ShuriZma
/
suyu
mirror of https://git.suyu.dev/suyu/suyu
1
0
Fork 1
Code Issues Packages Projects Releases Wiki Activity

remove TexelWeightParams

This commit is contained in:
Ameer J 2023-08-01 23:41:46 -04:00
parent 75ac7845ce
commit fd2051b401
1 changed files with 31 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

77
src/video_core/host_shaders/astc_decoder.comp
View File

@ -36,12 +36,6 @@ struct EncodingData {
uint data; uint data;
}; };
struct TexelWeightParams {
uvec2 size;
uint max_weight;
bool dual_plane;
};
layout(binding = BINDING_INPUT_BUFFER, std430) readonly restrict buffer InputBufferU32 { layout(binding = BINDING_INPUT_BUFFER, std430) readonly restrict buffer InputBufferU32 {
uvec4 astc_data[]; uvec4 astc_data[];
}; };
@ -1026,59 +1020,50 @@ bool IsError(uint mode) {
return false; return false;
} }
TexelWeightParams DecodeBlockInfo(uint mode) { uvec2 DecodeBlockSize(uint mode) {
TexelWeightParams params = TexelWeightParams(uvec2(0), 0, false);
uint A, B; uint A, B;
uint mode_layout = FindLayout(mode); switch (FindLayout(mode)) {
switch (mode_layout) {
case 0: case 0:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
B = (mode >> 7) & 0x3; B = (mode >> 7) & 0x3;
params.size = uvec2(B + 4, A + 2); return uvec2(B + 4, A + 2);
break;
case 1: case 1:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
B = (mode >> 7) & 0x3; B = (mode >> 7) & 0x3;
params.size = uvec2(B + 8, A + 2); return uvec2(B + 8, A + 2);
break;
case 2: case 2:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
B = (mode >> 7) & 0x3; B = (mode >> 7) & 0x3;
params.size = uvec2(A + 2, B + 8); return uvec2(A + 2, B + 8);
break;
case 3: case 3:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
B = (mode >> 7) & 0x1; B = (mode >> 7) & 0x1;
params.size = uvec2(A + 2, B + 6); return uvec2(A + 2, B + 6);
break;
case 4: case 4:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
B = (mode >> 7) & 0x1; B = (mode >> 7) & 0x1;
params.size = uvec2(B + 2, A + 2); return uvec2(B + 2, A + 2);
break;
case 5: case 5:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
params.size = uvec2(12, A + 2); return uvec2(12, A + 2);
break;
case 6: case 6:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
params.size = uvec2(A + 2, 12); return uvec2(A + 2, 12);
break;
case 7: case 7:
params.size = uvec2(6, 10); return uvec2(6, 10);
break;
case 8: case 8:
params.size = uvec2(10, 6); return uvec2(10, 6);
break;
case 9: case 9:
A = (mode >> 5) & 0x3; A = (mode >> 5) & 0x3;
B = (mode >> 9) & 0x3; B = (mode >> 9) & 0x3;
params.size = uvec2(A + 6, B + 6); return uvec2(A + 6, B + 6);
break;
default: default:
break; return uvec2(0);
} }
params.dual_plane = (mode_layout != 9) && ((mode & 0x400) != 0); }
uint DecodeMaxWeight(uint mode) {
const uint mode_layout = FindLayout(mode);
uint weight_index = (mode & 0x10) != 0 ? 1 : 0; uint weight_index = (mode & 0x10) != 0 ? 1 : 0;
if (mode_layout < 5) { if (mode_layout < 5) {
weight_index |= (mode & 0x3) << 1; weight_index |= (mode & 0x3) << 1;
@ -1089,14 +1074,11 @@ TexelWeightParams DecodeBlockInfo(uint mode) {
if ((mode_layout != 9) && ((mode & 0x200) != 0)) { if ((mode_layout != 9) && ((mode & 0x200) != 0)) {
weight_index += 6; weight_index += 6;
} }
params.max_weight = weight_index + 1; return weight_index + 1;
return params;
} }
void DecompressBlock(ivec3 coord) { void DecompressBlock(ivec3 coord) {
uint mode = StreamBits(11); uint mode = StreamBits(11);
const TexelWeightParams params = DecodeBlockInfo(mode);
if (IsError(mode)) { if (IsError(mode)) {
FillError(coord); FillError(coord);
return; return;
@ -1106,12 +1088,15 @@ void DecompressBlock(ivec3 coord) {
FillVoidExtentLDR(coord); FillVoidExtentLDR(coord);
return; return;
} }
if ((params.size.x > block_dims.x) || (params.size.y > block_dims.y)) { const uvec2 size_params = DecodeBlockSize(mode);
if ((size_params.x > block_dims.x) || (size_params.y > block_dims.y)) {
FillError(coord); FillError(coord);
