From 6f5007525c3b429dcecf254ba4bdaad3cbc03ac9 Mon Sep 17 00:00:00 2001
From: Zack Buhman <zack@buhman.org>
Date: Mon, 20 Oct 2025 21:09:52 -0500
Subject: [PATCH] shadertoy: assemble shadertoy.fs.asm from source

---
 drm/length.fs.asm       | 15 +++++++++++++++
 drm/length.fs.inc       | 28 ++++++++++++++++++++++++++++
 drm/shadertoy.c         | 27 +++++++++++++++------------
 drm/shadertoy.fs.asm    |  4 ++++
 drm/shadertoy.fs.inc    |  7 +++++++
 regs/us_disassemble2.py |  9 ++++++++-
 6 files changed, 77 insertions(+), 13 deletions(-)
 create mode 100644 drm/length.fs.asm
 create mode 100644 drm/length.fs.inc
 create mode 100644 drm/shadertoy.fs.asm
 create mode 100644 drm/shadertoy.fs.inc

diff --git a/drm/length.fs.asm b/drm/length.fs.asm
new file mode 100644
index 0000000..b5374d7
--- /dev/null
+++ b/drm/length.fs.asm
@@ -0,0 +1,15 @@
+src0.rgb = temp[0] :
+  temp[0].r    = DP3 src0.rg0 src0.rg0 ;
+
+src0.rgb = temp[0] :
+  temp[0].a    = RSQ |src0.r| ;
+
+OUT
+src0.a = temp[0] :
+                 RCP src0.a ,
+  out[0].r     = SOP ;
+
+OUT TEX_SEM_WAIT
+ :
+  out[0].a     = MAX src0.1 src0.1 ,
+  out[0].gb    = MAX src0.000 src0.000 ;
diff --git a/drm/length.fs.inc b/drm/length.fs.inc
new file mode 100644
index 0000000..2aae073
--- /dev/null
+++ b/drm/length.fs.inc
@@ -0,0 +1,28 @@
+0x00000800,
+0x08020000,
+0x08020080,
+0x00840420,
+0x00000000,
+0x00000001,
+
+0x00004000,
+0x08020000,
+0x08020080,
+0x00000000,
+0x0004000b,
+0x00000000,
+
+0x00008001,
+0x08020080,
+0x08020000,
+0x00000000,
+0x0000c00a,
+0x0000000a,
+
+0x00070005,
+0x08020080,
+0x08020080,
+0x00920490,
+0x00c18003,
+0x00000005,
+
diff --git a/drm/shadertoy.c b/drm/shadertoy.c
index a4fa851..eb216e7 100644
--- a/drm/shadertoy.c
+++ b/drm/shadertoy.c
@@ -244,9 +244,6 @@ int indirect_buffer()
   T0V(US_CONFIG
       , US_CONFIG__ZERO_TIMES_ANYTHING_EQUALS_ZERO(1)
       );
-  T0V(US_PIXSIZE
-      , US_PIXSIZE__PIX_SIZE(1)
-      );
   T0V(US_FC_CTRL, 0);
 
   T0V(FG_DEPTH_SRC, 0x00000000);
@@ -433,25 +430,31 @@ int indirect_buffer()
   // GA_US
   //////////////////////////////////////////////////////////////////////////////
 
+  const uint32_t fragment_shader[] = {
+    #include "shadertoy.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);
+  const int fragment_shader_instructions = fragment_shader_length / 6;
+  printf("fs instructions %d\n", fragment_shader_instructions);
+
+  T0V(US_PIXSIZE
+      , US_PIXSIZE__PIX_SIZE(1) // pixel shader stack frame size
+      );
+
   T0V(US_CODE_RANGE
       , US_CODE_RANGE__CODE_ADDR(0)
-      | US_CODE_RANGE__CODE_SIZE(0)
+      | US_CODE_RANGE__CODE_SIZE(fragment_shader_instructions - 1)
       );
   T0V(US_CODE_OFFSET
       , US_CODE_OFFSET__OFFSET_ADDR(0)
       );
   T0V(US_CODE_ADDR
       , US_CODE_ADDR__START_ADDR(0)
-      | US_CODE_ADDR__END_ADDR(0)
+      | US_CODE_ADDR__END_ADDR(fragment_shader_instructions - 1)
       );
 
-  const uint32_t fragment_shader[] = {
-    #include "../shader_examples/mesa/shadertoy.fs.txt"
-  };
-  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);
-
   T0V(GA_US_VECTOR_INDEX, 0x00000000);
   T0_ONE_REG(GA_US_VECTOR_DATA, fragment_shader_length - 1);
   for (int i = 0; i < fragment_shader_length; i++) {
diff --git a/drm/shadertoy.fs.asm b/drm/shadertoy.fs.asm
new file mode 100644
index 0000000..0620dbc
--- /dev/null
+++ b/drm/shadertoy.fs.asm
@@ -0,0 +1,4 @@
+OUT TEX_SEM_WAIT
+src0.rgb = temp[0] :
+  out[0].a    = MAX src0.1 src0.1 ,
+  out[0].rgb  = MAX src0.rg0 src0.rg0 ;
diff --git a/drm/shadertoy.fs.inc b/drm/shadertoy.fs.inc
new file mode 100644
index 0000000..baa663e
--- /dev/null
+++ b/drm/shadertoy.fs.inc
@@ -0,0 +1,7 @@
+0x00078005,
+0x08020000,
+0x08020080,
+0x00840420,
+0x00c18003,
+0x00000005,
+
diff --git a/regs/us_disassemble2.py b/regs/us_disassemble2.py
index 940c84e..61c54ac 100644
--- a/regs/us_disassemble2.py
+++ b/regs/us_disassemble2.py
@@ -299,6 +299,7 @@ def disassemble_alu(code, is_output):
     rgb_swizzle_sel, rgb_sels = disassemble_rgb_swizzle_sel(code)
     #print(", ".join([*rgb_swizzle_sel, *a_swizzle_sel]))
 
+    type = US_CMN_INST.TYPE(code)
     tex_sem_wait = US_CMN_INST.TEX_SEM_WAIT(code)
 
     _, a_op, _ = US_ALU_ALPHA_INST._ALPHA_OP(code)
@@ -309,8 +310,14 @@ def disassemble_alu(code, is_output):
 
     (a_out_str, a_temp_str), (rgb_out_str, rgb_temp_str) = disassemble_alu_dest(code)
 
+    tags = []
+    assert type in {0, 1}, type
+    if type == 1:
+        tags.append("OUT")
     if tex_sem_wait:
-        print("TEX_SEM_WAIT")
+        tags.append("TEX_SEM_WAIT")
+    if tags:
+        print(" ".join(tags))
 
     if not VERBOSE:
         a_swizzle_sel = a_swizzle_sel[:a_op_operands]