Add aliasing meta data to loads and stores
This commit is contained in:
parent
0b0ee2e3c2
commit
fc07a25306
22 changed files with 260 additions and 189 deletions
|
|
@ -13,58 +13,58 @@
|
|||
|
||||
void DrawWallCodegen::Generate(DrawWallVariant variant, bool fourColumns, SSAValue args, SSAValue thread_data)
|
||||
{
|
||||
dest = args[0][0].load();
|
||||
source[0] = args[0][1].load();
|
||||
source[1] = args[0][2].load();
|
||||
source[2] = args[0][3].load();
|
||||
source[3] = args[0][4].load();
|
||||
source2[0] = args[0][5].load();
|
||||
source2[1] = args[0][6].load();
|
||||
source2[2] = args[0][7].load();
|
||||
source2[3] = args[0][8].load();
|
||||
pitch = args[0][9].load();
|
||||
count = args[0][10].load();
|
||||
dest_y = args[0][11].load();
|
||||
texturefrac[0] = args[0][12].load();
|
||||
texturefrac[1] = args[0][13].load();
|
||||
texturefrac[2] = args[0][14].load();
|
||||
texturefrac[3] = args[0][15].load();
|
||||
texturefracx[0] = args[0][16].load();
|
||||
texturefracx[1] = args[0][17].load();
|
||||
texturefracx[2] = args[0][18].load();
|
||||
texturefracx[3] = args[0][19].load();
|
||||
iscale[0] = args[0][20].load();
|
||||
iscale[1] = args[0][21].load();
|
||||
iscale[2] = args[0][22].load();
|
||||
iscale[3] = args[0][23].load();
|
||||
textureheight[0] = args[0][24].load();
|
||||
textureheight[1] = args[0][25].load();
|
||||
textureheight[2] = args[0][26].load();
|
||||
textureheight[3] = args[0][27].load();
|
||||
light[0] = args[0][28].load();
|
||||
light[1] = args[0][29].load();
|
||||
light[2] = args[0][30].load();
|
||||
light[3] = args[0][31].load();
|
||||
srcalpha = args[0][32].load();
|
||||
destalpha = args[0][33].load();
|
||||
SSAShort light_alpha = args[0][34].load();
|
||||
SSAShort light_red = args[0][35].load();
|
||||
SSAShort light_green = args[0][36].load();
|
||||
SSAShort light_blue = args[0][37].load();
|
||||
SSAShort fade_alpha = args[0][38].load();
|
||||
SSAShort fade_red = args[0][39].load();
|
||||
SSAShort fade_green = args[0][40].load();
|
||||
SSAShort fade_blue = args[0][41].load();
|
||||
SSAShort desaturate = args[0][42].load();
|
||||
SSAInt flags = args[0][43].load();
|
||||
dest = args[0][0].load(true);
|
||||
source[0] = args[0][1].load(true);
|
||||
source[1] = args[0][2].load(true);
|
||||
source[2] = args[0][3].load(true);
|
||||
source[3] = args[0][4].load(true);
|
||||
source2[0] = args[0][5].load(true);
|
||||
source2[1] = args[0][6].load(true);
|
||||
source2[2] = args[0][7].load(true);
|
||||
source2[3] = args[0][8].load(true);
|
||||
pitch = args[0][9].load(true);
|
||||
count = args[0][10].load(true);
|
||||
dest_y = args[0][11].load(true);
|
||||
texturefrac[0] = args[0][12].load(true);
|
||||
texturefrac[1] = args[0][13].load(true);
|
||||
texturefrac[2] = args[0][14].load(true);
|
||||
texturefrac[3] = args[0][15].load(true);
|
||||
texturefracx[0] = args[0][16].load(true);
|
||||
texturefracx[1] = args[0][17].load(true);
|
||||
texturefracx[2] = args[0][18].load(true);
|
||||
texturefracx[3] = args[0][19].load(true);
|
||||
iscale[0] = args[0][20].load(true);
|
||||
iscale[1] = args[0][21].load(true);
|
||||
iscale[2] = args[0][22].load(true);
|
||||
iscale[3] = args[0][23].load(true);
|
||||
textureheight[0] = args[0][24].load(true);
|
||||
textureheight[1] = args[0][25].load(true);
|
