shadertoy: implement shadertoy_palette_fractal

drm/shadertoy.c

@@ -304,18 +304,18 @@ int indirect_buffer(float time)
       | SC_SCISSOR0__YS0(0)
       );
   T0V(SC_SCISSOR1
-      , SC_SCISSOR1__XS1(1600 - 1)
+      , SC_SCISSOR1__XS1(800 - 1)
-      | SC_SCISSOR1__YS1(1200 - 1)
+      | SC_SCISSOR1__YS1(600 - 1)
       );
 
   //////////////////////////////////////////////////////////////////////////////
   // VAP
   //////////////////////////////////////////////////////////////////////////////
 
-  T0Vf(VAP_VPORT_XSCALE,   800.0f);
+  T0Vf(VAP_VPORT_XSCALE,   400.0f);
-  T0Vf(VAP_VPORT_XOFFSET,  800.0f);
+  T0Vf(VAP_VPORT_XOFFSET,  400.0f);
-  T0Vf(VAP_VPORT_YSCALE,  -600.0f);
+  T0Vf(VAP_VPORT_YSCALE,  -300.0f);
-  T0Vf(VAP_VPORT_YOFFSET,  600.0f);
+  T0Vf(VAP_VPORT_YOFFSET,  300.0f);
   T0Vf(VAP_VPORT_ZSCALE,     0.5f);
   T0Vf(VAP_VPORT_ZOFFSET,    0.5f);
 
@@ -409,7 +409,7 @@ int indirect_buffer(float time)
   };
   const int vertex_shader_length = (sizeof (vertex_shader)) / (sizeof (vertex_shader[0]));
   assert(vertex_shader_length % 4 == 0);
-  printf("vs length %d\n", vertex_shader_length);
+  //printf("vs length %d\n", vertex_shader_length);
 
   T0_ONE_REG(VAP_PVS_VECTOR_DATA_REG_128, vertex_shader_length - 1);
   for (int i = 0; i < vertex_shader_length; i++) {
@@ -449,10 +449,17 @@ int indirect_buffer(float time)
 
   // fragment constants
 
+#define PI (3.14159274101257324219f)
+#define PI_2 (PI * 2.0f)
+#define I_PI_2 (1.0f / (PI_2))
+
   const float fragment_consts[] = {
-    time, 0, 0, 0,
+    time, 1.2, 0.01, 0.4,
+    PI_2, I_PI_2, 0, 0,
+    0.25, 0.40625, 0.5625, 0,
   };
   int fragment_consts_length = (sizeof (fragment_consts)) / (sizeof (fragment_consts[0]));
+  assert(fragment_consts_length % 4 == 0);
 
   T0V(GA_US_VECTOR_INDEX
       , GA_US_VECTOR_INDEX__INDEX(0)
@@ -465,16 +472,16 @@ int indirect_buffer(float time)
   // fragment code
 
   const uint32_t fragment_shader[] = {
-    #include "shadertoy_palette.fs.inc"
+    #include "shadertoy_palette_fractal.fs.inc"
   };
   const int fragment_shader_length = (sizeof (fragment_shader)) / (sizeof (fragment_shader[0]));
   assert(fragment_shader_length % 6 == 0);
-  printf("fs length %d\n", fragment_shader_length);
+  //printf("fs length %d\n", fragment_shader_length);
   const int fragment_shader_instructions = fragment_shader_length / 6;
-  printf("fs instructions %d\n", fragment_shader_instructions);
+  //printf("fs instructions %d\n", fragment_shader_instructions);
 
   T0V(US_PIXSIZE
-      , US_PIXSIZE__PIX_SIZE(2) // pixel shader stack frame size
+      , US_PIXSIZE__PIX_SIZE(3) // pixel shader stack frame size
       );
 
   T0V(US_CODE_RANGE
@@ -513,7 +520,7 @@ int indirect_buffer(float time)
     -1.0f,  1.0f, 0.0f
   };
   const int vertices_length = (sizeof (vertices)) / (sizeof (vertices[0]));
-  printf("vtx length %d\n", vertices_length);
+  //printf("vtx length %d\n", vertices_length);
   T3(_3D_DRAW_IMMD_2, (1 + vertices_length) - 1);
   ib[ix++].u32
     = VAP_VF_CNTL__PRIM_TYPE(4)
@@ -708,7 +715,7 @@ int main()
 #define D1GRPH_UPDATE__D1GRPH_SURFACE_UPDATE_PENDING (1 << 2)
 
     uint32_t d1crtc_double_buffer_control = rreg(rmmio, D1CRTC_DOUBLE_BUFFER_CONTROL);
-    printf("D1CRTC_DOUBLE_BUFFER_CONTROL: %08x\n", d1crtc_double_buffer_control);
+    //printf("D1CRTC_DOUBLE_BUFFER_CONTROL: %08x\n", d1crtc_double_buffer_control);
     assert(d1crtc_double_buffer_control == (1 << 8));
 
