Refactor: migrate pdf_parser.py to golang (#16323)

### What problem does this PR solve?

Http API based on onnx model.
pdf_parser.py to golang

### Type of change

- [x] Refactoring
This commit is contained in:
Jack
2026-06-25 20:16:16 +08:00
committed by GitHub
parent c7052f4dd1
commit 304d9e02bb
98 changed files with 24591 additions and 8 deletions

View File

@@ -0,0 +1,165 @@
// Package pdfium renders PDF pages using the system's libpdfium.so
// (bundled with pypdfium2). It exists solely to replace pdf_oxide's
// RenderPageRaw for use cases where image quality matters for downstream
// OCR/DLA — pdf_oxide still handles all text/char/table extraction.
package pdfium
/*
#cgo LDFLAGS: -L/home/shenyushi/cc-workspace/ragflow/.venv/lib/python3.13/site-packages/pypdfium2_raw -lpdfium -lm -lpthread -ldl
#cgo linux LDFLAGS: -Wl,-rpath,/home/shenyushi/cc-workspace/ragflow/.venv/lib/python3.13/site-packages/pypdfium2_raw
#include <stdint.h>
#include <stdlib.h>
typedef struct FPDF_DOCUMENT__ { int unused; } *FPDF_DOCUMENT;
typedef struct FPDF_PAGE__ { int unused; } *FPDF_PAGE;
typedef struct FPDF_BITMAP__ { int unused; } *FPDF_BITMAP;
extern void FPDF_InitLibrary(void);
extern FPDF_DOCUMENT FPDF_LoadMemDocument(const void* data_buf, int size, const char* password);
extern void FPDF_CloseDocument(FPDF_DOCUMENT document);
extern int FPDF_GetPageCount(FPDF_DOCUMENT document);
extern FPDF_PAGE FPDF_LoadPage(FPDF_DOCUMENT document, int page_index);
extern void FPDF_ClosePage(FPDF_PAGE page);
extern double FPDF_GetPageWidth(FPDF_PAGE page);
extern double FPDF_GetPageHeight(FPDF_PAGE page);
extern FPDF_BITMAP FPDFBitmap_Create(int width, int height, int alpha);
extern void FPDFBitmap_Destroy(FPDF_BITMAP bitmap);
extern void FPDF_RenderPageBitmap(FPDF_BITMAP bitmap, FPDF_PAGE page,
int start_x, int start_y, int size_x, int size_y,
int rotate, int flags);
extern void* FPDFBitmap_GetBuffer(FPDF_BITMAP bitmap);
extern int FPDFBitmap_GetWidth(FPDF_BITMAP bitmap);
extern int FPDFBitmap_GetHeight(FPDF_BITMAP bitmap);
extern int FPDFBitmap_GetStride(FPDF_BITMAP bitmap);
*/
import "C"
import (
"fmt"
"image"
"image/color"
"math"
"sync"
"unsafe"
)
var initOnce sync.Once
// pdfiumMu serializes all pdfium C API access. pdfium is NOT thread-safe —
// concurrent calls to FPDF_LoadPage / FPDF_RenderPageBitmap corrupt the
// global heap, causing SIGSEGV. See TestPdfiumConcurrentSafety.
var pdfiumMu sync.Mutex
// Init initializes the PDFium library. Safe to call multiple times.
func Init() { initOnce.Do(func() { C.FPDF_InitLibrary() }) }
// PageSize returns the page dimensions in PDF points (1/72 inch) as seen
// after rotation. For a page with /Rotate 90 on A4, this returns ~842×595
// (swapped from the MediaBox 595×842). The call is cheap — it opens the
