Depth raster: more improvements

Common/Math/CrossSIMD.h

@@ -111,6 +111,13 @@ struct Vec4F32 {
 		return Vec4F32{ _mm_and_ps(v, _mm_load_ps((float *)mask)) };
 	}
 
+	// Swaps the two lower elements. Useful for reversing triangles..
+	Vec4F32 SwapLowerElements() {
+		return Vec4F32{
+			_mm_shuffle_ps(v, v, _MM_SHUFFLE(3, 2, 0, 1))
+		};
+	}
+
 	inline Vec4F32 AsVec3ByMatrix44(const Mat4F32 &m) {
 		return Vec4F32{ _mm_add_ps(
 			_mm_add_ps(
@@ -210,8 +217,8 @@ struct Vec4S32 {
 	// Swaps the two lower elements, but NOT the two upper ones. Useful for reversing triangles..
 	// This is quite awkward on ARM64 :/ Maybe there's a better solution?
 	Vec4S32 SwapLowerElements() {
-		float32x2_t upper = vget_high_s32(v);
-		float32x2_t lowerSwapped = vrev64_s32(vget_low_s32(v));
+		int32x2_t upper = vget_high_s32(v);
+		int32x2_t lowerSwapped = vrev64_s32(vget_low_s32(v));
 		return Vec4S32{ vcombine_s32(lowerSwapped, upper) };
 	};
 
@@ -281,6 +288,13 @@ struct Vec4F32 {
 		return Vec4F32{ vsetq_lane_f32(0.0f, v, 3) };
 	}
 
+	// Swaps the two lower elements, but NOT the two upper ones. Useful for reversing triangles..
+	// This is quite awkward on ARM64 :/ Maybe there's a better solution?
+	Vec4F32 SwapLowerElements() {
+		float32x2_t lowerSwapped = vrev64_f32(vget_low_f32(v));
+		return Vec4F32{ vcombine_f32(lowerSwapped, vget_high_f32(v)) };
+	};
+
 	// One of many possible solutions. Sometimes we could also use vld4q_f32 probably..
 	static void Transpose(Vec4F32 &col0, Vec4F32 &col1, Vec4F32 &col2, Vec4F32 &col3) {
 #if PPSSPP_ARCH(ARM64_NEON)

GPU/Common/DepthRaster.cpp

@@ -81,10 +81,19 @@ struct Edge {
 	}
 };
 
+enum class TriangleResult {
+	OK,
+	NoPixels,
+	Backface,
+	TooSmall,
+};
+
+constexpr int MIN_TRI_AREA = 10;
+
 // Adapted from Intel's depth rasterizer example.
 // Started with the scalar version, will SIMD-ify later.
 // x1/y1 etc are the scissor rect.
-void DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2, int y2, const int *tx, const int *ty, const int *tz, ZCompareMode compareMode) {
+TriangleResult DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2, int y2, const int *tx, const int *ty, const float *tz, ZCompareMode compareMode) {
 	int tileStartX = x1;
 	int tileEndX = x2;
 
@@ -95,16 +104,12 @@ void DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2,
 	// Due to the many multiplications, we might want to do it in floating point as 32-bit integer muls
 	// are slow on SSE2.
 
-	// Convert to whole pixels for now. Later subpixel precision.
 	int v0x = tx[0];
 	int v0y = ty[0];
-	int v0z = tz[0];
 	int v1x = tx[1];
 	int v1y = ty[1];
-	int v1z = tz[1];
 	int v2x = tx[2];
 	int v2y = ty[2];
-	int v2z = tz[2];
 
 	// use fixed-point only for X and Y.  Avoid work for Z and W.
 	// We use 4x1 tiles for simplicity.
@@ -114,13 +119,16 @@ void DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2,
 	int maxY = std::min(std::max(std::max(v0y, v1y), v2y), tileEndY);
 	if (maxX == minX || maxY == minY) {
 		// No pixels, or outside screen.
-		return;
+		return TriangleResult::NoPixels;
 	}
 
-	// TODO: Cull really small triangles here.
+	// TODO: Cull really small triangles here - we can increase the threshold a bit probably.
 	int triArea = (v1y - v2y) * v0x + (v2x - v1x) * v0y + (v1x * v2y - v2x * v1y);
 	if (triArea <= 0) {
-		return;
+		return TriangleResult::Backface;
+	}
+	if (triArea < MIN_TRI_AREA) {
+		return TriangleResult::TooSmall;
 	}
 
 	float oneOverTriArea = 1.0f / (float)triArea;
@@ -132,20 +140,25 @@ void DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2,
 	Vec4S32 w2_row = e01.init(v0x, v0y, v1x, v1y, minX, minY);
 
