To automate high-throughput phenotyping for infield biomass sorghum morphological traits characterization, a capable 3D vision system that can overcome challenges imposed by field conditions including variable lighting, strong wind and extreme plant height is needed. Among all available 3D sensors, traditional stereo cameras offer a viable solution to obtaining high-resolution 3D point-cloud data with the use of high-accuracy (sub-pixel) stereo matching algorithms, which, however, are inevitably highly computational. This paper reports a GPU-based parallelized implementation of the PatchMatch Stereo algorithm which reconstructs highly slanted leaf and stalk surfaces of sorghum at high speed from high-resolution stereo image pairs. Our algorithm enhanced accuracy and smoothness by using L2 norm for color distance calculation instead of L1 norm and speeded up convergence by testing the plane of the lowest cost within a local window in addition to the original spatial propagation. To better handle textureless regions, after left-right consistency check, the disparity of an occluded pixel is assigned to that of a nearby non-occluded pixel with the most similar pattern. Some of these occluded pixels in textureless region would survive a following left-right consistency check. Therefore more valid pixels would exist in textureless regions for occlusion filling. Accuracy and performance were evaluated on Middlebury datasets as well as our sorghum datasets. It achieved a high ranking in Middlebury table of subpixel precision and revealed subtle details on leaf and stalk surfaces. The output disparity maps were used to estimate stalk diameters of different varieties and growth stages. The results showed high correlation to hand measurement.