1
0
mirror of https://github.com/godotengine/godot.git synced 2025-11-04 12:00:25 +00:00

Optimize CPU text shaping

This commit is contained in:
Mike Precup
2025-08-08 03:44:21 -07:00
parent 5950fca36c
commit 6fbb89df54
3 changed files with 133 additions and 40 deletions

View File

@@ -61,7 +61,8 @@ ScriptIterator::ScriptIterator(const String &p_string, int p_start, int p_length
} }
int paren_size = PAREN_STACK_DEPTH; int paren_size = PAREN_STACK_DEPTH;
ParenStackEntry *paren_stack = static_cast<ParenStackEntry *>(memalloc(paren_size * sizeof(ParenStackEntry))); ParenStackEntry starter_stack[PAREN_STACK_DEPTH];
ParenStackEntry *paren_stack = starter_stack;
int script_start; int script_start;
int script_end = p_start; int script_end = p_start;
@@ -78,7 +79,9 @@ ScriptIterator::ScriptIterator(const String &p_string, int p_start, int p_length
UChar32 n = (script_end + 1 < p_length) ? str[script_end + 1] : 0; UChar32 n = (script_end + 1 < p_length) ? str[script_end + 1] : 0;
UScriptCode sc = uscript_getScript(ch, &err); UScriptCode sc = uscript_getScript(ch, &err);
if (U_FAILURE(err)) { if (U_FAILURE(err)) {
memfree(paren_stack); if (paren_stack != starter_stack) {
memfree(paren_stack);
}
ERR_FAIL_MSG(u_errorName(err)); ERR_FAIL_MSG(u_errorName(err));
} }
if (is_emoji(ch, n)) { if (is_emoji(ch, n)) {
@@ -92,7 +95,11 @@ ScriptIterator::ScriptIterator(const String &p_string, int p_start, int p_length
if (unlikely(paren_sp >= paren_size)) { if (unlikely(paren_sp >= paren_size)) {
// If the stack is full, allocate more space to handle deeply nested parentheses. This is unlikely to happen with any real text. // If the stack is full, allocate more space to handle deeply nested parentheses. This is unlikely to happen with any real text.
paren_size += PAREN_STACK_DEPTH; paren_size += PAREN_STACK_DEPTH;
paren_stack = static_cast<ParenStackEntry *>(memrealloc(paren_stack, paren_size * sizeof(ParenStackEntry))); if (paren_stack == starter_stack) {
paren_stack = static_cast<ParenStackEntry *>(memalloc(paren_size * sizeof(ParenStackEntry)));
} else {
paren_stack = static_cast<ParenStackEntry *>(memrealloc(paren_stack, paren_size * sizeof(ParenStackEntry)));
}
} }
paren_stack[paren_sp].pair_index = ch; paren_stack[paren_sp].pair_index = ch;
paren_stack[paren_sp].script_code = script_code; paren_stack[paren_sp].script_code = script_code;
@@ -144,5 +151,7 @@ ScriptIterator::ScriptIterator(const String &p_string, int p_start, int p_length
script_ranges.push_back(rng); script_ranges.push_back(rng);
} while (script_end < p_length); } while (script_end < p_length);
memfree(paren_stack); if (paren_stack != starter_stack) {
memfree(paren_stack);
}
} }

