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

use ggml_qnn_tensor_writer for all parameters

Browse source
This commit is contained in:
hongruichen 2024-06-16 23:54:00 +08:00
parent a5679ddd8e
commit 5fe7b87ba1
1 changed files with 20 additions and 141 deletions
Download patch file Download diff file Expand all files Collapse all files

161
ggml-qnn.cpp
View file

@ -1991,7 +1991,6 @@ public:
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;
@ -2090,27 +2089,16 @@ static void ggml_qnn_add(ggml_backend_qnn_context * ctx, const ggml_tensor * src
qnn_instance * instance = nullptr;
std::string graph_name = "ggml_op_qnn_add";
Qnn_GraphHandle_t graph_handle = nullptr;
Qnn_Tensor_t * tensor_1 = nullptr;
Qnn_Param_t qnn_params[] = {};
enum ggml_op ggmlop = GGML_OP_ADD;
Qnn_DataType_t src1_qnn_type = QNN_DATATYPE_FLOAT_32;
CHECK_PARAMS(ctx, src0, src1, dst);
tensor_1 = (Qnn_Tensor_t *) src1->extra;
instance = ctx->instance;
QNN_INTERFACE_VER_TYPE qnn_raw_interface = ctx->raw_interface;
qnn_perf perf("ggml_qnn_add");
perf.start();
QNN_VER_PTR(*tensor_1)->type = QNN_TENSOR_TYPE_APP_WRITE;
src1_qnn_type = qnn_datatype_from_ggml_datatype(src1->type);
uint32_t dimensions_input_1[] = {
(uint32_t) src1->ne[0], (uint32_t) src1->ne[1], (uint32_t) src1->ne[2],
(uint32_t) src1->ne[3]};
std::string map_entry = std::string(ggml_op_name(ggmlop));
if (instance->_qnn_graph_map.find(map_entry) !=
instance->_qnn_graph_map.end()) {
@ -2119,8 +2107,6 @@ static void ggml_qnn_add(ggml_backend_qnn_context * ctx, const ggml_tensor * src
graph_handle = std::get<0>(graph_item);
}
uint32_t * tensor_1_dimensions = QNN_VER_PTR(*tensor_1)->dimensions;
if (!graph_initialized) {
graph_name = graph_name + "_" + std::to_string(ctx->threads) +
"_" + src0->name + "_" + src1->name;
@ -2178,49 +2164,21 @@ static void ggml_qnn_add(ggml_backend_qnn_context * ctx, const ggml_tensor * src
QNN_LOG_INFO("create qnn graph handle with graph name %s ok\n", graph_name.c_str());
}
if (ctx->device == QNN_BACKEND_NPU) {
QNN_VER_PTR(*tensor_1)->memType = QNN_TENSORMEMTYPE_MEMHANDLE;
QNN_VER_PTR(*tensor_1)->clientBuf= {.data=nullptr, .dataSize=0};
}
ggml_qnn_tensor_writer tensor_writer0(src0, graph_handle, ctx);
if (!tensor_writer0.is_valid()) {
goto failure;
}
error = qnn_raw_interface.tensorCreateGraphTensor(graph_handle, tensor_1);
if (QNN_SUCCESS != error) {
ggml_qnn_tensor_writer tensor_writer1(src1, graph_handle, ctx);
if (!tensor_writer1.is_valid()) {
QNN_LOG_INFO("error = %d\n", error);
goto failure;
}
ggml_qnn_tensor_reader tensor_reader(dst, graph_handle, ctx);
if (!tensor_writer0.is_valid()) {
if (!tensor_reader.is_valid()) {
goto failure;
}
QNN_VER_PTR(*tensor_1)->dimensions = dimensions_input_1;
QNN_VER_PTR(*tensor_1)->rank = qnn_get_ggml_tensor_rank(src1);
QNN_VER_PTR(*tensor_1)->dataType = src1_qnn_type;
