fix zenofx asm volatile invalid in omp context

AtomQu · Jul 15, 2021 · d720eb0 · d720eb0
1 parent 03681d2
commit d720eb0
Show file tree

Hide file tree

Showing 4 changed files with 160 additions and 18 deletions.
diff --git a/Makefile b/Makefile
@@ -1,4 +1,4 @@
-O=arts/testifelse.zsg
+O=arts/ZFXv2.zsg
 
 default: run
 
@@ -7,6 +7,10 @@ dist: all
 	./dist.sh
 
 all:
+	cmake -B build
+	make -C build -j `python -c 'from multiprocessing import cpu_count; print(cpu_count() * 2)'`
+
+release_all:
 	cmake -B build -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/tmp/tmp-install
 	make -C build -j `python -c 'from multiprocessing import cpu_count; print(cpu_count() * 2)'`
 

diff --git a/ZFX/include/zfx/x64.h b/ZFX/include/zfx/x64.h
@@ -19,13 +19,12 @@ struct Executable {
         float locals[SimdWidth * 256];
 
         void execute() {
-            auto entry = (void (*)())exec->mem;
             asm volatile (
-                "call *%0"
+                "call *(%%rax)"
                 :
-                : "" (entry)
-                , "c" ((uintptr_t)(void *)exec->consts)
-                , "d" ((uintptr_t)(void *)locals)
+                : "a" (&exec->mem)
+                , "c" ((uintptr_t)(void *)&exec->consts[0])
+                , "d" ((uintptr_t)(void *)&locals[0])
                 : "cc", "memory"
                 );
         }

diff --git a/arts/ZFXv2.zsg b/arts/ZFXv2.zsg
@@ -277,6 +277,20 @@
     "control"
    ]
   },
+  "BreakFor": {
+   "inputs": [
+    "FOR",
+    "SRC",
+    "COND"
+   ],
+   "outputs": [
+    "DST"
+   ],
+   "params": [],
+   "categories": [
+    "control"
+   ]
+  },
   "CachedOnce": {
    "inputs": [
     "input",
@@ -445,6 +459,93 @@
     "list"
    ]
   },
+  "ExtractSmallDict": {
+   "inputs": [
+    "dict",
+    "SRC",
+    "COND"
+   ],
+   "outputs": [
+    "obj0",
+    "obj1",
+    "obj2",
+    "obj3",
+    "DST"
+   ],
+   "params": [
+    [
+     "string",
+     "name0",
+     ""
+    ],
+    [
+     "string",
+     "name1",
+     ""
+    ],
+    [
+     "string",
+     "name2",
+     ""
+    ],
+    [
+     "string",
+     "name3",
+     ""
+    ]
+   ],
+   "categories": [
+    "dict"
+   ]
+  },
+  "FuncBegin": {
+   "inputs": [
+    "extraArgs",
+    "SRC",
+    "COND"
+   ],
+   "outputs": [
+    "args",
+    "FUNC",
+    "DST"
+   ],
+   "params": [],
+   "categories": [
+    "functional"
+   ]
+  },
+  "FuncCall": {
+   "inputs": [
+    "function",
+    "args",
+    "SRC",
+    "COND"
+   ],
+   "outputs": [
+    "rets",
+    "DST"
+   ],
+   "params": [],
+   "categories": [
+    "functional"
+   ]
+  },
+  "FuncEnd": {
+   "inputs": [
+    "rets",
+    "FUNC",
+    "SRC",
+    "COND"
+   ],
+   "outputs": [
+    "function",
+    "DST"
+   ],
+   "params": [],
+   "categories": [
+    "functional"
+   ]
+  },
   "GetFrameNum": {
    "inputs": [
     "SRC",
@@ -515,6 +616,23 @@
     "frame"
    ]
   },
+  "IfElse": {
+   "inputs": [
+    "true",
+    "false",
+    "cond",
+    "SRC",
+    "COND"
+   ],
+   "outputs": [
+    "result",
+    "DST"
+   ],
+   "params": [],
+   "categories": [
+    "control"
+   ]
+  },
   "ImportObjPrimitive": {
    "inputs": [
     "path",
@@ -620,7 +738,7 @@
     [
      "int",
      "value",
-     "0"
+     "42"
     ]
    ],
    "categories": [
@@ -1519,6 +1637,25 @@
     "primitive"
    ]
   },
+  "PrintMessage": {
+   "inputs": [
+    "SRC",
+    "COND"
+   ],
+   "outputs": [
+    "DST"
+   ],
+   "params": [
+    [
+     "string",
+     "message",
+     "hello"
+    ]
+   ],
+   "categories": [
+    "debug"
+   ]
+  },
   "PrintNumeric": {
    "inputs": [
     "value",