     // addresses were retrieved from /sys/kernel/debug/radeon_vram_mm

drm/shadertoy_palette_fractal.fs.asm

@@ -0,0 +1,37 @@
+-- CONST[0] = { time, 1.2, 0.01, 0.4 }
+-- CONST[1] = { PI_2, I_PI_2, 0, 0 },
+-- CONST[2] = { 0.25, 0.40625, 0.5625, 0 },
+
+-- temp[0] : { uv0.xy , _, l }
+-- temp[1] : { uv.xy  , _, d }
+-- temp[2] : final_color.xyzw
+-- temp[3] : {col.xyz , i }
+
+-- vec2 uv = uv0; // temp[1]
+src0.rgb = temp[0] : -- uv0
+  temp[1].rg  = MAX src0.rg_ src0.rg_ ;
+
+-- vec4 final_color = vec4(0, 0, 0, 1);
+:
+  temp[2].a   = MAX src0.1 src0.1 ,
+  temp[2].rgb = MAX src0.000 src0.000 ;
+
+-- i = 0;
+:
+  temp[3].a   = MAX src0.0 src0.0 ;
+
+--------------------------------------------------------------------------------
+-- loop start
+--------------------------------------------------------------------------------
+#include "shadertoy_palette_fractal_loop_inner.fs.asm"
+#include "shadertoy_palette_fractal_loop_inner.fs.asm"
+#include "shadertoy_palette_fractal_loop_inner.fs.asm"
+#include "shadertoy_palette_fractal_loop_inner.fs.asm"
+--------------------------------------------------------------------------------
+-- loop end
+--------------------------------------------------------------------------------
+
+OUT TEX_SEM_WAIT
+src0.rgb = temp[2] :
+  out[0].a     = MAX src0.1 src0.1 ,
+  out[0].rgb   = MAX src0.rgb src0.rgb ;

