sgilissen/llamaplayer
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

GPU/HW: Simplify 24-bit scanout and interlacing shader

Browse Source 
Has the added bonus of being faster (no downscale copy) and fixes the
edge case where the image is cropped in VRAM.
This commit is contained in:
Connor McLaughlin
2020-04-02 00:17:25 +10:00
parent 1d948a53e0
commit 2a6e04988f
5 changed files with 75 additions and 185 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
src/core/gpu_hw_shadergen.cpp
View File

@ -750,7 +750,7 @@ std::string GPU_HW_ShaderGen::GenerateDisplayFragmentShader(bool depth_24bit, bo
DefineMacro(ss, "INTERLACED", interlaced);
WriteCommonFunctions(ss);
DeclareUniformBuffer(ss, {"int3 u_base_coords"});
DeclareUniformBuffer(ss, {"int u_field_offset", "int u_vram_start_x"});
DeclareTexture(ss, "samp0", 0);
DeclareFragmentEntryPoint(ss, 0, 1, {}, true, false);
@ -759,55 +759,28 @@ std::string GPU_HW_ShaderGen::GenerateDisplayFragmentShader(bool depth_24bit, bo
int2 icoords = int2(v_pos.xy);
#if INTERLACED
if (((icoords.y - u_base_coords.z) & 1) != 0)
if (((icoords.y / RESOLUTION_SCALE) & 1) != u_field_offset)
discard;
#endif
#if DEPTH_24BIT
// compute offset in dwords from the start of the 24-bit values
int2 base = int2(u_base_coords.x, u_base_coords.y + icoords.y);
int xoff = int(icoords.x);
int dword_index = (xoff / 2) + (xoff / 4);
// relative to start of scanout
int relative_x = (icoords.x - u_vram_start_x) / RESOLUTION_SCALE;
icoords.x = u_vram_start_x + ((relative_x * 3) / 2) * RESOLUTION_SCALE;
// sample two adjacent dwords, or four 16-bit values as the 24-bit value will lie somewhere between these
uint s0 = RGBA8ToRGBA5551(LOAD_TEXTURE(samp0, int2(base.x + dword_index * 2 + 0, base.y), 0));
uint s1 = RGBA8ToRGBA5551(LOAD_TEXTURE(samp0, int2(base.x + dword_index * 2 + 1, base.y), 0));
uint s2 = RGBA8ToRGBA5551(LOAD_TEXTURE(samp0, int2(base.x + (dword_index + 1) * 2 + 0, base.y), 0));
uint s3 = RGBA8ToRGBA5551(LOAD_TEXTURE(samp0, int2(base.x + (dword_index + 1) * 2 + 1, base.y), 0));
// select the bit for this pixel depending on its offset in the 4-pixel block
uint r, g, b;
int block_offset = xoff & 3;
if (block_offset == 0)
{
r = s0 & 0xFFu;
g = s0 >> 8;
b = s1 & 0xFFu;
}
else if (block_offset == 1)
{
r = s1 >> 8;
g = s2 & 0xFFu;
b = s2 >> 8;
}
else if (block_offset == 2)
{
r = s1 & 0xFFu;
g = s1 >> 8;
b = s2 & 0xFFu;
}
else
{
r = s2 >> 8;
g = s3 & 0xFFu;
b = s3 >> 8;
}
// and normalize
o_col0 = float4(float(r) / 255.0, float(g) / 255.0, float(b) / 255.0, 1.0);
// load adjacent 16-bit texels
uint s0 = RGBA8ToRGBA5551(LOAD_TEXTURE(samp0, icoords, 0));
uint s1 = RGBA8ToRGBA5551(LOAD_TEXTURE(samp0, icoords + int2(RESOLUTION_SCALE, 0), 0));
// select which part of the combined 16-bit texels we are currently shading
uint s1s0 = ((s1 << 16) | s0) >> ((relative_x & 1) * 8);
// extract components and normalize
o_col0 = float4(float(s1s0 & 0xFFu) / 255.0, float((s1s0 >> 8u) & 0xFFu) / 255.0,
float((s1s0 >> 16u) & 0xFFu) / 255.0, 1.0);
#else
// load and return
o_col0 = LOAD_TEXTURE(samp0, u_base_coords.xy + icoords, 0);
o_col0 = LOAD_TEXTURE(samp0, icoords, 0);
#endif
}
)";