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

fix compiling error after merge latest master

Browse source
This commit is contained in:
hongruichen 2024-07-02 19:46:17 +08:00
parent 8b677d1b2f
commit 38f88d5fb1
2 changed files with 15 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

21
ggml/src/ggml-qnn.cpp
View file

@ -321,9 +321,7 @@ static bool ggml_qnn_can_handle_op(ggml_backend_qnn_context * ctx,
return true; return true;
} }
bool ggml_qnn_compute_forward(ggml_backend_qnn_context * ctx, bool ggml_qnn_compute_forward(ggml_backend_qnn_context * ctx, struct ggml_tensor * tensor) {
struct ggml_compute_params * params,
struct ggml_tensor * tensor) {
auto func = qnn::ggml_qnn_op_array()[tensor->op]; auto func = qnn::ggml_qnn_op_array()[tensor->op];
if (!func) { if (!func) {
QNN_LOG_WARN("unsupported op %d", tensor->op); QNN_LOG_WARN("unsupported op %d", tensor->op);
@ -515,13 +513,6 @@ GGML_CALL static size_t ggml_backend_qnn_buffer_type_get_max_size(ggml_backend_b
return (96 * 1024 * 1024); return (96 * 1024 * 1024);
} }
GGML_CALL static bool ggml_backend_qnn_buffer_type_supports_backend(
ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
GGML_UNUSED(buft);
return ggml_backend_is_qnn(backend) || ggml_backend_is_cpu(backend);
}
GGML_CALL static bool ggml_backend_qnn_buffer_is_host(ggml_backend_buffer_type_t buft) { GGML_CALL static bool ggml_backend_qnn_buffer_is_host(ggml_backend_buffer_type_t buft) {
GGML_UNUSED(buft); GGML_UNUSED(buft);
return true; return true;
@ -574,9 +565,6 @@ GGML_CALL static ggml_status ggml_backend_qnn_graph_compute(ggml_backend_t backe
ggml_backend_qnn_context * ctx = (ggml_backend_qnn_context *) backend->context; ggml_backend_qnn_context * ctx = (ggml_backend_qnn_context *) backend->context;
GGML_UNUSED(ctx); GGML_UNUSED(ctx);
ggml_compute_params params = {};
params.type = GGML_TASK_TYPE_COMPUTE;
params.ith = 0;
for (int i = 0; i < cgraph->n_nodes; i++) { for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i]; ggml_tensor * node = cgraph->nodes[i];
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE ||
@ -584,7 +572,7 @@ GGML_CALL static ggml_status ggml_backend_qnn_graph_compute(ggml_backend_t backe
node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) { node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue; continue;
} }
bool ok = ggml_qnn_compute_forward(ctx, &params, node); bool ok = ggml_qnn_compute_forward(ctx, node);
if (!ok) { if (!ok) {
QNN_LOG_DEBUG("error: op not supported %s (%s)\n", node->name, ggml_op_name(node->op)); QNN_LOG_DEBUG("error: op not supported %s (%s)\n", node->name, ggml_op_name(node->op));
} }
@ -616,9 +604,11 @@ static ggml_backend_i ggml_backend_qnn_interface = {
/* .synchronize = */ nullptr, /* .synchronize = */ nullptr,
/* .graph_plan_create = */ nullptr, /* .graph_plan_create = */ nullptr,
/* .graph_plan_free = */ nullptr, /* .graph_plan_free = */ nullptr,
/* .graph_plan_update = */ nullptr,
/* .graph_plan_compute = */ nullptr, /* .graph_plan_compute = */ nullptr,
/* .graph_compute = */ ggml_backend_qnn_graph_compute, /* .graph_compute = */ ggml_backend_qnn_graph_compute,
/* .supports_op = */ ggml_backend_qnn_supports_op, /* .supports_op = */ ggml_backend_qnn_supports_op,
/* .supports_buft = */ nullptr,
/* .offload_op = */ ggml_backend_qnn_offload_op, /* .offload_op = */ ggml_backend_qnn_offload_op,
/* .event_new = */ nullptr, /* .event_new = */ nullptr,
/* .event_free = */ nullptr, /* .event_free = */ nullptr,
@ -702,10 +692,9 @@ ggml_backend_buffer_type_t ggml_backend_qnn_buffer_type(size_t device) {
/* .get_alignment = */ ggml_backend_qnn_buffer_type_get_alignment, /* .get_alignment = */ ggml_backend_qnn_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_qnn_buffer_type_get_max_size, /* .get_max_size = */ ggml_backend_qnn_buffer_type_get_max_size,
/* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes /* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes
/* .supports_backend = */ ggml_backend_qnn_buffer_type_supports_backend,
/* .is_host = */ ggml_backend_qnn_buffer_is_host /* .is_host = */ ggml_backend_qnn_buffer_is_host
}, },
/* .context = */ & context, /* .context = */ &context,
}; };
} }
ggml_backend_qnn_buffer_type_initialized = true; ggml_backend_qnn_buffer_type_initialized = true;