drm/shadertoy_palette_fractal.fs.inc

@@ -0,0 +1,924 @@
+0x00001800,
+0x08020000,
+0x08020080,
+0x00e40720,
+0x00000000,
+0x00000015,
+
+0x00007800,
+0x08020080,
+0x08020080,
+0x00920490,
+0x00c18023,
+0x00000025,
+
+0x00004000,
+0x08020080,
+0x08020080,
+0x00000000,
+0x00810033,
+0x00000000,
+
+0x00001800,
+0x08020001,
+0x080200bc,
+0x00ed8720,
+0x00000000,
+0x00790010,
+
+0x00001800,
+0x08020001,
+0x08020080,
+0x00000720,
+0x00000000,
+0x00000019,
+
+0x00001800,
+0x08020001,
+0x0802c080,
+0x00fb0720,
+0x00000000,
+0x00f6d010,
+
+0x00004000,
+0x08020000,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x00300100,
+0x00000000,
+0x0060e010,
+0x1a000000,
+
+0x00004000,
+0x08020100,
+0x08000500,
+0x00000000,
+0x00600010,
+0x1a000000,
+
+0x00003800,
+0x8802c102,
+0x08020001,
+0x006d86d8,
+0x00000000,
+0x00223030,
+
+0x00003800,
+0x08020003,
+0x08020080,
+0x00000220,
+0x00000000,
+0x00000039,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x004406d8,
+0x00000000,
+0x00a21030,
+
+0x00000800,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000000d,
+0x0000003a,
+
+0x00001000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000400d,
+0x0000003a,
+
+0x00002000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000800d,
+0x0000003a,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x00440221,
+0x00000000,
+0x00221030,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0002c018,
+0x00000000,
+
+0x00004000,
+0x08020001,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08000400,
+0x00000000,
+0x0068c010,
+0x20000000,
+
+0x00004000,
+0x10020080,
+0x08034001,
+0x00000000,
+0x0068c010,
+0x04000000,
+
+0x00004000,
+0x08040480,
+0x0b020001,
+0x00000000,
+0x0028c010,
+0x1c000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c017,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x0802c001,
+0x00000000,
+0x00618010,
+0x5a000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x1800c01c,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0004c01a,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0048c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c019,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0028c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c018,
+0x00000000,
+
+0x00003800,
+0x00220003,
+0x08000480,
+0x006da220,
+0x00000000,
+0x00222020,
+
+0x00004000,
+0x08020080,
+0x08020003,
+0x00000000,
+0x00618030,
+0x30000000,
+
+0x00001800,
+0x08020001,
+0x080200bc,
+0x00ed8720,
+0x00000000,
+0x00790010,
+
+0x00001800,
+0x08020001,
+0x08020080,
+0x00000720,
+0x00000000,
+0x00000019,
+
+0x00001800,
+0x08020001,
+0x0802c080,
+0x00fb0720,
+0x00000000,
+0x00f6d010,
+
+0x00004000,
+0x08020000,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x00300100,
+0x00000000,
+0x0060e010,
+0x1a000000,
+
+0x00004000,
+0x08020100,
+0x08000500,
+0x00000000,
+0x00600010,
+0x1a000000,
+
+0x00003800,
+0x8802c102,
+0x08020001,
+0x006d86d8,
+0x00000000,
+0x00223030,
+
+0x00003800,
+0x08020003,
+0x08020080,
+0x00000220,
+0x00000000,
+0x00000039,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x004406d8,
+0x00000000,
+0x00a21030,
+
+0x00000800,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000000d,
+0x0000003a,
+
+0x00001000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000400d,
+0x0000003a,
+
+0x00002000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000800d,
+0x0000003a,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x00440221,
+0x00000000,
+0x00221030,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0002c018,
+0x00000000,
+
+0x00004000,
+0x08020001,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08000400,
+0x00000000,
+0x0068c010,
+0x20000000,
+
+0x00004000,
+0x10020080,
+0x08034001,
+0x00000000,
+0x0068c010,
+0x04000000,
+
+0x00004000,
+0x08040480,
+0x0b020001,
+0x00000000,
+0x0028c010,
+0x1c000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c017,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x0802c001,
+0x00000000,
+0x00618010,
+0x5a000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x1800c01c,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0004c01a,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0048c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c019,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0028c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c018,
+0x00000000,
+
+0x00003800,
+0x00220003,
+0x08000480,
+0x006da220,
+0x00000000,
+0x00222020,
+
+0x00004000,
+0x08020080,
+0x08020003,
+0x00000000,
