first milestone of t2x (#1)

* first milestone of t2x

* [Add nvidia_libdevice_bitcode]

* enable fifo and AOT (still have bugs in AOT)

* solve a bug in channel promotion

* rewrite channel promotion

* fix bugs

* fix a bug

* pass correctness tests

* [Cosmetic Revision]

* [Turn on WITH_EXCEPTIONS=1 for make by default. Revised README]

* [Add an intro video]

* [Revised README for GEMM]

* [Change all test.sh files to be executable]

* [Revised fft funct.h]

* [Make bash files executable]

* [Temporarily delete the intro video for being too large]

* [Add intro video again]

* [Add first slide image of the intro video]

* [Use URL for the intro video]

* [Remove embedding of video, which does not work]

* revise code and README

* Update README.md

* [Revised README for GEMM. Not fully done yet. To further revise]

* [Revise SGEMM README. Keep only AOT mode for simplicity. To verify on DevCloud]

* [Verified TINY emulation on DevCloud A10 1.2.1]

* [Recover gemm-run.cpp in aot]

* [Fix for security]

* [Revise GEMM README for more accurate instructions on DevCloud]

* fix a bug in insert_fpga_reg

* [Shared utilis for JIT runtime, AOT runtime, and roofline drawing. SGEMM perf test works]

* [Fix opencl.cpp: not to use SharedUtilsInC, which allocates and frees
 memory and sees to introduce issues with Halide]

* [Fix test/correctness/roofline]

* [Allow plotting rooflines without xserver]

* [Minor fixes to gemm perf test commands, etc.]

* [Misc tiny fixes]

Co-authored-by: xiaochen.hao <xiaochen.hao@stu.pku.edu.cn>
Co-authored-by: Hongbo Rong <hongbo.rong@intel.com>

.gitattributes

@@ -0,0 +1,2 @@
+
+intro.mp4 filter=lfs diff=lfs merge=lfs -text

ACKNOWLEDGEMENT.md

@@ -17,14 +17,14 @@ Many academic researchers have contributed to the project:
 
 Many Intel people have helped enabling the technology:
 
-+ System Software Group: Geoff Lowney, John C. Kreatsoulas, Nithin George, Paul Petersen, Gorge Powley, Daya Khudia, Charlotte Dryden, Adam Herr
++ System Software Group: Geoff Lowney, John C. Kreatsoulas, Nithin George, Paul Petersen, Gorge Powley, Carmen Badea, Daya Khudia, Adam Herr, Charlotte Dryden, Pablo Reble, Vishakha Agrawal, Mike Voss, Vasanth Tovinkere
 
 + Intel Labs: Christopher J. Hughes, Pradeep Dubey, Jim Held, Timothy Mattson, Sanket Tavarageri,  Kunal Banerjee, Bharat Kaul, Justin Gottschlich, Todd A. Anderson, Michael Beale
 
-+ Accelerated Computing Systems and Graphics Group: Hong Jiang, Kari Pulli, Lidong Xu, Fangwen Fu, Hongzheng Li, Sabareesh Ganapathy
++ Accelerated Computing Systems and Graphics Group: Hong Jiang, Lidong Xu, Kari Pulli, Fangwen Fu, Hongzheng Li, Sabareesh Ganapathy
 
 + GPU Software Engineering Group: Kai Yu Chen, Guei-Yuan Lueh, Yuting Yang
 
-+ Programmable Solution Group: Jose Alvarez, Bernhard Friebe, Mohamed Issa, Aravind Dasu, John Freeman, Davor Capalija, Tomasz Czajkowski
++ Programmable Solution Group: Jose Alvarez, Bernhard Friebe, Dan Prikster, Mohamed Issa, Aravind Dasu, John Freeman, Gordon Chiu, Davor Capalija, Tomasz Czajkowski, Andrei Hagiescu
 
