#version 450 // (Triforce) What the GPU drew into a small texture, written into the texture's storage in guest memory as the console's GPU // would have left it there (vk_texture.cpp, texture_end_recording). The image's levels and layers were copied into `lin`, // row by row; `sto` holds the storage as guest memory has it now; `map` says, for each unit of the storage (a unit is one // block: 1, 2, 4, 8 or 16 bytes), where in `lin` its bytes are, or 0xffffffff where the storage holds no texel (the padding of // the block-linear layout, which stays as it was). The map is made on the CPU with the same swizzle the uploads undo, so // this shader knows nothing of the layout. One invocation a 32-bit word of the storage: no two write the same word. layout(local_size_x = 64) in; layout(std430, binding = 0) readonly buffer Map { uint map[]; }; layout(std430, binding = 1) readonly buffer Lin { uint lin[]; }; layout(std430, binding = 2) buffer Sto { uint sto[]; }; layout(push_constant) uniform P { uint words; uint unit_shift; } pc; uint lin_byte(uint at) { return (lin[at >> 2] >> ((at & 3u) * 8u)) & 0xffu; } void main() { const uint w = gl_GlobalInvocationID.x; if (w >= pc.words) return; const uint unit_mask = (1u << pc.unit_shift) - 1u; uint v = sto[w]; for (uint b = 0u; b < 4u; b++) { const uint at = w * 4u + b; const uint from = map[at >> pc.unit_shift]; if (from == 0xffffffffu) continue; v = (v & ~(0xffu << (b * 8u))) | (lin_byte(from + (at & unit_mask)) << (b * 8u)); } sto[w] = v; }