+0x00618030,
+0x30000000,
+
+0x00001800,
+0x08020001,
+0x080200bc,
+0x00ed8720,
+0x00000000,
+0x00790010,
+
+0x00001800,
+0x08020001,
+0x08020080,
+0x00000720,
+0x00000000,
+0x00000019,
+
+0x00001800,
+0x08020001,
+0x0802c080,
+0x00fb0720,
+0x00000000,
+0x00f6d010,
+
+0x00004000,
+0x08020000,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x00300100,
+0x00000000,
+0x0060e010,
+0x1a000000,
+
+0x00004000,
+0x08020100,
+0x08000500,
+0x00000000,
+0x00600010,
+0x1a000000,
+
+0x00003800,
+0x8802c102,
+0x08020001,
+0x006d86d8,
+0x00000000,
+0x00223030,
+
+0x00003800,
+0x08020003,
+0x08020080,
+0x00000220,
+0x00000000,
+0x00000039,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x004406d8,
+0x00000000,
+0x00a21030,
+
+0x00000800,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000000d,
+0x0000003a,
+
+0x00001000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000400d,
+0x0000003a,
+
+0x00002000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000800d,
+0x0000003a,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x00440221,
+0x00000000,
+0x00221030,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0002c018,
+0x00000000,
+
+0x00004000,
+0x08020001,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08000400,
+0x00000000,
+0x0068c010,
+0x20000000,
+
+0x00004000,
+0x10020080,
+0x08034001,
+0x00000000,
+0x0068c010,
+0x04000000,
+
+0x00004000,
+0x08040480,
+0x0b020001,
+0x00000000,
+0x0028c010,
+0x1c000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c017,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x0802c001,
+0x00000000,
+0x00618010,
+0x5a000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x1800c01c,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0004c01a,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0048c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c019,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0028c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c018,
+0x00000000,
+
+0x00003800,
+0x00220003,
+0x08000480,
+0x006da220,
+0x00000000,
+0x00222020,
+
+0x00004000,
+0x08020080,
+0x08020003,
+0x00000000,
+0x00618030,
+0x30000000,
+
+0x00001800,
+0x08020001,
+0x080200bc,
+0x00ed8720,
+0x00000000,
+0x00790010,
+
+0x00001800,
+0x08020001,
+0x08020080,
+0x00000720,
+0x00000000,
+0x00000019,
+
+0x00001800,
+0x08020001,
+0x0802c080,
+0x00fb0720,
+0x00000000,
+0x00f6d010,
+
+0x00004000,
+0x08020000,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x00300100,
+0x00000000,
+0x0060e010,
+0x1a000000,
+
+0x00004000,
+0x08020100,
+0x08000500,
+0x00000000,
+0x00600010,
+0x1a000000,
+
+0x00003800,
+0x8802c102,
+0x08020001,
+0x006d86d8,
+0x00000000,
+0x00223030,
+
+0x00003800,
+0x08020003,
+0x08020080,
+0x00000220,
+0x00000000,
+0x00000039,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x004406d8,
+0x00000000,
+0x00a21030,
+
+0x00000800,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000000d,
+0x0000003a,
+
+0x00001000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000400d,
+0x0000003a,
+
+0x00002000,
+0x08020003,
+0x08020080,
+0x00000000,
+0x0000800d,
+0x0000003a,
+
+0x00003800,
+0x0802c003,
+0x08020080,
+0x00440221,
+0x00000000,
+0x00221030,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0002c018,
+0x00000000,
+
+0x00004000,
+0x08020001,
+0x08020080,
+0x00840420,
+0x00000001,
+0x00000001,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0004c00b,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08000400,
+0x00000000,
+0x0068c010,
+0x20000000,
+
+0x00004000,
+0x10020080,
+0x08034001,
+0x00000000,
+0x0068c010,
+0x04000000,
+
+0x00004000,
+0x08040480,
+0x0b020001,
+0x00000000,
+0x0028c010,
+0x1c000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c017,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x0802c001,
+0x00000000,
+0x00618010,
+0x5a000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x1800c01c,
+0x00000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0004c01a,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0048c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c019,
+0x00000000,
+
+0x00004000,
+0x08040080,
+0x08020001,
+0x00000000,
+0x0028c010,
+0x20000000,
+
+0x00004000,
+0x08020080,
+0x08020001,
+0x00000000,
+0x0000c018,
+0x00000000,
+
+0x00003800,
+0x00220003,
+0x08000480,
+0x006da220,
+0x00000000,
+0x00222020,
+
+0x00004000,
+0x08020080,
+0x08020003,
+0x00000000,
+0x00618030,
+0x30000000,
+
+0x00078005,
+0x08020002,
+0x08020080,
+0x00440220,
+0x00c18003,
+0x00000005,
+

