details This function performs 1D reflect padding on the input tensor.
* The amount of padding on each side is specified by parameters stored in `dst->op_params`.
* The operation reflects the values at the borders of the tensor to generate the padded output.
*
* This operation is optimized using the CANN backend for high-performance inference or training.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the padded result will be stored.
* dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`.
*/
void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Counts the number of equal elements in two ggml tensors using the CANN backend.
*
* @details This function performs an element-wise comparison between two input tensors,
* and counts the number of positions where the elements are equal. The result is
* stored in the destination tensor `dst` as a scalar.
*
* The operation is optimized using the CANN backend, making it suitable for
* high-performance inference or training scenarios.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the result will be stored.
* dst->op is expected to be `GGML_OP_COUNT_EQUAL`.
*/
void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies the Step activation function to a ggml tensor using the CANN backend.
*
* @details This function applies a step function element-wise to the input tensor, where
* each element is transformed to 1.0 if it is greater than 0, and 0.0 otherwise.
* The result is stored in the destination tensor `dst`.
*
* This operation is accelerated using the CANN backend to improve runtime performance.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the result will be stored.
* dst->op is expected to be `GGML_OP_STEP`.
*/
void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies a element-wise operation to two input tensors using the CANN
* backend.