// document and page, reads dimensions, then closes.
func PageSize(pdfData []byte, pageIdx int) (width, height float64, err error) {
Init()
pdfiumMu.Lock()
defer pdfiumMu.Unlock()
_, _, pw, ph, closeAll, err := openPage(pdfData, pageIdx)
if err != nil {
return 0, 0, err
}
closeAll()
return pw, ph, nil
}
// RenderPage renders a single page of a PDF to an *image.RGBA at the given DPI.
// pdfData is the raw PDF bytes, pageIdx is 0-based.
func RenderPage(pdfData []byte, pageIdx int, dpi float64) (*image.RGBA, error) {
Init()
pdfiumMu.Lock()
defer pdfiumMu.Unlock()
_, page, pw, ph, closeAll, err := openPage(pdfData, pageIdx)
if err != nil {
return nil, err
}
defer closeAll()
scale := dpi / 72.0
pxW := int(math.Round(pw * scale))
pxH := int(math.Round(ph * scale))
bitmap := C.FPDFBitmap_Create(C.int(pxW), C.int(pxH), 1) // 1 = RGBA
if bitmap == nil {
return nil, fmt.Errorf("pdfium: FPDFBitmap_Create(%d,%d) returned nil", pxW, pxH)
}
defer C.FPDFBitmap_Destroy(bitmap)
// Fill with opaque white before rendering, so transparent areas
// (e.g. outside crop box) are white rather than undefined.
stride := int(C.FPDFBitmap_GetStride(bitmap))
buf := C.FPDFBitmap_GetBuffer(bitmap)
pixels := (*[1 << 30]byte)(unsafe.Pointer(buf))[: pxH*stride : pxH*stride]
for i := range pixels {
pixels[i] = 255
}
// FPDF_ANNOT (0x01) — render annotations.
// LCD text AA (0x02) is left off; default text smoothing is sufficient.
C.FPDF_RenderPageBitmap(bitmap, page, 0, 0, C.int(pxW), C.int(pxH), 0, 0x01)
// pdfium outputs BGRA; convert to RGBA.
img := image.NewRGBA(image.Rect(0, 0, pxW, pxH))
for y := 0; y < pxH; y++ {
for x := 0; x < pxW; x++ {
off := y*stride + x*4
img.SetRGBA(x, y, color.RGBA{
R: pixels[off+2], // B
G: pixels[off+1], // G
B: pixels[off], // R
A: 255,
})
}
}
return img, nil
}
// openPage opens a document and page, returning post-rotation dimensions
// and a cleanup function. Callers must call closeAll() to free resources.
func openPage(pdfData []byte, pageIdx int) (
doc C.FPDF_DOCUMENT,
page C.FPDF_PAGE,
pw, ph float64,
closeAll func(),
err error,
) {
cData := C.CBytes(pdfData)
doc = C.FPDF_LoadMemDocument(unsafe.Pointer(cData), C.int(len(pdfData)), nil)
if doc == nil {
C.free(cData)
err = fmt.Errorf("pdfium: FPDF_LoadMemDocument returned nil")
return
}
page = C.FPDF_LoadPage(doc, C.int(pageIdx))
if page == nil {
C.FPDF_CloseDocument(doc)
C.free(cData)
err = fmt.Errorf("pdfium: FPDF_LoadPage(%d) returned nil", pageIdx)
return
}
pw = float64(C.FPDF_GetPageWidth(page))
ph = float64(C.FPDF_GetPageHeight(page))
if pw <= 0 || ph <= 0 {
C.FPDF_ClosePage(page)
C.FPDF_CloseDocument(doc)
C.free(cData)
err = fmt.Errorf("pdfium: invalid page dimensions %.1fx%.1f", pw, ph)
return
}
closeAll = func() {
C.FPDF_ClosePage(page)
C.FPDF_CloseDocument(doc)
C.free(cData)
}
return
}

