Enable depth raster for Armored Core by default, minor speedup

Common/Math/CrossSIMD.h

@@ -45,6 +45,17 @@ struct Vec4S32 {
 		};
 	}
 
+	// Reads 16 bits from both operands, produces a 32-bit result per lane.
+	// On SSE2, much faster than _mm_mullo_epi32_SSE2.
+	// On NEON though, it'll read the full 32 bits, so beware.
+	// See https://fgiesen.wordpress.com/2016/04/03/sse-mind-the-gap/.
+	Vec4S32 MulAsS16(Vec4S32 other) const {
+		// Note that we only need to mask one of the inputs, so we get zeroes - multiplying
+		// by zero is zero, so it doesn't matter what the upper halfword of each 32-bit word is
+		// in the other register.
+		return Vec4S32{ _mm_madd_epi16(v, _mm_and_si128(other.v, _mm_set1_epi32(0x0000FFFF))) };
+	}
+
 	Vec4S32 operator +(Vec4S32 other) const { return Vec4S32{ _mm_add_epi32(v, other.v) }; }
 	Vec4S32 operator -(Vec4S32 other) const { return Vec4S32{ _mm_sub_epi32(v, other.v) }; }
 	Vec4S32 operator |(Vec4S32 other) const { return Vec4S32{ _mm_or_si128(v, other.v) }; }
@@ -222,6 +233,9 @@ struct Vec4S32 {
 		return Vec4S32{ vcombine_s32(lowerSwapped, upper) };
 	};
 
+	// Warning: Unlike on x86, this is a full 32-bit multiplication.
+	Vec4S32 MulAsS16(Vec4S32 other) const { return Vec4S32{ vmulq_s32(v, other.v) }; }
+
 	Vec4S32 operator +(Vec4S32 other) const { return Vec4S32{ vaddq_s32(v, other.v) }; }
 	Vec4S32 operator -(Vec4S32 other) const { return Vec4S32{ vsubq_s32(v, other.v) }; }
 	Vec4S32 operator *(Vec4S32 other) const { return Vec4S32{ vmulq_s32(v, other.v) }; }
@@ -269,6 +283,7 @@ struct Vec4F32 {
 	Vec4F32 operator *(float f) const { return Vec4F32{ vmulq_f32(v, vdupq_n_f32(f)) }; }
 
 	Vec4F32 Mul(float f) const { return Vec4F32{ vmulq_f32(v, vdupq_n_f32(f)) }; }
+
 	Vec4F32 Recip() {
 		float32x4_t recip = vrecpeq_f32(v);
 		// Use a couple Newton-Raphson steps to refine the estimate.

GPU/Common/DepthRaster.cpp

@@ -60,9 +60,13 @@ void DepthRasterRect(uint16_t *dest, int stride, int x1, int y1, int x2, int y2,
 					ptr += 8;
 					w -= 8;
 				}
+				// Non-simd trailer.
+				while (w > 0) {
+					*ptr++ = depthValue;
+					w--;
+				}
 			}
 			break;
-			// TODO: Trailer
 		default:
 			// TODO
 			break;
@@ -111,7 +115,8 @@ 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.
-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) {
+template<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) {
 	int tileStartX = x1;
 	int tileEndX = x2;
 
@@ -190,7 +195,7 @@ TriangleResult DepthRasterTriangle(uint16_t *depthBuf, int stride, int x1, int y
 
 			Vec4U16 shortZ = Vec4U16::FromVec4F32(zs);
 
-			// TODO: Lift this switch out of the inner loop, or even out of the function with templating.
+			// This switch is on a templated constant, so should collapse away.
 			switch (compareMode) {
 			case ZCompareMode::Greater:
 				// To implement the greater/greater-than comparison, we can combine mask and max.
@@ -479,9 +484,25 @@ void DepthRasterScreenVerts(uint16_t *depth, int depthStride, GEPrimitiveType pr
 	case GE_PRIM_TRIANGLES:
 	{
 		int stats[4]{};
-		for (int i = 0; i < count; i += 3) {
-			TriangleResult result = DepthRasterTriangle(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i], comp);
-			stats[(int)result]++;
+		switch (comp) {
+		case ZCompareMode::Greater:
+			for (int i = 0; i < count; i += 3) {
+				TriangleResult result = DepthRasterTriangle<ZCompareMode::Greater>(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i]);
+				stats[(int)result]++;
+			}
+			break;
+		case ZCompareMode::Less:
+			for (int i = 0; i < count; i += 3) {
+				TriangleResult result = DepthRasterTriangle<ZCompareMode::Less>(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i]);
+				stats[(int)result]++;
+			}
+			break;
+		case ZCompareMode::Always:
+			for (int i = 0; i < count; i += 3) {
+				TriangleResult result = DepthRasterTriangle<ZCompareMode::Always>(depth, depthStride, x1, y1, x2, y2, &tx[i], &ty[i], &tz[i]);
+				stats[(int)result]++;
+			}
+			break;
 		}
 		gpuStats.numDepthRasterBackface += stats[(int)TriangleResult::Backface];
 		gpuStats.numDepthRasterNoPixels += stats[(int)TriangleResult::NoPixels];

assets/compat.ini

@@ -1636,21 +1636,21 @@ ULJM05216 = true
 NPJH50181 = true
 
 [SoftwareRasterDepth]
-# Midnight Club: LA Remix
+# Midnight Club: LA Remix (see #18625)
 ULUS10383 = true
 ULES01144 = true
 ULJS00180 = true
 ULJS00267 = true
 ULJM05904 = true
 NPJH50440 = true
 
-# Syphon Filter - Dark Mirror
+# Syphon Filter - Dark Mirror (light flares, see #10229)
 UCES00310 = true
 UCUS98641 = true
 UCUS98656 = true
 UCUS98656 = true  # Demo
 
-# Syphon Filter - Logan's Shadow
+# Syphon Filter - Logan's Shadow (light flares, see #10229)
 UCUS98606 = true
 UCES00710 = true
 NPUG80173 = true
@@ -1663,7 +1663,7 @@ SYPH04036 = true  # Prototype?
 NPUG80114 = true
 NPEG00004 = true
 
-# Wipeout Pure
+# Wipeout Pure (see #13344, sun lens flare)
 UCUS98612 = true
 UCJS10007 = true
 UCES00001 = true
@@ -1675,7 +1675,14 @@ UCES01184 = true
 UCUS98668 = true
 UCJP00174 = true
 
+# Armored Core: Silent Line (see #17597)
+ULJM05552 = true
+UCAS40289 = true
+NPUH10025 = true
+NPEH00047 = true
+
 [ReadbackDepth]
+# Obsolete, SoftwareRasterDepth is a far better solution.
 
 [BlockTransferDepth]
 # Iron Man - see issue #16530