if (ctx->device != QNN_BACKEND_NPU) {
QNN_VER_PTR(*tensor_1)->clientBuf = {src1->data,
qnn_get_ggml_tensor_data_size(src1)};
} else {
uint8_t * qnn_buffer_1 = nullptr;
qnn_instance * instance = ctx->instance;
qnn_buffer_1 = static_cast<uint8_t *>(instance->alloc_rpcmem(
ggml_nbytes(src1), 4));
if (nullptr == qnn_buffer_1) {
QNN_LOG_WARN("alloc rpcmem failure, %s\n", strerror(errno));
goto failure;
} else {
QNN_LOG_INFO("alloc rpcmem successfully\n");
}
instance->register_rpcmem(qnn_buffer_1, tensor_1);
memcpy(qnn_buffer_1, src1->data, ggml_nbytes(src1));
}
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_1};
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_writer1.get_qnn_tensor()};
Qnn_Tensor_t tensor_outputs[] = {*tensor_reader.get_qnn_tensor()};
Qnn_OpConfig_t op_config = {
(Qnn_OpConfigVersion_t) 1,
@ -2256,33 +2214,18 @@ static void ggml_qnn_add(ggml_backend_qnn_context * ctx, const ggml_tensor * src
goto failure;
}
auto graph_item = std::make_tuple(graph_handle, tensor_writer0.get_qnn_tensor(), tensor_1, tensor_reader.get_qnn_tensor());
auto graph_item = std::make_tuple(graph_handle,
tensor_writer0.get_qnn_tensor(),
tensor_writer1.get_qnn_tensor(),
tensor_reader.get_qnn_tensor());
instance->_qnn_graph_map[map_entry] = graph_item;
} else {
auto & graph_item = instance->_qnn_graph_map[map_entry];
ggml_qnn_tensor_writer tensor_writer0(src0, std::get<1>(graph_item), ctx);
tensor_1 = std::get<2>(graph_item);
ggml_qnn_tensor_writer tensor_writer1(src1, std::get<2>(graph_item), ctx);
ggml_qnn_tensor_reader tensor_reader(dst, std::get<3>(graph_item), ctx);
uint32_t dimensions_input_1[] = {
(uint32_t) src1->ne[0], (uint32_t) src1->ne[1],
(uint32_t) src1->ne[2], (uint32_t) src1->ne[3]};
QNN_VER_PTR(*tensor_1)->dimensions = dimensions_input_1;
QNN_VER_PTR(*tensor_1)->rank = qnn_get_ggml_tensor_rank(src1);
QNN_VER_PTR(*tensor_1)->dataType = src1_qnn_type;
if (ctx->device != QNN_BACKEND_NPU) {
QNN_VER_PTR(*tensor_1)->clientBuf = {src1->data,
qnn_get_ggml_tensor_data_size(src1)};
} else {
uint8_t * qnn_buffer_1 = static_cast<uint8_t *>(ctx->instance->get_rpcmem_from_memhandle(
QNN_VER_PTR(*tensor_1)->memHandle));
if (nullptr != qnn_buffer_1)
memcpy(qnn_buffer_1, src1->data, ggml_nbytes(src1));
}
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_1};
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_writer1.get_qnn_tensor()};
Qnn_Tensor_t tensor_outputs[] = {*tensor_reader.get_qnn_tensor()};
error = qnn_raw_interface.graphExecute(graph_handle,
tensor_inputs,2,
@ -2301,7 +2244,6 @@ static void ggml_qnn_add(ggml_backend_qnn_context * ctx, const ggml_tensor * src
failure:
if (QNN_SUCCESS != error) {
QNN_LOG_DEBUG("tensor1 name %s", QNN_TENSOR_GET_NAME(*tensor_1));
QNN_LOG_DEBUG("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64
