Skip to content

Commit

Permalink
GS/HW: Add ROV based rendering for DX11/DX12/Vulkan
Browse files Browse the repository at this point in the history
  • Loading branch information
@stenzek
stenzek committed May 12, 2024
1 parent 20dbcfd commit 3126952
Show file tree
Hide file tree
Showing 27 changed files with 1,521 additions and 285 deletions.
194 changes: 170 additions & 24 deletions bin/resources/shaders/dx11/tfx.fx
View file
Original file line number Diff line number Diff line change
Expand Up @@ -77,11 +77,16 @@
#define PS_NO_COLOR 0
#define PS_NO_COLOR1 0
#define PS_DATE 0
#define PS_ROV 0
#define PS_ZTST 0
#define PS_ZWE 0
#define PS_AFAIL 0
#endif

#define SW_BLEND (PS_BLEND_A || PS_BLEND_B || PS_BLEND_D)
#define SW_BLEND_NEEDS_RT (SW_BLEND && (PS_BLEND_A == 1 || PS_BLEND_B == 1 || PS_BLEND_C == 1 || PS_BLEND_D == 1))
#define SW_AD_TO_HW (PS_BLEND_C == 1 && PS_A_MASKED)
#define ROV_DEPTH (PS_ZTST != 0 || PS_ZWE != 0)

struct VS_INPUT
{
Expand Down Expand Up @@ -136,17 +141,26 @@ struct PS_OUTPUT
#endif
#endif
#endif
#if PS_ZCLAMP
#if PS_ZCLAMP && !ROV_DEPTH
float depth : SV_Depth;
#endif
};

Texture2D<float4> Texture : register(t0);
Texture2D<float4> Palette : register(t1);
#if !PS_ROV
Texture2D<float4> RtTexture : register(t2);
#endif
Texture2D<float> PrimMinTexture : register(t3);
SamplerState TextureSampler : register(s0);

#if PS_ROV
RasterizerOrderedTexture2D<unorm float4> rovRT : register(u0);
#if ROV_DEPTH
RasterizerOrderedTexture2D<float> rovDS : register(u1);
#endif
#endif

#ifdef DX12
cbuffer cb1 : register(b1)
#else
Expand All @@ -171,10 +185,73 @@ cbuffer cb1
float RcpScaleFactor;
};


#if PS_ROV

static float4 rovRTValue;
static uint4 rovFbMask;
static bool rovPixelTestResult;
static bool rovDepthWrite;

#define DISCARD return output

float4 sample_from_rt(int2 uv)
{
return rovRTValue;
}

void rov_read_rt(int2 uv)
{
rovRTValue = rovRT[uv];
rovPixelTestResult = true;
rovFbMask = FbMask;
rovDepthWrite = PS_ZWE != 0;
}

void rov_write_rt(int2 uv, float4 color)
{
if (rovPixelTestResult)
rovRT[uv] = color;
}

#if ROV_DEPTH

bool rov_depth_test(int2 uv, float z)
{
bool zpass = true;

#if PS_ZTST > 1
float ds_z = rovDS[uv];
#if PS_ZTST == 2
zpass = (z >= ds_z);
#elif PS_ZTST == 3
zpass = (z > ds_z);
#endif
#endif

if (zpass && rovDepthWrite)
rovDS[uv] = z;

