我試圖使用OpenCL在GPU上運行this代碼的部分代碼。我現在試圖運行處理YCbCr到RGB轉換的函數。OpenCL內核導致應用程序無限期地運行,並且僅在關閉我的IDE後停止
請注意,截至目前我沒有試圖優化GPU代碼。我只是想要一個與CPU上的輸出相同的輸出。
該功能最初是這樣寫的:
void YCbCr_to_ARGB(uint8_t *YCbCr_MCU[3], uint32_t *RGB_MCU, uint32_t nb_MCU_H, uint32_t nb_MCU_V)
{
uint8_t *MCU_Y, *MCU_Cb, *MCU_Cr;
int R, G, B;
uint32_t ARGB;
uint8_t index, i, j;
MCU_Y = YCbCr_MCU[0];
MCU_Cb = YCbCr_MCU[1];
MCU_Cr = YCbCr_MCU[2];
for (i = 0; i < 8 * nb_MCU_V; i++) {
for (j = 0; j < 8 * nb_MCU_H; j++) {
index = i * (8 * nb_MCU_H) + j;
R = (MCU_Cr[index] - 128) * 1.402f + MCU_Y[index];
B = (MCU_Cb[index] - 128) * 1.7772f + MCU_Y[index];
G = MCU_Y[index] - (MCU_Cb[index] - 128) * 0.34414f -
(MCU_Cr[index] - 128) * 0.71414f;
/* Saturate */
if (R > 255)
R = 255;
if (R < 0)
R = 0;
if (G > 255)
G = 255;
if (G < 0)
G = 0;
if (B > 255)
B = 255;
if (B < 0)
B = 0;
ARGB = ((R & 0xFF) << 16) | ((G & 0xFF) << 8) | (B & 0xFF);
// ARGB = 0xFF << 8;
RGB_MCU[(i * (8 * nb_MCU_H) + j)] = ARGB;
}
}
}
這個函數的變量聲明中main.c
以下列方式:
cl_uchar* YCbCr_MCU[3] = { NULL, NULL, NULL};
cl_uint* RGB_MCU = NULL;
內存這些變量在此被分配方式:
if (screen_init_needed == 1)
{
screen_init_needed = 0;
.....
.....
//Some code
for (index = 0 ; index < SOF_section.n ; index++) {
YCbCr_MCU[index] = malloc(MCU_sx * MCU_sy * max_ss_h * max_ss_v);
YCbCr_MCU_ds[index] = malloc(MCU_sx * MCU_sy * max_ss_h * max_ss_v);
}
RGB_MCU = malloc (MCU_sx * MCU_sy * max_ss_h * max_ss_v * sizeof(cl_int));
}
break;
}
我直接複製並將其粘貼到我的.cl
文件中,並做了一些小改動以使其與OpenCL標準一致。我修改的OpenCL代碼是這樣的:
__kernel void YCbCr_to_ARGB(__global uchar* YCbCr_MCU[3], __global uint* RGB_MCU, uint nb_MCU_H, uint nb_MCU_V)
{
__global uchar *MCU_Y, *MCU_Cb, *MCU_Cr;
int R, G, B;
uint ARGB;
uchar index, i, j;
MCU_Y = YCbCr_MCU[0];
MCU_Cb = YCbCr_MCU[1];
MCU_Cr = YCbCr_MCU[2];
//Same code as the first code snippet
......
......
......
}
當我建立和運行了我的.cl
文件我用上面的內核代碼的應用程序,我得到了錯誤。其中一個錯誤指出OpenCL不允許指向指針參數的指針。
爲了避開這些錯誤,我再次修改我的代碼看起來像這樣:
__kernel void YCbCr_to_ARGB(__global uchar YCbCr_MCU[3], __global uint* RGB_MCU, uint nb_MCU_H, uint nb_MCU_V)
{
__global uchar *MCU_Y, *MCU_Cb, *MCU_Cr;
int R, G, B;
uint ARGB;
uchar index, i, j;
MCU_Y = &YCbCr_MCU[0];
MCU_Cb = &YCbCr_MCU[1];
MCU_Cr = &YCbCr_MCU[2];
//Same code as the first code snippet
......
