gpu: fix 10-11-11 render-target storage

Co-authored-by: Çağatay Güçlü <33038942+defektu@users.noreply.github.com>
This commit is contained in:
nmzik
2026-10-04 21:18:19 +04:00
parent 07c84876a1
commit 5940e623d5
15 changed files with 319 additions and 88 deletions
@@ -0,0 +1,20 @@
// ColorTransform values match imageInfo.h. Preserve the floating-point bit fields exactly.
uvec2 transform_color(uvec2 value, uint transform, bool to_host) {
switch (transform) {
case 1u:
return uvec2((value.y & 0xffffu) | (value.x & 0xffff0000u),
(value.x & 0xffffu) | (value.y & 0xffff0000u));
case 2u: {
uint packed = value.x;
if (to_host) {
value.x = (packed >> 21u) | (((packed >> 10u) & 0x7ffu) << 11u) |
((packed & 0x3ffu) << 22u);
} else {
value.x = (packed >> 22u) | (((packed >> 11u) & 0x7ffu) << 10u) |
((packed & 0x7ffu) << 21u);
}
break;
}
}
return value;
}
@@ -1,4 +1,5 @@
#version 450
#extension GL_GOOGLE_include_directive : require
layout(local_size_x = 64) in;
@@ -19,17 +20,23 @@ layout(push_constant, std430) uniform Push {
uint tail_x;
uint tail_y;
uint tail;
uint color_transform;
uint to_host;
} params;
#include "gpu_tiler_color.inc"
void main() {
uint pixel = gl_GlobalInvocationID.x;
if (pixel >= params.width) {
return;
}
uint src = (params.src_base >> 2u) + pixel * 2u;
uint dst = (params.dst_base >> 2u) + pixel * 2u;
uint bg = input_buffer.data[src];
uint ra = input_buffer.data[src + 1u];
output_buffer.data[dst] = (ra & 0x0000ffffu) | (bg & 0xffff0000u);
output_buffer.data[dst + 1u] = (bg & 0x0000ffffu) | (ra & 0xffff0000u);
uint words = params.color_transform == 1u ? 2u : 1u;
uint src = (params.src_base >> 2u) + pixel * words;
uint dst = (params.dst_base >> 2u) + pixel * words;
uvec2 value = uvec2(input_buffer.data[src], 0u);
if (words == 2u) value.y = input_buffer.data[src + 1u];
value = transform_color(value, params.color_transform, params.to_host != 0u);
output_buffer.data[dst] = value.x;
if (words == 2u) output_buffer.data[dst + 1u] = value.y;
}
@@ -17,10 +17,19 @@ layout(set = 0, binding = 2, std140) uniform Push {
uint tail_x;
uint tail_y;
uint tail;
uint color_transform;
} params;
#include "gpu_tiler_color.inc"
void copy_element(uint src, uint dst) {
if (ELEMENT_BYTES >= 4) {
if ((ELEMENT_BYTES == 4u || ELEMENT_BYTES == 8u) && params.color_transform != 0u) {
uvec2 value = uvec2(input_buffer.data[src >> 2], 0u);
if (ELEMENT_BYTES == 8u) value.y = input_buffer.data[(src >> 2) + 1u];
value = transform_color(value, params.color_transform, TILE == 0u);
output_buffer.data[dst >> 2] = value.x;
if (ELEMENT_BYTES == 8u) output_buffer.data[(dst >> 2) + 1u] = value.y;
} else if (ELEMENT_BYTES >= 4) {
for (uint i = 0; i < ELEMENT_BYTES; i += 4) {
output_buffer.data[(dst + i) >> 2] = input_buffer.data[(src + i) >> 2];
}