thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

Reverted back to cpp

Browse source
This commit is contained in:
Alejandro Saucedo 2020-08-22 15:21:20 +01:00
parent 18fba90ec0
commit ce00048d8c
3 changed files with 26 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

5
shaders/glsl/opmult.comp
View file

@ -12,6 +12,10 @@ layout(binding = 2) buffer tensorOutput {
uint valuesOutput[ ];
};
layout(binding = 3) buffer tensorInvalid {
uint valuesInvalid[ ];
};
// TODO: Explore how to make layout inside shader dynamic
layout (local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
@ -24,6 +28,7 @@ void main()
valuesOutput[index] = 100 + index;
valuesRhs[index] = 100 + index;
valuesLhs[index] = 100 + index;
valuesInvalid[index] = 100 + index;
}

BIN
shaders/glsl/opmult.comp.spv
View file

Binary file not shown.

22
src/OpMult.tpp → src/OpMult.cpp
View file

@ -127,14 +127,34 @@ OpMult<tX, tY, tZ>::record()
vk::AccessFlagBits::eTransferRead,
vk::PipelineStageFlagBits::eComputeShader,
vk::PipelineStageFlagBits::eTransfer);
this->mTensorLHS->recordBufferMemoryBarrier(
vk::AccessFlagBits::eShaderWrite,
vk::AccessFlagBits::eTransferRead,
vk::PipelineStageFlagBits::eComputeShader,
vk::PipelineStageFlagBits::eTransfer);
this->mTensorRHS->recordBufferMemoryBarrier(
vk::AccessFlagBits::eShaderWrite,
vk::AccessFlagBits::eTransferRead,
vk::PipelineStageFlagBits::eComputeShader,
vk::PipelineStageFlagBits::eTransfer);
this->mTensorOutputStaging->recordCopyFrom(this->mTensorOutput);
this->mTensorOutputStaging->recordCopyFrom(this->mTensorLHS);
this->mTensorOutput->recordBufferMemoryBarrier(
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
this->mTensorLHS->recordBufferMemoryBarrier(
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
this->mTensorRHS->recordBufferMemoryBarrier(
vk::AccessFlagBits::eTransferWrite,
vk::AccessFlagBits::eHostRead,
vk::PipelineStageFlagBits::eTransfer,
vk::PipelineStageFlagBits::eHost);
}
template<uint32_t tX, uint32_t tY, uint32_t tZ>