EE Cache: Let's blow the roof off PCSX2

PCSX2 · Dec 23, 2024 · 3fb5be9 · 3fb5be9
1 parent ccaf224
commit 3fb5be9
Showing 1 changed file with 61 additions and 6 deletions.
diff --git a/pcsx2/vtlb.cpp b/pcsx2/vtlb.cpp
@@ -119,50 +119,105 @@ __noinline int CheckCache(u32 addr)
 	}
 
 	const size_t size = cachedTlbs.count;
-	const int stride = 4;
-
-	__m128i addr_vec = _mm_set1_epi32(addr);
+	const int stride = 8;
 
+	//__m128i addr_vec = _mm_set1_epi32(addr);
+	__m256i addr_vec = _mm256_set1_epi32(addr);
 	size_t i = 0;
 
 	for (; i + stride <= size; i += stride)
 	{
+		/*
 		__m128i pfn1_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&cachedTlbs.PFN1s[i]));
 		__m128i pfn0_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&cachedTlbs.PFN0s[i]));
 		__m128i mask_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&cachedTlbs.PageMasks[i]));
+		*/
+
+		__m256i pfn1_vec = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(&cachedTlbs.PFN1s[i]));
+		__m256i pfn0_vec = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(&cachedTlbs.PFN0s[i]));
+		__m256i mask_vec = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(&cachedTlbs.PageMasks[i]));
 
+		/*
 		__m128i cached1_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&cachedTlbs.CacheEnabled1[i]));
 		__m128i cached0_vec = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&cachedTlbs.CacheEnabled0[i]));
+		*/
 
+		__m256i cached1_vec = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(&cachedTlbs.CacheEnabled1[i]));
+		__m256i cached0_vec = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(&cachedTlbs.CacheEnabled0[i]));
+
+		/*
 		__m128i pfn1_end_vec = _mm_add_epi32(pfn1_vec, mask_vec);
 		__m128i pfn0_end_vec = _mm_add_epi32(pfn0_vec, mask_vec);
+		*/
+
+		__m256i pfn1_end_vec = _mm256_add_epi32(pfn1_vec, mask_vec);
+		__m256i pfn0_end_vec = _mm256_add_epi32(pfn0_vec, mask_vec);
 
+		/*
 		__m128i gteLowerBound0 = _mm_or_si128(
 			_mm_cmpgt_epi32(addr_vec, pfn0_vec),
 			_mm_cmpeq_epi32(addr_vec, pfn0_vec));
 		__m128i gteUpperBound0 = _mm_or_si128(
 			_mm_cmpgt_epi32(pfn0_end_vec, addr_vec),
 			_mm_cmpeq_epi32(pfn0_end_vec, addr_vec));
+		*/
 
+		__m256i gteLowerBound0 = _mm256_or_si256(
+			_mm256_cmpgt_epi32(addr_vec, pfn0_vec),
+			_mm256_cmpeq_epi32(addr_vec, pfn0_vec));
+
+		__m256i gteUpperBound0 = _mm256_or_si256(
+			_mm256_cmpgt_epi32(pfn0_end_vec, addr_vec),
+			_mm256_cmpeq_epi32(pfn0_end_vec, addr_vec));
+
+		/*
 		__m128i gteUpperBound1 = _mm_or_si128(
 			_mm_cmpgt_epi32(pfn1_end_vec, addr_vec),
 			_mm_cmpeq_epi32(pfn1_end_vec, addr_vec));
 		__m128i gteLowerBound1 = _mm_or_si128(
 			_mm_cmpgt_epi32(addr_vec, pfn1_vec),
 			_mm_cmpeq_epi32(addr_vec, pfn1_vec));
+			*/
 
+		__m256i gteUpperBound1 = _mm256_or_si256(
+			_mm256_cmpgt_epi32(pfn1_end_vec, addr_vec),
+			_mm256_cmpeq_epi32(pfn1_end_vec, addr_vec));
+
+		__m256i gteLowerBound1 = _mm256_or_si256(
+			_mm256_cmpgt_epi32(addr_vec, pfn1_vec),
+			_mm256_cmpeq_epi32(addr_vec, pfn1_vec));
+
+		/*
 		__m128i cmp0 = _mm_and_si128(gteLowerBound0, gteUpperBound0);
 		__m128i cmp1 = _mm_and_si128(gteLowerBound1, gteUpperBound1);
+		*/
+
+		__m256i cmp0 = _mm256_and_si256(gteLowerBound0, gteUpperBound0);
+		__m256i cmp1 = _mm256_and_si256(gteLowerBound1, gteUpperBound1);
 
-		cmp1 = _mm_and_si128(cmp1, cached1_vec);
-		cmp0 = _mm_and_si128(cmp0, cached0_vec);
 
-		__m128i cmp = _mm_or_si128(cmp1, cmp0);
+		//cmp1 = _mm_and_si128(cmp1, cached1_vec);
+		//cmp0 = _mm_and_si128(cmp0, cached0_vec);
 
+		cmp1 = _mm256_and_si256(cmp1, cached1_vec);
+		cmp0 = _mm256_and_si256(cmp0, cached0_vec);
+
+		//__m128i cmp = _mm_or_si128(cmp1, cmp0);
+
+		__m256i cmp = _mm256_or_si256(cmp1, cmp0);
+		//__m128i cmp = _mm_or_si128(cmp1, cmp0);
+
+		/*
 		if (!_mm_testz_si128(cmp, cmp))
 		{
 			return true;
 		}
+		*/
+
+		if (!_mm256_testz_si256(cmp, cmp))
+		{
+			return true;
+		}
 	}
 
 	for (; i < size; i++)