鹵化物：根據一堆輸入圖像（3D緩沖區）計算結果圖像（2D緩沖區）

Question

我是Halide的初學者。 我想從一堆輸入圖像中計算出一個結果圖像，我想使用GPU對其進行加速（當前，我使用CUDA作為目標）。 例如，我要計算4張圖像的atan2，如下所示：

Func result;
result(x,y) = Halide::atan2(images(x,y,3) - images(x,y,1), images(x,y,0) - images(x,y,2));

我對此有疑問。 它產生一個空結果數組或帶有偽造值的結果數組。

請注意，如果我修改函數以僅處理單個圖像並傳遞二維輸入緩沖區，它將產生正確的結果，但是一旦我將輸入數組變為三維並發送了N個圖像的堆棧，就會產生一些結果錯誤。

我的代碼中有錯誤嗎？ 當我傳遞3D緩沖區（帶有N張圖像）但僅在Halide函數中循環x和y時，我需要做些特別的事情嗎？

請參閱下面的完整代碼：

#include "Halide.h"
#include <stdio.h>
#include <math.h>
#include <string>
#include <fstream>
#include <sstream>

using namespace Halide;

void generatePhaseImage(Halide::Buffer<float> image, float phi)
{
    for (int y=0; y<image.height(); y++)
    {
        for (int x=0; x<image.width(); x++)
        {
            image(x,y) = 128.0f + 128.0f * sin((2.0f*M_PI * static_cast<float>(x) / 64.0f) + phi);
        }
    }
}    

void writeBinaryFile(const std::string& filename, const Buffer<float>& image)
{
    std::ofstream f(filename.c_str(), std::ios::binary);

    std::cout << "Writing image of height " << image.height() << " ";
    std::cout << "and width " << image.width() << std::endl;

    for (int i=0; i<image.height(); i++)
    {
        for (int j=0; j<image.width(); j++)
        {
            f.write(reinterpret_cast<const char*>(&image(i,j)), sizeof(float));
        }
    }
}

int main(int argc, char **argv) 
{
    Var x, y, c, i, ii, xo, yo, xi, yi;

    int h = 100;
    int w = 100;

    Buffer<float> images(w, h, 4);

    for (int i=0; i<4; i++)
    {
        float phi = i * (2*M_PI / 4.);        
        generatePhaseImage(images.sliced(2, i), phi);
    }

    Func phaseStepping;
    phaseStepping(x,y) = Halide::atan2( images(x,y,3) - images(x,y,1), images(x,y,0) - images(x,y,2));

    for (int i=0; i<4; i++)
    {
        std::stringstream ss;
        ss << "image" << i <<".bin";
        writeBinaryFile(ss.str(), images.sliced(2, i));
    }

    Target target = get_host_target();
    target.set_feature(Target::CUDA);
    target.set_feature(Target::Debug);

    phaseStepping.gpu_tile(x, y, xo, yo, xi, yi, 8, 8);
    phaseStepping.compile_jit(target);

    Buffer<float> result(w,h);
    phaseStepping.realize(result);

