Scalarize error after loop in realign_weights_decimated

solidpixel · solidpixel · commit 7b1b3bcb4663 · 2022-04-08T23:10:14.000+01:00
diff --git a/Source/astcenc_block_sizes.cpp b/Source/astcenc_block_sizes.cpp
@@ -953,7 +953,7 @@ static void construct_block_size_descriptor_2d(
 
 			if (is_dual_plane)
 			{
-				dm.ref_2_planes= 1;
+				dm.ref_2_planes = 1;
 			}
 			else
 			{
@@ -1081,7 +1081,7 @@ static void construct_block_size_descriptor_3d(
 				bsd.decimation_modes[decimation_mode_count].maxprec_1plane = static_cast<int8_t>(maxprec_1plane);
 				bsd.decimation_modes[decimation_mode_count].maxprec_2planes = static_cast<int8_t>(maxprec_2planes);
 				bsd.decimation_modes[decimation_mode_count].ref_1_plane = maxprec_1plane == -1 ? 0 : 1;
-				bsd.decimation_modes[decimation_mode_count].ref_2_planes= maxprec_2planes == -1 ? 0 : 1;
+				bsd.decimation_modes[decimation_mode_count].ref_2_planes = maxprec_2planes == -1 ? 0 : 1;
 				decimation_mode_count++;
 			}
 		}
diff --git a/Source/astcenc_compress_symbolic.cpp b/Source/astcenc_compress_symbolic.cpp
@@ -318,19 +318,20 @@ static bool realign_weights_decimated(
 				vfloat4 color = color_base + color_offset * plane_weight;
 
 				vfloat4 orig_color   = blk.texel(texel);
-				vfloat4 error_weight = blk.channel_weight;
 
 				vfloat4 color_diff      = color - orig_color;
 				vfloat4 color_up_diff   = color_diff + color_offset * plane_up_weight;
 				vfloat4 color_down_diff = color_diff + color_offset * plane_down_weight;
-				current_errorv += dot(color_diff      * color_diff,      error_weight);
-				up_errorv      += dot(color_up_diff   * color_up_diff,   error_weight);
-				down_errorv    += dot(color_down_diff * color_down_diff, error_weight);
+
+				current_errorv += color_diff * color_diff;
+				up_errorv      += color_up_diff * color_up_diff;
+				down_errorv    += color_down_diff * color_down_diff;
 			}
 
-			float current_error = current_errorv.lane<0>();
-			float up_error = up_errorv.lane<0>();
-			float down_error = down_errorv.lane<0>();
+			vfloat4 error_weight = blk.channel_weight;
+			float current_error = hadd_s(current_errorv * error_weight);
+			float up_error = hadd_s(up_errorv * error_weight);
+			float down_error = hadd_s(down_errorv * error_weight);
 
 			// Check if the prev or next error is better, and if so use it
 			if ((up_error < current_error) && (up_error < down_error))

Original file line number	Diff line number	Diff line change
`@@ -953,7 +953,7 @@ static void construct_block_size_descriptor_2d(`
`953`	`953`
`954`	`954`	`if (is_dual_plane)`
`955`	`955`	`{`
`956`		`- dm.ref_2_planes= 1;`
	`956`	`+ dm.ref_2_planes = 1;`
`957`	`957`	`}`
`958`	`958`	`else`
`959`	`959`	`{`
`@@ -1081,7 +1081,7 @@ static void construct_block_size_descriptor_3d(`
`1081`	`1081`	`bsd.decimation_modes[decimation_mode_count].maxprec_1plane = static_cast<int8_t>(maxprec_1plane);`
`1082`	`1082`	`bsd.decimation_modes[decimation_mode_count].maxprec_2planes = static_cast<int8_t>(maxprec_2planes);`
`1083`	`1083`	`bsd.decimation_modes[decimation_mode_count].ref_1_plane = maxprec_1plane == -1 ? 0 : 1;`
`1084`		`- bsd.decimation_modes[decimation_mode_count].ref_2_planes= maxprec_2planes == -1 ? 0 : 1;`
	`1084`	`+ bsd.decimation_modes[decimation_mode_count].ref_2_planes = maxprec_2planes == -1 ? 0 : 1;`
`1085`	`1085`	`decimation_mode_count++;`
`1086`	`1086`	`}`
`1087`	`1087`	`}`