 	// Prepare to interpolate Z
-	Vec4F32 zz0 = Vec4F32::Splat((float)v0z);
-	Vec4F32 zz1 = Vec4F32::Splat((float)(v1z - v0z) * oneOverTriArea);
-	Vec4F32 zz2 = Vec4F32::Splat((float)(v2z - v0z) * oneOverTriArea);
+	Vec4F32 zz0 = Vec4F32::Splat(tz[0]);
+	Vec4F32 zz1 = Vec4F32::Splat((tz[1] - tz[0]) * oneOverTriArea);
+	Vec4F32 zz2 = Vec4F32::Splat((tz[2] - tz[0]) * oneOverTriArea);
+
+	Vec4F32 zdeltaX = zz1 * Vec4F32FromS32(e20.oneStepX) + zz2 * Vec4F32FromS32(e01.oneStepX);
+	Vec4F32 zdeltaY = zz1 * Vec4F32FromS32(e20.oneStepY) + zz2 * Vec4F32FromS32(e01.oneStepY);
+	Vec4F32 zrow = zz0 + Vec4F32FromS32(w1_row) * zz1 + Vec4F32FromS32(w2_row) * zz2;
 
 	// Rasterize
-	for (int y = minY; y <= maxY; y += Edge::stepYSize, w0_row += e12.oneStepY, w1_row += e20.oneStepY, w2_row += e01.oneStepY) {
+	for (int y = minY; y <= maxY; y += Edge::stepYSize, w0_row += e12.oneStepY, w1_row += e20.oneStepY, w2_row += e01.oneStepY, zrow += zdeltaY) {
 		// Barycentric coordinates at start of row
 		Vec4S32 w0 = w0_row;
 		Vec4S32 w1 = w1_row;
 		Vec4S32 w2 = w2_row;
+		Vec4F32 zs = zrow;
 
 		uint16_t *rowPtr = depthBuf + stride * y;
 
-		for (int x = minX; x <= maxX; x += Edge::stepXSize, w0 += e12.oneStepX, w1 += e20.oneStepX, w2 += e01.oneStepX) {
+		for (int x = minX; x <= maxX; x += Edge::stepXSize, w0 += e12.oneStepX, w1 += e20.oneStepX, w2 += e01.oneStepX, zs += zdeltaX) {
 			// If p is on or inside all edges for any pixels,
 			// render those pixels.
 			Vec4S32 signCalc = w0 | w1 | w2;
@@ -157,9 +170,7 @@ void DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2,
 			Vec4U16 shortMaskInv = SignBits32ToMaskU16(signCalc);
 			// Now, the mask has 1111111 where we should preserve the contents of the depth buffer.
 
-			// Compute the Z value for all four pixels.
-			// float depth = zz[0] + beta * zz[1] + gamma * zz[2];
-			Vec4U16 shortZ = Vec4U16::FromVec4F32(zz0 + Vec4F32FromS32(w1) * zz1 + Vec4F32FromS32(w2) * zz2);
+			Vec4U16 shortZ = Vec4U16::FromVec4F32(zs);
 
 			// TODO: Lift this switch out of the inner loop, or even out of the function with templating.
 			switch (compareMode) {
@@ -180,6 +191,7 @@ void DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2,
 			}
 		}
 	}
+	return TriangleResult::OK;
 }
 
 void DecodeAndTransformForDepthRaster(float *dest, const float *worldviewproj, const void *vertexData, int indexLowerBound, int indexUpperBound, VertexDecoder *dec, u32 vertTypeID) {
@@ -249,7 +261,7 @@ void ConvertPredecodedThroughForDepthRaster(float *dest, const void *decodedVert
 	}
 }
 
-int DepthRasterClipIndexedRectangles(int *tx, int *ty, int *tz, const float *transformed, const uint16_t *indexBuffer, int count) {
+int DepthRasterClipIndexedRectangles(int *tx, int *ty, float *tz, const float *transformed, const uint16_t *indexBuffer, int count) {
 	// TODO: On ARM we can do better by keeping these in lanes instead of splatting.
 	// However, hard to find a common abstraction.
 	const Vec4F32 viewportX = Vec4F32::Splat(gstate.getViewportXCenter());
@@ -289,20 +301,21 @@ int DepthRasterClipIndexedRectangles(int *tx, int *ty, int *tz, const float *tra
 		y *= recipW;
 		z *= recipW;
 
