From 5df88fc1aab1feb92e60a6b14eec01f3b56e85b2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Sat, 21 Dec 2024 11:33:37 +0100
Subject: [PATCH] Convert the rect implementation to CrossSIMD

---
 Common/Math/CrossSIMD.h    | 20 ++++++++++++++++++++
 GPU/Common/DepthRaster.cpp | 33 ++-------------------------------
 2 files changed, 22 insertions(+), 31 deletions(-)

diff --git a/Common/Math/CrossSIMD.h b/Common/Math/CrossSIMD.h
index 982c859439..83ac18add4 100644
--- a/Common/Math/CrossSIMD.h
+++ b/Common/Math/CrossSIMD.h
@@ -153,6 +153,16 @@ struct Vec4U16 {
 	Vec4U16 CompareLT(Vec4U16 other) { return Vec4U16{ _mm_cmplt_epu16(v, other.v) }; }
 };
 
+struct Vec8U16 {
+	__m128i v;
+
+	static Vec8U16 Zero() { return Vec8U16{ _mm_setzero_si128() }; }
+	static Vec8U16 Splat(uint16_t value) { return Vec8U16{ _mm_set1_epi16((int16_t)value) }; }
+
+	static Vec8U16 Load(const uint16_t *mem) { return Vec8U16{ _mm_loadu_si128((__m128i *)mem) }; }
+	void Store(uint16_t *mem) { _mm_storeu_si128((__m128i *)mem, v); }
+};
+
 Vec4U16 SignBits32ToMaskU16(Vec4S32 v) {
 	__m128i temp = _mm_srai_epi32(v.v, 31);
 	return Vec4U16 {
@@ -342,6 +352,16 @@ Vec4U16 AndNot(Vec4U16 a, Vec4U16 inverted) {
 	return Vec4U16{ vand_u16(a.v, vmvn_u16(inverted.v)) };
 }
 
+struct Vec8U16 {
+	uint16x8_t v;
+
+	static Vec8U16 Zero() { return Vec8U16{ vdupq_n_u16(0) }; }
+	static Vec8U16 Splat(uint16_t value) { return Vec8U16{ vdupq_n_u16(value) }; }
+
+	static Vec8U16 Load(const uint16_t *mem) { return Vec8U16{ vld1q_u16(mem) }; }
+	void Store(uint16_t *mem) { vst1q_u16(mem, v); }
+};
+
 #else
 
 struct Vec4S32 {
diff --git a/GPU/Common/DepthRaster.cpp b/GPU/Common/DepthRaster.cpp
index 7075ebd9f7..4443974059 100644
--- a/GPU/Common/DepthRaster.cpp
+++ b/GPU/Common/DepthRaster.cpp
@@ -28,43 +28,17 @@ void DepthRasterRect(uint16_t *dest, int stride, int x1, int y1, int x2, int y2,
 		return;
 	}
 
-#if PPSSPP_ARCH(SSE2)
-	__m128i valueX8 = _mm_set1_epi16(depthValue);
-	for (int y = y1; y < y2; y++) {
-		__m128i *ptr = (__m128i *)(dest + stride * y + x1);
-		int w = x2 - x1;
-		switch (compareMode) {
-		case ZCompareMode::Always:
-			if (depthValue == 0) {
-				memset(ptr, 0, w * 2);
-			} else {
-				while (w >= 8) {
-					_mm_storeu_si128(ptr, valueX8);
-					ptr++;
-					w -= 8;
-				}
-			}
-			break;
-			// TODO: Trailer
-		default:
-			// TODO
-			break;
-		}
-	}
-
-#elif PPSSPP_ARCH(ARM64_NEON)
-	uint16x8_t valueX8 = vdupq_n_u16(depthValue);
+	Vec8U16 valueX8 = Vec8U16::Splat(depthValue);
 	for (int y = y1; y < y2; y++) {
 		uint16_t *ptr = (uint16_t *)(dest + stride * y + x1);
 		int w = x2 - x1;
-
 		switch (compareMode) {
 		case ZCompareMode::Always:
 			if (depthValue == 0) {
 				memset(ptr, 0, w * 2);
 			} else {
 				while (w >= 8) {
-					vst1q_u16(ptr, valueX8);
+					valueX8.Store(ptr);
 					ptr += 8;
 					w -= 8;
 				}
@@ -76,9 +50,6 @@ void DepthRasterRect(uint16_t *dest, int stride, int x1, int y1, int x2, int y2,
 			break;
 		}
 	}
-#else
-	// Do nothing for now
-#endif
 }
 
 alignas(16) static const int zero123[4]  = {0, 1, 2, 3};