return; return;
} }
const uint num_partitions = StreamBits(2) + 1; const uint num_partitions = StreamBits(2) + 1;
if (num_partitions > 4 || (num_partitions == 4 && params.dual_plane)) { const uint mode_layout = FindLayout(mode);
const bool dual_plane = (mode_layout != 9) && ((mode & 0x400) != 0);
if (num_partitions > 4 || (num_partitions == 4 && dual_plane)) {
FillError(coord); FillError(coord);
return; return;
} }
@ -1127,7 +1112,8 @@ void DecompressBlock(ivec3 coord) {
base_cem = StreamBits(6); base_cem = StreamBits(6);
} }
const uint base_mode = base_cem & 3; const uint base_mode = base_cem & 3;
const uint weight_bits = GetPackedBitSize(params.size, params.dual_plane, params.max_weight); const uint max_weight = DecodeMaxWeight(mode);
const uint weight_bits = GetPackedBitSize(size_params, dual_plane, max_weight);
uint remaining_bits = 128 - weight_bits - total_bitsread; uint remaining_bits = 128 - weight_bits - total_bitsread;
uint extra_cem_bits = 0; uint extra_cem_bits = 0;
if (base_mode > 0) { if (base_mode > 0) {
@ -1146,7 +1132,7 @@ void DecompressBlock(ivec3 coord) {
} }
} }
remaining_bits -= extra_cem_bits; remaining_bits -= extra_cem_bits;
const uint plane_selector_bits = params.dual_plane ? 2 : 0; const uint plane_selector_bits = dual_plane ? 2 : 0;
remaining_bits -= plane_selector_bits; remaining_bits -= plane_selector_bits;
if (remaining_bits > 128) { if (remaining_bits > 128) {
// Bad data, more remaining bits than 4 bytes // Bad data, more remaining bits than 4 bytes
@ -1198,7 +1184,6 @@ void DecompressBlock(ivec3 coord) {
// This decode phase should at most push 32 elements into the vector // This decode phase should at most push 32 elements into the vector
result_vector_max_index = 32; result_vector_max_index = 32;
// uvec4 color_values[8];
uint colvals_index = 0; uint colvals_index = 0;
DecodeColorValues(color_endpoint_mode, num_partitions, color_data_bits); DecodeColorValues(color_endpoint_mode, num_partitions, color_data_bits);
for (uint i = 0; i < num_partitions; i++) { for (uint i = 0; i < num_partitions; i++) {
@ -1226,13 +1211,13 @@ void DecompressBlock(ivec3 coord) {
result_limit_reached = false; result_limit_reached = false;
// The limit for the Unquantize phase, avoids decoding more data than needed. // The limit for the Unquantize phase, avoids decoding more data than needed.
result_vector_max_index = params.size.x * params.size.y; result_vector_max_index = size_params.x * size_params.y;
if (params.dual_plane) { if (dual_plane) {
result_vector_max_index *= 2; result_vector_max_index *= 2;
} }
DecodeIntegerSequence(params.max_weight, GetNumWeightValues(params.size, params.dual_plane)); DecodeIntegerSequence(max_weight, GetNumWeightValues(size_params, dual_plane));
UnquantizeTexelWeights(params.size, params.dual_plane); UnquantizeTexelWeights(size_params, dual_plane);
for (uint j = 0; j < block_dims.y; j++) { for (uint j = 0; j < block_dims.y; j++) {
for (uint i = 0; i < block_dims.x; i++) { for (uint i = 0; i < block_dims.x; i++) {
uint local_partition = 0; uint local_partition = 0;
@ -1247,7 +1232,7 @@ void DecompressBlock(ivec3 coord) {
const uint vector_index = weight_offset % 4; const uint vector_index = weight_offset % 4;
const uint primary_weight = unquantized_texel_weights[array_index][vector_index]; const uint primary_weight = unquantized_texel_weights[array_index][vector_index];
uvec4 weight_vec = uvec4(primary_weight); uvec4 weight_vec = uvec4(primary_weight);
if (params.dual_plane) { if (dual_plane) {
const uint secondary_weight_offset = (j * block_dims.x + i) + ARRAY_NUM_ELEMENTS; const uint secondary_weight_offset = (j * block_dims.x + i) + ARRAY_NUM_ELEMENTS;
const uint secondary_array_index = secondary_weight_offset / 4; const uint secondary_array_index = secondary_weight_offset / 4;
const uint secondary_vector_index = secondary_weight_offset % 4; const uint secondary_vector_index = secondary_weight_offset % 4;