vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

add ggml_qnn_tensor_binder

Browse source
This commit is contained in:
hongruichen 2024-06-14 18:52:54 +08:00
parent 5e18cdc268
commit 6c68adc1d9
1 changed files with 110 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

110
ggml-qnn.cpp
View file

@ -1959,6 +1959,116 @@ static bool ggml_qnn_can_handle_op(ggml_backend_qnn_context * ctx,
return true;
}
template <Qnn_TensorType_t _tensorType>
class ggml_qnn_tensor_binder
{
public:
ggml_qnn_tensor_binder(const ggml_tensor *tensor, ggml_backend_qnn_context * ctx, Qnn_GraphHandle_t graph_handle)
: _tensor(tensor)
, _qnn_tensor(reinterpret_cast<Qnn_Tensor_t *>(tensor->extra))
, _context(ctx) {
_old_dimensions = QNN_VER_PTR(*_qnn_tensor)->dimensions;
const auto qnn_data_type = qnn_datatype_from_ggml_datatype(tensor->type);
const bool is_npu = ctx->device == QNN_BACKEND_NPU;
QNN_VER_PTR(*_qnn_tensor)->type = _tensorType;
if (is_npu) {
QNN_VER_PTR(*_qnn_tensor)->memType = QNN_TENSORMEMTYPE_MEMHANDLE;
QNN_VER_PTR(*_qnn_tensor)->clientBuf= {.data=nullptr, .dataSize=0};
}
auto err = ctx->raw_interface.tensorCreateGraphTensor(graph_handle, _qnn_tensor);
if (err != QNN_SUCCESS) {
QNN_LOG_INFO("error = %d\n", err);
_context = nullptr;
return;
}
_dimensions[0] = (uint32_t)tensor->ne[0];
_dimensions[1] = (uint32_t)tensor->ne[1];
_dimensions[2] = (uint32_t)tensor->ne[2];
_dimensions[3] = (uint32_t)tensor->ne[3];
QNN_VER_PTR(*_qnn_tensor)->dimensions = _dimensions;
QNN_VER_PTR(*_qnn_tensor)->rank = qnn_get_ggml_tensor_rank(tensor);
QNN_VER_PTR(*_qnn_tensor)->dataType = qnn_data_type;
if (is_npu) {
qnn_instance * instance = ctx->instance;
uint8_t *qnn_buffer = static_cast<uint8_t *>(instance->alloc_rpcmem(
ggml_nbytes(tensor), 4)); // TODO: should we get the align param from device here?
if (!qnn_buffer) {
QNN_LOG_WARN("alloc rpcmem failure, %s\n", strerror(errno));
_context = nullptr;
return;
} else {
QNN_LOG_INFO("alloc rpcmem successfully\n");
}
instance->register_rpcmem(qnn_buffer, _qnn_tensor);
if (_tensorType == QNN_TENSOR_TYPE_APP_WRITE || _tensorType == QNN_TENSOR_TYPE_APP_READWRITE) {
memcpy(qnn_buffer, tensor->data, ggml_nbytes(tensor));
}
} else {
QNN_VER_PTR(*_qnn_tensor)->clientBuf = {tensor->data,
qnn_get_ggml_tensor_data_size(tensor)};
}
}
ggml_qnn_tensor_binder(const ggml_tensor *tensor, Qnn_Tensor_t *qnn_tensor, ggml_backend_qnn_context * ctx)
: _tensor(tensor)
, _qnn_tensor(qnn_tensor)
, _context(ctx) {
_old_dimensions = QNN_VER_PTR(*_qnn_tensor)->dimensions;
const auto qnn_data_type = qnn_datatype_from_ggml_datatype(tensor->type);
const bool is_npu = ctx->device == QNN_BACKEND_NPU;
_dimensions[0] = (uint32_t)tensor->ne[0];
_dimensions[1] = (uint32_t)tensor->ne[1];
_dimensions[2] = (uint32_t)tensor->ne[2];
_dimensions[3] = (uint32_t)tensor->ne[3];
QNN_VER_PTR(*_qnn_tensor)->dimensions = _dimensions;
QNN_VER_PTR(*_qnn_tensor)->rank = qnn_get_ggml_tensor_rank(tensor);
QNN_VER_PTR(*_qnn_tensor)->dataType = qnn_data_type;
if (is_npu) {
uint8_t * qnn_buffer = static_cast<uint8_t *>(ctx->instance->get_rpcmem_from_memhandle(
QNN_VER_PTR(*_qnn_tensor)->memHandle));
if (qnn_buffer) {
memcpy(qnn_buffer, tensor->data, ggml_nbytes(tensor));
} else {
QNN_LOG_WARN("can't find rpcmem from qnn mem handle\n");
}
} else {
QNN_VER_PTR(*_qnn_tensor)->clientBuf = {tensor->data,
qnn_get_ggml_tensor_data_size(tensor)};
}
}
~ggml_qnn_tensor_binder() {
if (_context && _context->device == QNN_BACKEND_NPU &&
(_tensorType == QNN_TENSOR_TYPE_APP_READWRITE || _tensorType == QNN_TENSOR_TYPE_APP_READ)) {
uint8_t * qnn_buffer = static_cast<uint8_t *>(_context->instance->get_rpcmem_from_memhandle(
QNN_VER_PTR(*_qnn_tensor)->memHandle));
memcpy(_tensor->data, qnn_buffer, ggml_nbytes(_tensor));
}
QNN_VER_PTR(*_qnn_tensor)->dimensions = _old_dimensions;
}
private:
const ggml_tensor *_tensor;
Qnn_Tensor_t *_qnn_tensor;
ggml_backend_qnn_context *_context;
uint32_t *_old_dimensions;
uint32_t _dimensions[4] = {};
ggml_qnn_tensor_binder(const ggml_qnn_tensor_binder&) = delete;
ggml_qnn_tensor_binder(ggml_qnn_tensor_binder&&) = delete;
void operator=(const ggml_qnn_tensor_binder&) = delete;
void operator=(ggml_qnn_tensor_binder&&) = delete;
};
//TODO: this function can be removed later because there are duplicated codes with ggml_qnn_mul_mat
// keep it for illustrate how to implement a specified GGMPL OP using QNN API + QNN RPC
static void ggml_qnn_add(ggml_backend_qnn_context * ctx, const ggml_tensor * src0,