24
ggml/src/ggml-qnn/backend-ops.cpp
View file

@ -8,21 +8,17 @@
static bool qnn_is_valid_params(ggml_backend_qnn_context* ctx, const ggml_tensor* src0, static bool qnn_is_valid_params(ggml_backend_qnn_context* ctx, const ggml_tensor* src0,
const ggml_tensor* src1, ggml_tensor* dst) { const ggml_tensor* src1, ggml_tensor* dst) {
if ((nullptr == ctx) || (nullptr == src0) || (nullptr == src1) || (nullptr == dst)) { if (!ctx || !src0 || !src1 || !dst) {
QNN_LOG_WARN("invalid params\n"); QNN_LOG_WARN("invalid params\n");
return false; return false;
} }
qnn::qnn_instance* instance = nullptr; auto* instance = ctx->instance;
Qnn_Tensor_t* tensor_0 = nullptr; auto* tensor0 = src0->extra;
Qnn_Tensor_t* tensor_1 = nullptr; auto* tensor1 = src1->extra;
Qnn_Tensor_t* tensor_2 = nullptr; auto* tensor2 = dst->extra;
tensor_0 = (Qnn_Tensor_t*)src0->extra; if (!instance || !tensor0 || !tensor1 || !tensor2) {
tensor_1 = (Qnn_Tensor_t*)src1->extra; QNN_LOG_WARN("invalid tensors\n");
tensor_2 = (Qnn_Tensor_t*)dst->extra;
instance = ctx->instance;
if ((nullptr == instance) || (nullptr == tensor_0) || (nullptr == tensor_1) || (nullptr == tensor_2)) {
QNN_LOG_WARN("invalid params\n");
return false; return false;
} }
@ -60,7 +56,7 @@ static void ggml_qnn_add(ggml_backend_qnn_context* ctx, const ggml_tensor* src0,
qnn::qnn_perf perf("ggml_qnn_add"); qnn::qnn_perf perf("ggml_qnn_add");
perf.start(); perf.start();
std::string map_entry = std::string(ggml_op_name(ggmlop)); std::string map_entry(ggml_op_name(ggmlop));
if (instance->_qnn_graph_map.find(map_entry) != if (instance->_qnn_graph_map.find(map_entry) !=
instance->_qnn_graph_map.end()) { instance->_qnn_graph_map.end()) {
graph_initialized = true; graph_initialized = true;
@ -141,8 +137,8 @@ static void ggml_qnn_add(ggml_backend_qnn_context* ctx, const ggml_tensor* src0,
goto failure; goto failure;
} }
Qnn_Tensor_t tensor_inputs[] = { *tensor_input0.get_qnn_tensor(), *tensor_input1.get_qnn_tensor() }; Qnn_Tensor_t tensor_inputs[] = { *tensor_input0.get_qnn_tensor(), *tensor_input1.get_qnn_tensor() };
Qnn_Tensor_t tensor_outputs[] = { *tensor_output.get_qnn_tensor() }; Qnn_Tensor_t tensor_outputs[] = { *tensor_output.get_qnn_tensor() };
Qnn_OpConfig_t op_config = { Qnn_OpConfig_t op_config = {
(Qnn_OpConfigVersion_t)1, (Qnn_OpConfigVersion_t)1,
.v1 = {"ggml_op_add", .v1 = {"ggml_op_add",