Latte: decode indices with NEON on aarch64

cemu-project · Jul 20, 2024 · 10bed1a · 10bed1a
1 parent e6e65af
commit 10bed1a
Showing 1 changed file with 120 additions and 6 deletions.
diff --git a/src/Cafe/HW/Latte/Core/LatteIndices.cpp b/src/Cafe/HW/Latte/Core/LatteIndices.cpp
@@ -5,6 +5,8 @@
 
 #if defined(ARCH_X86_64) && defined(__GNUC__)
 #include <immintrin.h>
+#elif defined(__aarch64__)
+#include <arm_neon.h>
 #endif
 
 struct  
@@ -480,6 +482,114 @@ void LatteIndices_fastConvertU32_AVX2(const void* indexDataInput, void* indexDat
 	indexMax = std::max(indexMax, _maxIndex);
 	indexMin = std::min(indexMin, _minIndex);
 }
+#elif defined(__aarch64__)
+
+void LatteIndices_fastConvertU16_NEON(const void* indexDataInput, void* indexDataOutput, uint32 count, uint32& indexMin, uint32& indexMax)
+{
+	const uint16* indicesU16BE = (const uint16*)indexDataInput;
+	uint16* indexOutput = (uint16*)indexDataOutput;
+	sint32 count8 = count >> 3;
+	sint32 countRemaining = count & 7;
+
+	if (count8)
+	{
+		uint16x8_t mMin = vdupq_n_u16(0xFFFF);
+		uint16x8_t mMax = vdupq_n_u16(0x0000);
+		uint16x8_t mTemp;
+		uint16x8_t* mRawIndices = (uint16x8_t*) indicesU16BE;
+		indicesU16BE += count8 * 8;
+		uint16x8_t* mOutputIndices = (uint16x8_t*) indexOutput;
+		indexOutput += count8 * 8;
+
+		while (count8--)
+		{
+			mTemp = vld1q_u16((uint16*)mRawIndices);
+			mRawIndices++;
+			mTemp = vrev16q_u8(mTemp);
+			mMin = vminq_u16(mMin, mTemp);
+			mMax = vmaxq_u16(mMax, mTemp);
+			vst1q_u16((uint16*)mOutputIndices, mTemp);
+			mOutputIndices++;
+		}
+
+		uint16* mMaxU16 = (uint16*)&mMax;
+		uint16* mMinU16 = (uint16*)&mMin;
+
+		for (int i = 0; i < 8; ++i) {
+			indexMax = std::max(indexMax, (uint32)mMaxU16[i]);
+			indexMin = std::min(indexMin, (uint32)mMinU16[i]);
+		}
+	}
+	// process remaining indices
+	uint32 _minIndex = 0xFFFFFFFF;
+	uint32 _maxIndex = 0;
+	for (sint32 i = countRemaining; (--i) >= 0;)
+	{
+		uint16 idx = _swapEndianU16(*indicesU16BE);
+		*indexOutput = idx;
+		indexOutput++;
+		indicesU16BE++;
+		_maxIndex = std::max(_maxIndex, (uint32)idx);
+		_minIndex = std::min(_minIndex, (uint32)idx);
+	}
+	// update min/max
+	indexMax = std::max(indexMax, _maxIndex);
+	indexMin = std::min(indexMin, _minIndex);
+}
+
+void LatteIndices_fastConvertU32_NEON(const void* indexDataInput, void* indexDataOutput, uint32 count, uint32& indexMin, uint32& indexMax)
+{
+	const uint32* indicesU32BE = (const uint32*)indexDataInput;
+	uint32* indexOutput = (uint32*)indexDataOutput;
+	sint32 count8 = count >> 2;
+	sint32 countRemaining = count & 3;
+
+	if (count8)
+	{
+		uint32x4_t mMin = vdupq_n_u32(0xFFFFFFFF);
+		uint32x4_t mMax = vdupq_n_u32(0x00000000);
+		uint32x4_t mTemp;
+		uint32x4_t* mRawIndices = (uint32x4_t*) indicesU32BE;
+		indicesU32BE += count8 * 4;
+		uint32x4_t* mOutputIndices = (uint32x4_t*) indexOutput;
+		indexOutput += count8 * 4;
+
+		while (count8--)
+		{
+			mTemp = vld1q_u32((uint32*)mRawIndices);
+			mRawIndices++;
+			mTemp = vrev32q_u8(mTemp);
+			mMin = vminq_u32(mMin, mTemp);
+			mMax = vmaxq_u32(mMax, mTemp);
+			vst1q_u32((uint32*)mOutputIndices, mTemp);
+			mOutputIndices++;
+		}
+
+		uint32* mMaxU32 = (uint32*)&mMax;
+		uint32* mMinU32 = (uint32*)&mMin;
+
+		for (int i = 0; i < 4; ++i) {
+			indexMax = std::max(indexMax, mMaxU32[i]);
+			indexMin = std::min(indexMin, mMinU32[i]);
+		}
+	}
+	// process remaining indices
+	uint32 _minIndex = 0xFFFFFFFF;
+	uint32 _maxIndex = 0;
+	for (sint32 i = countRemaining; (--i) >= 0;)
+	{
+		uint32 idx = _swapEndianU32(*indicesU32BE);
+		*indexOutput = idx;
+		indexOutput++;
+		indicesU32BE++;
+		_maxIndex = std::max(_maxIndex, idx);
+		_minIndex = std::min(_minIndex, idx);
+	}
+	// update min/max
+	indexMax = std::max(indexMax, _maxIndex);
+	indexMin = std::min(indexMin, _minIndex);
+}
+
 #endif
 
 template<typename T>
@@ -663,27 +773,31 @@ void LatteIndices_decode(const void* indexData, LatteIndexType indexType, uint32
 	{
 		if (indexType == LatteIndexType::U16_BE)
 		{
-            #if defined(ARCH_X86_64)
+#if defined(ARCH_X86_64)
 			if (g_CPUFeatures.x86.avx2)
 				LatteIndices_fastConvertU16_AVX2(indexData, indexOutputPtr, count, indexMin, indexMax);
 			else if (g_CPUFeatures.x86.sse4_1 && g_CPUFeatures.x86.ssse3)
 				LatteIndices_fastConvertU16_SSE41(indexData, indexOutputPtr, count, indexMin, indexMax);
 			else
 				LatteIndices_convertBE<uint16>(indexData, indexOutputPtr, count, indexMin, indexMax);
-            #else
+#elif defined(__aarch64__)
+			LatteIndices_fastConvertU16_NEON(indexData, indexOutputPtr, count, indexMin, indexMax);
+#else
 			LatteIndices_convertBE<uint16>(indexData, indexOutputPtr, count, indexMin, indexMax);            
-            #endif
+#endif
 		}
 		else if (indexType == LatteIndexType::U32_BE)
 		{
-            #if defined(ARCH_X86_64)
+#if defined(ARCH_X86_64)
 			if (g_CPUFeatures.x86.avx2)
 				LatteIndices_fastConvertU32_AVX2(indexData, indexOutputPtr, count, indexMin, indexMax);
 			else
 				LatteIndices_convertBE<uint32>(indexData, indexOutputPtr, count, indexMin, indexMax);
-            #else
+#elif defined(__aarch64__)
+			LatteIndices_fastConvertU32_NEON(indexData, indexOutputPtr, count, indexMin, indexMax);
+#else
 			LatteIndices_convertBE<uint32>(indexData, indexOutputPtr, count, indexMin, indexMax);            
-            #endif
+#endif
 		}
 		else if (indexType == LatteIndexType::U16_LE)
 		{