GPU execution of a compiled tensor Graph via generated compute shaders. 更多...

#include <GpuBackend.h>

类
struct	Impl

Public 成员函数
	~GpuProgram ()

std::vector< float >	run (const std::vector< const float * > &feeds) const
	feeds[slot] must point to `placeholderSize(slot)` floats. Returns the output buffer.

静态 Public 成员函数
static GpuProgram *	tryBuild (const Graph &graph, const OptimizedGraph &opt, int outputNode)

详细描述

GPU execution of a compiled tensor Graph via generated compute shaders.

Built once per CompiledFunction (shapes are static at compile time):

every fused group is lowered by KernelGen into a fully specialized GLSL kernel (shapes baked as constants, no per-dispatch shape work);
a static memory plan reuses arena buffers whenever node lifetimes don't overlap (AITemplate-style memory planning);
rank-2 matmuls are autotuned between a naive thread-per-output kernel and a shared-memory 16x16 tiled kernel.

tryBuild() never throws: it returns nullptr when Vulkan/gpgpu isn't available or a group cannot be lowered, in which case the caller falls back to the CPU interpreter.

在文件 GpuBackend.h 第 26 行定义.

构造及析构函数说明

◆ ~GpuProgram()

eve::tensor::GpuProgram::~GpuProgram ( )

在文件 GpuBackend.cpp 第 166 行定义.

成员函数说明

◆ run()

std::vector< float > eve::tensor::GpuProgram::run ( const std::vector< const float * > & feeds ) const

feeds[slot] must point to placeholderSize(slot) floats. Returns the output buffer.

在文件 GpuBackend.cpp 第 310 行定义.

◆ tryBuild()

GpuProgram * eve::tensor::GpuProgram::tryBuild	(	const Graph &	graph,
		const OptimizedGraph &	opt,
		int	outputNode
	)

static

在文件 GpuBackend.cpp 第 168 行定义.

引用了 eve::tensor::Alias, eve::tensor::Const, eve::tensor::GraphNode::constBytes, eve::tensor::GraphNode::constScales, eve::tensor::generateKernel(), eve::tensor::generateMatMulVariant(), graph, eve::tensor::OptimizedGraph::groupOrder, eve::tensor::OptimizedGraph::groups, eve::tensor::KernelSpec::groupsX2, eve::tensor::KernelSpec::groupsY2, id, impl, eve::tensor::GpuProgram::Impl::GroupRuntime::inputs, eve::tensor::MatMul, eve::tensor::OptimizedGraph::nodeSlot, eve::tensor::GpuProgram::Impl::GroupRuntime::output, eve::tensor::GpuProgram::Impl::GroupRuntime::pass1, eve::tensor::KernelSpec::pass1, eve::tensor::GpuProgram::Impl::GroupRuntime::pass2, eve::tensor::KernelSpec::pass2, eve::tensor::Placeholder, eve::tensor::GpuProgram::Impl::GroupRuntime::qScales, eve::tensor::GraphNode::rank, s, eve::tensor::OptimizedGraph::slotSize, eve::tensor::GpuProgram::Impl::GroupRuntime::spec, eve::tensor::GpuProgram::Impl::GroupRuntime::stats, eve::tensor::KernelSpec::statsCount, eve::tensor::KernelSpec::statsSize , 以及 eve::tensor::KernelSpec::twoPass.

被这些函数引用 eve::tensor::CompiledFunction::fromFunc().

该类的文档由以下文件生成:

src/modules/tensor/GpuBackend.h
src/modules/tensor/GpuBackend.cpp

类

Public 成员函数