github/duckstation
1
0
Fork 0
mirror of https://github.com/stenzek/duckstation.git synced 2025-06-07 12:05:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

GPU/HW: Vectorize flipped sprite handling

Browse Source
This commit is contained in:
Stenzek 2024-12-29 16:54:13 +10:00
parent 1a211e0a21
commit b7832e609f
No known key found for this signature in database
2 changed files with 70 additions and 87 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

153
src/core/gpu_hw.cpp
View File

@ -2101,30 +2101,33 @@ ALWAYS_INLINE_RELEASE void GPU_HW::DrawBatchVertices(BatchRenderMode render_mode
} }
} }
ALWAYS_INLINE_RELEASE void GPU_HW::ComputeUVPartialDerivatives(const BatchVertex* vertices, float* dudx, float* dudy,
float* dvdx, float* dvdy, float* xy_area, s32* uv_area)
{
const float v01x = vertices[1].x - vertices[0].x;
const float v01y = vertices[1].y - vertices[0].y;
const float v12x = vertices[2].x - vertices[1].x;
const float v12y = vertices[2].y - vertices[1].y;
const float v23x = vertices[0].x - vertices[2].x;
const float v23y = vertices[0].y - vertices[2].y;
const float v0u = static_cast<float>(vertices[0].u);
const float v0v = static_cast<float>(vertices[0].v);
const float v1u = static_cast<float>(vertices[1].u);
const float v1v = static_cast<float>(vertices[1].v);
const float v2u = static_cast<float>(vertices[2].u);
const float v2v = static_cast<float>(vertices[2].v);
*dudx = -v01y * v2u - v12y * v0u - v23y * v1u;
*dvdx = -v01y * v2v - v12y * v0v - v23y * v1v;
*dudy = v01x * v2u + v12x * v0u + v23x * v1u;
*dvdy = v01x * v2v + v12x * v0v + v23x * v1v;
*xy_area = v12x * v23y - v12y * v23x;
*uv_area = (vertices[1].u - vertices[0].u) * (vertices[2].v - vertices[0].v) -
(vertices[2].u - vertices[0].u) * (vertices[1].v - vertices[0].v);
}
ALWAYS_INLINE_RELEASE void GPU_HW::HandleFlippedQuadTextureCoordinates(const GPUBackendDrawCommand* cmd, ALWAYS_INLINE_RELEASE void GPU_HW::HandleFlippedQuadTextureCoordinates(const GPUBackendDrawCommand* cmd,
BatchVertex* vertices) BatchVertex* vertices)
{ {
// Taken from beetle-psx gpu_polygon.cpp
// For X/Y flipped 2D sprites, PSX games rely on a very specific rasterization behavior. If U or V is decreasing in X
// or Y, and we use the provided U/V as is, we will sample the wrong texel as interpolation covers an entire pixel,
// while PSX samples its interpolation essentially in the top-left corner and splats that interpolant across the
// entire pixel. While we could emulate this reasonably well in native resolution by shifting our vertex coords by
// 0.5, this breaks in upscaling scenarios, because we have several samples per native sample and we need NN rules to
// hit the same UV every time. One approach here is to use interpolate at offset or similar tricks to generalize the
// PSX interpolation patterns, but the problem is that vertices sharing an edge will no longer see the same UV (due to
// different plane derivatives), we end up sampling outside the intended boundary and artifacts are inevitable, so the
// only case where we can apply this fixup is for "sprites" or similar which should not share edges, which leads to
// this unfortunate code below.
// It might be faster to do more direct checking here, but the code below handles primitives in any order and
// orientation, and is far more SIMD-friendly if needed.
const float abx = vertices[1].x - vertices[0].x;
const float aby = vertices[1].y - vertices[0].y;
const float bcx = vertices[2].x - vertices[1].x;
const float bcy = vertices[2].y - vertices[1].y;
const float cax = vertices[0].x - vertices[2].x;
const float cay = vertices[0].y - vertices[2].y;
// Hack for Wild Arms 2: The player sprite is drawn one line at a time with a quad, but the bottom V coordinates // Hack for Wild Arms 2: The player sprite is drawn one line at a time with a quad, but the bottom V coordinates
// are set to a large distance from the top V coordinate. When upscaling, this means that the coordinate is // are set to a large distance from the top V coordinate. When upscaling, this means that the coordinate is