 + Intel FPGA DevCloud: Lawrence Landis, Jimmy Tran

Halide/Makefile

@@ -1,3 +1,4 @@
+
 # 'make' builds libHalide.a, the internal test suite, and runs the internal test suite
 # 'make run_tests' builds and runs all the end-to-end tests in the test subdirectory
 # 'make {error,performance}_foo' builds and runs test/{...}/foo.cpp for any
@@ -136,7 +137,7 @@ ifeq ($(OS), Windows_NT)
 else
     WITH_INTROSPECTION ?= not-empty
 endif
-WITH_EXCEPTIONS ?=
+WITH_EXCEPTIONS ?= 1
 WITH_LLVM_INSIDE_SHARED_LIBHALIDE ?= not-empty
 
 WITH_V8 ?=
@@ -796,6 +797,7 @@ T2S_DIR  = $(ROOT_DIR)/../t2s
 T2S_SOURCE_FILES = \
   AutorunKernels.cpp \
   BuildCallRelation.cpp \
+  ChannelPromotion.cpp \
   CheckFuncConstraints.cpp \
   CheckRecursiveCalls.cpp \
   CombineChannels.cpp \
@@ -814,14 +816,15 @@ T2S_SOURCE_FILES = \
   MinimizeShregs.cpp \
   NoIfSimplify.cpp \
   Overlay.cpp \
+  PatternMatcher.cpp \
   Place.cpp \
   PreprocessBeforeLower.cpp \
   ScatterAndBuffer.cpp \
   SliceExprTree.cpp \
   SpaceTimeTransform.cpp \
+  Stensor.cpp \
   StructType.cpp \
-  Utilities.cpp \
-  roofline.cpp
+  Utilities.cpp
 
 T2S_HEADER_FILES = \
   AutorunKernels.h \
@@ -841,14 +844,15 @@ T2S_HEADER_FILES = \
   MinimizeShregs.h \
   NoIfSimplify.h \
   Overlay.h \
+  PatternMatcher.h \
   Place.h \
   PreprocessBeforeLower.h \
   ScatterAndBuffer.h \
   SliceExprTree.h \
   SpaceTimeTransform.h \
+  Stensor.h \
   StructType.h \
-  Utilities.h \
-  roofline.h
+  Utilities.h
 
 OBJECTS += $(T2S_SOURCE_FILES:%.cpp=$(BUILD_DIR)/t2s/%.o)
 HEADERS += $(T2S_HEADER_FILES:%.h=$(T2S_DIR)/src/%.h)

Halide/apps/fft/complex.h

@@ -68,6 +68,9 @@ inline Halide::Expr im(Halide::Expr x) {
 inline ComplexExpr conj(ComplexExpr z) {
     return ComplexExpr(re(z), -im(z));
 }
+inline Halide::Expr conj(Halide::Expr x) {
+    return x;
+}
 
 // Unary negation.
 inline ComplexExpr operator-(ComplexExpr z) {
@@ -106,6 +109,10 @@ inline ComplexExpr operator/(ComplexExpr a, Halide::Expr b) {
     return ComplexExpr(re(a) / b, im(a) / b);
 }
 