-		Vec4S32 screen[3];
+		Vec4S32 screen[2];
+		Vec4F32 depth;
 		screen[0] = Vec4S32FromF32((x * viewportScaleX + viewportX) - offsetX);
 		screen[1] = Vec4S32FromF32((y * viewportScaleY + viewportY) - offsetY);
-		screen[2] = Vec4S32FromF32((z * viewportScaleZ + viewportZ).Clamp(0.0f, 65535.0f));
+		depth = (z * viewportScaleZ + viewportZ).Clamp(0.0f, 65535.0f);
 
 		screen[0].Store(tx + outCount);
 		screen[1].Store(ty + outCount);
-		screen[2].Store(tz + outCount);
+		depth.Store(tz + outCount);
 		outCount += 2;
 	}
 	return outCount;
 }
 
-int DepthRasterClipIndexedTriangles(int *tx, int *ty, int *tz, const float *transformed, const uint16_t *indexBuffer, int count) {
+int DepthRasterClipIndexedTriangles(int *tx, int *ty, float *tz, const float *transformed, const uint16_t *indexBuffer, int count) {
 	bool cullEnabled = gstate.isCullEnabled();
 	GECullMode cullMode = gstate.getCullMode();
 
@@ -351,14 +364,14 @@ int DepthRasterClipIndexedTriangles(int *tx, int *ty, int *tz, const float *tran
 		y *= recipW;
 		z *= recipW;
 
-		Vec4S32 screen[3];
+		Vec4S32 screen[2];
 		screen[0] = Vec4S32FromF32((x * viewportScaleX + viewportX) - offsetX);
 		screen[1] = Vec4S32FromF32((y * viewportScaleY + viewportY) - offsetY);
-		screen[2] = Vec4S32FromF32((z * viewportScaleZ + viewportZ).Clamp(0.0f, 65535.0f));
+		Vec4F32 depth = (z * viewportScaleZ + viewportZ).Clamp(0.0f, 65535.0f);
 
 		screen[0].Store(tx + outCount);
 		screen[1].Store(ty + outCount);
-		screen[2].Store(tz + outCount);
+		depth.Store(tz + outCount);
 		outCount += 3;
 
 		if (!cullEnabled) {
@@ -371,25 +384,25 @@ int DepthRasterClipIndexedTriangles(int *tx, int *ty, int *tz, const float *tran
 
 			screen[0].SwapLowerElements().Store(tx + outCount);
 			screen[1].SwapLowerElements().Store(ty + outCount);
-			screen[2].SwapLowerElements().Store(tz + outCount);
+			depth.SwapLowerElements().Store(tz + outCount);
 			outCount += 3;
 		}
 	}
 	return outCount;
 }
 
-void DepthRasterConvertTransformed(int *tx, int *ty, int *tz, const float *transformed, const uint16_t *indexBuffer, int count) {
+void DepthRasterConvertTransformed(int *tx, int *ty, float *tz, const float *transformed, const uint16_t *indexBuffer, int count) {
 	// TODO: This is basically a transpose, or AoS->SoA conversion. There may be fast ways.
 	for (int i = 0; i < count; i++) {
 		const float *pos = transformed + indexBuffer[i] * 4;
 		tx[i] = (int)pos[0];
 		ty[i] = (int)pos[1];
-		tz[i] = (u16)pos[2];
+		tz[i] = pos[2];  // clamp?
 	}
 }
 
 // Rasterizes screen-space vertices.
-void DepthRasterScreenVerts(uint16_t *depth, int depthStride, GEPrimitiveType prim, int x1, int y1, int x2, int y2, const int *tx, const int *ty, const int *tz, int count) {
+void DepthRasterScreenVerts(uint16_t *depth, int depthStride, GEPrimitiveType prim, int x1, int y1, int x2, int y2, const int *tx, const int *ty, const float *tz, int count) {
 	// Prim should now be either TRIANGLES or RECTs.
 	_dbg_assert_(prim == GE_PRIM_RECTANGLES || prim == GE_PRIM_TRIANGLES);
 
