Human-based optimization of LLM-provided code: +10% in rendition performance

Author: tyoma

src/agge/blenders_arm.cpp

@@ -38,45 +38,21 @@ namespace agge
 
 		void blender_solid_color::blend4(pixel *pixels, uint16x8_t color_u16, uint16x8_t alpha_u16, unsigned int covers_packed)
 		{
-			uint8x8_t covers = vdup_n_u8(0);
-			covers = vset_lane_u8(static_cast<uint8_t>(covers_packed >> 0), covers, 0);
-			covers = vset_lane_u8(static_cast<uint8_t>(covers_packed >> 8), covers, 1);
-			covers = vset_lane_u8(static_cast<uint8_t>(covers_packed >> 16), covers, 2);
-			covers = vset_lane_u8(static_cast<uint8_t>(covers_packed >> 24), covers, 3);
+			const uint16x4_t covers_u16 = vget_low_u16(vshlq_n_u16(vmovl_u8(vcreate_u8(covers_packed)), 8));
+			const int16x4_t alpha = vreinterpret_s16_u16(vshrn_n_u32(vmull_u16(covers_u16, vget_low_u16(alpha_u16)), 16));
 
-			const uint16x4_t covers_u16 = vget_low_u16(vshlq_n_u16(vmovl_u8(covers), 8));
-			const uint16x4_t alpha = vshrn_n_u32(vmull_u16(covers_u16, vget_low_u16(alpha_u16)), 16);
-
-			const uint16_t a0 = vget_lane_u16(alpha, 0);
-			const uint16_t a1 = vget_lane_u16(alpha, 1);
-			const uint16_t a2 = vget_lane_u16(alpha, 2);
-			const uint16_t a3 = vget_lane_u16(alpha, 3);
-
-			const int16x8_t alpha10 = vreinterpretq_s16_u16(vcombine_u16(vdup_n_u16(a0), vdup_n_u16(a1)));
-			const int16x8_t alpha32 = vreinterpretq_s16_u16(vcombine_u16(vdup_n_u16(a2), vdup_n_u16(a3)));
+			const int16x8_t alpha10 = vcombine_s16(vdup_lane_s16(alpha, 0), vdup_lane_s16(alpha, 1));
+			const int16x8_t alpha32 = vcombine_s16(vdup_lane_s16(alpha, 2), vdup_lane_s16(alpha, 3));
 
 			const uint8x16_t source_u8 = vld1q_u8(reinterpret_cast<uint8_t *>(pixels));
-			const int16x8_t source10 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(source_u8)));
-			const int16x8_t source32 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(source_u8)));
-
-			const int16x8_t d10 = vshlq_n_s16(vsubq_s16(source10, vreinterpretq_s16_u16(color_u16)), 2);
-			const int16x8_t d32 = vshlq_n_s16(vsubq_s16(source32, vreinterpretq_s16_u16(color_u16)), 2);
-
-			const int32x4_t m10l = vmull_s16(vget_low_s16(d10), vget_low_s16(alpha10));
-			const int32x4_t m10h = vmull_s16(vget_high_s16(d10), vget_high_s16(alpha10));
-			const int32x4_t m32l = vmull_s16(vget_low_s16(d32), vget_low_s16(alpha32));
-			const int32x4_t m32h = vmull_s16(vget_high_s16(d32), vget_high_s16(alpha32));
-
-			const int16x8_t c10 = vcombine_s16(vshrn_n_s32(m10l, 16), vshrn_n_s32(m10h, 16));
-			const int16x8_t c32 = vcombine_s16(vshrn_n_s32(m32l, 16), vshrn_n_s32(m32h, 16));
-
-			const int16x8_t r10 = vsubq_s16(source10, c10);
-			const int16x8_t r32 = vsubq_s16(source32, c32);
+			int16x8_t source10 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(source_u8)));
+			int16x8_t source32 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(source_u8)));
 
-			const uint8x8_t result10 = vqmovun_s16(r10);
-			const uint8x8_t result32 = vqmovun_s16(r32);
+			// source -= ((source - color) << 2 * alpha) >> 16
+			source10 = vsubq_s16(source10, vqdmulhq_s16(vshlq_n_s16(vsubq_s16(source10, vreinterpretq_s16_u16(color_u16)), 1), alpha10));
+			source32 = vsubq_s16(source32, vqdmulhq_s16(vshlq_n_s16(vsubq_s16(source32, vreinterpretq_s16_u16(color_u16)), 1), alpha32));
 
-			vst1q_u8(reinterpret_cast<uint8_t *>(pixels), vcombine_u8(result10, result32));
+			vst1q_u8(reinterpret_cast<uint8_t *>(pixels), vcombine_u8(vqmovun_s16(source10), vqmovun_s16(source32)));
 		}
 	}
 }