matrix_ops.md

Matrix Operations (GPU)

Overview

GPU-accelerated matrix operations for graph algorithms. Uses shared-memory tiling and CUDA kernels.

Source

core/task/gpu_task/matrix_ops.cuh

---

Matmult

Computes C = A × B on GPU.

Parameter	Type	Description
A	float*	Matrix A (row-major), shape (m × k)
B	float*	Matrix B (row-major), shape (k × n)
C	float*	Output (row-major), shape (m × n)
m	int	Rows of A and C
n	int	Columns of B and C
k	int	Columns of A / rows of B

Example:

#include "core/task/gpu_task/matrix_ops.cuh"
using sics::matrixgraph::core::task::MatrixOps;

auto* task = new MatrixOps();
task->Matmult(A, B, C, m, k, n);
delete task;

---

Activate

Applies ReLU activation in-place on a GPU array.

Parameter	Type	Description
A	float*	Array (row-major)
n	int	Number of elements

Example:

auto* task = new MatrixOps();
task->Activate(A, m * n);
delete task;

---

Usage with Unified Buffer

#include "core/data_structures/unified_buffer.cuh"
#include "core/data_structures/host_buffer.cuh"
#include "core/task/gpu_task/matrix_ops.cuh"

using sics::matrixgraph::core::task::MatrixOps;
using UnifiedOwnedBufferFloat =
    sics::matrixgraph::core::data_structures::UnifiedOwnedBuffer<float>;

UnifiedOwnedBufferFloat buf_A, buf_B, buf_C;
// ... init buffers ...

auto* task = new MatrixOps();
task->Matmult(buf_A.GetPtr(), buf_B.GetPtr(), buf_C.GetPtr(), m, k, n);
delete task;