Eden/eden
1
0
Fork 0
mirror of https://git.eden-emu.dev/eden-emu/eden synced 2026-02-04 02:51:18 +01:00
Code Issues Packages Projects Releases Wiki Activity

[astc] Optimizations of astc decoding compute shader

Browse Source 
1. The Select2DPartition function was recalculating Hash52 and generating 8 random seeds for every single pixel in a block.
Now it precomputes its and stores it in a table.
2. Instead of Float math use integer math
3. Pr 3437 changes
This commit is contained in:
wildcard
2026-02-01 20:44:20 +01:00
parent 5113f503d1
commit d2965ed0b1
1 changed files with 180 additions and 144 deletions
Download Patch File Download Diff File Expand all files Collapse all files

324
src/video_core/host_shaders/astc_decoder.comp
View File

@@ -83,6 +83,12 @@ int result_index = 0;
uint result_vector_max_index;
bool result_limit_reached = false;
// avoid intermediate result_vector storage during color decode phase
bool write_color_values = false;
uint color_values_direct[32];
uint color_out_index = 0;
uint color_num_values = 0;
// EncodingData helpers
uint Encoding(EncodingData val) {
return bitfieldExtract(val.data, 0, 8);
@@ -114,9 +120,110 @@ EncodingData CreateEncodingData(uint encoding, uint num_bits, uint bit_val, uint
return EncodingData(((encoding) << 0u) | ((num_bits) << 8u) |
((bit_val) << 16u) | ((quint_trit_val) << 24u));
}
uint ReplicateBitTo9(uint bit);
uint FastReplicateTo8(uint value, uint num_bits);
void EmitColorValue(EncodingData val) {
// write directly to color_values_direct[]
const uint encoding = Encoding(val);
const uint bitlen = NumBits(val);
const uint bitval = BitValue(val);
if (encoding == JUST_BITS) {
color_values_direct[++color_out_index] = FastReplicateTo8(bitval, bitlen);
return;
}
uint A = ReplicateBitTo9((bitval & 1));
uint B = 0, C = 0, D = QuintTritValue(val);
if (encoding == TRIT) {
switch (bitlen) {
case 1:
C = 204;
break;
case 2: {
C = 93;
const uint b = (bitval >> 1) & 1;
B = (b << 8) | (b << 4) | (b << 2) | (b << 1);
break;
}
case 3: {
C = 44;
const uint cb = (bitval >> 1) & 3;
B = (cb << 7) | (cb << 2) | cb;
break;
}
case 4: {
C = 22;
const uint dcb = (bitval >> 1) & 7;
B = (dcb << 6) | dcb;
break;
}
case 5: {
C = 11;
const uint edcb = (bitval >> 1) & 0xF;
B = (edcb << 5) | (edcb >> 2);
break;
}
case 6: {
C = 5;
const uint fedcb = (bitval >> 1) & 0x1F;
B = (fedcb << 4) | (fedcb >> 4);
break;
}
}
} else { // QUINT
switch (bitlen) {
case 1:
C = 113;
break;
case 2: {
C = 54;
const uint b = (bitval >> 1) & 1;
B = (b << 8) | (b << 3) | (b << 2);
break;
}
case 3: {
C = 26;
const uint cb = (bitval >> 1) & 3;
B = (cb << 7) | (cb << 1) | (cb >> 1);
break;
}
case 4: {
C = 13;
const uint dcb = (bitval >> 1) & 7;
B = (dcb << 6) | (dcb >> 1);
break;
}
case 5: {
C = 6;
const uint edcb = (bitval >> 1) & 0xF;
B = (edcb << 5) | (edcb >> 3);
break;
}
}
}
uint T = (D * C) + B;
T ^= A;
T = (A & 0x80) | (T >> 2);
color_values_direct[++color_out_index] = T;
}
void ResultEmplaceBack(EncodingData val) {
if (write_color_values) {
if (color_out_index >= color_num_values) {
// avoid decoding more than needed by this phase
result_limit_reached = true;
return;
}
EmitColorValue(val);
return;
}
if (result_index >= result_vector_max_index) {
// Alert callers to avoid decoding more than needed by this phase
result_limit_reached = true;
@@ -196,33 +303,36 @@ uint Hash52(uint p) {
p ^= p >> 17;
return p;
}
struct PartitionTable {
uint s1, s2, s3, s4, s5, s6, s7, s8;
uint rnum;
bool small_block;
};