+inline ComplexExpr operator/(ComplexExpr a, ComplexExpr b) {
+    return a * conj(b) / (re(b)*re(b) + im(b)*im(b));
+}
+
 // Compute exp(j*x)
 inline ComplexExpr expj(Halide::Expr x) {
     return ComplexExpr(Halide::cos(x), Halide::sin(x));
@@ -116,9 +123,14 @@ inline ComplexExpr sum(ComplexExpr z, const std::string &s = "sum") {
     return ComplexExpr(Halide::sum(re(z), s + "_re"),
                        Halide::sum(im(z), s + "_im"));
 }
+
 inline ComplexExpr select(Halide::Expr c, ComplexExpr t, ComplexExpr f) {
-    return ComplexExpr(Halide::select(c, re(t), re(f)),
-                       Halide::select(c, im(t), im(f)));
+    Halide::Expr re_part=Halide::select(c, re(t), re(f));
+    Halide::Expr im_part=Halide::select(c, im(t), im(f));
+    return ComplexExpr(re_part, im_part);
+}
+inline ComplexExpr select(Halide::Expr c, ComplexExpr t) {
+    return ComplexExpr(Halide::select(c, re(t)), Halide::select(c, im(t)));
 }
 inline ComplexExpr select(Halide::Expr c1, ComplexExpr t1,
                           Halide::Expr c2, ComplexExpr t2,
@@ -136,5 +148,18 @@ inline ComplexExpr cast(Halide::Type type, ComplexExpr z) {
 inline ComplexExpr likely(ComplexExpr z) {
     return ComplexExpr(Halide::likely(re(z)), Halide::likely(im(z)));
 }
+template<typename T>
+inline T select(Halide::Expr c, ComplexExpr t, FuncRefT<T> f){
+    return select(c,t,f.toT());
+}
+template<typename T>
+inline T select(Halide::Expr c, FuncRefT<T> t, ComplexExpr f){
+    return select(c,t.toT(),f);
+}
+template<typename T>
+inline T select(Halide::Expr c, FuncRefT<T> t, FuncRefT<T> f){
+    return select(c,t.toT(),f.toT());
+}
+
 
 #endif

Halide/apps/fft/funct.h

@@ -5,7 +5,9 @@
 #include <vector>
 
 #include "Halide.h"
+#include "../../../../Halide/src/Func.h"
 
+struct ComplexExpr;
 template<typename T>
 class FuncRefT : public T {
     Halide::FuncRef untyped;
@@ -34,14 +36,23 @@ class FuncRefT : public T {
     Stage operator/=(T x) {
         return untyped = T(Tuple(untyped)) / x;
     }
+    T toT(){
+        if(typeid(ComplexExpr)==typeid(T)){
+            return ComplexExpr(Halide::Call::make(untyped.function(), untyped.arguments(), 0), Halide::Call::make(untyped.function(), untyped.arguments(), 1));
+        }
+    }
 };
 
+
 template<typename T>
 class FuncT : public Halide::Func {
 public:
     typedef Halide::Var Var;
     typedef Halide::Expr Expr;
     typedef Halide::Func Func;
+    typedef Halide::Place Place;
+    typedef Halide::Type Type;
+    typedef Halide::FuncRef FuncRef;
 
     explicit FuncT(const std::string &name)
         : Func(name) {
@@ -54,10 +65,26 @@ class FuncT : public Halide::Func {
     explicit FuncT(Func f)
         : Func(f) {
     }
+
+    explicit FuncT(Place place)
+        : Func(place) {
+    }
+    explicit FuncT(const std::string &name, Place place)
+        : Func(name, place) {
+    }
+
     explicit FuncT(Halide::Internal::Function f)
         : Func(f) {
     }
 
+    explicit FuncT(const std::vector<Type> &return_types, const std::vector<Var> &args, Place place)
+        : Func(return_types, args, place) {
+    }
+
+    explicit FuncT(const std::string &name, const std::vector<Type> &return_types, const std::vector<Var> &args, Place place)
+        : Func(name, return_types, args, place) {
+    }
+
     template<typename... Args>
     FuncRefT<T> operator()(Args &&... args) const {
         return Func::operator()(std::forward<Args>(args)...);
@@ -78,7 +105,7 @@ class FuncT : public Halide::Func {
 // there is one.
 template<typename T>
 T operator-(FuncRefT<T> x) {
-    return -static_cast<T>(x);
+    return -x.toT();
 }
 template<typename T>
 T operator~(FuncRefT<T> x) {
@@ -87,43 +114,64 @@ T operator~(FuncRefT<T> x) {
 