// interpolated between these two values, result in out-of-bounds sampling. At native, it's fine, because at the // interpolated between these two values, result in out-of-bounds sampling. At native, it's fine, because at the
@ -2143,63 +2146,47 @@ ALWAYS_INLINE_RELEASE void GPU_HW::HandleFlippedQuadTextureCoordinates(const GPU
vertices[3].v = vertices[0].v; vertices[3].v = vertices[0].v;
} }
// Compute static derivatives, just assume W is uniform across the primitive and that the plane equation remains the // Handle interpolation differences between PC GPUs and the PSX GPU. The first pixel on each span/scanline is given
// same across the quad. (which it is, there is no Z.. yet). // the initial U/V coordinate without any further interpolation on the PSX GPU, in contrast to PC GPUs. This results
const float dudx = -aby * static_cast<float>(vertices[2].u) - bcy * static_cast<float>(vertices[0].u) - // in oversampling on the right edge, so compensate by offsetting the left (right in texture space) UV.
cay * static_cast<float>(vertices[1].u); alignas(VECTOR_ALIGNMENT) float pd[4];
const float dvdx = -aby * static_cast<float>(vertices[2].v) - bcy * static_cast<float>(vertices[0].v) - float xy_area;
cay * static_cast<float>(vertices[1].v); s32 uv_area;
const float dudy = +abx * static_cast<float>(vertices[2].u) + bcx * static_cast<float>(vertices[0].u) + ComputeUVPartialDerivatives(vertices, &pd[0], &pd[1], &pd[2], &pd[3], &xy_area, &uv_area);
cax * static_cast<float>(vertices[1].u); if (xy_area == 0.0f || uv_area == 0)
const float dvdy = +abx * static_cast<float>(vertices[2].v) + bcx * static_cast<float>(vertices[0].v) +
cax * static_cast<float>(vertices[1].v);
const float area = bcx * cay - bcy * cax;
// Detect and reject any triangles with 0 size texture area
const s32 texArea = (vertices[1].u - vertices[0].u) * (vertices[2].v - vertices[0].v) -
(vertices[2].u - vertices[0].u) * (vertices[1].v - vertices[0].v);
// Shouldn't matter as degenerate primitives will be culled anyways.
if (area == 0.0f || texArea == 0)
return; return;
// Use floats here as it'll be faster than integer divides. const GSVector4 pd_area = GSVector4::load<true>(pd) / GSVector4(xy_area);
const float rcp_area = 1.0f / area; const GSVector4 neg_pd = (pd_area < GSVector4::zero());
const float dudx_area = dudx * rcp_area; const GSVector4 zero_pd = (pd_area == GSVector4::zero());
const float dudy_area = dudy * rcp_area; const int mask = (neg_pd.mask() | (zero_pd.mask() << 4));
const float dvdx_area = dvdx * rcp_area;
const float dvdy_area = dvdy * rcp_area;
const bool neg_dudx = dudx_area < 0.0f;
const bool neg_dudy = dudy_area < 0.0f;
const bool neg_dvdx = dvdx_area < 0.0f;
const bool neg_dvdy = dvdy_area < 0.0f;
const bool zero_dudx = dudx_area == 0.0f;
const bool zero_dudy = dudy_area == 0.0f;
const bool zero_dvdx = dvdx_area == 0.0f;
const bool zero_dvdy = dvdy_area == 0.0f;
// If we have negative dU or dV in any direction, increment the U or V to work properly with nearest-neighbor in // Addressing the 8-bit status code above.
// this impl. If we don't have 1:1 pixel correspondence, this creates a slight "shift" in the sprite, but we static constexpr int NEG_DUDX = 0x1;
// guarantee that we don't sample garbage at least. Overall, this is kinda hacky because there can be legitimate, static constexpr int NEG_DUDY = 0x2;
// rare cases where 3D meshes hit this scenario, and a single texel offset can pop in, but this is way better than static constexpr int NEG_DVDX = 0x4;
// having borked 2D overall. static constexpr int NEG_DVDY = 0x8;
// static constexpr int ZERO_DUDX = 0x10;
// TODO: If perf becomes an issue, we can probably SIMD the 8 comparisons above, static constexpr int ZERO_DUDY = 0x20;
// create an 8-bit code, and use a LUT to get the offsets. static constexpr int ZERO_DVDX = 0x40;
// Case 1: U is decreasing in X, but no change in Y. static constexpr int ZERO_DVDY = 0x80;
// Case 2: U is decreasing in Y, but no change in X.
// Case 3: V is decreasing in X, but no change in Y. // Flipped horizontal sprites: negative dudx+zero dudy or negative dudy+zero dudx.
