Fix compute fuzzing

tape-drive/assemble_trace.py

@@ -10,6 +10,8 @@ SHADERS_OUT = "shaders_out"
 SHADERS_IN = "src/shaders"
 TARGET_ENV = "vulkan1.3"
 
+DUMP_DISASSEMBLIES = True
+
 if os.path.isdir(SHADERS_OUT):
     shutil.rmtree(SHADERS_OUT)
 os.mkdir(SHADERS_OUT)
@@ -55,7 +57,7 @@ for shader, ty, usage in replacement_shaders:
             "glslangValidator",
             "--spirv-val",
             "--spirv-dis",
-            "-gVS",
+            "-g",
             "-S",
             ty,
             "--target-env",
@@ -70,8 +72,10 @@ for shader, ty, usage in replacement_shaders:
         print(result.stdout)
         sys.exit(result.returncode)
     asm = result.stdout
-    with open(f"{SHADERS_OUT}/{shader}.old.spv-dis", "w") as f:
+
-        f.write(asm)
+    if DUMP_DISASSEMBLIES:
+        with open(f"{SHADERS_OUT}/{shader}.stage0.spv-dis", "w") as f:
+            f.write(asm)
 
     asm = re.sub(r"(?ms)%scene_vf4_ =.*?OpFunctionEnd", "", asm)
     asm = re.sub(r"%scene_vf4_", "%10000", asm)
@@ -80,8 +84,10 @@ for shader, ty, usage in replacement_shaders:
     asm = re.sub(r"(?ms)%gradient_scene_vf4_ =.*?OpFunctionEnd", "", asm)
     asm = re.sub(r"%gradient_scene_vf4_", "%30000", asm)
 
-    with open(f"{SHADERS_OUT}/{shader}.spv-dis", "w") as f:
+    if DUMP_DISASSEMBLIES:
-        f.write(asm)
+        with open(f"{SHADERS_OUT}/{shader}.stage1.spv-dis", "w") as f:
+            f.write(asm)
+
     result = subprocess.run(
         [
             "spirv-as",

tape-drive/src/main.rs

@@ -5,6 +5,7 @@ const DUMP_SPV_TO_FILE: bool = false;
 const PIPELINE_CACHING: bool = false;
 const IGNORE_STENCIL: bool = true;
 const VSYNC: bool = true;
+const COMPUTE_FUZZING: bool = false;
 
 const DEFAULT_SUBDIVISION: u32 = 16;
 const MAXIMUM_SUBDIVISION: u32 = 64;
@@ -18,7 +19,7 @@ use std::{
     fs::{File, remove_file, rename},
     io::{Cursor, Read, Write},
     path::{Path, PathBuf},
-    simd::f32x8,
+    simd::{f32x8, num::SimdFloat},
     sync::{Arc, Mutex, atomic::AtomicBool, mpmc, mpsc},
     thread::{self, JoinHandle},
     time::Instant,
@@ -211,6 +212,7 @@ struct App {
     uniform_buffer_allocator: Arc<Mutex<SubbufferAllocator>>,
     block_enable_allocator: Arc<Mutex<SubbufferAllocator>>,
     host_visible_allocator: Arc<Mutex<SubbufferAllocator>>,
+    compute_visible_allocator: Arc<Mutex<SubbufferAllocator>>,
     pipeline_cache: Arc<PipelineCache>,
     draw_gui: bool,
     gstate: GState,
@@ -538,6 +540,16 @@ impl App {
             },
         )));
 
+        let compute_visible_allocator = Arc::new(Mutex::new(SubbufferAllocator::new(
+            memory_allocator.clone(),
+            SubbufferAllocatorCreateInfo {
+                buffer_usage: BufferUsage::STORAGE_BUFFER,
+                memory_type_filter: MemoryTypeFilter::HOST_SEQUENTIAL_WRITE
+                    | MemoryTypeFilter::PREFER_DEVICE,
+                ..Default::default()
+            },
+        )));
+
         let memory_allocator: Arc<dyn MemoryAllocator> = memory_allocator;
 
         let pipeline_cache = get_pipeline_cache(device.clone());
@@ -669,6 +681,7 @@ impl App {
             descriptor_set_allocator,
             command_buffer_allocator,
             uniform_buffer_allocator,
+            compute_visible_allocator,
             host_visible_allocator,
             block_enable_allocator,
             pipeline_cache,
@@ -1871,7 +1884,7 @@ impl App {
                         if index < 0 {
                             let csg = csg.clone();
 