uint Select2DPartition(uint seed, uint x, uint y, uint partition_count) {
if ((block_dims.y * block_dims.x) < 32) {
x <<= 1;
y <<= 1;
}
PartitionTable GetPartitionTable(uint seed, uint partition_count) {
PartitionTable pt;
pt.small_block = (block_dims.y * block_dims.x) < 32;
seed += (partition_count - 1) * 1024;
uint rnum = Hash52(uint(seed));
pt.rnum = rnum;
const uint rnum = Hash52(uint(seed));
uint seed1 = uint(rnum & 0xF);
uint seed2 = uint((rnum >> 4) & 0xF);
uint seed3 = uint((rnum >> 8) & 0xF);
uint seed4 = uint((rnum >> 12) & 0xF);
uint seed5 = uint((rnum >> 16) & 0xF);
uint seed6 = uint((rnum >> 20) & 0xF);
uint seed7 = uint((rnum >> 24) & 0xF);
uint seed8 = uint((rnum >> 28) & 0xF);
seed1 = (seed1 * seed1);
seed2 = (seed2 * seed2);
seed3 = (seed3 * seed3);
seed4 = (seed4 * seed4);
seed5 = (seed5 * seed5);
seed6 = (seed6 * seed6);
seed7 = (seed7 * seed7);
seed8 = (seed8 * seed8);
uint seed1 = (rnum & 0xF);
seed1 *= seed1;
uint seed2 = (rnum >> 4) & 0xF;
seed2 *= seed2;
uint seed3 = (rnum >> 8) & 0xF;
seed3 *= seed3;
uint seed4 = (rnum >> 12) & 0xF;
seed4 *= seed4;
uint seed5 = (rnum >> 16) & 0xF;
seed5 *= seed5;
uint seed6 = (rnum >> 20) & 0xF;
seed6 *= seed6;
uint seed7 = (rnum >> 24) & 0xF;
seed7 *= seed7;
uint seed8 = (rnum >> 28) & 0xF;
seed8 *= seed8;
uint sh1, sh2;
if ((seed & 1) > 0) {
@@ -232,31 +342,37 @@ uint Select2DPartition(uint seed, uint x, uint y, uint partition_count) {
sh1 = (partition_count == 3) ? 6 : 5;
sh2 = (seed & 2) > 0 ? 4 : 5;
}
seed1 >>= sh1;
seed2 >>= sh2;
seed3 >>= sh1;
seed4 >>= sh2;
seed5 >>= sh1;
seed6 >>= sh2;
seed7 >>= sh1;
seed8 >>= sh2;
uint a = seed1 * x + seed2 * y + (rnum >> 14);
uint b = seed3 * x + seed4 * y + (rnum >> 10);
uint c = seed5 * x + seed6 * y + (rnum >> 6);
uint d = seed7 * x + seed8 * y + (rnum >> 2);
pt.s1 = seed1 >> sh1;
pt.s2 = seed2 >> sh2;
pt.s3 = seed3 >> sh1;
pt.s4 = seed4 >> sh2;
pt.s5 = seed5 >> sh1;
pt.s6 = seed6 >> sh2;
pt.s7 = seed7 >> sh1;
pt.s8 = seed8 >> sh2;
return pt;
}
uint SelectPartition(PartitionTable pt, uint x, uint y, uint partition_count) {
if (pt.small_block) {
x <<= 1;
y <<= 1;
}
uint a = pt.s1 * x + pt.s2 * y + (pt.rnum >> 14);
uint b = pt.s3 * x + pt.s4 * y + (pt.rnum >> 10);
uint c = pt.s5 * x + pt.s6 * y + (pt.rnum >> 6);
uint d = pt.s7 * x + pt.s8 * y + (pt.rnum >> 2);
a &= 0x3F;
b &= 0x3F;
c &= 0x3F;
d &= 0x3F;
if (partition_count < 4) {
d = 0;
}
if (partition_count < 3) {
c = 0;
}
if (partition_count < 4) d = 0;
if (partition_count < 3) c = 0;
if (a >= b && a >= c && a >= d) {
return 0;
@@ -457,7 +573,7 @@ void DecodeIntegerSequence(uint max_range, uint num_values) {
}
}
void DecodeColorValues(uvec4 modes, uint num_partitions, uint color_data_bits, out uint color_values[32]) {
void DecodeColorValues(uvec4 modes, uint num_partitions, uint color_data_bits) {
uint num_values = 0;
for (uint i = 0; i < num_partitions; i++) {
num_values += ((modes[i] >> 2) + 1) << 1;
@@ -471,104 +587,21 @@ void DecodeColorValues(uvec4 modes, uint num_partitions, uint color_data_bits, o
break;
}
}
DecodeIntegerSequence(range - 1, num_values);
uint out_index = 0;
for (int itr = 0; itr < result_index; ++itr) {
if (out_index >= num_values) {
break;
}
const EncodingData val = GetEncodingFromVector(itr);
const uint encoding = Encoding(val);
const uint bitlen = NumBits(val);
const uint bitval = BitValue(val);
uint A = 0, B = 0, C = 0, D = 0;
A = ReplicateBitTo9((bitval & 1));
switch (encoding) {
case JUST_BITS:
color_values[++out_index] = FastReplicateTo8(bitval, bitlen);