drm/shadertoy_palette_fractal_loop_inner.fs.asm

@@ -0,0 +1,142 @@
+-- uv = uv * 1.5;
+src0.rgb = temp[1] , -- uv
+src0.a = float(60) : -- 1.5
+  temp[1].rg  = MAD src0.rg_ src0.aa_ src0.00_ ;
+
+-- uv = fract(uv);
+src0.rgb = temp[1] : -- uv
+  temp[1].rg  = FRC src0.rg_ ;
+
+-- uv = uv - 0.5;
+src0.rgb = temp[1] ,
+src1.a = float(48) : -- 0.5
+  temp[1].rg  = MAD src0.rg_ src0.11_ -src1.aa_ ;
+
+-- l = length(uv0);
+src0.rgb = temp[0] : -- uv0
+                 DP3 src0.rg0 src0.rg0 ,
+  temp[0].a    = DP ;
+src0.a   = temp[0] :
+  temp[0].a    = RSQ |src0.a| ;
+src0.a   = temp[0] :
+  temp[0].a    = RCP src0.a ;
+
+-- d = i * 0.4 + l;
+src0.a = const[0] , -- 0.4
+src1.a = temp[0]  , -- l
+src2.a = temp[3]  : -- i
+  temp[1].a   = MAD src2.a src0.a src1.a ;
+
+-- d = time * 0.4 + d;
+src0.a = const[0]   , -- 0.4
+src1.a = temp[1]    , -- d
+src0.rgb = const[0] : -- time (r)
+  temp[1].a   = MAD src0.r src0.a src1.a ;
+
+--------------------------------------------------------------------------------
+-- start of 'palette' function
+--------------------------------------------------------------------------------
+
+-- v = d + (vec3(0.25, 0.40625, 0.5625) + 0.5)
+src0.a = temp[1]     , -- d
+src0.rgb = const[2]  , -- vec3(0.25, 0.40625, 0.5625)
+src1.rgb = float(48) , -- 0.5
+srcp.rgb = add       : -- (vec3(0.25, 0.40625, 0.5625) + 0.5)
+  temp[3].rgb = MAD src0.111 src0.aaa srcp.rgb ;
+
+-- v = frac(v)
+src0.rgb = temp[3] : -- v
+  temp[3].rgb = FRC src0.rgb ;
+
+src0.rgb = temp[3]   , -- v
+src1.rgb = float(48) : -- 0.5
+  temp[3].rgb = MAD src0.111 src0.rgb -src1.rgb ;
+
+-- v = cos(v)
+src0.rgb = temp[3] : -- v
+              COS src0.r ,
+  temp[3].r = SOP ;
+src0.rgb = temp[3] : -- v
+              COS src0.g ,
+  temp[3].g = SOP ;
+src0.rgb = temp[3] : -- v
+              COS src0.b ,
+  temp[3].b = SOP ;
+
+-- col = vec3(0.5, 0.5, 0.5) * v + vec3(0.5, 0.5, 0.5)
+src0.rgb = temp[3]   , -- v
+src1.rgb = float(48) : -- 0.5
+  temp[3].rgb = MAD src1.rgb src0.rgb src1.rgb;
+
+--------------------------------------------------------------------------------
+-- end of 'palette' function
+--------------------------------------------------------------------------------
+
+-- d = ex2(-l);
+src0.a = temp[0] : -- l
+  temp[1].a   = EX2 -src0.a ;
+
+-- l = length(uv);
+src0.rgb = temp[1] : -- uv
+                 DP3 src0.rg0 src0.rg0 ,
+  temp[0].a    = DP ;
+src0.a   = temp[0] :
+  temp[0].a    = RSQ |src0.a| ;
+src0.a   = temp[0] :
+  temp[0].a    = RCP src0.a ;
+
+-- d = l * d;
+src0.a = temp[0] , -- l
+src1.a = temp[1] : -- d
+  temp[1].a   = MAD src0.a src1.a src0.0 ;
+
+-- d = d * 8.0 + time;
+src0.a   = temp[1]   , -- d
+src1.a   = float(80) , -- 8.0
+src2.rgb = const[0]  : -- time (r)
+  temp[1].a   = MAD src0.a src1.a src2.r ;
+
+-- d = 0.125 * sin(d); <OMOD>
+-- d = d * 0.159154936671257019043 + 0.5; // 48
+src0.a   = temp[1]   , -- d
+src1.rgb = const[1]  , -- I_PI_2 (g)
+src2.a   = float(48) : -- 0.5
+  temp[1].a  = MAD src0.a src1.g src2.a ;
+-- d = fract(d);
+src0.a   = temp[1] : -- d
+  temp[1].a  = FRC src0.a ;
+-- d = d - 0.5;
+src0.a = temp[1]   , -- d
+src1.a = float(48) : -- 0.5
+  temp[1].a = MAD src0.1 src0.a -src1.a ;
+-- d = 0.125 * sin(d * PI_2);
+src0.a = temp[1] :
+  temp[1].a = 0.125 * SIN src0.a ;
+
+-- d = 1.0 / abs(d);
+src0.a = temp[1] : -- d
+  temp[1].a = RCP |src0.a|;
+
+-- d = 0.01 * d;
+src0.a = temp[1]    , -- d
+src1.rgb = const[0] : -- 0.01 (b)
+  temp[1].a = MAD src0.a src1.b src0.0 ;
+
+-- d = pow(d, 1.2);
+src0.a = temp[1] : -- d
+  temp[1].a = LN2 src0.a ;
+src0.a = temp[1] ,
+src1.rgb = const[0] : -- 1.2 (g)
+  temp[1].a = MAD src0.a src1.g src0.0 ;
+src0.a = temp[1] :
+  temp[1].a = EX2 src0.a ;
+
+-- final_color = col * d + final_color
+src0.rgb = temp[3] , -- col
+src1.a   = temp[1] , -- d
+src2.rgb = temp[2] : -- final_color
+  temp[2].rgb = MAD src0.rgb src1.aaa src2.rgb ;
+
+-- i = i + 1
+src0.a = temp[3] :
+  temp[3].a   = MAD src0.1 src0.a src0.1 ;

regs/assembler/error.py

@@ -1,8 +1,10 @@
 import sys
 
-def print_error(filename, buf, e):
+def print_error(e):
     assert len(e.args) == 2, e
     message, token = e.args
+    with open(token.filename, 'rb') as f:
+        buf = f.read()
     lines = buf.splitlines()
     line = lines[token.line - 1]
 