// Case 4: V is decreasing in Y, but no change in X. if ((mask & (NEG_DUDX | ZERO_DUDY)) == (NEG_DUDX | ZERO_DUDY) ||
if ((neg_dudx && zero_dudy) || (neg_dudy && zero_dudx)) (mask & (NEG_DUDY | ZERO_DUDX)) == (NEG_DUDY | ZERO_DUDX))
{ {
GL_INS_FMT("Horizontal flipped sprite detected at {},{}", vertices[0].x, vertices[0].y);
vertices[0].u++; vertices[0].u++;
vertices[1].u++; vertices[1].u++;
vertices[2].u++; vertices[2].u++;
vertices[3].u++; vertices[3].u++;
} }
if ((neg_dvdx && zero_dvdy) || (neg_dvdy && zero_dvdx)) // Flipped vertical sprites: negative dvdx+zero dvdy or negative dvdy+zero dvdx.
if ((mask & (NEG_DVDX | ZERO_DVDY)) == (NEG_DVDX | ZERO_DVDY) ||
(mask & (NEG_DVDY | ZERO_DVDX)) == (NEG_DVDY | ZERO_DVDX))
{ {
GL_INS_FMT("Vertical flipped sprite detected at {},{}", vertices[0].x, vertices[0].y);
vertices[0].v++; vertices[0].v++;
vertices[1].v++; vertices[1].v++;
vertices[2].v++; vertices[2].v++;
@ -2208,32 +2195,24 @@ ALWAYS_INLINE_RELEASE void GPU_HW::HandleFlippedQuadTextureCoordinates(const GPU
// 2D polygons should have zero change in V on the X axis, and vice versa. // 2D polygons should have zero change in V on the X axis, and vice versa.
if (m_allow_sprite_mode) if (m_allow_sprite_mode)
SetBatchSpriteMode(cmd, zero_dudy && zero_dvdx); {
const bool is_sprite = (mask & (ZERO_DVDX | ZERO_DUDY)) == (ZERO_DVDX | ZERO_DUDY);
SetBatchSpriteMode(cmd, is_sprite);
}
} }
bool GPU_HW::IsPossibleSpritePolygon(const BatchVertex* vertices) const bool GPU_HW::IsPossibleSpritePolygon(const BatchVertex* vertices) const
{ {
const float abx = vertices[1].x - vertices[0].x; float dudx, dudy, dvdx, dvdy, xy_area;
const float aby = vertices[1].y - vertices[0].y; s32 uv_area;
const float bcx = vertices[2].x - vertices[1].x; ComputeUVPartialDerivatives(vertices, &dudx, &dudy, &dvdx, &dvdy, &xy_area, &uv_area);
const float bcy = vertices[2].y - vertices[1].y; if (xy_area == 0.0f || uv_area == 0)
const float cax = vertices[0].x - vertices[2].x;
const float cay = vertices[0].y - vertices[2].y;
const float dvdx = -aby * static_cast<float>(vertices[2].v) - bcy * static_cast<float>(vertices[0].v) -
cay * static_cast<float>(vertices[1].v);
const float dudy = +abx * static_cast<float>(vertices[2].u) + bcx * static_cast<float>(vertices[0].u) +
cax * static_cast<float>(vertices[1].u);
const float area = bcx * cay - bcy * cax;
const s32 texArea = (vertices[1].u - vertices[0].u) * (vertices[2].v - vertices[0].v) -
(vertices[2].u - vertices[0].u) * (vertices[1].v - vertices[0].v);
// Doesn't matter.
if (area == 0.0f || texArea == 0)
return m_batch.sprite_mode; return m_batch.sprite_mode;
const float rcp_area = 1.0f / area; // Could vectorize this, but it's not really worth it as we're only checking two partial derivatives.
const bool zero_dudy = ((dudy * rcp_area) == 0.0f); const float rcp_xy_area = 1.0f / xy_area;
const bool zero_dvdx = ((dvdx * rcp_area) == 0.0f); const bool zero_dudy = ((dudy * rcp_xy_area) == 0.0f);
const bool zero_dvdx = ((dvdx * rcp_xy_area) == 0.0f);
return (zero_dudy && zero_dvdx); return (zero_dudy && zero_dvdx);
} }

4
src/core/gpu_hw.h
View File

@ -250,6 +250,10 @@ private:
/// Expands a line into two triangles. /// Expands a line into two triangles.
void DrawLine(const GSVector4 bounds, u32 col0, u32 col1, float depth); void DrawLine(const GSVector4 bounds, u32 col0, u32 col1, float depth);
/// Computes partial derivatives and area for the given triangle. Needed for sprite/line detection.
static void ComputeUVPartialDerivatives(const BatchVertex* vertices, float* dudx, float* dudy, float* dvdx,
float* dvdy, float* xy_area, s32* uv_area);
/// Handles quads with flipped texture coordinate directions. /// Handles quads with flipped texture coordinate directions.
void HandleFlippedQuadTextureCoordinates(const GPUBackendDrawCommand* cmd, BatchVertex* vertices); void HandleFlippedQuadTextureCoordinates(const GPUBackendDrawCommand* cmd, BatchVertex* vertices);
bool IsPossibleSpritePolygon(const BatchVertex* vertices) const; bool IsPossibleSpritePolygon(const BatchVertex* vertices) const;