From 3ccb01b09cff4d32f6ae4d5234c47c53e8b43f3d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Sun, 22 Dec 2024 10:02:29 +0100
Subject: [PATCH] Depth Raster: Add more stats, discard triangles with < 10 px
 area. Still good results.

---
 GPU/Common/DepthRaster.cpp | 33 +++++++++++++++++++++++----------
 GPU/GPU.h                  |  8 ++++++--
 GPU/GPUCommonHW.cpp        |  6 ++++--
 3 files changed, 33 insertions(+), 14 deletions(-)

diff --git a/GPU/Common/DepthRaster.cpp b/GPU/Common/DepthRaster.cpp
index eb4485a75f..f9e3f3b165 100644
--- a/GPU/Common/DepthRaster.cpp
+++ b/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.
-bool 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) {
+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;
 
@@ -110,13 +119,16 @@ bool 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 false;
+		return TriangleResult::NoPixels;
 	}
 
 	// 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 false;
+		return TriangleResult::Backface;
+	}
+	if (triArea < MIN_TRI_AREA) {
+		return TriangleResult::TooSmall;
 	}
 
 	float oneOverTriArea = 1.0f / (float)triArea;
@@ -179,7 +191,7 @@ bool DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y1, int x2,
 			}
 		}
 	}
-	return true;
+	return TriangleResult::OK;
 }
 
 void DecodeAndTransformForDepthRaster(float *dest, const float *worldviewproj, const void *vertexData, int indexLowerBound, int indexUpperBound, VertexDecoder *dec, u32 vertTypeID) {
@@ -448,14 +460,15 @@ void DepthRasterScreenVerts(uint16_t *depth, int depthStride, GEPrimitiveType pr
 		break;
 	case GE_PRIM_TRIANGLES:
 	{
-		int culled = 0;
+		int stats[4]{};
 		for (int i = 0; i < count; i += 3) {
-			if (!DepthRasterTriangle(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i], comp)) {
-				culled++;
-			}
+			TriangleResult result = DepthRasterTriangle(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i], comp);
+			stats[(int)result]++;
 		}
-		gpuStats.numDepthRasterCulls += culled;
-		gpuStats.numDepthRasterPrims += count / 3;
+		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:
diff --git a/GPU/GPU.h b/GPU/GPU.h
index ef83235775..9b98e86344 100644
--- a/GPU/GPU.h
+++ b/GPU/GPU.h
@@ -110,7 +110,9 @@ struct GPUStatistics {
 		msProcessingDisplayLists = 0;
 		msRasterizingDepth = 0.0f;
 		numDepthRasterPrims = 0;
-		numDepthRasterCulls = 0;
+		numDepthRasterBackface = 0;
+		numDepthRasterNoPixels = 0;
+		numDepthRasterTooSmall = 0;
 		vertexGPUCycles = 0;
 		otherGPUCycles = 0;
 	}
@@ -154,7 +156,9 @@ struct GPUStatistics {
 	int vertexGPUCycles;
 	int otherGPUCycles;
 	int numDepthRasterPrims;
-	int numDepthRasterCulls;
+	int numDepthRasterBackface;
+	int numDepthRasterNoPixels;
+	int numDepthRasterTooSmall;
 	// Flip count. Doesn't really belong here.
 	int numFlips;
 };
diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index d8db488251..552a38179c 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -1776,7 +1776,7 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		"replacer: tracks %d references, %d unique textures\n"
 		"Cpy: depth %d, color %d, reint %d, blend %d, self %d\n"
 		"GPU cycles: %d (%0.1f per vertex)\n"
-		"Depth raster: %0.2f ms, %d prim, %d prims culled\n%s",
+		"Depth raster: %0.2f ms, %d prim, %d nopix, %d small, %d backface\n%s",
 		gpuStats.msProcessingDisplayLists * 1000.0f,
 		gpuStats.numDrawSyncs,
 		gpuStats.numListSyncs,
@@ -1815,7 +1815,9 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		vertexAverageCycles,
 		gpuStats.msRasterizingDepth * 1000.0,
 		gpuStats.numDepthRasterPrims,
-		gpuStats.numDepthRasterCulls,
+		gpuStats.numDepthRasterNoPixels,
+		gpuStats.numDepthRasterTooSmall,
+		gpuStats.numDepthRasterBackface,
 		debugRecording_ ? "(debug-recording)" : ""
 	);
 }