View File

@@ -0,0 +1,241 @@
package pdfium
import (
"image"
"math"
"os"
"path/filepath"
"sync"
"testing"
)
// testdataDir points at the shared test-pdf directory.
var testdataDir = filepath.Join("..", "parser", "testdata", "pdfs")
func readPDF(t *testing.T, name string) []byte {
t.Helper()
data, err := os.ReadFile(filepath.Join(testdataDir, name))
if err != nil {
t.Fatalf("read %s: %v", name, err)
}
return data
}
func TestRenderPage_EnglishSimple(t *testing.T) {
data := readPDF(t, "01_english_simple.pdf")
img, err := RenderPage(data, 0, 72)
if err != nil {
t.Fatal(err)
}
b := img.Bounds()
t.Logf("01_english_simple.pdf @ 72 DPI: %dx%d", b.Dx(), b.Dy())
if b.Dx() <= 0 || b.Dy() <= 0 {
t.Errorf("expected non-zero dimensions, got %dx%d", b.Dx(), b.Dy())
}
// Must not be pure white (text should be present).
if isPureWhite(img) {
t.Error("rendered page is pure white — expected text content")
}
}
func TestRenderPage_ChineseSimple(t *testing.T) {
data := readPDF(t, "02_chinese_simple.pdf")
img, err := RenderPage(data, 0, 72)
if err != nil {
t.Fatal(err)
}
b := img.Bounds()
t.Logf("02_chinese_simple.pdf @ 72 DPI: %dx%d", b.Dx(), b.Dy())
if b.Dx() <= 0 || b.Dy() <= 0 {
t.Errorf("expected non-zero dimensions, got %dx%d", b.Dx(), b.Dy())
}
if isPureWhite(img) {
t.Error("rendered page is pure white — expected text content")
}
}
func TestRenderPage_MultiPage(t *testing.T) {
data := readPDF(t, "03_multipage.pdf")
// Render both pages.
for pg := 0; pg < 2; pg++ {
img, err := RenderPage(data, pg, 72)
if err != nil {
t.Fatalf("page %d: %v", pg, err)
}
b := img.Bounds()
t.Logf("03_multipage.pdf page %d @ 72 DPI: %dx%d", pg, b.Dx(), b.Dy())
if b.Dx() <= 0 || b.Dy() <= 0 {
t.Errorf("page %d: expected non-zero dimensions", pg)
}
}
}
func TestRenderPage_OutOfRange(t *testing.T) {
data := readPDF(t, "01_english_simple.pdf")
_, err := RenderPage(data, 99, 72)
if err == nil {
t.Error("expected error for out-of-range page index")
}
}
func TestRenderPage_InvalidPDF(t *testing.T) {
_, err := RenderPage([]byte("not a pdf"), 0, 72)
if err == nil {
t.Error("expected error for invalid PDF data")
}
}
func TestRenderPage_EmptyData(t *testing.T) {
_, err := RenderPage(nil, 0, 72)
if err == nil {
t.Error("expected error for nil data")
}
_, err = RenderPage([]byte{}, 0, 72)
if err == nil {
t.Error("expected error for empty data")
}
}
func TestRenderPage_DPI(t *testing.T) {
data := readPDF(t, "01_english_simple.pdf")
// Higher DPI → larger image.
low, err := RenderPage(data, 0, 72)
if err != nil {
t.Fatal(err)
}
high, err := RenderPage(data, 0, 144)
if err != nil {
t.Fatal(err)
}
lw, lh := low.Bounds().Dx(), low.Bounds().Dy()
hw, hh := high.Bounds().Dx(), high.Bounds().Dy()
t.Logf("72 DPI: %dx%d 144 DPI: %dx%d", lw, lh, hw, hh)
if hw < lw*2-2 || hw > lw*2+2 {
t.Errorf("144 DPI width %d not ≈ 2× 72 DPI width %d", hw, lw)
}
if hh < lh*2-2 || hh > lh*2+2 {
t.Errorf("144 DPI height %d not ≈ 2× 72 DPI height %d", hh, lh)
}
}
func TestRenderPage_AllTestPDFs(t *testing.T) {
entries, err := os.ReadDir(testdataDir)