return zpass;
}

#endif

#else

#define DISCARD discard

float4 sample_from_rt(int2 uv)
{
return RtTexture.Load(int3(uv, 0));
}

#endif

float4 sample_c(float2 uv, float uv_w)
{
#if PS_TEX_IS_FB == 1
return RtTexture.Load(int3(int2(uv * WH.zw), 0));
return sample_from_rt(int2(uv * WH.zw));
#elif PS_REGION_RECT == 1
return Texture.Load(int3(int2(uv), 0));
#else
Expand Down Expand Up @@ -378,7 +455,7 @@ float4x4 sample_4p(uint4 u)
int fetch_raw_depth(int2 xy)
{
#if PS_TEX_IS_FB == 1
float4 col = RtTexture.Load(int3(xy, 0));
float4 col = sample_from_rt(xy);
#else
float4 col = Texture.Load(int3(xy, 0));
#endif
Expand All @@ -388,7 +465,7 @@ int fetch_raw_depth(int2 xy)
float4 fetch_raw_color(int2 xy)
{
#if PS_TEX_IS_FB == 1
return RtTexture.Load(int3(xy, 0));
return sample_from_rt(xy);
#else
return Texture.Load(int3(xy, 0));
#endif
Expand Down Expand Up @@ -792,12 +869,16 @@ float4 ps_color(PS_INPUT input)
return C;
}

void ps_fbmask(inout float4 C, float2 pos_xy)
void ps_fbmask(inout float4 C, int2 pos_xy)
{
if (PS_FBMASK)
{
float4 RT = trunc(RtTexture.Load(int3(pos_xy, 0)) * 255.0f + 0.1f);
C = (float4)(((uint4)C & ~FbMask) | ((uint4)RT & FbMask));
float4 RT = trunc(sample_from_rt(pos_xy) * 255.0f + 0.1f);
#if PS_ROV
C = (float4)(((uint4)C & ~rovFbMask) | ((uint4)RT & rovFbMask));
#else
C = (float4)(((uint4)C & ~FbMask) | ((uint4)RT & FbMask));
#endif
}
}

Expand Down Expand Up @@ -850,7 +931,7 @@ void ps_color_clamp_wrap(inout float3 C)
}
}

void ps_blend(inout float4 Color, inout float4 As_rgba, float2 pos_xy)
void ps_blend(inout float4 Color, inout float4 As_rgba, int2 pos_xy)
{
float As = As_rgba.a;

Expand All @@ -864,7 +945,7 @@ void ps_blend(inout float4 Color, inout float4 As_rgba, float2 pos_xy)
return;
}

float4 RT = SW_BLEND_NEEDS_RT ? RtTexture.Load(int3(pos_xy, 0)) : (float4)0.0f;
float4 RT = SW_BLEND_NEEDS_RT ? sample_from_rt(pos_xy) : (float4)0.0f;

if (PS_SHUFFLE && SW_BLEND_NEEDS_RT)
{
Expand Down Expand Up @@ -980,23 +1061,57 @@ void ps_blend(inout float4 Color, inout float4 As_rgba, float2 pos_xy)
}
}

#if PS_ROV && !ROV_DEPTH
[earlydepthstencil]
#endif

PS_OUTPUT ps_main(PS_INPUT input)
{
int2 input_xy = int2(input.p.xy);
float input_z = PS_ZCLAMP ? min(input.p.z, MaxDepthPS) : input.p.z;

#if PS_ROV
rov_read_rt(input_xy);
#endif

PS_OUTPUT output;
#if PS_ZCLAMP && !ROV_DEPTH
output.depth = input_z;
#endif

float4 C = ps_color(input);
bool atst_pass = atst(C);

#if PS_AFAIL == 0 // KEEP or ATST off
if (!atst_pass)
discard;
{
#if PS_ROV
if (PS_AFAIL == 0)
{
DISCARD;
}
else if (PS_AFAIL == 1) // FB_ONLY
{
rovDepthWrite = false;
}
else if (PS_AFAIL == 2) // ZB_ONLY
{
rovFbMask = 0xFF;
}
else if (PS_AFAIL == 3) // RGB_ONLY
{
rovFbMask.a = 0xFF;
rovDepthWrite = false;
}
#elif PS_AFAIL == 0
DISCARD;
#endif

PS_OUTPUT output;
}

if (PS_SCANMSK & 2)
{
// fail depth test on prohibited lines
if ((int(input.p.y) & 1) == (PS_SCANMSK & 1))
discard;
if ((input_xy.y & 1) == (PS_SCANMSK & 1))
DISCARD;
}