diff --git a/projects/ZenoFX/pw.cpp b/projects/ZenoFX/pw.cpp
@@ -7,7 +7,8 @@
 #include <zfx/x64.h>
 #include <cassert>
 
-static zfx::Compiler<zfx::x64::Program> compiler;
+static zfx::Compiler compiler;
+static zfx::x64::Assembler assembler;
 
 struct Buffer {
     float *base = nullptr;
@@ -17,7 +18,7 @@ struct Buffer {
 };
 
 static void vectors_wrangle
-    ( zfx::Program<zfx::x64::Program> *prog
+    ( zfx::x64::Executable *exec
     , std::vector<Buffer> const &chs
     ) {
     if (chs.size() == 0)
@@ -28,20 +29,20 @@ static void vectors_wrangle
     }
 
     #pragma omp parallel for
-    for (int i = 0; i < size - prog->SimdWidth + 1; i += prog->SimdWidth) {
-        auto ctx = prog->make_context();
+    for (int i = 0; i < size - exec->SimdWidth + 1; i += exec->SimdWidth) {
+        auto ctx = exec->make_context();
         for (int j = 0; j < chs.size(); j++) {
-            for (int k = 0; k < prog->SimdWidth; k++)
+            for (int k = 0; k < exec->SimdWidth; k++)
                 ctx.channel(j)[k] = chs[j].base[chs[j].stride * (i + k)];
         }
         ctx.execute();
         for (int j = 0; j < chs.size(); j++) {
-            for (int k = 0; k < prog->SimdWidth; k++)
+            for (int k = 0; k < exec->SimdWidth; k++)
                  chs[j].base[chs[j].stride * (i + k)] = ctx.channel(j)[k];
         }
     }
-    for (int i = size / prog->SimdWidth * prog->SimdWidth; i < size; i++) {
-        auto ctx = prog->make_context();
+    for (int i = size / exec->SimdWidth * exec->SimdWidth; i < size; i++) {
+        auto ctx = exec->make_context();
         for (int j = 0; j < chs.size(); j++) {
             ctx.channel(j)[0] = chs[j].base[chs[j].stride * i];
         }
@@ -57,7 +58,7 @@ struct ParticlesWrangle : zeno::INode {
         auto prim = get_input<zeno::PrimitiveObject>("prim");
         auto code = get_input<zeno::StringObject>("zfxCode")->get();
 
-        zfx::Options opts;
+        zfx::Options opts(zfx::Options::for_x64);
         for (auto const &[key, attr]: prim->m_attrs) {
             int dim = std::visit([] (auto const &v) {
                 using T = std::decay_t<decltype(v[0])>;
@@ -96,6 +97,7 @@ struct ParticlesWrangle : zeno::INode {
         }
 
         auto prog = compiler.compile(code, opts);
+        auto exec = assembler.assemble(prog->assembly);
 
         std::vector<float> pars(prog->params.size());
         for (int i = 0; i < pars.size(); i++) {
@@ -106,7 +108,7 @@ struct ParticlesWrangle : zeno::INode {
                 parnames.end(), std::pair{name, dimid});
             auto value = parvals.at(it - parnames.begin());
             printf("(valued %f)\n", value);
-            prog->parameter(prog->param_id(name, dimid)) = value;
+            exec->parameter(prog->param_id(name, dimid)) = value;
         }
 
         std::vector<Buffer> chs(prog->symbols.size());
@@ -123,7 +125,7 @@ struct ParticlesWrangle : zeno::INode {
             }, attr);
             chs[i] = iob;
         }
-        vectors_wrangle(prog, chs);
+        vectors_wrangle(exec, chs);
 
         set_output("prim", std::move(prim));
     }