......
......
}
當我建立並再次運行應用程序,我沒有得到任何錯誤。這促使我寫這個內核的主機代碼。
它看起來像這樣:
color_kernel= clCreateKernel(program, "YCbCr_to_ARGB", &ret);
//YCbCr_MCU for YCbCrtoARGB
cl_mem colorMCU_GPU= clCreateBuffer(context, CL_MEM_READ_WRITE, 3 * sizeof(cl_uchar), NULL, &ret);
//rgb_MCU for YCbCrtoARGB
cl_mem RGB_GPU= clCreateBuffer(context, CL_MEM_READ_WRITE, sizeof(cl_uint), NULL, &ret);
我所謂的內核參數正是原函數被調用的main.c
。我執行的其餘步驟此內核通過以下方式:
if(color&&(SOF_section.n>1)
{
ret = clEnqueueWriteBuffer(command_queue, colorMCU_GPU, CL_TRUE, 0, 3 * sizeof(cl_uchar), YCbCr_MCU, 0, NULL, NULL);
ret = clEnqueueWriteBuffer(command_queue, RGB_GPU, CL_TRUE, 0, sizeof(cl_uint), RGB_MCU, 0, NULL, NULL);
ret = clSetKernelArg(color_kernel, 0, sizeof(cl_mem), (void *)&colorMCU_GPU);
ret |= clSetKernelArg(color_kernel, 1, sizeof(cl_mem), (void *)&RGB_GPU);
ret = clSetKernelArg(color_kernel, 2, sizeof(cl_uint), (void *)&max_ss_h);
ret |= clSetKernelArg(color_kernel, 3, sizeof(cl_uint), (void *)&max_ss_v);
ret = clEnqueueTask(command_queue, color_kernel, 0, NULL, NULL);
ret = clEnqueueReadBuffer(command_queue, RGB_GPU, CL_TRUE, 0, sizeof(cl_uint), RGB_MCU, 0, NULL, NULL);
//YCbCr_to_ARGB(YCbCr_MCU, RGB_MCU, max_ss_h, max_ss_v);
我運行,並利用這些參數,生成代碼,該代碼保持無限期運行後(這樣做的輸出應該是一個影片剪輯上運行有了這個代碼,我只能看到一個黑屏)。我必須關閉Eclipse並重新打開它,以便在此之後對代碼進行其他更改。
是什麼導致程序行爲像這樣?無論如何要安全地在GPU上運行此功能?
更新:
我跟着安德斯Cedronius的意見,改變了以下列方式我的內核代碼:
__kernel void YCbCr_to_ARGB(__global uchar YCbCr_MCU[3], __global uint* RGB_MCU, uint nb_MCU_H, uint nb_MCU_V)
{
printf("Doing color conversion\n");
__global uchar *MCU_Y, *MCU_Cb, *MCU_Cr;
int R, G, B;
uint ARGB;
uchar index, i, j;
i= get_global_id(0);
j= get_global_id(1);
MCU_Y = &YCbCr_MCU[0];
MCU_Cb = &YCbCr_MCU[1];
MCU_Cr = &YCbCr_MCU[2];
if (i < 8 * nb_MCU_V && j < 8 * nb_MCU_H)
{
index = i * (8 * nb_MCU_H) + j;
R = (MCU_Cr[index] - 128) * 1.402f + MCU_Y[index];
B = (MCU_Cb[index] - 128) * 1.7772f + MCU_Y[index];
G = MCU_Y[index] - (MCU_Cb[index] - 128) * 0.34414f -
(MCU_Cr[index] - 128) * 0.71414f;
/* Saturate */
if (R > 255)
R = 255;
if (R < 0)
R = 0;
if (G > 255)
G = 255;
if (G < 0)
G = 0;
if (B > 255)
B = 255;
if (B < 0)
B = 0;
ARGB = ((R & 0xFF) << 16) | ((G & 0xFF) << 8) | (B & 0xFF);
// ARGB = 0xFF << 8;
RGB_MCU[(i * (8 * nb_MCU_H) + j)] = ARGB;
}
printf("Finished color conversion\n");