// Must be done before alpha correction
Expand All @@ -1010,7 +1125,7 @@ PS_OUTPUT ps_main(PS_INPUT input)
float4 alpha_blend = (float4)0.0f;
if (SW_AD_TO_HW)
{
float4 RT = PS_RTA_CORRECTION ? trunc(RtTexture.Load(int3(input.p.xy, 0)) * 128.0f + 0.1f) : trunc(RtTexture.Load(int3(input.p.xy, 0)) * 255.0f + 0.1f);
float4 RT = PS_RTA_CORRECTION ? trunc(sample_from_rt(input_xy) * 128.0f + 0.1f) : trunc(sample_from_rt(input_xy) * 255.0f + 0.1f);
alpha_blend = (float4)(RT.a / 128.0f);
}
else
Expand All @@ -1030,12 +1145,40 @@ PS_OUTPUT ps_main(PS_INPUT input)
if (C.a < A_one) C.a += A_one;
}

#if PS_DATE < 10 && (((PS_DATE & 3) == 1 || (PS_DATE & 3) == 2))

#if PS_WRITE_RG == 1
// Pseudo 16 bits access.
float rt_a = sample_from_rt(input_xy).g;
#else
float rt_a = sample_from_rt(input_xy).a;
#endif

#if (PS_DATE & 3) == 1
// DATM == 0: Pixel with alpha equal to 1 will failed
bool bad = (127.5f / 255.0f) < rt_a;
#elif (PS_DATE & 3) == 2
// DATM == 1: Pixel with alpha equal to 0 will failed
bool bad = rt_a < (127.5f / 255.0f);
#endif

if (bad) {
#if PS_ROV || PS_DATE >= 5
DISCARD;
#else
return;
#endif
}

#endif


#if PS_DATE == 3
// Note gl_PrimitiveID == stencil_ceil will be the primitive that will update
// the bad alpha value so we must keep it.
int stencil_ceil = int(PrimMinTexture.Load(int3(input.p.xy, 0)));
int stencil_ceil = int(PrimMinTexture.Load(int3(input_xy, 0)));
if (int(input.primid) > stencil_ceil)
discard;
DISCARD;
#endif

// Get first primitive that will write a failling alpha value
Expand All @@ -1053,7 +1196,7 @@ PS_OUTPUT ps_main(PS_INPUT input)
#else
// Not primid DATE setup

ps_blend(C, alpha_blend, input.p.xy);
ps_blend(C, alpha_blend, input_xy);

if (PS_SHUFFLE)
{
Expand Down Expand Up @@ -1142,7 +1285,7 @@ PS_OUTPUT ps_main(PS_INPUT input)
// Color clamp/wrap needs to be done after sw blending and dithering
ps_color_clamp_wrap(C.rgb);

ps_fbmask(C, input.p.xy);
ps_fbmask(C, input_xy);

#if PS_AFAIL == 3 // RGB_ONLY
// Use alpha blend factor to determine whether to update A.
Expand All @@ -1157,12 +1300,15 @@ PS_OUTPUT ps_main(PS_INPUT input)
#endif
#endif // !PS_NO_COLOR

#endif // PS_DATE != 1/2

#if PS_ZCLAMP
output.depth = min(input.p.z, MaxDepthPS);
#if PS_ROV && ROV_DEPTH
if (rov_depth_test(input_xy, input_z))
rov_write_rt(input_xy, C / float4(255.0f, 255.0f, 255.0f, PS_RTA_CORRECTION ? 128.0f : 255.0f));
#elif PS_ROV
rov_write_rt(input_xy, C / float4(255.0f, 255.0f, 255.0f, PS_RTA_CORRECTION ? 128.0f : 255.0f));
#endif

#endif // PS_DATE != 1/2

return output;
}

Expand Down
Loading

0 comments on commit 3126952

Please sign in to comment.