Add Cython-accelerated row-batch decoding

.circleci/config.yml

@@ -41,8 +41,10 @@ jobs:
     steps:
       - checkout
       - run:
-          name: Install make
-          command: dnf install make -y
+          name: Install build tools
+          command: |
+            PYVER=$(python3 -c 'import sys; print(f"{sys.version_info.major}.{sys.version_info.minor}")')
+            dnf install -y make gcc "python${PYVER}-devel"
       - run:
           name: Install pip
           command: |

.gitignore

@@ -9,6 +9,9 @@
 
 *.so
 
+# Cython-generated C source (rebuilt from .pyx by setup.py)
+pynuodb/_fetch.c
+
 /.virttemp
 /.testtemp
 

MANIFEST.in

@@ -1 +1,2 @@
-include README.rst LICENSE
+include README.rst LICENSE
+recursive-include pynuodb *.pyx *.pxd *.pxi *.c *.h

Makefile

@@ -32,6 +32,8 @@ SUDO ?= sudo -n
 NUODB_HOME ?= /opt/nuodb
 
 _INSTALL_CMD =	$(PIP) install '.[crypto]'
+_BUILD_EXT_CMD = $(PIP) install 'setuptools>=40.8.0' 'Cython>=3.0' \
+			&& $(PYTHON) setup.py build_ext --inplace
 _VERIFY_CMD =	$(NUODB_HOME)/bin/nuocmd show domain
 _PYTEST_CMD =	$(MKDIR) $(ARTIFACTDIR) $(RESULTSDIR) \
 		&& TMPDIR='$(TMPDIR)' PATH="$(NUODB_HOME)/bin:$$PATH" \
@@ -43,11 +45,13 @@ all:
 
 install:
 	$(_INSTALL_CMD)
+	$(_BUILD_EXT_CMD)
 
 check: mypy pylint fulltest
 
 fulltest:
 	$(_INSTALL_CMD)
+	$(_BUILD_EXT_CMD)
 	$(PIP) install -r test_requirements.txt
 	$(_VERIFY_CMD)
 	$(_PYTEST_CMD)