@@ -438,17 +451,26 @@ void DepthRasterScreenVerts(uint16_t *depth, int depthStride, GEPrimitiveType pr
 	switch (prim) {
 	case GE_PRIM_RECTANGLES:
 		for (int i = 0; i < count; i += 2) {
-			uint16_t z = tz[i + 1];  // depth from second vertex
+			uint16_t z = (uint16_t)tz[i + 1];  // depth from second vertex
 			// TODO: Should clip coordinates to the scissor rectangle.
 			// We remove the subpixel information here.
 			DepthRasterRect(depth, depthStride, tx[i], ty[i], tx[i + 1], ty[i + 1], z, comp);
 		}
+		gpuStats.numDepthRasterPrims += count / 2;
 		break;
 	case GE_PRIM_TRIANGLES:
+	{
+		int stats[4]{};
 		for (int i = 0; i < count; i += 3) {
-			DepthRasterTriangle(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i], comp);
+			TriangleResult result = DepthRasterTriangle(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i], comp);
+			stats[(int)result]++;
 		}
+		gpuStats.numDepthRasterBackface += stats[(int)TriangleResult::Backface];
+		gpuStats.numDepthRasterNoPixels += stats[(int)TriangleResult::NoPixels];
+		gpuStats.numDepthRasterTooSmall += stats[(int)TriangleResult::TooSmall];
+		gpuStats.numDepthRasterPrims += stats[(int)TriangleResult::OK];
 		break;
+	}
 	default:
 		_dbg_assert_(false);
 	}

GPU/Common/DepthRaster.h

@@ -17,13 +17,13 @@ struct DepthScreenVertex {
 class VertexDecoder;
 struct TransformedVertex;
 
-int DepthRasterClipIndexedTriangles(int *tx, int *ty, int *tz, const float *transformed, const uint16_t *indexBuffer, int count);
-int DepthRasterClipIndexedRectangles(int *tx, int *ty, int *tz, const float *transformed, const uint16_t *indexBuffer, int count);
+int DepthRasterClipIndexedTriangles(int *tx, int *ty, float *tz, const float *transformed, const uint16_t *indexBuffer, int count);
+int DepthRasterClipIndexedRectangles(int *tx, int *ty, float *tz, const float *transformed, const uint16_t *indexBuffer, int count);
 void DecodeAndTransformForDepthRaster(float *dest, const float *worldviewproj, const void *vertexData, int indexLowerBound, int indexUpperBound, VertexDecoder *dec, u32 vertTypeID);
 void TransformPredecodedForDepthRaster(float *dest, const float *worldviewproj, const void *decodedVertexData, VertexDecoder *dec, int count);
 void ConvertPredecodedThroughForDepthRaster(float *dest, const void *decodedVertexData, VertexDecoder *dec, int count);
-void DepthRasterConvertTransformed(int *tx, int *ty, int *tz, const float *transformed, const uint16_t *indexBuffer, int count);
+void DepthRasterConvertTransformed(int *tx, int *ty, float *tz, const float *transformed, const uint16_t *indexBuffer, int count);
 
 // void DepthRasterConvertTransformed(int *tx, int *ty, int *tz, GEPrimitiveType prim, const TransformedVertex *transformed, int count);
 
-void DepthRasterScreenVerts(uint16_t *depth, int depthStride, GEPrimitiveType prim, int x1, int y1, int x2, int y2, const int *tx, const int *ty, const int *tz, int count);
+void DepthRasterScreenVerts(uint16_t *depth, int depthStride, GEPrimitiveType prim, int x1, int y1, int x2, int y2, const int *tx, const int *ty, const float *tz, int count);

GPU/Common/DrawEngineCommon.cpp

@@ -23,6 +23,7 @@
 #include "Common/LogReporting.h"
 #include "Common/Math/SIMDHeaders.h"
 #include "Common/Math/lin/matrix4x4.h"
+#include "Common/TimeUtil.h"
 #include "Core/System.h"
 #include "Core/Config.h"
 #include "GPU/Common/DrawEngineCommon.h"
@@ -914,6 +915,7 @@ inline void ComputeFinalProjMatrix(float *worldviewproj) {
 }
 
 void DrawEngineCommon::DepthRasterTransform(GEPrimitiveType prim, VertexDecoder *dec, uint32_t vertTypeID, int vertexCount) {
+
 	switch (prim) {
 	case GE_PRIM_INVALID:
 	case GE_PRIM_KEEP_PREVIOUS:
@@ -929,6 +931,8 @@ void DrawEngineCommon::DepthRasterTransform(GEPrimitiveType prim, VertexDecoder
 		return;
 	}
 
+	TimeCollector collectStat(&gpuStats.msRasterizingDepth, coreCollectDebugStats);
+
 	float worldviewproj[16];
 	ComputeFinalProjMatrix(worldviewproj);
 
