|
|
|
@@ -98,10 +98,10 @@ void main() |
|
|
|
{ |
|
|
|
ivec4 x4 = i4; |
|
|
|
|
|
|
|
vr = image1d_ld4(bottom_blob_1d, x4.r); |
|
|
|
vg = image1d_ld4(bottom_blob_1d, x4.g); |
|
|
|
vb = image1d_ld4(bottom_blob_1d, x4.b); |
|
|
|
va = image1d_ld4(bottom_blob_1d, x4.a); |
|
|
|
vr = image1d_ld4(bottom_blob_1d, x4.r / 4); |
|
|
|
vg = image1d_ld4(bottom_blob_1d, x4.g / 4); |
|
|
|
vb = image1d_ld4(bottom_blob_1d, x4.b / 4); |
|
|
|
va = image1d_ld4(bottom_blob_1d, x4.a / 4); |
|
|
|
|
|
|
|
lane4 = x4 % 4; |
|
|
|
} |
|
|
|
@@ -110,10 +110,10 @@ void main() |
|
|
|
ivec4 y4 = i4 / psc(w); |
|
|
|
ivec4 x4 = i4 % psc(w); |
|
|
|
|
|
|
|
vr = image2d_ld4(bottom_blob_2d, ivec2(x4.r, y4.r)); |
|
|
|
vg = image2d_ld4(bottom_blob_2d, ivec2(x4.g, y4.g)); |
|
|
|
vb = image2d_ld4(bottom_blob_2d, ivec2(x4.b, y4.b)); |
|
|
|
va = image2d_ld4(bottom_blob_2d, ivec2(x4.a, y4.a)); |
|
|
|
vr = image2d_ld4(bottom_blob_2d, ivec2(x4.r, y4.r / 4)); |
|
|
|
vg = image2d_ld4(bottom_blob_2d, ivec2(x4.g, y4.g / 4)); |
|
|
|
vb = image2d_ld4(bottom_blob_2d, ivec2(x4.b, y4.b / 4)); |
|
|
|
va = image2d_ld4(bottom_blob_2d, ivec2(x4.a, y4.a / 4)); |
|
|
|
|
|
|
|
lane4 = y4 % 4; |
|
|
|
} |
|
|
|
@@ -125,10 +125,10 @@ void main() |
|
|
|
ivec4 y4 = i4 % size / psc(w); |
|
|
|
ivec4 x4 = i4 % size % psc(w); |
|
|
|
|
|
|
|
vr = image3d_ld4(bottom_blob_3d, ivec3(x4.r, y4.r, z4.r)); |
|
|
|
vg = image3d_ld4(bottom_blob_3d, ivec3(x4.g, y4.g, z4.g)); |
|
|
|
vb = image3d_ld4(bottom_blob_3d, ivec3(x4.b, y4.b, z4.b)); |
|
|
|
va = image3d_ld4(bottom_blob_3d, ivec3(x4.a, y4.a, z4.a)); |
|
|
|
vr = image3d_ld4(bottom_blob_3d, ivec3(x4.r, y4.r, z4.r / 4)); |
|
|
|
vg = image3d_ld4(bottom_blob_3d, ivec3(x4.g, y4.g, z4.g / 4)); |
|
|
|
vb = image3d_ld4(bottom_blob_3d, ivec3(x4.b, y4.b, z4.b / 4)); |
|
|
|
va = image3d_ld4(bottom_blob_3d, ivec3(x4.a, y4.a, z4.a / 4)); |
|
|
|
|
|
|
|
lane4 = z4 % 4; |
|
|
|
} |
|
|
|
|