pynuodb/_codes.pxi

@@ -0,0 +1,47 @@
+# Wire-protocol type codes, mirrored from protocol.py as C ints so
+# decode_next_batch's per-cell range checks in _fetch.pyx don't pay for a
+# Python attribute lookup on every column.
+from . import protocol as _protocol
+
+cdef int NULL_V                  = _protocol.NULL
+cdef int TRUE_V                  = _protocol.TRUE
+cdef int FALSE_V                 = _protocol.FALSE
+cdef int INTMINUS10              = _protocol.INTMINUS10
+cdef int INT0                    = _protocol.INT0  # inline value 0: end-of-batch marker
+cdef int INT31                   = _protocol.INT31
+cdef int INTLEN0                 = _protocol.INTLEN0  # == INT31; codes 52-59 carry 1-8 byte integers
+cdef int INTLEN8                 = _protocol.INTLEN8
+cdef int SCALEDLEN0              = _protocol.SCALEDLEN0  # base for 1-8 byte scaled decimals (61-68)
+cdef int SCALEDLEN8              = _protocol.SCALEDLEN8
+cdef int UTF8COUNT0              = _protocol.UTF8COUNT0  # base for length-prefixed strings (69-72)
+cdef int UTF8COUNT1              = _protocol.UTF8COUNT1
+cdef int UTF8COUNT4              = _protocol.UTF8COUNT4
+cdef int OPAQUECOUNT0            = _protocol.OPAQUECOUNT0  # base for length-prefixed binary (73-76)
+cdef int OPAQUECOUNT1            = _protocol.OPAQUECOUNT1
+cdef int OPAQUECOUNT4            = _protocol.OPAQUECOUNT4
+cdef int DOUBLELEN0              = _protocol.DOUBLELEN0  # 77 == double 0.0; 78-85 carry 1-8 byte doubles
+cdef int DOUBLELEN8              = _protocol.DOUBLELEN8
+cdef int MILLISECLEN0            = _protocol.MILLISECLEN0  # base for 1-8 byte millisecond timestamps
+cdef int MILLISECLEN8            = _protocol.MILLISECLEN8
+cdef int NANOSECLEN0             = _protocol.NANOSECLEN0  # base for 1-8 byte nanosecond timestamps
+cdef int NANOSECLEN8             = _protocol.NANOSECLEN8
+cdef int TIMELEN0                = _protocol.TIMELEN0  # base for 1-4 byte ms-since-midnight
+cdef int TIMELEN4                = _protocol.TIMELEN4
+cdef int UTF8LEN0                = _protocol.UTF8LEN0  # base for 0-39 byte inline-length strings
+cdef int UTF8LEN39               = _protocol.UTF8LEN39
+cdef int OPAQUELEN0              = _protocol.OPAQUELEN0  # base for 0-39 byte inline-length binary
+cdef int OPAQUELEN39             = _protocol.OPAQUELEN39
+cdef int BLOBLEN0                = _protocol.BLOBLEN0  # base for 0-4 byte length-prefixed BLOB
+cdef int BLOBLEN4                = _protocol.BLOBLEN4
+cdef int CLOBLEN0                = _protocol.CLOBLEN0  # base for 0-4 byte length-prefixed CLOB
+cdef int CLOBLEN4                = _protocol.CLOBLEN4
+cdef int UUID_C                  = _protocol.UUID
+cdef int SCALEDDATELEN0          = _protocol.SCALEDDATELEN0  # 201-208 carry 1-8 byte scaled dates
+cdef int SCALEDDATELEN8          = _protocol.SCALEDDATELEN8
+cdef int SCALEDTIMELEN0          = _protocol.SCALEDTIMELEN0  # 209-216 carry 1-8 byte scaled times
+cdef int SCALEDTIMELEN8          = _protocol.SCALEDTIMELEN8
+cdef int SCALEDTIMESTAMPLEN0     = _protocol.SCALEDTIMESTAMPLEN0  # 217-224 carry 1-8 byte scaled timestamps
+cdef int SCALEDTIMESTAMPLEN8     = _protocol.SCALEDTIMESTAMPLEN8
+# No LEN1..LEN8 range: 234-240 belong to ARRAYLEN/SCALEDCOUNT3/DEBUGBARRIER.
+cdef int SCALEDTIMESTAMPNOTZLEN0 = _protocol.SCALEDTIMESTAMPNOTZLEN0
+cdef int SCALEDTIMESTAMPNOTZ     = _protocol.SCALEDTIMESTAMPNOTZ