if err != nil {
t.Skipf("testdata dir not found: %v", err)
}
for _, e := range entries {
if e.IsDir() || filepath.Ext(e.Name()) != ".pdf" {
continue
}
data, err := os.ReadFile(filepath.Join(testdataDir, e.Name()))
if err != nil {
t.Errorf("%s: read: %v", e.Name(), err)
continue
}
img, err := RenderPage(data, 0, 72)
if err != nil {
t.Errorf("%s: RenderPage: %v", e.Name(), err)
continue
}
b := img.Bounds()
if b.Dx() <= 0 || b.Dy() <= 0 {
t.Errorf("%s: zero dimensions %dx%d", e.Name(), b.Dx(), b.Dy())
}
t.Logf("%s: %dx%d", e.Name(), b.Dx(), b.Dy())
}
}
func isPureWhite(img image.Image) bool {
b := img.Bounds()
for y := b.Min.Y; y < b.Max.Y; y++ {
for x := b.Min.X; x < b.Max.X; x++ {
r, g, b, _ := img.At(x, y).RGBA()
// RGBA() returns premultiplied values in [0, 65535].
if r>>8 < 250 || g>>8 < 250 || b>>8 < 250 {
return false
}
}
}
return true
}
func TestPageSize(t *testing.T) {
// Non-rotated A4: expect ~595×842
data := readPDF(t, "rotate_0.pdf")
w, h, err := PageSize(data, 0)
if err != nil {
t.Fatal(err)
}
if w < 500 || w > 700 || h < 700 || h > 900 {
t.Errorf("rotate_0.pdf: got %.1f×%.1f, want ~595×842", w, h)
}
t.Logf("rotate_0.pdf: %.1f×%.1f pts", w, h)
// Rotate=90 A4: expect swapped ~842×595
data90 := readPDF(t, "rotate_90.pdf")
w90, h90, err := PageSize(data90, 0)
if err != nil {
t.Fatal(err)
}
if w90 < 700 || w90 > 950 || h90 < 500 || h90 > 700 {
t.Errorf("rotate_90.pdf: got %.1f×%.1f, want ~842×595 (swapped)", w90, h90)
}
t.Logf("rotate_90.pdf: %.1f×%.1f pts (post-rotation)", w90, h90)
// Verify dimensions ARE swapped relative to Rotate=0
if math.Abs(w-w90) < 50 {
t.Errorf("Rotate=90 width %.1f not significantly different from Rotate=0 width %.1f — rotation not reflected?", w90, w)
}
if math.Abs(w-h90) > 2 || math.Abs(h-w90) > 2 {
t.Errorf("Rotate=90 dimensions (%.1f×%.1f) are not swapped from Rotate=0 (%.1f×%.1f)", w90, h90, w, h)
}
// Invalid page index
_, _, err = PageSize(data, 999)
if err == nil {
t.Error("expected error for out-of-range page")
}
// Empty data
_, _, err = PageSize([]byte{}, 0)
if err == nil {
t.Error("expected error for empty PDF data")
}
}
// TestPdfiumConcurrentSafety verifies that the pdfiumMu mutex prevents
// SIGSEGV from concurrent pdfium access. Without the mutex, 10 goroutines
// calling PageSize/RenderPage simultaneously causes heap corruption within
// milliseconds (empirically proven). If this test completes without
// crashing, the mutex is working.
func TestPdfiumConcurrentSafety(t *testing.T) {
data := readPDF(t, "01_english_simple.pdf")
const goroutines = 10
const iterations = 3
var wg sync.WaitGroup
for i := 0; i < goroutines; i++ {
wg.Add(1)
go func() {
defer wg.Done()
for j := 0; j < iterations; j++ {
if _, _, err := PageSize(data, 0); err != nil {
t.Errorf("PageSize: %v", err)
return
}
if img, err := RenderPage(data, 0, 72); err != nil {
t.Errorf("RenderPage: %v", err)
return
} else if img.Bounds().Dx() <= 0 {
t.Error("RenderPage returned zero-width image")
return
}
}
}()
}
wg.Wait()
// Reaching here without SIGSEGV = mutex is effective.
}