-                            let alloc = self.host_visible_allocator.lock().unwrap();
+                            let alloc = self.compute_visible_allocator.lock().unwrap();
                             let input: Subbuffer<[fuzz_cs::InputData]> = alloc
                                 .allocate_slice((LOCAL_WIDTH * GLOBAL_WIDTH).into())
                                 .unwrap();
@@ -1885,7 +1898,7 @@ impl App {
                                 for i in 0..4 {
                                     val.point[i] = rand::random_range(-1000.0..1000.0);
                                     val.gradient[i] = rand::random_range(-1000.0..1000.0);
-                                    let midpoint = rand::random_range(-1000.0..1000.0);
+                                    let midpoint = rand::random_range(-999.0..999.0);
                                     val.interval_low[i] = rand::random_range(-1000.0..midpoint);
                                     val.interval_high[i] = rand::random_range(midpoint..1000.0);
                                 }
@@ -1961,7 +1974,7 @@ impl App {
                 .into_iter()
                 .zip(output.as_chunks::<8>().0)
             {
-                println!("{output:?}");
+                //println!("{output:?}");
                 let point_val = f32x8::from_slice(&output.map(|o| *o.point));
                 let point_expected = PointInterpreter::new(&csg.parts).scene(
                     f32x8::from_slice(&input.map(|i| i.point[0])),
@@ -1970,7 +1983,7 @@ impl App {
                     f32x8::from_slice(&input.map(|i| i.point[3])),
                 );
 
-                if point_expected != point_val {
+                if (point_expected - point_val).abs() > f32x8::splat(0.001) {
                     println!(
                         "ERROR: point expected {:?}, got {:?}",
                         point_expected, point_val
@@ -2000,10 +2013,18 @@ impl App {
                     ),
                 );
 
-                if interval_expected != interval_val {
+                if (interval_expected.upper() - interval_val.upper()).abs() > f32x8::splat(0.001) {
                     println!(
-                        "ERROR: interval expected {:?}, got {:?}",
+                        "ERROR: interval upper expected {:?}, got {:?}",
-                        interval_expected, interval_val
+                        interval_expected.upper(),
+                        interval_val.upper()
+                    );
+                }
+                if (interval_expected.lower() - interval_val.lower()).abs() > f32x8::splat(0.001) {
+                    println!(
+                        "ERROR: interval lower expected {:?}, got {:?}",
+                        interval_expected.lower(),
+                        interval_val.lower()
                     );
                 }
             }
@@ -2048,7 +2069,9 @@ impl App {
 
         let mut work_for_later = vec![];
 
-        self.compute_fuzz(rcx, &mut work_for_later);
+        if COMPUTE_FUZZING {
+            self.compute_fuzz(rcx, &mut work_for_later);
+        }
 
         if self.draw_gui {
             gui_up(&mut rcx.gui, &mut self.gstate);

tape-load/src/spirv_compilers/gradient.rs

@@ -12,15 +12,17 @@ pub(crate) fn compile_gradient_function(
     types: SpirVTypes,
     function_id: Option<spirv::Word>,
 ) {
+    let gradient_fn_type = b.type_function(types.vec4, vec![types.vec4p]);
+    let jit_string = b.string("Gradient JIT");
     let _scene = b
         .begin_function(
-            types.float,
+            types.vec4,
             function_id,
             //spirv::FunctionControl::DONT_INLINE
             spirv::FunctionControl::INLINE
                 | spirv::FunctionControl::PURE
                 | spirv::FunctionControl::CONST,
-            types.gradient_fn_type,
+            gradient_fn_type,
         )
         .unwrap();
     let pos_p = b.function_parameter(types.vec4p).unwrap();
@@ -35,7 +37,7 @@ pub(crate) fn compile_gradient_function(
         use SSAOpcode::*;
         use rspirv::dr::Operand::IdRef;
 