" x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi)\n",
src0->name, src0->type, ggml_type_name(src0->type),
@ -2319,8 +2261,6 @@ failure:
dst->nb[1], dst->nb[2]);
}
QNN_VER_PTR(*tensor_1)->dimensions = tensor_1_dimensions;
perf.info();
}
@ -2343,30 +2283,16 @@ static void ggml_qnn_mul_mat(ggml_backend_qnn_context * ctx,
qnn_instance * instance = nullptr;
std::string graph_name = "ggml_op_qnn_mul_mat";
Qnn_GraphHandle_t graph_handle = nullptr;
Qnn_Tensor_t * tensor_1 = nullptr;
Qnn_Param_t qnn_params[] = {};
enum ggml_op ggmlop = GGML_OP_MUL_MAT;
Qnn_DataType_t src1_qnn_type = QNN_DATATYPE_FLOAT_32;
CHECK_PARAMS(ctx, src0, src1, dst);
tensor_1 = (Qnn_Tensor_t *) src1->extra;
instance = ctx->instance;
QNN_INTERFACE_VER_TYPE qnn_raw_interface = ctx->raw_interface;
qnn_perf perf("ggml_qnn_mul_mat");
perf.start();
tensor_1 = (Qnn_Tensor_t *) src1->extra;
instance = ctx->instance;
QNN_VER_PTR(*tensor_1)->type = QNN_TENSOR_TYPE_APP_WRITE;
src1_qnn_type = qnn_datatype_from_ggml_datatype(src1->type);
uint32_t dimensions_input_1[] = {
(uint32_t) src1->ne[0], (uint32_t) src1->ne[1], (uint32_t) src1->ne[2],
(uint32_t) src1->ne[3]};
std::string map_entry = std::string(ggml_op_name(ggmlop));
if (instance->_qnn_graph_map.find(map_entry) !=
instance->_qnn_graph_map.end()) {
@ -2375,8 +2301,6 @@ static void ggml_qnn_mul_mat(ggml_backend_qnn_context * ctx,
graph_handle = std::get<0>(graph_item);
}
uint32_t * tensor_1_dimensions = QNN_VER_PTR(*tensor_1)->dimensions;
//TODO: for scenarios of quantized data in src0
// pass-1: dequantize src0 to FP32
// pass-2: dq-src0 * src1
@ -2436,49 +2360,20 @@ static void ggml_qnn_mul_mat(ggml_backend_qnn_context * ctx,
goto failure;
}
if (ctx->device == QNN_BACKEND_NPU) {
QNN_VER_PTR(*tensor_1)->memType = QNN_TENSORMEMTYPE_MEMHANDLE;
QNN_VER_PTR(*tensor_1)->clientBuf= {.data=nullptr, .dataSize=0};
}
ggml_qnn_tensor_writer tensor_writer0(src0, graph_handle, ctx);
if (!tensor_writer0.is_valid()) {
goto failure;
}
error = qnn_raw_interface.tensorCreateGraphTensor(graph_handle, tensor_1);
if (QNN_SUCCESS != error) {
QNN_LOG_INFO("error = %d\n", error);
ggml_qnn_tensor_writer tensor_writer1(src1, graph_handle, ctx);
if (!tensor_writer1.is_valid()) {
goto failure;
}
ggml_qnn_tensor_reader tensor_reader(dst, graph_handle, ctx);
if (!tensor_writer0.is_valid()) {
if (!tensor_reader.is_valid()) {
goto failure;
}
QNN_VER_PTR(*tensor_1)->dimensions = dimensions_input_1;
QNN_VER_PTR(*tensor_1)->rank = qnn_get_ggml_tensor_rank(src1);
QNN_VER_PTR(*tensor_1)->dataType = src1_qnn_type;
if (ctx->device != QNN_BACKEND_NPU) {
QNN_VER_PTR(*tensor_1)->clientBuf = {src1->data,
qnn_get_ggml_tensor_data_size(src1)};
} else {
uint8_t * qnn_buffer_1 = nullptr;
qnn_instance * instance = ctx->instance;
qnn_buffer_1 = static_cast<uint8_t *>(instance->alloc_rpcmem(
