fix(asm/wasm): oscillator phase was causing rounding errors once large enough

gopher had fixed this, but we foolishly removed it. reintroducing fix, although this could be optional only for those who really care. ultimate size optimizers could still want to get rid of it.

templates/amd64-386/sources.asm

@@ -172,20 +172,31 @@ su_op_oscillat_normalize_note:
     fmul    dword [{{.Float 0.000092696138 | .Use}}]   ; // st0 is now frequency
 su_op_oscillat_normalized:
     fadd    dword [{{.WRK}}]
-    fst     dword [{{.WRK}}]
-    fadd    dword [{{.Input "oscillator" "phase"}}]
 {{- if .SupportsParamValue "oscillator" "type" .Sample}}
     test    al, byte 0x80
     jz      short su_op_oscillat_not_sample
+    fst     dword [{{.WRK}}]  ; for samples, we store the phase without mod(p,1)
+{{- if or (.SupportsParamValueOtherThan "oscillator" "phase" 0) (.SupportsModulation "oscillator" "phase")}}
+    fadd    dword [{{.Input "oscillator" "phase"}}]
+{{- end}}
     {{.Call "su_oscillat_sample"}}
     jmp     su_op_oscillat_shaping ; skip the rest to avoid color phase normalization and colorloading
 su_op_oscillat_not_sample:
 {{- end}}
-    fld1
-    fadd    st1, st0
+    fld1                     ; we need to take mod(p,1) so the frequency does not drift as the float
+    fadd    st1, st0         ; make no mistake: without this, there is audible drifts in oscillator pitch
+    fxch                     ; as the actual period changes once the phase becomes too big
+    fprem                    ; we actually computed mod(p+1,1) instead of mod(p,1) as the fprem takes mod
+    fstp    st1              ; towards zero
+    fst     dword [{{.WRK}}] ; store back the updated phase
+{{- if or (.SupportsParamValueOtherThan "oscillator" "phase" 0) (.SupportsModulation "oscillator" "phase")}}
+    fadd    dword [{{.Input "oscillator" "phase"}}]
+    fld1                    ; this is a bit stupid, but we need to take mod(x,1) again after phase modulations
+    fadd    st1, st0        ; as the actual oscillator functions expect x in [0,1]
     fxch
     fprem
     fstp    st1
+{{- end}}
     fld     dword [{{.Input "oscillator" "color"}}]               ; // c      p
     ; every oscillator test included if needed
 {{- if .SupportsParamValue "oscillator" "type" .Sine}}

templates/wasm/sources.wat

@@ -121,27 +121,34 @@
 {{- end}}
     (f32.store ;; update phase
         (global.get $WRK)
-        ;; Transpose calculation starts
-        (f32.div
-            (call $inputSigned (i32.const {{.InputNumber "oscillator" "transpose"}}))
-            (f32.const 0.015625)
-        ) ;; scale back to 0 - 128
-        (f32.add (local.get $detune)) ;; add detune. detune is -1 to 1 so can detune a full note up or down at max
-        (f32.add (select
-            (f32.const 0)
-            (f32.convert_i32_u (i32.load (global.get $voice)))
-            (i32.and (local.get $flags) (i32.const 0x8))
-        ))  ;; if lfo is not enabled, add the note number to it
-        (f32.mul (f32.const 0.0833333)) ;; /12, in full octaves
-        (call $pow2)
-        (f32.mul (select
-            (f32.const 0.000038) ;; pretty random scaling constant to get LFOs into reasonable range. Historical reasons, goes all the way back to 4klang
-            (f32.const 0.000092696138) ;; scaling constant to get middle-C to where it should be
-            (i32.and (local.get $flags) (i32.const 0x8))
-        ))
-        (f32.add (f32.load (global.get $WRK))) ;; add the current phase of the oscillator
+        (local.tee $phase
+            (f32.sub
+                (local.tee $phase
+                    ;; Transpose calculation starts
+                    (f32.div
+                        (call $inputSigned (i32.const {{.InputNumber "oscillator" "transpose"}}))
+                        (f32.const 0.015625)
+                    ) ;; scale back to 0 - 128
+                    (f32.add (local.get $detune)) ;; add detune. detune is -1 to 1 so can detune a full note up or down at max
+                    (f32.add (select
+                        (f32.const 0)
+                        (f32.convert_i32_u (i32.load (global.get $voice)))
+                        (i32.and (local.get $flags) (i32.const 0x8))
+                    ))  ;; if lfo is not enabled, add the note number to it
+                    (f32.mul (f32.const 0.0833333)) ;; /12, in full octaves
+                    (call $pow2)
+                    (f32.mul (select
+                        (f32.const 0.000038) ;; pretty random scaling constant to get LFOs into reasonable range. Historical reasons, goes all the way back to 4klang
+                        (f32.const 0.000092696138) ;; scaling constant to get middle-C to where it should be
+                        (i32.and (local.get $flags) (i32.const 0x8))
+                    ))
+                    (f32.add (f32.load (global.get $WRK))) ;; add the current phase of the oscillator
+                )
+                (f32.floor (local.get $phase))
+            )
+        )
     )
-    (f32.add (f32.load (global.get $WRK)) (call $input (i32.const {{.InputNumber "oscillator" "phase"}})))
+    (f32.add (local.get $phase) (call $input (i32.const {{.InputNumber "oscillator" "phase"}})))
     (local.set $phase (f32.sub (local.tee $phase) (f32.floor (local.get $phase)))) ;; phase = phase mod 1.0
     (local.set $color (call $input (i32.const {{.InputNumber "oscillator" "color"}})))
 {{- if .SupportsParamValue "oscillator" "type" .Sine}}

tests/wasm_test_renderer.es6

@@ -85,7 +85,7 @@ if (process.argv.length <= 3) {
         }
         errorCount++
     }
-    if (errorCount > 100) {
+    if (errorCount > 200) {
         console.error("got different buffer than expected. First error at: "+(firstErrorPos/2|0)+(firstErrorPos%1," right"," left"));
         return 1;
     }