Fengttt cl (#18212)

## What type of PR is this?

- [ ] API-change
- [ ] BUG
- [ ] Improvement
- [ ] Documentation
- [x] Feature
- [ ] Test and CI
- [ ] Code Refactoring

## Which issue(s) this PR fixes:

issue #18213

## What this PR does / why we need it:

Added a generic call method, which can let function framework to issue
generic RPC calls to a cl_host.
Only a special cl_host, "CGO", is implemented at this time. CGO is
treated as an special cl_host so there
is no need for memcpy or network RPC. Function calls can have a
cl_runtime, this PR support "C", and "CUDA".

l2distance of f32/f64 version are used as example to show how to offload
the computing to "C" or "CUDA".
For "C", the performance is roughly the same as pure go (so you may ask
why should we do this ...). Same
for vecf32(128) and 10%-15% faster on vecf32(960).

"CUDA" as a demo, works. Build and tested on 4090, wsl. The PR has all
the infrastructure to send data from
database to GPU, run GPU kernel and return result. CUDA shines when some
real computing intensive workload.
Deep network, bitcoin, image, ...   All ready to plugin.

To build: 
A simple `make clean; make` will build "CGO" cl_host and "C" cl_runtime.
`MO_CL_CUDA=1 make` will build "CUDA" cl_runtime. You must have cuda
installed already.

Author: fengttt

.gitignore

@@ -4,6 +4,7 @@
 *.exe
 *.lib
 *.swp
+*.fatbin
 bin/
 y.go
 *.output

Makefile

@@ -110,8 +110,13 @@ THIRDPARTIES_INSTALL_DIR=$(ROOT_DIR)/thirdparties/install
 RACE_OPT :=
 DEBUG_OPT :=
 CGO_DEBUG_OPT :=
+
+ifeq ($(MO_CL_CUDA),1)
+	CUDA_LDFLAGS := -L/usr/local/cuda/lib64/stubs -lcuda -L/usr/local/cuda/lib64 -lcudart -lstdc++
+endif
+
 CGO_OPTS :=CGO_CFLAGS="-I$(THIRDPARTIES_INSTALL_DIR)/include"
-GOLDFLAGS=-ldflags="-extldflags '-L$(THIRDPARTIES_INSTALL_DIR)/lib -Wl,-rpath,\$${ORIGIN}/lib' $(VERSION_INFO)"
+GOLDFLAGS=-ldflags="-extldflags '$(CUDA_LDFLAGS) -L$(THIRDPARTIES_INSTALL_DIR)/lib -Wl,-rpath,\$${ORIGIN}/lib' $(VERSION_INFO)"
 TAGS :=
 
 ifeq ("$(UNAME_S)","darwin")

cgo/Makefile

@@ -1,12 +1,32 @@
 DEBUG_OPT :=
-OPT_LV := -O3
+
+# Yeah, fast math.  We want it to be fast, for all xcall, 
+# IEEE compliance should not be an issue.
+OPT_LV := -O3 -ffast-math
 CFLAGS=-std=c99 -g ${OPT_LV} -Wall -Werror
-OBJS=mo.o arith.o compare.o logic.o
+OBJS=mo.o arith.o compare.o logic.o xcall.o
+CUDA_OBJS=
+
+ifeq ($(MO_CL_CUDA),1)
+	CC = /usr/local/cuda/bin/nvcc 
+	CFLAGS = -ccbin g++ -m64 --shared -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_89,code=sm_89 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_90,code=compute_90
+	CFLAGS += -DMO_CL_CUDA
+	CUDA_OBJS += cuda/cuda.o
+	CUDA_LDFLAGS := -L/usr/local/cuda/lib64/stubs -lcuda -L/usr/local/cuda/lib64 -lcudart -lstdc++
+endif
 
 all: libmo.a
 
-libmo.a: $(OBJS)
-	ar -rcs libmo.a *.o
+libmo.a: $(OBJS) 
+ifeq ($(MO_CL_CUDA),1)
+	make -C cuda
+endif
+	ar -rcs libmo.a $(OBJS) $(CUDA_OBJS)
+
+# 
+#	$(CC) -o libmo.a $(OBJS) $(CUDA_OBJS) $(CUDA_LDFLAGS)
+
+
 
 .PHONY: debug
 debug: override OPT_LV := -O0
@@ -15,4 +35,5 @@ debug: all
 
 .PHONY: clean
 clean:
-	rm -f *.o *.a
+	rm -f *.o *.a *.so
+	make -C cuda clean

cgo/bitmap.h

@@ -17,7 +17,6 @@
 #ifndef _BITMAP_H_
 #define _BITMAP_H_
 
-#include "mo.h"
 #include <string.h>
 
 /*
@@ -46,6 +45,9 @@ static inline uint64_t bitmap_pos2mask(uint64_t pos) {
 }
 
 static inline bool bitmap_test(uint64_t *p, uint64_t pos) {
+    if (p == NULL) {
+        return false;
+    }
     return (p[bitmap_pos2idx(pos)] & bitmap_pos2mask(pos)) != 0;
 }
 

cgo/cuda/.gitignore

@@ -0,0 +1,2 @@
+mytest
+*.fatbin