@@ -11,7 +13,7 @@ def print_error(filename, buf, e):
     col_pointer = '^' * len(token.lexeme)
     RED = "\033[0;31m"
     DEFAULT = "\033[0;0m"
-    print(f'File: "{filename}", line {token.line}, column {token.col}\n', file=sys.stderr)
+    print(f'File: "{token.filename}", line {token.line}, column {token.col}\n', file=sys.stderr)
     sys.stderr.write('    ')
     wrote_default = False
     for i, c in enumerate(line.decode('utf-8')):

regs/assembler/fs/__main__.py

@@ -10,8 +10,8 @@ from assembler.fs.validator import validate_instruction
 from assembler.fs.emitter import emit_instruction
 from assembler.error import print_error
 
-def frontend_inner(buf):
+def frontend_inner(filename, buf):
-    lexer = Lexer(buf, find_keyword, emit_newlines=False, minus_is_token=True)
+    lexer = Lexer(filename, buf, find_keyword, emit_newlines=False, minus_is_token=True)
     tokens = list(lexer.lex_tokens())
     parser = Parser(tokens)
     for ins_ast in parser.instructions():
@@ -22,15 +22,15 @@ def frontend_inner(buf):
 
 def frontend(filename, buf):
     try:
-        yield from frontend_inner(buf)
+        yield from frontend_inner(filename, buf)
     except LexerError as e:
-        print_error(filename, buf, e)
+        print_error(e)
         raise
     except ParserError as e:
-        print_error(filename, buf, e)
+        print_error(e)
         raise
     except ValidatorError as e:
-        print_error(filename, buf, e)
+        print_error(e)
         raise
 
 if __name__ == "__main__":

regs/assembler/fs/alu_emitter.py

@@ -17,6 +17,9 @@ def emit_alpha_op(code, alpha_op):
     if alpha_op.dest.omask is not None:
         US_CMN_INST.ALPHA_OMASK(code, alpha_op.dest.omask.value)
 
+    # omod
+    US_ALU_ALPHA_INST.OMOD(code, alpha_op.omod.value)
+
     # opcode
     US_ALU_ALPHA_INST.ALPHA_OP(code, alpha_op.opcode.value)
 
@@ -59,6 +62,9 @@ def emit_rgb_op(code, rgb_op):
     if rgb_op.dest.omask is not None:
         US_CMN_INST.RGB_OMASK(code, rgb_op.dest.omask.value)
 
+    # omod
+    US_ALU_RGB_INST.OMOD(code, rgb_op.omod.value)
+
     # opcode
     US_ALU_RGBA_INST.RGB_OP(code, rgb_op.opcode.value)
 

regs/assembler/fs/alu_validator.py

@@ -110,6 +110,26 @@ class Swizzle(IntEnum):
     one = 6
     unused = 7
 
+class Omod(IntEnum):
+    mul_1 = 0
+    mul_2 = 1
+    mul_4 = 2
+    mul_8 = 3
+    div_2 = 4
+    div_4 = 5
+    div_8 = 6
+    disable = 7
+
+omod_lexemes = OrderedDict([
+    ((b"1", b"0"), Omod.mul_1),
+    ((b"2", b"0"), Omod.mul_2),
+    ((b"4", b"0"), Omod.mul_4),
+    ((b"8", b"0"), Omod.mul_8),
+    ((b"0", b"5"), Omod.div_2),
+    ((b"0", b"25"), Omod.div_4),
+    ((b"0", b"125"), Omod.div_8),
+])
+
 @dataclass
 class SwizzleSel:
     src: SwizzleSelSrc
@@ -119,12 +139,14 @@ class SwizzleSel:
 @dataclass
 class AlphaOperation:
     dest: AlphaDest
+    omod: Omod
     opcode: AlphaOp
     sels: list[SwizzleSel]
 
 @dataclass
 class RGBOperation:
     dest: RGBDest
+    omod: Omod
     opcode: RGBOp
     sels: list[SwizzleSel]
 
@@ -413,14 +435,27 @@ def validate_instruction_operation_sels(swizzle_sels, is_alpha):
         sels.append(SwizzleSel(src, swizzle, mod))
     return sels
 
