adding spirv interpreter results table dump

.gitignore

@@ -14,3 +14,4 @@ scripts/__pycache__/
 *.html
 *.pyc
 *.spv
+*_table_dump.txt

build.zig

@@ -299,6 +299,8 @@ fn addMultithreadedCTS(b: *std.Build, target: std.Build.ResolvedTarget, impl: *c
     run.addArg(b.fmt("{s}{s}", .{ cache_path, mustpass_path }));
     run.addArg("--output");
     run.addArg("./cts");
+    run.addArg("--timeout");
+    run.addArg("300");
     if (jobs_count) |count| {
         run.addArg(b.fmt("-j{d}", .{count}));
     }

build.zig.zon

@@ -59,8 +59,8 @@
             .lazy = true,
         },
         .SPIRV_Interpreter = .{
-            .url = "git+https://git.kbz8.me/kbz_8/SPIRV-Interpreter#fe47277468c694b10153282761efd3d6f90d8f6a",
+            .url = "git+https://git.kbz8.me/kbz_8/SPIRV-Interpreter#569d7fda01340f2fc195d868df7d3fc583f900bd",
-            .hash = "SPIRV_Interpreter-0.0.1-ajmpn2u-AwAPdrXzLzxHPezx3Kkb63hkecdY0oPEJ2ad",
+            .hash = "SPIRV_Interpreter-0.0.1-ajmpn3PFAwD7uWTP4W8ULr8mLWlV5BiqvfOIsLyYjDij",
         },
     },
 

src/soft/device/ComputeRoutines.zig

@@ -29,11 +29,19 @@ device: *SoftDevice,
 state: *PipelineState,
 batch_size: usize,
 
+invocation_index: std.atomic.Value(usize),
+
+early_dump: ?u32,
+final_dump: ?u32,
+
 pub fn init(device: *SoftDevice, state: *PipelineState) Self {
     return .{
         .device = device,
         .state = state,
         .batch_size = 0,
+        .invocation_index = .init(0),
+        .early_dump = std.process.parseEnvVarInt(lib.DUMP_EARLY_RESULT_TABLE_ENV_NAME, u32, 10) catch null,
+        .final_dump = std.process.parseEnvVarInt(lib.DUMP_FINAL_RESULT_TABLE_ENV_NAME, u32, 10) catch null,
     };
 }
 
@@ -51,6 +59,8 @@ pub fn dispatch(self: *Self, group_count_x: u32, group_count_y: u32, group_count
 
     const invocations_per_workgroup = spv_module.local_size_x * spv_module.local_size_y * spv_module.local_size_z;
 
+    self.invocation_index.store(0, .monotonic);
+
     var wg: std.Thread.WaitGroup = .{};
     for (0..@min(self.batch_size, group_count)) |batch_id| {
         if (std.process.hasEnvVarConstant(lib.SINGLE_THREAD_COMPUTE_EXECUTION_ENV_NAME)) {
@@ -128,12 +138,19 @@ inline fn run(data: RunData) !void {
         });
 
         for (0..data.invocations_per_workgroup) |i| {
+            const invocation_index = data.self.invocation_index.fetchAdd(1, .monotonic);
+
             try setupSubgroupBuiltins(data.self, rt, .{
                 @as(u32, @intCast(group_x)),
                 @as(u32, @intCast(group_y)),
                 @as(u32, @intCast(group_z)),
             }, i);
 
+            if (data.self.early_dump != null and data.self.early_dump.? == invocation_index) {
+                @branchHint(.cold);
+                try dumpResultsTable(allocator, rt, true);
+            }
+
             rt.callEntryPoint(allocator, entry) catch |err| switch (err) {
                 // Some errors can be ignored
                 SpvRuntimeError.OutOfBounds,
@@ -141,11 +158,29 @@ inline fn run(data: RunData) !void {
                 => {},
                 else => return err,
             };
+
+            if (data.self.final_dump != null and data.self.final_dump.? == invocation_index) {
+                @branchHint(.cold);
+                try dumpResultsTable(allocator, rt, false);
+            }
+
             try rt.flushDescriptorSets(allocator);
         }
     }
 }
 
+inline fn dumpResultsTable(allocator: std.mem.Allocator, rt: *spv.Runtime, is_early: bool) !void {
+    @branchHint(.cold);
+    const file = try std.fs.cwd().createFile(
+        std.fmt.comptimePrint("{s}_compute_result_table_dump.txt", .{if (is_early) "early" else "final"}),
+        .{ .truncate = true },
+    );
+    defer file.close();
+    var buffer = [_]u8{0} ** 1024;
+    var writer = file.writer(buffer[0..]);
+    try rt.dumpResultsTable(allocator, &writer.interface);
+}
+
 fn writeDescriptorSets(self: *Self, rt: *spv.Runtime) !void {
     sets: for (self.state.sets[0..], 0..) |set, set_index| {
         if (set == null)

src/soft/lib.zig

@@ -41,7 +41,9 @@ pub const DRIVER_VERSION = vk.makeApiVersion(0, 0, 0, 1);
 pub const DEVICE_ID = 0x600DCAFE;
 
 pub const NO_SHADER_SIMD_ENV_NAME = "STROLL_SOFT_NO_SIMD";
-pub const SINGLE_THREAD_COMPUTE_EXECUTION_ENV_NAME = "STROLL_SINGLE_THREAD_COMPUTE_EXECUTION";
+pub const SINGLE_THREAD_COMPUTE_EXECUTION_ENV_NAME = "STROLL_SOFT_SINGLE_THREAD_COMPUTE_EXECUTION";
+pub const DUMP_EARLY_RESULT_TABLE_ENV_NAME = "STROLL_SOFT_DUMP_EARLY_RESULT_TABLE_FOR_INVOCATION";
+pub const DUMP_FINAL_RESULT_TABLE_ENV_NAME = "STROLL_SOFT_DUMP_FINAL_RESULT_TABLE_FOR_INVOCATION";
 
 /// Generic system memory.
 pub const MEMORY_TYPE_GENERIC_BIT = 0;