BatchVisualObservations() is not optimized. #927
Description
BatchVisualObservations() in CoreBrainInternal is very slow.
I think There are couple of reasons why:
- Based on profiling data, textures[b].height is very slow for some reason.
- pixel colors are not access in a continuous memory pattern.
The code I am using is like this, not sure if it is what you want but it is much faster
this List<Texture2D> textures, bool blackAndWhite)
{
int batchSize = textures.Count;
int width = textures[0].width;
int height = textures[0].height;
int pixels = 0;
if (blackAndWhite)
pixels = 1;
else
pixels = 3;
float[,,,] result = new float[batchSize, height, width, pixels];
float[] resultTemp = new float[batchSize * height * width * pixels];
int hwp = height * width * pixels;
int wp = width * pixels;
for (int b = 0; b < batchSize; b++)
{
Color32[] cc = textures[b].GetPixels32();
for (int h = height - 1; h >= 0; h--)
{
for (int w = 0; w < width; w++)
{
Color32 currentPixel = cc[(height - h - 1) * width + w];
if (!blackAndWhite)
{
// For Color32, the r, g and b values are between
// 0 and 255.
/*result[b, height - h - 1, w, 0] =
currentPixel.r / 255.0f;
result[b, height - h - 1, w, 1] =
currentPixel.g / 255.0f;
result[b, height - h - 1, w, 2] =
currentPixel.b / 255.0f;*/
resultTemp[b * hwp + h * wp + w * pixels] = currentPixel.r / 255.0f;
resultTemp[b * hwp + h * wp + w * pixels + 1] = currentPixel.g / 255.0f;
resultTemp[b * hwp + h * wp + w * pixels + 2] = currentPixel.b / 255.0f;
/*result[b,h,w,0] = currentPixel.r / 255.0f;
result[b, h, w, 1] = currentPixel.g / 255.0f;
result[b, h, w, 2] = currentPixel.b / 255.0f;*/
}
else
{
/*result[b, height - h - 1, w, 0] =
(currentPixel.r + currentPixel.g + currentPixel.b)
/ 3;*/
resultTemp[b * hwp + h * wp + w * pixels] =
(currentPixel.r + currentPixel.g + currentPixel.b)
/ 3;
/*result[b, h, w, 0] =
(currentPixel.r + currentPixel.g + currentPixel.b)
/ 3;*/
}
}
}
}
Buffer.BlockCopy(resultTemp, 0, result, 0, batchSize * height * width * pixels * sizeof(float));
return result;
}`