+def validate_omod_operation(operation):
+    omod = Omod.mul_1
+    if operation.omod != None:
+        integer, decimal = operation.omod
+        key = (integer.lexeme, decimal.lexeme)
+        if key not in omod_lexemes:
+            valid_omods = b", ".join(b".".join(key) for key in omod_lexemes.keys()).decode('utf-8')
+            raise ValidatorError(f"invalid omod, expected one of [{valid_omods}]", integer)
+        omod = omod_lexemes[key]
+    return omod
+
 def validate_alpha_instruction_operation(operation):
     dest = validate_instruction_operation_dest(operation.dest_addr_swizzles,
                                                mask_lookup=alpha_masks,
                                                type_cls=AlphaDest)
+    omod = validate_omod_operation(operation)
     opcode = alpha_op_kws[operation.opcode_keyword.keyword]
     sels = validate_instruction_operation_sels(operation.swizzle_sels, is_alpha=True)
     return AlphaOperation(
         dest,
+        omod,
         opcode,
         sels
     )
@@ -429,10 +464,12 @@ def validate_rgb_instruction_operation(operation):
     dest = validate_instruction_operation_dest(operation.dest_addr_swizzles,
                                                mask_lookup=rgb_masks,
                                                type_cls=RGBDest)
+    omod = validate_omod_operation(operation)
     opcode = rgb_op_kws[operation.opcode_keyword.keyword]
     sels = validate_instruction_operation_sels(operation.swizzle_sels, is_alpha=False)
     return RGBOperation(
         dest,
+        omod,
         opcode,
         sels
     )

regs/assembler/fs/parser.py

@@ -35,6 +35,7 @@ class ALUSwizzleSel:
 @dataclass
 class ALUOperation:
     dest_addr_swizzles: list[DestAddrSwizzle]
+    omod: tuple[Token, Token]
     opcode_keyword: Token
     swizzle_sels: list[ALUSwizzleSel]
 
@@ -115,6 +116,15 @@ class Parser(BaseParser):
             return token.keyword in opcode_keywords
         return False
 
+    def alu_is_omod(self):
+        is_omod = (
+            self.match(TT.identifier, offset=0)
+            and self.match(TT.dot, offset=1)
+            and self.match(TT.identifier, offset=2)
+            and self.match(TT.star, offset=3)
+        )
+        return is_omod
+
     def alu_is_neg(self):
         result = self.match(TT.minus)
         if result:
@@ -154,9 +164,17 @@ class Parser(BaseParser):
 
     def alu_operation(self):
         dest_addr_swizzles = []
-        while not self.alu_is_opcode():
+        while not (self.alu_is_opcode() or self.alu_is_omod()):
             dest_addr_swizzles.append(self.dest_addr_swizzle())
 
+        omod = None
+        if self.alu_is_omod():
+            omod_integer = self.consume(TT.identifier, "expected omod decimal identifier")
+            self.consume(TT.dot, "expected omod decimal dot")
+            omod_decimal = self.consume(TT.identifier, "expected omod decimal identifier")
+            self.consume(TT.star, "expected omod star")
+            omod = (omod_integer, omod_decimal)
+
         opcode_keyword = self.consume(TT.keyword, "expected opcode keyword")
 
         swizzle_sels = []
@@ -165,6 +183,7 @@ class Parser(BaseParser):
 
         return ALUOperation(
             dest_addr_swizzles,
+            omod,
             opcode_keyword,
             swizzle_sels
         )

regs/assembler/lexer.py

@@ -2,6 +2,7 @@ from dataclasses import dataclass
 from enum import Enum, auto
 from itertools import chain
 from typing import Union, Any
+from os import path
 
 DEBUG = True
 
@@ -21,9 +22,11 @@ class TT(Enum):
     bar = auto()
     comma = auto()
     minus = auto()
+    star = auto()
 
 @dataclass
 class Token:
+    filename: str
     start_ix: int
     line: int
     col: int
@@ -42,7 +45,10 @@ class LexerError(Exception):
     pass
 
 class Lexer:
-    def __init__(self, buf: memoryview, find_keyword,
+    def __init__(self,
+                 filename: str,
+                 buf: memoryview,
+                 find_keyword,
                  emit_newlines=False,
                  minus_is_token=False):
         self.start_ix = 0
@@ -53,6 +59,8 @@ class Lexer:
         self.find_keyword = find_keyword
         self.emit_newlines = emit_newlines
         self.minus_is_token = minus_is_token
+        self.filename = filename
+        self.nested_lexer = None
 
     def at_end_p(self):
         return self.current_ix >= len(self.buf)
@@ -73,7 +81,7 @@ class Lexer:
         return self.buf[self.current_ix]
 
     def pos(self):
-        return self.start_ix, self.line, self.col - (self.current_ix - self.start_ix)
+        return self.filename, self.start_ix, self.line, self.col - (self.current_ix - self.start_ix)
 
     def identifier(self):
         while not self.at_end_p() and self.peek() in identifier_characters:
@@ -84,8 +92,26 @@ class Lexer:
         else:
             return Token(*self.pos(), TT.identifier, self.lexeme(), None)
 
