Convolutional Neural Network Operations

GPU kernels for convolution and pooling operations using ShaderM monad. Supports 2D convolution, depthwise convolution, and pooling layers.

structure Hesper.NN.Conv.Conv2DConfig : Type

Configuration for 2D convolution

• batch : Nat

  Input dimensions: [batch, height, width, channels]

• inputHeight : Nat
• inputWidth : Nat
• inputChannels : Nat
• kernelHeight : Nat

  Kernel dimensions

• kernelWidth : Nat
• outputChannels : Nat

  Output channels

• stride : Nat

  Stride

• padding : Nat

  Padding

• workgroupSize : Nat

  Workgroup size

def Hesper.NN.Conv.instInhabitedConv2DConfig.default : Conv2DConfig

Equations

• One or more equations did not get rendered due to their size.

instance Hesper.NN.Conv.instInhabitedConv2DConfig : Inhabited Conv2DConfig

Equations

• Hesper.NN.Conv.instInhabitedConv2DConfig = { default := Hesper.NN.Conv.instInhabitedConv2DConfig.default }

def Hesper.NN.Conv.instReprConv2DConfig.repr : Conv2DConfig → Nat → Std.Format

Equations

• One or more equations did not get rendered due to their size.

instance Hesper.NN.Conv.instReprConv2DConfig : Repr Conv2DConfig

Equations

• Hesper.NN.Conv.instReprConv2DConfig = { reprPrec := Hesper.NN.Conv.instReprConv2DConfig.repr }

def Hesper.NN.Conv.Conv2DConfig.outputHeight (c : Conv2DConfig) : Nat

Output height after convolution

Equations

• c.outputHeight = (c.inputHeight + 2 * c.padding - c.kernelHeight) / c.stride + 1

def Hesper.NN.Conv.Conv2DConfig.outputWidth (c : Conv2DConfig) : Nat

Output width after convolution

Equations

• c.outputWidth = (c.inputWidth + 2 * c.padding - c.kernelWidth) / c.stride + 1

def Hesper.NN.Conv.Conv2DConfig.numWorkgroups (c : Conv2DConfig) : Nat × Nat × Nat

Number of workgroups

Equations

• c.numWorkgroups = ((c.outputWidth + c.workgroupSize - 1) / c.workgroupSize, (c.outputHeight + c.workgroupSize - 1) / c.workgroupSize, c.outputChannels)

structure Hesper.NN.Conv.PoolingConfig : Type

Configuration for pooling operations

• batch : Nat

  Input dimensions: [batch, height, width, channels]

• inputHeight : Nat
• inputWidth : Nat
• channels : Nat
• poolHeight : Nat

  Pool dimensions

• poolWidth : Nat
• stride : Nat

  Stride

• workgroupSize : Nat

  Workgroup size

def Hesper.NN.Conv.instInhabitedPoolingConfig.default : PoolingConfig

Equations

• One or more equations did not get rendered due to their size.

instance Hesper.NN.Conv.instInhabitedPoolingConfig : Inhabited PoolingConfig

Equations

• Hesper.NN.Conv.instInhabitedPoolingConfig = { default := Hesper.NN.Conv.instInhabitedPoolingConfig.default }

instance Hesper.NN.Conv.instReprPoolingConfig : Repr PoolingConfig

Equations

• Hesper.NN.Conv.instReprPoolingConfig = { reprPrec := Hesper.NN.Conv.instReprPoolingConfig.repr }

def Hesper.NN.Conv.instReprPoolingConfig.repr : PoolingConfig → Nat → Std.Format

Equations

• One or more equations did not get rendered due to their size.

def Hesper.NN.Conv.PoolingConfig.outputHeight (c : PoolingConfig) : Nat

Output height after pooling

Equations

• c.outputHeight = (c.inputHeight - c.poolHeight) / c.stride + 1

def Hesper.NN.Conv.PoolingConfig.outputWidth (c : PoolingConfig) : Nat

Output width after pooling

Equations

• c.outputWidth = (c.inputWidth - c.poolWidth) / c.stride + 1

def Hesper.NN.Conv.PoolingConfig.numWorkgroups (c : PoolingConfig) : Nat × Nat × Nat

Number of workgroups

Equations

• c.numWorkgroups = ((c.outputWidth + c.workgroupSize - 1) / c.workgroupSize, (c.outputHeight + c.workgroupSize - 1) / c.workgroupSize, c.channels)

def Hesper.NN.Conv.maxPoolingKernel (config : PoolingConfig) : WGSL.Monad.ShaderM Unit

Max pooling kernel using ShaderM monad

Equations

• One or more equations did not get rendered due to their size.

def Hesper.NN.Conv.avgPoolingKernel (config : PoolingConfig) : WGSL.Monad.ShaderM Unit

Average pooling kernel using ShaderM monad

Equations

• One or more equations did not get rendered due to their size.

def Hesper.NN.Conv.conv2DKernel (config : Conv2DConfig) : WGSL.Monad.ShaderM Unit

2D convolution kernel using ShaderM monad (NHWC format)

Equations

• One or more equations did not get rendered due to their size.

def Hesper.NN.Conv.depthwiseConv2DKernel (config : Conv2DConfig) : WGSL.Monad.ShaderM Unit

Depthwise convolution kernel using ShaderM monad (for MobileNets)

Equations

• One or more equations did not get rendered due to their size.

def Hesper.NN.Conv.generateConv2DShaderFromMonad (config : Conv2DConfig) : String

Generate WGSL shader for 2D convolution

Equations

• Hesper.NN.Conv.generateConv2DShaderFromMonad config = Hesper.WGSL.CodeGen.generateWGSL "main" { x := config.workgroupSize, y := config.workgroupSize } [] [] (Hesper.NN.Conv.conv2DKernel config)

def Hesper.NN.Conv.generateDepthwiseConv2DShaderFromMonad (config : Conv2DConfig) : String

Generate WGSL shader for depthwise convolution

Equations

• One or more equations did not get rendered due to their size.

def Hesper.NN.Conv.generateMaxPoolingShaderFromMonad (config : PoolingConfig) : String

Generate WGSL shader for max pooling

Equations

• One or more equations did not get rendered due to their size.

def Hesper.NN.Conv.generateAvgPoolingShaderFromMonad (config : PoolingConfig) : String

Generate WGSL shader for average pooling

Equations

• One or more equations did not get rendered due to their size.