Matmul.cpp

Go to the documentation of this file.

// ----------------------------------------------------------------------------
// -                        CloudViewer: www.cloudViewer.org                  -
// ----------------------------------------------------------------------------
// Copyright (c) 2018-2024 www.cloudViewer.org
// SPDX-License-Identifier: MIT
// ----------------------------------------------------------------------------
 
#include "cloudViewer/core/linalg/Matmul.h"
 
#include <unordered_map>
 
#include "cloudViewer/core/CUDAUtils.h"
 
namespace cloudViewer {
namespace core {
 
void Matmul(const Tensor& A, const Tensor& B, Tensor& output) {
    AssertTensorDevice(B, A.GetDevice());
    AssertTensorDtype(B, A.GetDtype());
 
    const Device device = A.GetDevice();
    const Dtype dtype_original = A.GetDtype();
    Dtype dtype;
 
    if (dtype_original != core::Float32 && dtype_original != core::Float64) {
        utility::LogDebug("Converting to Float32 dtype to from {}.",
                          dtype_original.ToString());
        dtype = core::Float32;
    } else {
        dtype = dtype_original;
    }
 
    // Check shapes
    SizeVector A_shape = A.GetShape();
    SizeVector B_shape = B.GetShape();
 
    if (A_shape.size() != 2) {
        utility::LogError("Tensor A must be 2D, but got {}D.", A_shape.size());
    }
    if (B_shape.size() != 1 && B_shape.size() != 2) {
        utility::LogError(
                "Tensor B must be 1D (vector) or 2D (matrix), but got {}D.",
                B_shape.size());
    }
    if (A_shape[1] != B_shape[0]) {
        utility::LogError("Tensor A columns {} mismatch with Tensor B rows {}.",
                          A_shape[1], B_shape[0]);
    }
 
    // Dispatch to backends
    int64_t m = A_shape[0];
    int64_t k = A_shape[1];
    int64_t n = B_shape.size() == 2 ? B_shape[1] : 1;
 
    if (m == 0 || k == 0 || n == 0) {
        utility::LogError(
                "Tensor shapes should not contain dimensions with zero.");
    }
 
    Tensor A_contiguous = A.Contiguous().To(dtype);
    Tensor B_contiguous = B.Contiguous().To(dtype);
    void* A_data = A_contiguous.GetDataPtr();
    void* B_data = B_contiguous.GetDataPtr();
 
    output = Tensor::Empty({m, n}, dtype, device);
    void* C_data = output.GetDataPtr();
 
    if (device.IsSYCL()) {
#ifdef BUILD_SYCL_MODULE
        MatmulSYCL(B_data, A_data, C_data, n, k, m, dtype, device);
#else
        utility::LogError("Unimplemented device.");
#endif
    } else if (device.IsCUDA()) {
#ifdef BUILD_CUDA_MODULE
        CUDAScopedDevice scoped_device(device);
        MatmulCUDA(B_data, A_data, C_data, n, k, m, dtype, device);
#else
        utility::LogError("Unimplemented device.");
#endif
    } else {
        MatmulCPU(B_data, A_data, C_data, n, k, m, dtype);
    }
 
    output = output.To(dtype_original);
};
 
}  // namespace core
}  // namespace cloudViewer