+    def include(self, filename):
+        dirname = path.dirname(self.filename)
+        new_filename = path.join(dirname, filename.decode('utf-8'))
+        with open(new_filename, 'rb') as f:
+            buf = f.read()
+        self.nested_lexer = Lexer(new_filename,
+                                  buf,
+                                  find_keyword=self.find_keyword,
+                                  emit_newlines=self.emit_newlines,
+                                  minus_is_token=self.minus_is_token)
+
     def lex_token(self):
         while True:
+            if self.nested_lexer is not None:
+                token = self.nested_lexer.lex_token()
+                if token.type is TT.eof:
+                    self.nested_lexer = None
+                else:
+                    return token
+
             self.start_ix = self.current_ix
 
             if self.at_end_p():
@@ -112,6 +138,8 @@ class Lexer:
                 return Token(*self.pos(), TT.semicolon, self.lexeme())
             elif c == ord(','):
                 return Token(*self.pos(), TT.comma, self.lexeme())
+            elif c == ord('*'):
+                return Token(*self.pos(), TT.star, self.lexeme())
             elif c == ord('-') and self.peek() == ord('-'):
                 self.advance()
                 while not self.at_end_p() and self.peek() != ord('\n'):
@@ -119,8 +147,24 @@ class Lexer:
             elif self.minus_is_token and c == ord('-'):
                 return Token(*self.pos(), TT.minus, self.lexeme())
             elif c == ord('#'):
-                while not self.at_end_p() and self.peek() != ord('\n'):
+                for c in b"include":
+                    o = self.advance()
+                    if o != c:
+                        token = Token(*self.pos(), None, self.lexeme())
+                        raise LexerError(f"unexpected character at line:{self.line} col:{self.col}, expected `#include`", token)
+                while self.peek() == ord(' '):
                     self.advance()
+                self.start_ix = self.current_ix
+                quote = self.advance()
+                if quote != ord('"'):
+                    token = Token(*self.pos(), None, self.lexeme())
+                    raise LexerError(f"unexpected character at line:{self.line} col:{self.col}, expected `\"`", token)
+                self.start_ix = self.current_ix
+                while self.peek() != ord('"'):
+                    self.advance()
+                filename = self.lexeme()
+                assert self.advance() == ord('"')
+                self.include(filename)
             elif c == ord(' ') or c == ord('\r') or c == ord('\t'):
                 pass
             elif c == ord('\n'):

regs/assembler/parser.py

@@ -22,8 +22,8 @@ class BaseParser:
         self.current_ix += 1
         return token
 
-    def match(self, token_type):
+    def match(self, token_type, offset=0):
-        token = self.peek()
+        token = self.peek(offset)
         return token.type == token_type
 
     def match_keyword(self, keyword):

regs/assembler/vs/__main__.py

@@ -22,8 +22,8 @@ out[0].xz    = VE_MAD    input[0].-y-_-0-_ temp[0].x_0_      temp[0].y_0_
 out[0].yw    = VE_MAD    input[0]._x_0     temp[0]._x_0      temp[0]._z_1
 """
 
-def frontend_inner(buf):
+def frontend_inner(filename, buf):
-    lexer = Lexer(buf, find_keyword)
+    lexer = Lexer(filename, buf, find_keyword)
     tokens = list(lexer.lex_tokens())
     parser = Parser(tokens)
     for ins in parser.instructions():
@@ -37,15 +37,15 @@ def frontend_inner(buf):
 
 def frontend(filename, buf):
     try:
-        yield from frontend_inner(buf)
+        yield from frontend_inner(filename, buf)
     except ParserError as e:
-        print_error(input_filename, buf, e)
+        print_error(e)
         raise
     except LexerError as e:
-        print_error(input_filename, buf, e)
+        print_error(e)
         raise
     except ValidatorError as e:
-        print_error(filename, buf, e)
+        print_error(e)
         raise
 
 if __name__ == "__main__":