break;
case TRIT: {
D = QuintTritValue(val);
switch (bitlen) {
case 1:
C = 204;
break;
case 2: {
C = 93;
const uint b = (bitval >> 1) & 1;
B = (b << 8) | (b << 4) | (b << 2) | (b << 1);
break;
}
case 3: {
C = 44;
const uint cb = (bitval >> 1) & 3;
B = (cb << 7) | (cb << 2) | cb;
break;
}
case 4: {
C = 22;
const uint dcb = (bitval >> 1) & 7;
B = (dcb << 6) | dcb;
break;
}
case 5: {
C = 11;
const uint edcb = (bitval >> 1) & 0xF;
B = (edcb << 5) | (edcb >> 2);
break;
}
case 6: {
C = 5;
const uint fedcb = (bitval >> 1) & 0x1F;
B = (fedcb << 4) | (fedcb >> 4);
break;
}
}
break;
}
case QUINT: {
D = QuintTritValue(val);
switch (bitlen) {
case 1:
C = 113;
break;
case 2: {
C = 54;
const uint b = (bitval >> 1) & 1;
B = (b << 8) | (b << 3) | (b << 2);
break;
}
case 3: {
C = 26;
const uint cb = (bitval >> 1) & 3;
B = (cb << 7) | (cb << 1) | (cb >> 1);
break;
}
case 4: {
C = 13;
const uint dcb = (bitval >> 1) & 7;
B = (dcb << 6) | (dcb >> 1);
break;
}
case 5: {
C = 6;
const uint edcb = (bitval >> 1) & 0xF;
B = (edcb << 5) | (edcb >> 3);
break;
}
}
break;
}
}
if (encoding != JUST_BITS) {
uint T = (D * C) + B;
T ^= A;
T = (A & 0x80) | (T >> 2);
color_values[++out_index] = T;
}
// Decode directly into color_values_direct[]
write_color_values = true;
color_out_index = 0;
color_num_values = num_values;
for (uint i = 0; i < 32; ++i) {
color_values_direct[i] = 0;
}
DecodeIntegerSequence(range - 1, num_values);
write_color_values = false;
}
ivec2 BitTransferSigned(int a, int b) {
ivec2 transferred;
transferred.y = b >> 1;
@@ -730,7 +763,7 @@ uint UnquantizeTexelWeight(EncodingData val) {
uint encoding = Encoding(val), bitlen = NumBits(val), bitval = BitValue(val);
if (encoding == JUST_BITS) {
return (bitlen >= 1 && bitlen <= 5)
? uint(floor(0.5f + float(bitval) * 64.0f / float((1 << bitlen) - 1)))
? ((bitval * 64) + ((1 << bitlen) - 1) / 2) / ((1 << bitlen) - 1)
: FastReplicateTo6(bitval, bitlen);
} else if (encoding == TRIT || encoding == QUINT) {
uint B = 0, C = 0, D = 0;
@@ -1069,13 +1102,12 @@ void DecompressBlock(ivec3 coord) {
uvec4 endpoints0[4];
uvec4 endpoints1[4];
{
// This decode phase should at most push 32 elements into the vector
result_vector_max_index = 32;
uint color_values[32];
// Decode directly into color_values_direct[] (no intermediate result_vector storage)
result_limit_reached = false;
uint colvals_index = 0;
DecodeColorValues(color_endpoint_mode, num_partitions, color_data_bits, color_values);
DecodeColorValues(color_endpoint_mode, num_partitions, color_data_bits);
for (uint i = 0; i < num_partitions; i++) {
ComputeEndpoints(endpoints0[i], endpoints1[i], color_endpoint_mode[i], color_values,
ComputeEndpoints(endpoints0[i], endpoints1[i], color_endpoint_mode[i], color_values_direct,
colvals_index);
}
}
@@ -1106,11 +1138,15 @@ void DecompressBlock(ivec3 coord) {
DecodeIntegerSequence(max_weight, GetNumWeightValues(size_params, dual_plane));
UnquantizeTexelWeights(size_params, dual_plane);
PartitionTable pt;
if (num_partitions > 1) {
pt = GetPartitionTable(partition_index, num_partitions);
}
for (uint j = 0; j < block_dims.y; j++) {
for (uint i = 0; i < block_dims.x; i++) {
uint local_partition = 0;
if (num_partitions > 1) {
local_partition = Select2DPartition(partition_index, i, j, num_partitions);
local_partition = SelectPartition(pt, i, j, num_partitions);
}
const uvec4 C0 = ReplicateByteTo16(endpoints0[local_partition]);
const uvec4 C1 = ReplicateByteTo16(endpoints1[local_partition]);