-        b.line(types.jit_string, line as u32, 0);
+        b.line(jit_string, line as u32, 0);
 
         fn input_resolve(
             float: u32,

tape-load/src/spirv_compilers/interval.rs

@@ -13,6 +13,8 @@ pub(crate) fn compile_interval_function(
     types: SpirVTypes,
     function_id: Option<spirv::Word>,
 ) {
+    let interval_fn_type = b.type_function(types.vec2, vec![types.vec4p, types.vec4p]);
+    let jit_string = b.string("Interval JIT");
     let _scene = b
         .begin_function(
             types.vec2,
@@ -21,7 +23,7 @@ pub(crate) fn compile_interval_function(
             spirv::FunctionControl::INLINE
                 | spirv::FunctionControl::PURE
                 | spirv::FunctionControl::CONST,
-            types.interval_fn_type,
+            interval_fn_type,
         )
         .unwrap();
     let pos_low = b.function_parameter(types.vec4p).unwrap();
@@ -38,7 +40,7 @@ pub(crate) fn compile_interval_function(
         use SSAOpcode::*;
         use rspirv::dr::Operand::IdRef;
 
-        b.line(types.jit_string, line as u32, 0);
+        b.line(jit_string, line as u32, 0);
 
         fn input_resolve(
             float: u32,

tape-load/src/spirv_compilers/mod.rs

@@ -4,17 +4,13 @@ pub(crate) mod point;
 
 #[derive(Debug, Clone, Copy)]
 pub(crate) struct SpirVTypes {
-    pub glsl:             u32,
+    pub glsl:   u32,
-    pub void:             u32,
+    pub void:   u32,
-    pub float:            u32,
+    pub float:  u32,
-    pub bool:             u32,
+    pub bool:   u32,
-    pub choice:           u32,
+    pub choice: u32,
-    pub vec2:             u32,
+    pub vec2:   u32,
-    pub vec3:             u32,
+    pub vec3:   u32,
-    pub vec4:             u32,
+    pub vec4:   u32,
-    pub vec4p:            u32,
+    pub vec4p:  u32,
-    pub point_fn_type:    u32,
-    pub interval_fn_type: u32,
-    pub gradient_fn_type: u32,
-    pub jit_string:       u32,
 }

tape-load/src/spirv_compilers/point.rs

@@ -12,6 +12,8 @@ pub(crate) fn compile_point_function(
     types: SpirVTypes,
     function_id: Option<spirv::Word>,
 ) {
+    let point_fn_type = b.type_function(types.float, vec![types.vec4p]);
+    let jit_string = b.string("Point JIT");
     let _scene = b
         .begin_function(
             types.float,
@@ -20,7 +22,7 @@ pub(crate) fn compile_point_function(
             spirv::FunctionControl::INLINE
                 | spirv::FunctionControl::PURE
                 | spirv::FunctionControl::CONST,
-            types.point_fn_type,
+            point_fn_type,
         )
         .unwrap();
     let pos_p = b.function_parameter(types.vec4p).unwrap();
@@ -35,7 +37,7 @@ pub(crate) fn compile_point_function(
         use SSAOpcode::*;
         use rspirv::dr::Operand::IdRef;
 
-        b.line(types.jit_string, line as u32, 0);
+        b.line(jit_string, line as u32, 0);
 
         fn input_resolve(
             float: u32,

tape-load/src/ssa.rs

@@ -467,7 +467,7 @@ impl SSATape {
             b
         };
         let glsl = if with_module {
-            1
+            4
         } else {
             b.ext_inst_import("GLSL.std.450")
         };
@@ -480,10 +480,6 @@ impl SSATape {
         let vec3 = b.type_vector(float, 3);
         let vec4 = b.type_vector(float, 4);
         let vec4p = b.type_pointer(None, spirv::StorageClass::Function, vec4);
-        let point_fn_type = b.type_function(float, vec![vec4p]);
-        let interval_fn_type = b.type_function(vec2, vec![vec4p, vec4p]);
-        let gradient_fn_type = b.type_function(vec4, vec![vec4p]);
-        let jit_string = b.string("JIT");
 
         const POINT_FN_LOC: u32 = 10000;
         const INTERVAL_FN_LOC: u32 = 20000;
@@ -499,10 +495,6 @@ impl SSATape {
             vec3,
             vec4,
             vec4p,
-            point_fn_type,
-            interval_fn_type,
-            gradient_fn_type,
-            jit_string,
         };
 
         // Manually fix the header bounds
@@ -561,7 +553,7 @@ impl SSATape {
         if DUMP_SPV_DIS_TO_FILE {
             std::fs::write(
                 format!(
-                    "{}.spv-dis",
+                    "spv-dis/{}.spv-dis",
                     humantime::format_rfc3339(std::time::SystemTime::now())
                 ),
                 module.disassemble(),