}
現在調用內核我的主機代碼如下所示:
color_kernel= clCreateKernel(program, "YCbCr_to_ARGB", &ret);
我以下面的方式設置工作大小和內核參數:
ret = clEnqueueWriteBuffer(command_queue, colorMCU_GPU, CL_TRUE, 0, 3*sizeof(cl_uchar), YCbCr_MCU, 0, NULL, NULL);
chk(ret, "clEnqueueWriteBuffer");
ret = clEnqueueWriteBuffer(command_queue, RGB_GPU, CL_TRUE, 0, sizeof(cl_uint), RGB_MCU, 0, NULL, NULL);
chk(ret, "clEnqueueWriteBuffer");
ret = clSetKernelArg(color_kernel, 0, sizeof(cl_mem), (void *)&colorMCU_GPU);
ret |= clSetKernelArg(color_kernel, 1, sizeof(cl_mem), (void *)&RGB_GPU);
ret = clSetKernelArg(color_kernel, 2, sizeof(cl_uint), (void *)&max_ss_h);
ret |= clSetKernelArg(color_kernel, 3, sizeof(cl_uint), (void *)&max_ss_v);
size_t itemColor[2] = {1, 1};
ret = clEnqueueNDRangeKernel(command_queue, kernel, 2, NULL, itemColor, NULL, 0, NULL, NULL);
chk(ret, "clEnqueueNDRange");
ret = clEnqueueReadBuffer(command_queue, RGB_GPU, CL_TRUE, 0, sizeof(cl_uint), RGB_MCU, 0, NULL, NULL);
clFinish(command_queue);
我運行了這段代碼,我不再有黑屏。但是,「YCbCr到RGB」的內核現在還沒有被識別。 即使我的printf評論沒有顯示在輸出控制檯上。這就像我的代碼沒有顏色轉換功能。
更新:
我沒有在命令EnqueueNDRangeKernel
改變了我的內核的名稱。我改了名字,現在printf語句出現在控制檯上。但是,我仍然沒有得到正確的輸出結果。
size_t itemColor[2] = {1, 1};
ret = clEnqueueNDRangeKernel(command_queue, color_kernel, 2, NULL, itemColor, NULL, 0, NULL, NULL);
chk(ret, "clEnqueueNDRange");
clFinish(command_queue);
更新:
我也跟着pmdj的建議,修改了自己的內核代碼。現在看起來是這樣的:
__kernel void YCbCr_to_ARGB(__global uchar* Y_GPU, __global uchar* Cb_GPU, __global uchar* Cr_GPU, __global uint* RGB_MCU, uint nb_MCU_H, uint nb_MCU_V)
{
__global uchar *MCU_Y, *MCU_Cb, *MCU_Cr;
int R, G, B;
uint ARGB;
uchar index, i, j;
unsigned char iid= get_global_id(0);
unsigned char jid= get_global_id(1);
// MCU_Y = &YCbCr_MCU[0];
// MCU_Cb = &YCbCr_MCU[1];
// MCU_Cr = &YCbCr_MCU[2];
MCU_Y= Y_GPU;
MCU_Cb= Cb_GPU;
MCU_Cr= Cr_GPU;
if (iid <= (8 * nb_MCU_V) && jid <= (8 * nb_MCU_H))
{
index = iid * (8 * nb_MCU_H) + jid;
R = (MCU_Cr[index] - 128) * 1.402f + MCU_Y[index];
B = (MCU_Cb[index] - 128) * 1.7772f + MCU_Y[index];
G = MCU_Y[index] - (MCU_Cb[index] - 128) * 0.34414f -
(MCU_Cr[index] - 128) * 0.71414f;
/* Saturate */
if (R > 255)
R = 255;
if (R < 0)
R = 0;
if (G > 255)
G = 255;
if (G < 0)
G = 0;
if (B > 255)
B = 255;
if (B < 0)
B = 0;
ARGB = ((R & 0xFF) << 16) | ((G & 0xFF) << 8) | (B & 0xFF);
