Removed bug calling start/end during sequence submit phase

src/Algorithm.cpp

@@ -49,7 +49,7 @@ Algorithm::createParameters(std::vector<std::shared_ptr<Tensor>>& tensorParams)
     // TODO: Explore design for having multiple descriptor pool sizes
     std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
         vk::DescriptorPoolSize(vk::DescriptorType::eStorageBuffer,
-                               1 // Descriptor count
+                               static_cast<uint32_t>(tensorParams.size()) // Descriptor count
                                )
     };
 
@@ -99,6 +99,11 @@ Algorithm::createParameters(std::vector<std::shared_ptr<Tensor>>& tensorParams)
     this->mDevice->allocateDescriptorSets(&descriptorSetAllocateInfo,
                                           this->mDescriptorSet.get());
 
+    std::vector<vk::DescriptorBufferInfo> descriptorBufferInfos;
+    for (size_t i = 0; i < tensorParams.size(); i++) {
+        descriptorBufferInfos.push_back(tensorParams[i]->constructDescriptorBufferInfo());
+    }
+    // TODO: Explore design exposing the destination array element
     std::vector<vk::WriteDescriptorSet> computeWriteDescriptorSets;
     for (size_t i = 0; i < tensorParams.size(); i++) {
 

src/OpMult.cpp

@@ -109,6 +109,7 @@ OpMult<tX, tY, tZ>::record()
 {
     SPDLOG_DEBUG("Kompute OpMult record called");
 
+    // Barrier to ensure the data is finished writing to buffer memory
     this->mTensorLHS->recordBufferMemoryBarrier(
         vk::AccessFlagBits::eHostWrite,
         vk::AccessFlagBits::eShaderRead,
@@ -122,6 +123,7 @@ OpMult<tX, tY, tZ>::record()
 
     this->mAlgorithm->recordDispatch(this->mX, this->mY, this->mZ);
 
+    // Barrier to ensure the shader code is executed before buffer read
     this->mTensorOutput->recordBufferMemoryBarrier(
         vk::AccessFlagBits::eShaderWrite,
         vk::AccessFlagBits::eTransferRead,
@@ -140,6 +142,7 @@ OpMult<tX, tY, tZ>::record()
 
     this->mTensorOutputStaging->recordCopyFrom(this->mTensorLHS);
 
+    // Buffer to ensure wait until data is copied to staging buffer
     this->mTensorOutput->recordBufferMemoryBarrier(
         vk::AccessFlagBits::eTransferWrite,
         vk::AccessFlagBits::eHostRead,

src/OpMult.hpp

@@ -52,5 +52,5 @@ class OpMult : public OpBase
 } // End namespace kp
 
 // Including implemenation for template class
-#include "OpMult.tpp"
+#include "OpMult.cpp"
 

src/Sequence.cpp

@@ -97,11 +97,6 @@ Sequence::eval()
 {
     SPDLOG_DEBUG("Kompute sequence compute recording EVAL");
 
-    bool toggleSingleRecording = !this->mRecording;
-    if (toggleSingleRecording) {
-        this->begin();
-    }
-
     const vk::PipelineStageFlags waitStageMask =
       vk::PipelineStageFlagBits::eTransfer;
     vk::SubmitInfo submitInfo(
@@ -116,10 +111,6 @@ Sequence::eval()
     this->mDevice->waitForFences(1, &fence, VK_TRUE, UINT64_MAX);
     this->mDevice->destroy(fence);
 
-    if (toggleSingleRecording) {
-        this->end();
-    }
-
     // TODO: Explore whether moving postSubmit calls to a separate sequence
     // function that is explicitly called by the manager
     for (size_t i = 0; i < this->mOperations.size(); i++) {

src/Tensor.cpp

@@ -58,7 +58,7 @@ Tensor::init(std::shared_ptr<vk::PhysicalDevice> physicalDevice,
              std::vector<uint32_t> data)
 {
     SPDLOG_DEBUG(
-      "Kompute Tensor running init with physicalDevice and logical device");
+      "Kompute Tensor running init with Vulkan params and data size: {}", data.size());
 
     this->mPhysicalDevice = physicalDevice;
     this->mDevice = device;