 template<typename T>
 T operator+(FuncRefT<T> a, T b) {
-    return static_cast<T>(a) + b;
+    return a.toT()+b;
 }
 template<typename T>
 T operator-(FuncRefT<T> a, T b) {
-    return static_cast<T>(a) - b;
+    return a.toT()-b;
 }
 template<typename T>
 T operator*(FuncRefT<T> a, T b) {
-    return static_cast<T>(a) * b;
+    return a.toT()*b;
 }
 template<typename T>
 T operator/(FuncRefT<T> a, T b) {
-    return static_cast<T>(a) / b;
+    return a.toT()/b;
 }
 template<typename T>
 T operator%(FuncRefT<T> a, T b) {
-    return static_cast<T>(a) % b;
+    return a.toT()%b;
 }
 template<typename T>
 T operator+(T a, FuncRefT<T> b) {
-    return a + static_cast<T>(b);
+    return a+b.toT();
 }
 template<typename T>
 T operator-(T a, FuncRefT<T> b) {
-    return a - static_cast<T>(b);
+    return a-b.toT();
 }
 template<typename T>
 T operator*(T a, FuncRefT<T> b) {
-    return a * static_cast<T>(b);
+    return a*b.toT();
 }
 template<typename T>
 T operator/(T a, FuncRefT<T> b) {
-    return a / static_cast<T>(b);
+    return a/b.toT();
 }
 template<typename T>
 T operator%(T a, FuncRefT<T> b) {
-    return a % static_cast<T>(b);
+    return a%b.toT();
+}
+
+template<typename T>
+T operator+(FuncRefT<T> a, FuncRefT<T> b) {
+    return a.toT()+b.toT();
+}
+template<typename T>
+T operator-(FuncRefT<T> a, FuncRefT<T> b) {
+    return a.toT()-b.toT();
+}
+template<typename T>
+T operator*(FuncRefT<T> a, FuncRefT<T> b) {
+    return a.toT()*b.toT();
+}
+template<typename T>
+T operator/(FuncRefT<T> a, FuncRefT<T> b) {
+    return a.toT()/b.toT();
+}
+template<typename T>
+T operator%(FuncRefT<T> a, FuncRefT<T> b) {
+    return a.toT()%b.toT();
 }
 
 template<typename T>
@@ -175,4 +223,4 @@ Halide::Expr operator>(T a, FuncRefT<T> b) {
     return a > static_cast<T>(b);
 }
 
-#endif
+#endif

Halide/src/CodeGen_C.cpp

@@ -1597,7 +1597,7 @@ class GatherKernelInfo : public IRVisitor {
     CodeGen_C* parent;
 public:
     vector<string> kernel_names;
-    
+
     GatherKernelInfo() {}
 
     void visit(const For *op) override {
@@ -1715,7 +1715,7 @@ void CodeGen_C::compile(const Module &input) {
             f.body.accept(&g);
         }
 
-        stream << "int MAX_DEVICES = 4;\n" 
+        stream << "int MAX_DEVICES = 4;\n"
                << "int NUM_QUEUES_TO_CREATE = " << g.kernel_names.size() << ";\n"
                << "int NUM_KERNELS_TO_CREATE = " << g.kernel_names.size() << ";\n"
                << "cl_int status;\n"
@@ -1724,13 +1724,13 @@ void CodeGen_C::compile(const Module &input) {
                << "cl_device_id devices[4];\n"
                << "int current_kernel = 0;\n"
                << "cl_kernel kernel[" << g.kernel_names.size() << "];\n\n";
-        
+
         stream << "const char *kernel_name[] = {\n";
         for (auto name : g.kernel_names) {
             stream << "    \"" << name << "\",\n";
         }
         stream << "};\n";
-        
+
     }
 
     for (const auto &b : input.buffers()) {
@@ -2803,7 +2803,7 @@ void CodeGen_C::visit(const For *op) {
                 stream << "sizeof(cl_mem), "
                        << "(void *)&((device_handle *)_halide_buffer_get_device(" << print_name(arg.name + ".buffer") << "))->mem";
             } else {
-                stream << "sizeof(" << print_type(arg.type) << "), " 
+                stream << "sizeof(" << print_type(arg.type) << "), "
                        << "(void *)&" << arg.name;
             }
             stream << ");\n"