    result.copy_to_host();  
    writeBinaryFile("result.bin", result);
}

代碼的輸出如下所示：

Writing image of height 100 and width 100
Writing image of height 100 and width 100
Writing image of height 100 and width 100
Writing image of height 100 and width 100
Entering Pipeline phaseStepping
 Input Buffer images: buffer(0, 0x0, 0x564812eefd80, 0, float32, {0, 100, 1}, {0, 100, 100}, {0, 4, 10000})
 Input (void *) __user_context: 0x7ffff380ac68
 Output Buffer phaseStepping: buffer(0, 0x0, 0x56481325b200, 0, float32, {0, 100, 1}, {0, 100, 100})
CUDA: halide_cuda_initialize_kernels (user_context: 0x0, state_ptr: 0x7f76cbb2b000, ptx_src: 0x7f76cbb28340, size: 6919
    load_libcuda (user_context: 0x0)
    Loaded CUDA runtime library: libcuda.so
    Got device 0
      GeForce GTX 560
      total memory: 959 MB
      max threads per block: 1024
      warp size: 32
      max block size: 1024 1024 64
      max grid size: 65535 65535 65535
      max shared memory per block: 49152
      max constant memory per block: 65536
      compute capability 2.1
      cuda cores: 7 x 48 = 48
    cuCtxCreate 0 -> 0x5648134ab8d0(3020)
    cuModuleLoadData 0x7f76cbb28340, 6919 -> 0x56481386f850
    Time: 1.857735e+00 ms
halide_copy_to_device 0x564812e02a68, host: 0x56481325b200, dev: 0, host_dirty: 0, dev_dirty: 0
halide_device_malloc: 0x564812e02a68 interface 0x7f76cbb340f0 host: 0x56481325b200, dev: 0, host_dirty: 0, dev_dirty:0 buf current interface: 0x0
CUDA: halide_cuda_device_malloc (user_context: 0x7ffff380ac68, buf: 0x564812e02a68)
    allocating buffer(0, 0x0, 0x56481325b200, 0, float32, {0, 100, 1}, {0, 100, 100})
    cuMemAlloc 40000 -> 0x501700000
    Time: 1.549260e-01 ms
halide_copy_to_device 0x564812dc6418, host: 0x564812eefd80, dev: 0, host_dirty: 0, dev_dirty: 0
halide_device_malloc: 0x564812dc6418 interface 0x7f76cbb340f0 host: 0x564812eefd80, dev: 0, host_dirty: 0, dev_dirty:0 buf current interface: 0x0
CUDA: halide_cuda_device_malloc (user_context: 0x7ffff380ac68, buf: 0x564812dc6418)
    allocating buffer(0, 0x0, 0x564812eefd80, 0, float32, {0, 100, 1}, {0, 100, 100}, {0, 4, 10000})
    cuMemAlloc 160000 -> 0x501800000
    Time: 1.099330e-01 ms
CUDA: halide_cuda_run (user_context: 0x7ffff380ac68, entry: kernel_phaseStepping_s0_y_yo___block_id_y, blocks: 13x13x1, threads: 8x8x1, shmem: 0
Got context.
Got module 0x56481386f850
Got function 0x56481387b2b0
    halide_cuda_run 0 4 [0x6400000064 ...] 0
    halide_cuda_run 1 4 [0x64 ...] 0
    halide_cuda_run 2 4 [0x0 ...] 0
    halide_cuda_run 3 4 [0x6400000000 ...] 0
    halide_cuda_run 4 4 [0x4000000064 ...] 0
    halide_cuda_run 5 8 [0x501800000 ...] 1
    halide_cuda_run 6 8 [0x501700000 ...] 1
    halide_cuda_run translated arg5 [0x501800000 ...]
    halide_cuda_run translated arg6 [0x501700000 ...]
    Time: 4.394600e-02 ms
Exiting Pipeline phaseStepping
halide_copy_to_host 0x564812e02a68
copy_to_host_already_locked 0x564812e02a68 dev_dirty is true
CUDA: halide_cuda_copy_to_host (user_context: 0x0, buf: 0x564812e02a68)
c.extent[0] = 100
c.extent[1] = 100
    cuMemcpyDtoH 0x501700000 -> 0x56481325b200, 40000 bytes
    Time: 2.062520e-01 ms
Writing image of height 100 and width 100
halide_device_free: 0x564812e02a68 buf dev 21498953728 interface 0x7f76cbb340f0
CUDA: halide_cuda_device_free (user_context: 0x0, buf: 0x564812e02a68)
    cuMemFree 0x501700000
    Time: 7.846700e-02 ms
halide_device_free: 0x564812dc6418 buf dev 21500002304 interface 0x7f76cbb340f0
CUDA: halide_cuda_device_free (user_context: 0x0, buf: 0x564812dc6418)
    cuMemFree 0x501800000
    Time: 8.416100e-02 ms

Answer 1

根據更新后的問題下的評論，我相當確定問題是輸入Buffer沒有被標記為主機臟，因此不會將其復制到GPU。 這在緩沖區邏輯中有點bug，但是我不確定我們是否能夠修復它。 顯式調用Buffer上的“ set_host_dirty”應該可以解決該問題。