shadertoy_circle.fs.asm : use presubtract to reduce instruction count

drm/shadertoy_circle.fs.asm

@@ -1,29 +1,23 @@
-# CONSTS[0] = (-0.1, 0, 0, 0)
+# CONST[0] = (-0.1, 0, 0, 0)
 
 # d = length(uv)
-src0.rgb = temp[0]  :
+src0.rgb = temp[0] :
   temp[0].r    = DP3 src0.rg0 src0.rg0 ;
-src0.rgb = temp[0]  :
+src0.rgb = temp[0] :
   temp[0].a    = RSQ |src0.r| ;
-src0.a   = temp[0]  :
+src0.a   = temp[0] :
   temp[0].a    = RCP src0.a ;
 
-# d -= 0.5
+# d = abs(d - 0.5) * 1 + -0.1
-src0.a   = temp[0]  :
+src0.rgb = float(48), # 0.5
-  temp[0].r    = MAD src0.a00 src0.100 -src0.h00 ;
+src1.rgb = temp[0],   # d
+src2.rgb = const[0],  # -0.1
+srcp.rgb = sub :      # (src1.rgb - src0.rgb)
+  temp[0].r    = MAD |srcp.r00| src0.100 src2.r00 ;
 
-# d = abs(d) * 1 + -0.1
+# d = (d >= 0.0) ? 1.0 : 0.0
-src0.rgb = temp[0] , src1.rgb = const[0] :
+# out.rgba = vec4(d, 0, 0, 1)
-  temp[0].r    = MAD |src0.r00| src0.100 src1.r00 ;
-
-# out.r = (d >= 0.0) ? 1.0 : 0.0
-OUT
-src0.rgb = temp[0]  :
-  out[0].r    = CMP src0.100 src0.000 src0.r00 ;
-
-# out.a = 1
-# out.gb = vec2(0, 0)
 OUT TEX_SEM_WAIT
- :
+src0.rgb = temp[0] :
   out[0].a    = MAX src0.1 src0.1 ,
-  out[0].gb   = MAX src0.000 src0.000 ;
+  out[0].rgb  = CMP src0.100 src0.000 src0.r00 ;

drm/shadertoy_circle.fs.inc

@@ -20,30 +20,9 @@
 0x00000000,
 
 0x00000800,
+0x500000b0,
 0x08020080,
-0x08020000,
+0x00931483,
-0x0093048c,
 0x00000000,
-0x00c94000,
+0x00482000,
-
-0x00000800,
-0x08040000,
-0x08020080,
-0x00931480,
-0x00000000,
-0x00481000,
-
-0x00008001,
-0x08020000,
-0x08020080,
-0x00920498,
-0x00000000,
-0x00480008,
-
-0x00070005,
-0x08020080,
-0x08020080,
-0x00920490,
-0x00c18003,
-0x00000005,
 

regs/us_disassemble2.py

@@ -1,11 +1,11 @@
 import sys
-from os import path
+from os import path, environ
 import parse_bits
 from collections import OrderedDict
 from functools import partial
 from pprint import pprint
 
-VERBOSE = False
+VERBOSE = environ.get("VERBOSE", "false").lower() == "true"
 
 class BaseRegister:
     def get(self, code, *, code_ix, descriptor):
@@ -324,7 +324,13 @@ def disassemble_alu(code, is_output):
         rgb_swizzle_sel = rgb_swizzle_sel[:rgb_op_operands]
 
         a_sources = set(a_sels)
+        if 3 in a_sources:
+            a_sources.add(0)
+            a_sources.add(1)
         rgb_sources = set(rgb_sels)
+        if 3 in rgb_sources:
+            rgb_sources.add(0)
+            rgb_sources.add(1)
         a_addr_strs = [s for i, s in enumerate(a_addr_strs) if i in a_sources]
         rgb_addr_strs = [s for i, s in enumerate(rgb_addr_strs) if i in rgb_sources]