ggml_nbytes(src1), 4));
if (nullptr == qnn_buffer_1) {
QNN_LOG_WARN("alloc rpcmem failure, %s\n", strerror(errno));
goto failure;
} else {
QNN_LOG_INFO("alloc rpcmem successfully\n");
}
instance->register_rpcmem(qnn_buffer_1, tensor_1);
memcpy(qnn_buffer_1, src1->data, ggml_nbytes(src1));
}
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_1};
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_writer1.get_qnn_tensor()};
Qnn_Tensor_t tensor_outputs[] = {*tensor_reader.get_qnn_tensor()};
Qnn_OpConfig_t op_config = {
(Qnn_OpConfigVersion_t) 1,
@ -2514,32 +2409,18 @@ static void ggml_qnn_mul_mat(ggml_backend_qnn_context * ctx,
goto failure;
}
auto graph_item = std::make_tuple(graph_handle, tensor_writer0.get_qnn_tensor(), tensor_1, tensor_reader.get_qnn_tensor());
auto graph_item = std::make_tuple(graph_handle,
tensor_writer0.get_qnn_tensor(),
tensor_writer1.get_qnn_tensor(),
tensor_reader.get_qnn_tensor());
instance->_qnn_graph_map[map_entry] = graph_item;
} else {
auto & graph_item= instance->_qnn_graph_map[map_entry];
ggml_qnn_tensor_writer tensor_writer0(src0, std::get<1>(graph_item), ctx);
tensor_1 = std::get<2>(graph_item);
ggml_qnn_tensor_writer tensor_writer1(src1, std::get<2>(graph_item), ctx);
ggml_qnn_tensor_reader tensor_reader(dst, std::get<3>(graph_item), ctx);
uint32_t dimensions_input_1[] = {
(uint32_t) src1->ne[0], (uint32_t) src1->ne[1],
(uint32_t) src1->ne[2], (uint32_t) src1->ne[3]};
QNN_VER_PTR(*tensor_1)->dimensions = dimensions_input_1;
QNN_VER_PTR(*tensor_1)->rank = qnn_get_ggml_tensor_rank(src1);
QNN_VER_PTR(*tensor_1)->dataType = src1_qnn_type;
if (ctx->device != QNN_BACKEND_NPU) {
QNN_VER_PTR(*tensor_1)->clientBuf = {src1->data,
qnn_get_ggml_tensor_data_size(src1)};
} else {
uint8_t * qnn_buffer_1 = static_cast<uint8_t *>(ctx->instance->get_rpcmem_from_memhandle(
QNN_VER_PTR(*tensor_1)->memHandle));
if (nullptr != qnn_buffer_1)
memcpy(qnn_buffer_1, src1->data, ggml_nbytes(src1));
}
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_1};
Qnn_Tensor_t tensor_inputs[] = {*tensor_writer0.get_qnn_tensor(), *tensor_writer1.get_qnn_tensor()};
Qnn_Tensor_t tensor_outputs[] = {*tensor_reader.get_qnn_tensor()};
error = qnn_raw_interface.graphExecute(graph_handle,
tensor_inputs, 2,
@ -2558,7 +2439,6 @@ static void ggml_qnn_mul_mat(ggml_backend_qnn_context * ctx,
failure:
if (QNN_SUCCESS != error) {
QNN_LOG_DEBUG("tensor1 name %s", QNN_TENSOR_GET_NAME(*tensor_1));
QNN_LOG_DEBUG("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64
" x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi)\n",
src0->name, src0->type, ggml_type_name(src0->type),
@ -2575,7 +2455,6 @@ failure:
dst->ne[1], dst->ne[2], dst->nb[0], dst->nb[1], dst->nb[2]);
}
QNN_VER_PTR(*tensor_1)->dimensions = tensor_1_dimensions;
perf.info();
}