||||
textureheight[2] = args[0][26].load(true);
|
||||
textureheight[3] = args[0][27].load(true);
|
||||
light[0] = args[0][28].load(true);
|
||||
light[1] = args[0][29].load(true);
|
||||
light[2] = args[0][30].load(true);
|
||||
light[3] = args[0][31].load(true);
|
||||
srcalpha = args[0][32].load(true);
|
||||
destalpha = args[0][33].load(true);
|
||||
SSAShort light_alpha = args[0][34].load(true);
|
||||
SSAShort light_red = args[0][35].load(true);
|
||||
SSAShort light_green = args[0][36].load(true);
|
||||
SSAShort light_blue = args[0][37].load(true);
|
||||
SSAShort fade_alpha = args[0][38].load(true);
|
||||
SSAShort fade_red = args[0][39].load(true);
|
||||
SSAShort fade_green = args[0][40].load(true);
|
||||
SSAShort fade_blue = args[0][41].load(true);
|
||||
SSAShort desaturate = args[0][42].load(true);
|
||||
SSAInt flags = args[0][43].load(true);
|
||||
shade_constants.light = SSAVec4i(light_blue.zext_int(), light_green.zext_int(), light_red.zext_int(), light_alpha.zext_int());
|
||||
shade_constants.fade = SSAVec4i(fade_blue.zext_int(), fade_green.zext_int(), fade_red.zext_int(), fade_alpha.zext_int());
|
||||
shade_constants.desaturate = desaturate.zext_int();
|
||||
|
||||
thread.core = thread_data[0][0].load();
|
||||
thread.num_cores = thread_data[0][1].load();
|
||||
thread.pass_start_y = thread_data[0][2].load();
|
||||
thread.pass_end_y = thread_data[0][3].load();
|
||||
thread.core = thread_data[0][0].load(true);
|
||||
thread.num_cores = thread_data[0][1].load(true);
|
||||
thread.pass_start_y = thread_data[0][2].load(true);
|
||||
thread.pass_end_y = thread_data[0][3].load(true);
|
||||
|
||||
is_simple_shade = (flags & DrawWallArgs::simple_shade) == SSAInt(DrawWallArgs::simple_shade);
|
||||
is_nearest_filter = (flags & DrawWallArgs::nearest_filter) == SSAInt(DrawWallArgs::nearest_filter);
|
||||
|
|
@ -118,7 +118,7 @@ void DrawWallCodegen::Loop(DrawWallVariant variant, bool fourColumns, bool isSim
|
|||
|
||||
if (fourColumns)
|
||||
{
|
||||
SSAVec16ub bg = dest[offset].load_unaligned_vec16ub();
|
||||
SSAVec16ub bg = dest[offset].load_unaligned_vec16ub(false);
|
||||
SSAVec8s bg0 = SSAVec8s::extendlo(bg);
|
||||
SSAVec8s bg1 = SSAVec8s::extendhi(bg);
|
||||
SSAVec4i bgcolors[4] =
|
||||
|
|
@ -138,12 +138,12 @@ void DrawWallCodegen::Loop(DrawWallVariant variant, bool fourColumns, bool isSim
|
|||
}
|
||||
else
|
||||
{
|
||||
SSAVec4i bgcolor = dest[offset].load_vec4ub();
|
||||
SSAVec4i bgcolor = dest[offset].load_vec4ub(false);
|
||||
SSAVec4i color = Blend(Shade(Sample(frac[0], 0, isNearestFilter), 0, isSimpleShade), bgcolor, variant);
|
||||
dest[offset].store_vec4ub(color);
|
||||
}
|
||||
|
||||
stack_index.store(index + 1);
|
||||
stack_index.store(index.add(SSAInt(1), true, true));
|
||||
for (int i = 0; i < numColumns; i++)
|
||||
stack_frac[i].store(frac[i] + fracstep[i]);
|
||||
loop.end_block();
|
||||
|
|
@ -155,7 +155,7 @@ SSAVec4i DrawWallCodegen::Sample(SSAInt frac, int index, bool isNearestFilter)
|
|||
if (isNearestFilter)
|
||||
{
|
||||
SSAInt sample_index = ((frac >> FRACBITS) * textureheight[index]) >> FRACBITS;
|
||||
return source[index][sample_index * 4].load_vec4ub();
|
||||
return source[index][sample_index * 4].load_vec4ub(false);
|
||||
}
|
||||
else
|
||||
{
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue