More NEON

GPU/Common/DrawEngineCommon.cpp

@@ -442,11 +442,12 @@ bool DrawEngineCommon::TestBoundingBoxFast(const void *vdata, int vertexCount, u
 			__m128 pos = _mm_mul_ps(_mm_cvtepi32_ps(bits), scaleFactor);
 			_mm_storeu_ps(verts + i * 3, pos);  // TODO: use stride 4 to avoid clashing writes?
 		}
-#elif PPSSPP_ARCH(ARM_NEON)
+#elif 0 && PPSSPP_ARCH(ARM_NEON)
+		__m128 scaleFactor = vdupq_n_f32(1.0f / 32768.0f);
 		for (int i = 0; i < vertexCount; i++) {
 			const int16_t *dataPtr = ((const int16_t *)((const s8 *)vdata + i * stride + offset));
 			int32x4_t data = vmovl_s16(vld1_s16(dataPtr));
-			float32x4_t pos = vcvtq_n_s32_f32(data, 15);  // This does the division by 32768.0f, effectively.
+			float32x4_t pos = vmulq_f32(scaleFactor, vcvtq_s32_f32(data));  // This does the division by 32768.0f, effectively.
 			vst1q_f32(verts + i * 3, pos);
 		}
 #else
@@ -522,28 +523,26 @@ bool DrawEngineCommon::TestBoundingBoxFast(const void *vdata, int vertexCount, u
 	uint32x4_t inside = vdupq_n_u32(0);
 	for (int i = 0; i < vertexCount; i++) {
 		const float *pos = verts + i * vertStride;
+		float32x4_t objpos = vld1q_f32(pos);
 		float32x4_t worldpos = vaddq_f32(
 			vaddq_f32(
-				vmulq_f32(worldX, vdupq_n_u32(pos[0])),
-				vmulq_f32(worldY, vdupq_n_u32(pos[1]))
+				vmulq_laneq_f32(worldX, objpos, 0),
+				vmulq_laneq_f32(worldY, objpos, 1)
 			),
 			vaddq_f32(
-				vmulq_f32(worldZ, vdupq_n_u32(pos[2])),
+				vmulq_laneq_f32(worldZ, objpos, 2),
 				worldW
 			)
 		);
 		// OK, now we check it against the four planes.
 		// This is really curiously similar to a matrix multiplication (well, it is one).
-		float32x4_t posX = vdupq_laneq_f32(worldpos, 0);
-		float32x4_t posY = vdupq_laneq_f32(worldpos, 1);
-		float32x4_t posZ = vdupq_laneq_f32(worldpos, 2);
 		float32x4_t planeDist = vaddq_f32(
 			vaddq_f32(
-				vmulq_f32(planeX, posX),
-				vmulq_f32(planeY, posY)
+				vmulq_laneq_f32(planeX, worldpos, 0),
+				vmulq_laneq_f32(planeY, worldpos, 1)
 			),
 			vaddq_f32(
-				vmulq_f32(planeZ, posZ),
+				vmulq_laneq_f32(planeZ, worldpos, 2),
 				planeW
 			)
 		);