tree fb2911bb0522d8ab3ce7aa9d7dd14c2b7e1023e6
parent c4129446f0d4739c7fd576cac6692534af0a4225
author Lei Zhang <antiagainst@google.com> 1606860799 -0500
committer GitHub <noreply@github.com> 1606860799 -0500
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJfxr//CRBK7hj4Ov3rIwAAdHIIAJqjZvVA6Qa4V7PbbEEMAN1i
 V5LBuY8Bg4bhXb6qB2APETMcibPJIbDbh4HhfY+ZJiQqe6Uavtt5HmV0VrrTu/EG
 5wuiW2FJPW3RwGqSbLmTjIIa9//6H4Rlv7nsqs3Dl2XoLQz1zehlL3Z8xyu5Afu1
 Wjnz/nwEfyI3t9V+QmdvMPXMK+KIZHm/KGq89amcXytdL83X0V1ZGeSVTlxTP8Qr
 r6qJajAVE96CVSXQrafV2lMQ54nEse5eB/+UYw6Hy540663Wzh3cJHk43nBVuzs4
 tQn+pmkAHOQS+kZrVY6my8yuc3eWakrYWY+V/gs+UJsV12t7Dp7AeV3pDspwoeY=
 =cUKu
 -----END PGP SIGNATURE-----
 

[spirv] Add a better tiled and vectorized convolution path (#3990)

This commit tiles a 2-D convolution along the output height, width,
and channel dimensions. The tiled convolution is distributed to
workgroup of size 32x8x1, with each group containing 8x2x1
invocations. This is just the initial configuration for Mali to
get started. After tiling, we further vectorize the convolution
for each GPU invocation.