pynuodb/_cutil.h

@@ -0,0 +1,135 @@
+#ifndef PYNUODB_CUTIL_H
+#define PYNUODB_CUTIL_H
+/* C support for _fetch.pyx's decode loop: raw-PyObject* value builders for
+   the row tuple, and big-endian integer/double readers for the wire
+   protocol. */
+
+#include <Python.h>
+#include <string.h>
+#include <stdint.h>
+
+#if defined(_MSC_VER)
+#include <stdlib.h>
+#define NUODB_INLINE static __inline
+#define NUODB_LIKELY(x) (x)
+#define NUODB_UNLIKELY(x) (x)
+#define NUODB_BSWAP16(x) _byteswap_ushort(x)
+#define NUODB_BSWAP32(x) _byteswap_ulong(x)
+#define NUODB_BSWAP64(x) _byteswap_uint64(x)
+#else
+#define NUODB_INLINE static inline
+#define NUODB_LIKELY(x) __builtin_expect(!!(x), 1)
+#define NUODB_UNLIKELY(x) __builtin_expect(!!(x), 0)
+#define NUODB_BSWAP16(x) __builtin_bswap16(x)
+#define NUODB_BSWAP32(x) __builtin_bswap32(x)
+#define NUODB_BSWAP64(x) __builtin_bswap64(x)
+#endif
+
+/* Build a value as a raw PyObject* for decode_next_batch to steal into a
+   row tuple via _pynuodb_tuple_steal. Don't replace these with Cython's
+   cpython.* pxd bindings: those take/return `object`, not `PyObject*`,
+   which would route the value through Cython's own refcounting first. */
+NUODB_INLINE void _pynuodb_tuple_steal(PyObject *t, Py_ssize_t i, PyObject *o)
+{
+    PyTuple_SET_ITEM(t, i, o);
+}
+
+NUODB_INLINE PyObject *_pynuodb_long_from_long(long v)
+{
+    return PyLong_FromLong(v);
+}
+
+NUODB_INLINE PyObject *_pynuodb_long_from_longlong(long long v)
+{
+    return PyLong_FromLongLong(v);
+}
+
+NUODB_INLINE PyObject *_pynuodb_decode_utf8(const char *s, Py_ssize_t n)
+{
+    return PyUnicode_DecodeUTF8(s, n, NULL);
+}
+
+NUODB_INLINE PyObject *_pynuodb_incref(PyObject *o)
+{
+    Py_INCREF(o);
+    return o;
+}
+
+NUODB_INLINE PyObject *_pynuodb_none(void) { return Py_None; }
+NUODB_INLINE PyObject *_pynuodb_true(void) { return Py_True; }
+NUODB_INLINE PyObject *_pynuodb_false(void) { return Py_False; }
+
+NUODB_INLINE unsigned long long _pynuodb_be_u64(const unsigned char *p, int n) 
+{
+    switch (n)
+    {
+        case 0: return 0;
+        case 1: return p[0];
+        case 2: {
+            uint16_t x;
+            memcpy(&x, p, 2);
+            return NUODB_BSWAP16(x);
+        }
+        case 4: {
+            uint32_t x;
+            memcpy(&x, p, 4);
+            return NUODB_BSWAP32(x);
+        }
+        case 8: {
+            uint64_t x;
+            memcpy(&x, p, 8);
+            return NUODB_BSWAP64(x);
+        }
+        default: {
+            unsigned long long v = 0;
+            for (int i = 0; i < n; i++) {
+                v = (v << 8) | p[i];
+            }
+            return v;
+        }
+    }
+}
+
+/* Sign-extends the n-byte big-endian unsigned value from _pynuodb_be_u64*/
+NUODB_INLINE long long _pynuodb_be_i64(const unsigned char *p, int n)
+{
+    if (n <= 0) {
+        return 0;
+    }
+    unsigned long long v = _pynuodb_be_u64(p, n);
+    unsigned long long m = (n < 8) ? (1ULL << ((n << 3) - 1)) : 0x8000000000000000ULL;
+    return (long long)((v ^ m) - m);
+}
+
+NUODB_INLINE double _pynuodb_be_double(const unsigned char *p, int n)
+{
+    unsigned char buf[8] = {0};
+    if (n < 0) {
+        n = 0;
+    } else if (n > 8) {
+        n = 8;
+    }
+    memcpy(buf, p, n);
+    uint64_t v = _pynuodb_be_u64(buf, 8);
+    double d;
+    memcpy(&d, &v, 8);
+    return d;
+}
+
+/* Try to make boundary checking fast through inline and marking the happy path as likely*/
+NUODB_INLINE int _pynuodb_avail_ok(Py_ssize_t pos, Py_ssize_t need, Py_ssize_t n)
+{
+    return NUODB_LIKELY(need >= 0 && pos <= n - need);
+}
+
+NUODB_INLINE PyObject *_pynuodb_pylong_be_signed(const unsigned char *p, Py_ssize_t n)
+{
+    if (NUODB_UNLIKELY(n > 8)) {
+        PyErr_SetString(PyExc_OverflowError,
+            "_pynuodb_pylong_be_signed: width > 8 bytes not supported");
+        return NULL;
+    }
+    return PyLong_FromLongLong(_pynuodb_be_i64(p, (int)n));
+}
+
+#endif /* PYNUODB_CUTIL_H */