2 ms·
It seems like the author has very muddled notions of the separation between task parallelism and data parallelism. He talks about examples that could be data pa
by KirinDave 17y ago
It seems like the author has very muddled notions of the separation between task parallelism and data parallelism. He talks about examples that could be data parallelism but then uses terminology like it's task parallelism. His section on race conditions is likewise muddied at best.
His basic argument at the top seems correct if you limit it to data parallelism. But if you read the rest of the article the author doesn't seem to follow the thread of his own argument.
At its core the notion that data parallelism is easier to debug than control parallelism is obvious. It is, and its often hardware supported. But unfortunately data parallelism is harder to write and often requires skill in math, so it's not used nearly so much.