From 8123cc27dbe1903b37fffc672f55d0ac5e12c471 Mon Sep 17 00:00:00 2001
From: TellowKrinkle <tellowkrinkle@gmail.com>
Date: Wed, 8 Sep 2021 22:38:01 -0500
Subject: [PATCH] GS: Improved SW renderer debug stats

---
 pcsx2/GS/Renderers/Common/GSFunctionMap.h | 38 ++++++++++++++---------
 pcsx2/GS/Renderers/SW/GSRasterizer.cpp    |  7 +++--
 2 files changed, 28 insertions(+), 17 deletions(-)

diff --git a/pcsx2/GS/Renderers/Common/GSFunctionMap.h b/pcsx2/GS/Renderers/Common/GSFunctionMap.h
index 93d9f09f4e..59726b0ada 100644
--- a/pcsx2/GS/Renderers/Common/GSFunctionMap.h
+++ b/pcsx2/GS/Renderers/Common/GSFunctionMap.h
@@ -19,6 +19,7 @@
 #include "GS/GSCodeBuffer.h"
 
 #include "GS/Renderers/SW/GSScanlineEnvironment.h"
+#include "common/General.h"
 
 #include <xbyak/xbyak_util.h>
 
@@ -28,12 +29,11 @@ class GSFunctionMap
 protected:
 	struct ActivePtr
 	{
-		uint64 frame, frames;
+		uint64 frame, frames, prims;
 		uint64 ticks, actual, total;
 		VALUE f;
 	};
 
-	std::unordered_map<KEY, VALUE> m_map;
 	std::unordered_map<KEY, ActivePtr*> m_map_active;
 
 	ActivePtr* m_active;
@@ -64,15 +64,13 @@ public:
 		}
 		else
 		{
-			auto i = m_map.find(key);
-
 			ActivePtr* p = new ActivePtr();
 
 			memset(p, 0, sizeof(*p));
 
 			p->frame = (uint64)-1;
 
-			p->f = i != m_map.end() ? i->second : GetDefaultFunction(key);
+			p->f = GetDefaultFunction(key);
 
 			m_map_active[key] = p;
 
@@ -92,6 +90,7 @@ public:
 				m_active->frames++;
 			}
 
+			m_active->prims++;
 			m_active->ticks += ticks;
 			m_active->actual += actual;
 			m_active->total += total;
@@ -114,25 +113,36 @@ public:
 			}
 		}
 
+		uint64_t million = 1000 * 1000;
+		uint64_t billion = 1000 * million;
+		uint64_t tps = GetTickFrequency();
+
 		printf("GS stats\n");
 
+		printf("      key      | frames | prims |       runtime      |          pixels\n");
+		printf("               |        |  #/f  |   pct   µs/f ns/px |    #/f   #/prim overdraw\n");
+
 		for (const auto& i : m_map_active)
 		{
 			KEY key = i.first;
 			ActivePtr* p = i.second;
 
-			if (p->frames && ttpf)
+			if (p->frames && p->actual && ttpf)
 			{
-				uint64 tpp = p->actual > 0 ? p->ticks / p->actual : 0;
-				uint64 tpf = p->frames > 0 ? p->ticks / p->frames : 0;
-				uint64 ppf = p->frames > 0 ? p->actual / p->frames : 0;
+				uint64 tpp = p->ticks  / p->actual;
+				uint64 tpf = p->ticks  / p->frames;
+				uint64 ppf = p->actual / p->frames;
 
-				printf("[%014llx]%c %6.2f%% %5.2f%% f %4llu t %12llu p %12llu w %12lld tpp %4llu tpf %9llu ppf %9llu\n",
-					(uint64)key, m_map.find(key) == m_map.end() ? '*' : ' ',
-					(float)(tpf * 10000 / 34000000) / 100,
+				printf("%014llx | %6llu | %5llu | %5.2f%% %6llu %4llu | %8llu %6llu %5.2f%%\n",
+					(uint64)key,
+					p->frames,
+					p->prims / p->frames,
 					(float)(tpf * 10000 / ttpf) / 100,
-					p->frames, p->ticks, p->actual, p->total - p->actual,
-					tpp, tpf, ppf);
+					(tpf * million) / tps,
+					(tpp * billion) / tps,
+					ppf,
+					p->actual / p->prims,
+					(float)((p->total - p->actual) * 10000 / p->total) / 100);
 			}
 		}
 	}
diff --git a/pcsx2/GS/Renderers/SW/GSRasterizer.cpp b/pcsx2/GS/Renderers/SW/GSRasterizer.cpp
index ba6b4052c6..c3cd313a73 100644
--- a/pcsx2/GS/Renderers/SW/GSRasterizer.cpp
+++ b/pcsx2/GS/Renderers/SW/GSRasterizer.cpp
@@ -17,6 +17,7 @@
 
 #include "PrecompiledHeader.h"
 #include "GSRasterizer.h"
+#include "common/General.h"
 
 int GSRasterizerData::s_counter = 0;
 
@@ -138,7 +139,7 @@ void GSRasterizer::Draw(GSRasterizerData* data)
 	m_pixels.actual = 0;
 	m_pixels.total = 0;
 
-	data->start = __rdtsc();
+	data->start = GetCPUTicks();
 
 	m_ds->BeginDraw(data);
 
@@ -244,7 +245,7 @@ void GSRasterizer::Draw(GSRasterizerData* data)
 
 	data->pixels = m_pixels.actual;
 
-	uint64 ticks = __rdtsc() - data->start;
+	uint64 ticks = GetCPUTicks() - data->start;
 
 	m_pixels.sum += m_pixels.actual;
 
@@ -1146,7 +1147,7 @@ void GSRasterizer::Flush(const GSVertexSW* vertex, const uint32* index, const GS
 void GSRasterizer::DrawScanline(int pixels, int left, int top, const GSVertexSW& scan)
 {
 	m_pixels.actual += pixels;
-	m_pixels.total += ((left + pixels + (PIXELS_PER_LOOP - 1)) & ~(PIXELS_PER_LOOP - 1)) - (left & (PIXELS_PER_LOOP - 1));
+	m_pixels.total += ((left + pixels + (PIXELS_PER_LOOP - 1)) & ~(PIXELS_PER_LOOP - 1)) - (left & ~(PIXELS_PER_LOOP - 1));
 	//m_pixels.total += ((left + pixels + (PIXELS_PER_LOOP - 1)) & ~(PIXELS_PER_LOOP - 1)) - left;
 
 	ASSERT(m_pixels.actual <= m_pixels.total);