Support inferences with multiple inputs and outputs
Build flatbuffers library.
Update network class to extract IFM and OFM dimensions from the tflite
file.
Update the uapi and core apis to support up to 16 IFM and OFM buffers
per inference.
Change-Id: I2f2f177aa4c2d5f9f50f23eb33c44e01ec2cbe09
diff --git a/driver_library/src/ethosu.cpp b/driver_library/src/ethosu.cpp
index 6b2b3b1..39e39f0 100644
--- a/driver_library/src/ethosu.cpp
+++ b/driver_library/src/ethosu.cpp
@@ -16,6 +16,8 @@
* limitations under the License.
*/
+#include "autogen/tflite_schema.hpp"
+
#include <ethosu.hpp>
#include <uapi/ethosu.h>
@@ -43,11 +45,50 @@
return ret;
}
+
+/****************************************************************************
+ * TFL micro helpers
+ ****************************************************************************/
+
+size_t getShapeSize(const flatbuffers::Vector<int32_t> *shape)
+{
+ size_t size = 1;
+
+ for (auto it = shape->begin(); it != shape->end(); ++it)
+ {
+ size *= *it;
+ }
+
+ return size;
+}
+
+vector<size_t> getSubGraphDims(const tflite::SubGraph *subgraph, const flatbuffers::Vector<int32_t> *tensorMap)
+{
+ vector<size_t> dims;
+
+ for (auto index = tensorMap->begin(); index != tensorMap->end(); ++index)
+ {
+ auto tensor = subgraph->tensors()->Get(*index);
+ size_t size = getShapeSize(tensor->shape());
+
+ if (size > 0)
+ {
+ dims.push_back(size);
+ }
+ }
+
+ return dims;
+}
+
}
namespace EthosU
{
+/****************************************************************************
+ * Exception
+ ****************************************************************************/
+
Exception::Exception(const char *msg) :
msg(msg)
{}
@@ -60,6 +101,10 @@
return msg.c_str();
}
+/****************************************************************************
+ * Device
+ ****************************************************************************/
+
Device::Device(const char *device)
{
fd = open(device, O_RDWR | O_NONBLOCK);
@@ -79,6 +124,10 @@
return eioctl(fd, cmd, data);
}
+/****************************************************************************
+ * Buffer
+ ****************************************************************************/
+
Buffer::Buffer(Device &device, const size_t capacity) :
fd(-1),
dataPtr(nullptr),
@@ -121,7 +170,6 @@
ethosu_uapi_buffer uapi;
uapi.offset = offset;
uapi.size = size;
-
eioctl(fd, ETHOSU_IOCTL_BUFFER_SET, static_cast<void *>(&uapi));
}
@@ -144,15 +192,29 @@
return fd;
}
+/****************************************************************************
+ * Network
+ ****************************************************************************/
+
Network::Network(Device &device, shared_ptr<Buffer> &buffer) :
fd(-1),
buffer(buffer)
{
+ // Create buffer handle
ethosu_uapi_network_create uapi;
-
uapi.fd = buffer->getFd();
-
fd = device.ioctl(ETHOSU_IOCTL_NETWORK_CREATE, static_cast<void *>(&uapi));
+
+ // Create model handle
+ const tflite::Model *model = tflite::GetModel(reinterpret_cast<void *>(buffer->data()));
+
+ // Get input dimensions for first subgraph
+ auto *subgraph = *model->subgraphs()->begin();
+ ifmDims = getSubGraphDims(subgraph, subgraph->inputs());
+
+ // Get output dimensions for last subgraph
+ subgraph = *model->subgraphs()->rbegin();
+ ofmDims = getSubGraphDims(subgraph, subgraph->outputs());
}
Network::~Network()
@@ -165,30 +227,83 @@
return eioctl(fd, cmd, data);
}
-std::shared_ptr<Buffer> Network::getBuffer()
+shared_ptr<Buffer> Network::getBuffer()
{
return buffer;
}
-Inference::Inference(std::shared_ptr<Network> &network, std::shared_ptr<Buffer> &ifmBuffer, std::shared_ptr<Buffer> &ofmBuffer) :
- fd(-1),
- network(network),
- ifmBuffer(ifmBuffer),
- ofmBuffer(ofmBuffer)
+const std::vector<size_t> &Network::getIfmDims() const
{
- ethosu_uapi_inference_create uapi;
-
- uapi.ifm_fd = ifmBuffer->getFd();
- uapi.ofm_fd = ofmBuffer->getFd();
-
- fd = network->ioctl(ETHOSU_IOCTL_INFERENCE_CREATE, static_cast<void *>(&uapi));
+ return ifmDims;
}
+size_t Network::getIfmSize() const
+{
+ size_t size = 0;
+
+ for (auto s: ifmDims)
+ {
+ size += s;
+ }
+
+ return size;
+}
+
+const std::vector<size_t> &Network::getOfmDims() const
+{
+ return ofmDims;
+}
+
+size_t Network::getOfmSize() const
+{
+ size_t size = 0;
+
+ for (auto s: ofmDims)
+ {
+ size += s;
+ }
+
+ return size;
+}
+
+/****************************************************************************
+ * Inference
+ ****************************************************************************/
+
Inference::~Inference()
{
close(fd);
}
+void Inference::create()
+{
+ ethosu_uapi_inference_create uapi;
+
+ if (ifmBuffers.size() > ETHOSU_FD_MAX)
+ {
+ throw Exception("IFM buffer overflow");
+ }
+
+ if (ofmBuffers.size() > ETHOSU_FD_MAX)
+ {
+ throw Exception("OFM buffer overflow");
+ }
+
+ uapi.ifm_count = 0;
+ for (auto it: ifmBuffers)
+ {
+ uapi.ifm_fd[uapi.ifm_count++] = it->getFd();
+ }
+
+ uapi.ofm_count = 0;
+ for (auto it: ofmBuffers)
+ {
+ uapi.ofm_fd[uapi.ofm_count++] = it->getFd();
+ }
+
+ fd = network->ioctl(ETHOSU_IOCTL_INFERENCE_CREATE, static_cast<void *>(&uapi));
+}
+
void Inference::wait(int timeoutSec)
{
pollfd pfd;
@@ -214,19 +329,19 @@
return fd;
}
-std::shared_ptr<Network> Inference::getNetwork()
+shared_ptr<Network> Inference::getNetwork()
{
return network;
}
-std::shared_ptr<Buffer> Inference::getIfmBuffer()
+vector<shared_ptr<Buffer>> &Inference::getIfmBuffers()
{
- return ifmBuffer;
+ return ifmBuffers;
}
-std::shared_ptr<Buffer> Inference::getOfmBuffer()
+vector<shared_ptr<Buffer>> &Inference::getOfmBuffers()
{
- return ofmBuffer;
+ return ofmBuffers;
}
}