happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix compiling error at new hexagon sdk

This commit is contained in:
hongruichen 2025-07-15 10:17:52 +08:00
parent 4a3a87409b
commit 9a43a23e0b
4 changed files with 46 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
ggml/src/ggml-qnn/npu/CMakeLists.txt
View File

@ -10,6 +10,15 @@ else()
message(FATAL_ERROR "HEXAGON_SDK_ROOT not defined")
endif()
if(${CMAKE_SYSTEM_NAME} MATCHES "Android")
set(PREBUILT_LIB_DIR "android_aarch64")
elseif(${CMAKE_SYSTEM_NAME} MATCHES "Linux")
set(PREBUILT_LIB_DIR "UbuntuARM_aarch64")
elseif(${CMAKE_SYSTEM_NAME} MATCHES "Windows")
# Windows
set(PREBUILT_LIB_DIR "windows_aarch64")
endif()
if(HEXAGON_SDK_ROOT)
include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
else()
@ -138,21 +147,13 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Android|Linux|Windows")
link_options(hexagon-npu-host)
if(${CMAKE_SYSTEM_NAME} MATCHES "Android")
set(PREBUILT_LIB_DIR "android_aarch64")
elseif(${CMAKE_SYSTEM_NAME} MATCHES "Linux")
set(PREBUILT_LIB_DIR "UbuntuARM_aarch64")
else()
# Windows
set(PREBUILT_LIB_DIR "windows_aarch64")
endif()
choose_dsprpc("3" dsprpc) # cdsprpc
link_custom_library(hexagon-npu-host ${dsprpc})
cmake_host_system_information(RESULT BUILD_CPU_COUNT QUERY NUMBER_OF_PHYSICAL_CORES)
add_dsp_targets_for_host(hexagon-npu-host "v73" ${BUILD_CPU_COUNT})
add_dsp_targets_for_host(hexagon-npu-host "v75" ${BUILD_CPU_COUNT})
add_dsp_targets_for_host(hexagon-npu-host "v79" ${BUILD_CPU_COUNT})
list(APPEND NPU_RUNTIME_LIBS "${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonApp")
list(APPEND NPU_RUNTIME_LIBS "${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonAppLE")
@ -249,6 +250,7 @@ else()
target_link_libraries(hexagon_npu_skel
${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc++abi.a
${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc++.a
${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc.a
)
set_target_properties(hexagon_npu_skel PROPERTIES OUTPUT_NAME "hexagon_npu_skel_${HEXAGON_ARCH}")
target_link_libraries(hexagon_npu_skel qprintf_static)

1
ggml/src/ggml-qnn/npu/device/op_mul_mat.cpp
View File

@ -3,7 +3,6 @@
#include "thread_pool.hpp" // TODO: remove this dependency
#include "type_traits.hpp"
#include "vec_ops.hpp"
#include "vtcm_mem.hpp"
namespace {

5
ggml/src/ggml-qnn/npu/device/thread_pool.hpp
View File

@ -2,6 +2,7 @@
#include <qurt.h>
#include <array>
#include <atomic>
#include <cstdint>
#include <memory>
@ -237,8 +238,8 @@ template <size_t _ThreadCount> class thread_pool {
DEVICE_LOG_DEBUG("thread_func_impl.end: %zu", param->tidx);
}
std::atomic_bool _thread_exit = false;
std::array<qurt_thread_ptr, kMaxSubThreadCount> _threads;
std::atomic_bool _thread_exit = false;
std::array<qurt_thread_ptr, kMaxSubThreadCount> _threads = {};
qurt_barrier_t _pending = {};
qurt_barrier_t _completed = {};
thread_params _thread_params[kMaxThreadCount] = {};

37
ggml/src/ggml-qnn/npu/device/vtcm_mem.hpp
View File

@ -1,5 +1,6 @@
#pragma once
#include <HAP_compute_res.h>
#include <HAP_vtcm_mgr.h>
#include "util.hpp"
@ -9,15 +10,28 @@ namespace hexagon {
class vtcm_mem {
public:
explicit vtcm_mem(size_t size, bool single_page) {
constexpr const unsigned int kTimeoutUs = 10000; // 10ms timeout
size_t avail_size = single_page ? get_avail_page_size() : get_avail_block_size();
if (size > avail_size) {
DEVICE_LOG_ERROR("Requested VTCM size %zu exceeds available size %zu\n", size, avail_size);
return;
}
_vtcm_mem = HAP_request_VTCM((unsigned int) size, single_page ? 1 : 0);
compute_res_attr_t compute_res;
HAP_compute_res_attr_init(&compute_res);
HAP_compute_res_attr_set_serialize(&compute_res, false);
HAP_compute_res_attr_set_vtcm_param(&compute_res, size, single_page ? 1 : 0);
_vtcm_context_id = HAP_compute_res_acquire(&compute_res, kTimeoutUs); // 10ms timeout
if (_vtcm_context_id == 0) {
DEVICE_LOG_ERROR("Failed to acquire VTCM context: %zu bytes, timeout %zu us\n", size, kTimeoutUs);
return;
}
_vtcm_mem = HAP_compute_res_attr_get_vtcm_ptr(&compute_res);
if (_vtcm_mem == nullptr) {
DEVICE_LOG_ERROR("Failed to allocate VTCM memory: %zu bytes\n", size);
DEVICE_LOG_ERROR("Failed to allocate VTCM memory: %zu bytes, timeout %zu us\n", size, kTimeoutUs);
return;
}
@ -26,7 +40,18 @@ class vtcm_mem {
}
explicit vtcm_mem(size_t size, bool single_page, size_t timeout_us) {
_vtcm_mem = HAP_request_async_VTCM((unsigned int) size, single_page ? 1 : 0, (unsigned int) timeout_us);
compute_res_attr_t compute_res;
HAP_compute_res_attr_init(&compute_res);
HAP_compute_res_attr_set_serialize(&compute_res, false);
HAP_compute_res_attr_set_vtcm_param(&compute_res, size, single_page ? 1 : 0);
_vtcm_context_id = HAP_compute_res_acquire(&compute_res, timeout_us);
if (_vtcm_context_id == 0) {
DEVICE_LOG_ERROR("Failed to acquire VTCM context: %zu bytes, timeout %zu us\n", size, timeout_us);
return;
}
_vtcm_mem = HAP_compute_res_attr_get_vtcm_ptr(&compute_res);
if (_vtcm_mem == nullptr) {
DEVICE_LOG_ERROR("Failed to allocate VTCM memory: %zu bytes, timeout %zu us\n", size, timeout_us);
return;
@ -37,8 +62,8 @@ class vtcm_mem {
}
~vtcm_mem() {
if (is_valid()) {
auto ret = HAP_release_VTCM(_vtcm_mem);
if (_vtcm_context_id != 0) {
auto ret = HAP_compute_res_release(_vtcm_context_id);
if (ret != AEE_SUCCESS) {
DEVICE_LOG_ERROR("Failed to release VTCM memory: %d\n", ret);
}
@ -95,6 +120,8 @@ class vtcm_mem {
void * _vtcm_mem = nullptr;
size_t _vtcm_size = 0;
unsigned int _vtcm_context_id = 0;
DISABLE_COPY_AND_MOVE(vtcm_mem);
};