wkhere / fractx / 37864715501

56%

push

github

fix arm64 differencies by disabling fused FMADD

Note: FMADD is actually doing less rounding (only add the end of
mult+add), while separate mult & add each with rounding give less
accurate results. But, the issue is that the output differs
on different architectures and that's a no for now.

Here are the assembly diffs:

-               cy := f.Bounds.Y0 + float64(y)*dy
+               cy := f.Bounds.Y0 + float64(float64(y)*dy)
   0x3ee94              fd400c02                FMOVD 24(R0), F2
   0x3ee98              9e620063                SCVTFD R3, F3
-  0x3ee9c              1f430822                FMADDD F3, F2, F1, F2
-  0x3eea0              fd002be2                FMOVD F2, 80(RSP)
-  0x3eea4              d2800004                MOVD $0, R4
+  0x3ee9c              1e610863                FMULD F1, F3, F3
+  0x3eea0              1e632842                FADDD F3, F2, F2
+  0x3eea4              fd002be2                FMOVD F2, 80(RSP)
+  0x3eea8              d2800004                MOVD $0, R4

                for x := 0; x < f.Size.W; x++ {
-  0x3ef00              f9400005                MOVD (R0), R5
-  0x3ef04              eb05009f                CMP R5, R4
-  0x3ef08              54fffbca                BGE -34(PC)
-                       i := iter(f.Bounds.X0+float64(x)*dx, cy, f.MaxI)
-  0x3ef0c              fd400803                FMOVD 16(R0), F3
-  0x3ef10              9e620084                SCVTFD R4, F4
-  0x3ef14              1f440c03                FMADDD F4, F3, F0, F3
-  0x3ef18              f9401805                MOVD 48(R0), R5
-  0x3ef1c              1e604064                FMOVD F3, F4
-               cy := f.Bounds.Y0 + float64(y)*dy
-  0x3ef20              1e604045                FMOVD F2, F5
-  0x3ef24              d2800006                MOVD $0, R6
+  0x3ef04              f9400005                MOVD (R0), R5
+  0x3ef08              eb05009f                CMP R5, R4
+  0x3ef0c              54fffbaa                BGE -35(P... (continued)

4 of 4 new or added lines in 1 file covered. (100.0%)

97 of 172 relevant lines covered (56.4%)

0.62 hits per line