@@ -953,7 +957,7 @@ void DrawEngineCommon::DepthRasterTransform(GEPrimitiveType prim, VertexDecoder
 
 	int *tx = depthScreenVerts_;
 	int *ty = depthScreenVerts_ + DEPTH_SCREENVERTS_COMPONENT_COUNT;
-	int *tz = depthScreenVerts_ + DEPTH_SCREENVERTS_COMPONENT_COUNT * 2;
+	float *tz = (float *)(depthScreenVerts_ + DEPTH_SCREENVERTS_COMPONENT_COUNT * 2);
 
 	// Clip and triangulate using the index buffer.
 	int outVertCount = DepthRasterClipIndexedTriangles(tx, ty, tz, depthTransformed_, decIndex_, vertexCount);
@@ -962,7 +966,7 @@ void DrawEngineCommon::DepthRasterTransform(GEPrimitiveType prim, VertexDecoder
 		for (int i = outVertCount; i < ((outVertCount + 16) & ~15); i++) {
 			tx[i] = 0;
 			ty[i] = 0;
-			tz[i] = 0;
+			tz[i] = 0.0f;
 		}
 	}
 
@@ -972,6 +976,8 @@ void DrawEngineCommon::DepthRasterTransform(GEPrimitiveType prim, VertexDecoder
 }
 
 void DrawEngineCommon::DepthRasterPredecoded(GEPrimitiveType prim, const void *inVerts, int numDecoded, VertexDecoder *dec, int vertexCount) {
+	TimeCollector collectStat(&gpuStats.msRasterizingDepth, coreCollectDebugStats);
+
 	switch (prim) {
 	case GE_PRIM_INVALID:
 	case GE_PRIM_KEEP_PREVIOUS:
@@ -987,7 +993,7 @@ void DrawEngineCommon::DepthRasterPredecoded(GEPrimitiveType prim, const void *i
 
 	int *tx = depthScreenVerts_;
 	int *ty = depthScreenVerts_ + DEPTH_SCREENVERTS_COMPONENT_COUNT;
-	int *tz = depthScreenVerts_ + DEPTH_SCREENVERTS_COMPONENT_COUNT * 2;
+	float *tz = (float *)(depthScreenVerts_ + DEPTH_SCREENVERTS_COMPONENT_COUNT * 2);
 
 	int outVertCount = 0;
 
@@ -996,6 +1002,9 @@ void DrawEngineCommon::DepthRasterPredecoded(GEPrimitiveType prim, const void *i
 		DepthRasterConvertTransformed(tx, ty, tz, depthTransformed_, decIndex_, vertexCount);
 		outVertCount = vertexCount;
 	} else {
+		if (dec->VertexType() & (GE_VTYPE_WEIGHT_MASK | GE_VTYPE_MORPHCOUNT_MASK)) {
+			return;
+		}
 		float worldviewproj[16];
 		ComputeFinalProjMatrix(worldviewproj);
 		TransformPredecodedForDepthRaster(depthTransformed_, worldviewproj, decoded_, dec, numDecoded);
@@ -1018,7 +1027,7 @@ void DrawEngineCommon::DepthRasterPredecoded(GEPrimitiveType prim, const void *i
 		for (int i = outVertCount; i < ((outVertCount + 16) & ~15); i++) {
 			tx[i] = 0;
 			ty[i] = 0;
-			tz[i] = 0;
+			tz[i] = 0.0f;
 		}
 	}
 	DepthRasterScreenVerts((uint16_t *)Memory::GetPointerWrite(gstate.getDepthBufRawAddress() | 0x04000000), gstate.DepthBufStride(),

GPU/Common/FramebufferManagerCommon.cpp

@@ -1866,6 +1866,8 @@ void FramebufferManagerCommon::ResizeFramebufFBO(VirtualFramebuffer *vfb, int w,
 	char tag[128];
 	size_t len = FormatFramebufferName(vfb, tag, sizeof(tag));
 
+	gpuStats.numFBOsCreated++;
+
 	vfb->fbo = draw_->CreateFramebuffer({ vfb->renderWidth, vfb->renderHeight, 1, GetFramebufferLayers(), msaaLevel_, true, tag });
 	if (Memory::IsVRAMAddress(vfb->fb_address) && vfb->fb_stride != 0) {
 		NotifyMemInfo(MemBlockFlags::ALLOC, vfb->fb_address, vfb->BufferByteSize(RASTER_COLOR), tag, len);