RGB_MCU[(iid * (8 * nb_MCU_H) + jid)] = ARGB;
}
}
在主機代碼,我創建了4個新的變量分配內存:
Y_ForGPU= (cl_uchar *)malloc(MCU_sx * MCU_sy * max_ss_h * max_ss_v);
Cb_ForGPU= (cl_uchar *)malloc(MCU_sx * MCU_sy * max_ss_h * max_ss_v);
Cr_ForGPU= (cl_uchar *)malloc(MCU_sx * MCU_sy * max_ss_h * max_ss_v);
//Now will do it for RGB
RGB_testing= (cl_uint *)malloc (MCU_sx * MCU_sy * max_ss_h * max_ss_v * sizeof(cl_int));
我以下列方式創建緩衝區:
cl_mem for_Y= clCreateBuffer(context, CL_MEM_READ_WRITE| CL_MEM_COPY_HOST_PTR, (MCU_sx * MCU_sy * max_ss_h * max_ss_v), Y_ForGPU, &ret);
cl_mem for_Cb= clCreateBuffer(context, CL_MEM_READ_WRITE| CL_MEM_COPY_HOST_PTR, (MCU_sx * MCU_sy * max_ss_h * max_ss_v), Cb_ForGPU , &ret);
cl_mem for_Cr= clCreateBuffer(context, CL_MEM_READ_WRITE| CL_MEM_COPY_HOST_PTR, (MCU_sx * MCU_sy * max_ss_h * max_ss_v), Cr_ForGPU, &ret);
//rgb_MCU for YCbCrtoARGB
cl_mem RGB_GPU= clCreateBuffer(context, CL_MEM_READ_WRITE, (MCU_sx * MCU_sy * max_ss_h * max_ss_v * sizeof(cl_int)), NULL, &ret);
然後我設置內核參數,執行內核並將計算的數據發送回主機:
ret = clSetKernelArg(color_kernel, 0, sizeof(cl_mem), &for_Y);
ret |= clSetKernelArg(color_kernel, 1, sizeof(cl_mem), &for_Cb);
ret |= clSetKernelArg(color_kernel, 2, sizeof(cl_mem), &for_Cr);
ret |= clSetKernelArg(color_kernel, 3, sizeof(cl_mem), &RGB_GPU);
ret |= clSetKernelArg(color_kernel, 4, sizeof(cl_uint), &max_ss_h);
ret |= clSetKernelArg(color_kernel, 5, sizeof(cl_uint), &max_ss_v);
const size_t itemColor[2] = {100, 100};
ret = clEnqueueNDRangeKernel(command_queue, color_kernel, 2, NULL, itemColor, NULL, 0, NULL, NULL);
clFinish(command_queue);
//Copy result to the host
ret = clEnqueueReadBuffer(command_queue, RGB_GPU, CL_TRUE, 0, (MCU_sx * MCU_sy * max_ss_h * max_ss_v * sizeof(cl_int)), RGB_testing, 0, NULL, NULL);
但是,現在我的代碼突然終止。爲什麼會發生這種情況?
更新:
我的代碼現在正在工作。這些問題可能是由於指針的差異而發生的。我設置Y,Cb,Cr和RGB變量(我創建的)等於主機代碼中的原始變量。
//---Setting color variables equal to array elements----//
Y_ForGPU= YCbCr_MCU[0];
Cb_ForGPU= YCbCr_MCU[1];
Cr_ForGPU= YCbCr_MCU[2];
//----RGB is being assigned value-----//
RGB_testing= RGB_MCU;
你不應該在內核程序中有內循環。內核程序應該只對單個「元素」執行數學YUV-> RGB,然後使用clEnqueueNDRangeKernel將工作人員分割爲GPU。 –
@AndersCedronius我已更新我的代碼以包含您的建議。但是,我的內核現在還沒有被識別。我在上面的問題中包含了細節。 –
不要只是猜測的事情。如果你想做一些有意義的事情,拿起一本書並開始閱讀,那麼你將無法獲得任何其他地方。 – Jovasa