View File

@@ -4275,7 +4275,13 @@ bool TextServerAdvanced::_font_is_script_supported(const RID &p_font_rid, const
Vector2i size = _get_size(fd, 16); Vector2i size = _get_size(fd, 16);
FontForSizeAdvanced *ffsd = nullptr; FontForSizeAdvanced *ffsd = nullptr;
ERR_FAIL_COND_V(!_ensure_cache_for_size(fd, size, ffsd), false); ERR_FAIL_COND_V(!_ensure_cache_for_size(fd, size, ffsd), false);
return fd->supported_scripts.has(hb_tag_from_string(p_script.ascii().get_data(), -1)); char ascii_script[] = { ' ', ' ', ' ', ' ' };
for (int i = 0; i < MIN(4, p_script.size()); i++) {
if (p_script[i] <= 0x7f) {
ascii_script[i] = p_script[i];
}
}
return fd->supported_scripts.has(hb_tag_from_string(ascii_script, -1));
} }
} }
@@ -5294,6 +5300,14 @@ bool TextServerAdvanced::_shape_substr(ShapedTextDataAdvanced *p_new_sd, const S
int32_t bidi_run_start = _convert_pos(p_sd, ov_start + start + _bidi_run_start); int32_t bidi_run_start = _convert_pos(p_sd, ov_start + start + _bidi_run_start);
int32_t bidi_run_end = _convert_pos(p_sd, ov_start + start + _bidi_run_start + _bidi_run_length); int32_t bidi_run_end = _convert_pos(p_sd, ov_start + start + _bidi_run_start + _bidi_run_length);
bool cache_valid = false;
int cached_font_size = -1;
RID cached_font_rid = RID();
double cached_font_ascent = 0;
double cached_font_descent = 0;
double cached_font_top_spacing = 0;
double cached_font_bottom_spacing = 0;
p_new_sd->glyphs.reserve(p_new_sd->glyphs.size() + MIN(sd_size, bidi_run_end - bidi_run_start));
for (int j = 0; j < sd_size; j++) { for (int j = 0; j < sd_size; j++) {
int col_key_off = (sd_glyphs[j].start == sd_glyphs[j].end) ? 1 : 0; int col_key_off = (sd_glyphs[j].start == sd_glyphs[j].end) ? 1 : 0;
if ((sd_glyphs[j].start >= bidi_run_start) && (sd_glyphs[j].end <= bidi_run_end - col_key_off)) { if ((sd_glyphs[j].start >= bidi_run_start) && (sd_glyphs[j].end <= bidi_run_end - col_key_off)) {
@@ -5346,20 +5360,32 @@ bool TextServerAdvanced::_shape_substr(ShapedTextDataAdvanced *p_new_sd, const S
} else { } else {
if (gl.font_rid.is_valid()) { if (gl.font_rid.is_valid()) {
if (p_new_sd->orientation == ORIENTATION_HORIZONTAL) { if (p_new_sd->orientation == ORIENTATION_HORIZONTAL) {
p_new_sd->ascent = MAX(p_new_sd->ascent, MAX(_font_get_ascent(gl.font_rid, gl.font_size) + _font_get_spacing(gl.font_rid, SPACING_TOP), -gl.y_off)); if (!cache_valid || cached_font_rid != gl.font_rid || cached_font_size != gl.font_size) {
p_new_sd->descent = MAX(p_new_sd->descent, MAX(_font_get_descent(gl.font_rid, gl.font_size) + _font_get_spacing(gl.font_rid, SPACING_BOTTOM), gl.y_off)); cache_valid = true;
cached_font_rid = gl.font_rid;
cached_font_size = gl.font_size;
cached_font_ascent = _font_get_ascent(gl.font_rid, gl.font_size);
cached_font_descent = _font_get_descent(gl.font_rid, gl.font_size);
cached_font_top_spacing = _font_get_spacing(gl.font_rid, SPACING_TOP);
cached_font_bottom_spacing = _font_get_spacing(gl.font_rid, SPACING_BOTTOM);
}
p_new_sd->ascent = MAX(p_new_sd->ascent, MAX(cached_font_ascent + cached_font_top_spacing, -gl.y_off));
p_new_sd->descent = MAX(p_new_sd->descent, MAX(cached_font_descent + cached_font_bottom_spacing, gl.y_off));
} else { } else {
p_new_sd->ascent = MAX(p_new_sd->ascent, Math::round(_font_get_glyph_advance(gl.font_rid, gl.font_size, gl.index).x * 0.5)); double glyph_advance = Math::round(_font_get_glyph_advance(gl.font_rid, gl.font_size, gl.index).x * 0.5);
p_new_sd->descent = MAX(p_new_sd->descent, Math::round(_font_get_glyph_advance(gl.font_rid, gl.font_size, gl.index).x * 0.5)); p_new_sd->ascent = MAX(p_new_sd->ascent, glyph_advance);
p_new_sd->descent = MAX(p_new_sd->descent, glyph_advance);
} }
} else if (p_new_sd->preserve_invalid || (p_new_sd->preserve_control && is_control(ch[gl.start - p_sd->start]))) { } else if (p_new_sd->preserve_invalid || (p_new_sd->preserve_control && is_control(ch[gl.start - p_sd->start]))) {
// Glyph not found, replace with hex code box. // Glyph not found, replace with hex code box.
if (p_new_sd->orientation == ORIENTATION_HORIZONTAL) { if (p_new_sd->orientation == ORIENTATION_HORIZONTAL) {
p_new_sd->ascent = MAX(p_new_sd->ascent, get_hex_code_box_size(gl.font_size, gl.index).y * 0.85); double box_size = get_hex_code_box_size(gl.font_size, gl.index).y;
p_new_sd->descent = MAX(p_new_sd->descent, get_hex_code_box_size(gl.font_size, gl.index).y * 0.15); p_new_sd->ascent = MAX(p_new_sd->ascent, box_size * 0.85);
p_new_sd->descent = MAX(p_new_sd->descent, box_size * 0.15);
} else { } else {
p_new_sd->ascent = MAX(p_new_sd->ascent, Math::round(get_hex_code_box_size(gl.font_size, gl.index).x * 0.5)); double box_size = Math::round(get_hex_code_box_size(gl.font_size, gl.index).x * 0.5);
p_new_sd->descent = MAX(p_new_sd->descent, Math::round(get_hex_code_box_size(gl.font_size, gl.index).x * 0.5)); p_new_sd->ascent = MAX(p_new_sd->ascent, box_size);
p_new_sd->descent = MAX(p_new_sd->descent, box_size);
} }
} }
p_new_sd->width += gl.advance * gl.repeat; p_new_sd->width += gl.advance * gl.repeat;
@@ -6649,10 +6675,9 @@ UBreakIterator *TextServerAdvanced::_create_line_break_iterator_for_locale(const
return ubrk_clone(bi, r_err); return ubrk_clone(bi, r_err);
} }
void TextServerAdvanced::_shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_start, int64_t p_end, hb_script_t p_script, hb_direction_t p_direction, TypedArray<RID> p_fonts, int64_t p_span, int64_t p_fb_index, int64_t p_prev_start, int64_t p_prev_end, RID p_prev_font) { void TextServerAdvanced::_shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_start, int64_t p_end, hb_script_t p_script, hb_direction_t p_direction, FontPriorityList &p_fonts, int64_t p_span, int64_t p_fb_index, int64_t p_prev_start, int64_t p_prev_end, RID p_prev_font) {
RID f; RID f;
int fs = p_sd->spans[p_span].font_size; int fs = p_sd->spans[p_span].font_size;
if (p_fb_index >= 0 && p_fb_index < p_fonts.size()) { if (p_fb_index >= 0 && p_fb_index < p_fonts.size()) {
// Try font from list. // Try font from list.
f = p_fonts[p_fb_index]; f = p_fonts[p_fb_index];
@@ -6697,7 +6722,7 @@ void TextServerAdvanced::_shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_star
} }
bool found = false; bool found = false;
for (int j = 0; j <= p_fonts.size(); j++) { for (uint32_t j = 0; j <= p_fonts.size(); j++) {
RID f_rid; RID f_rid;
if (j == p_fonts.size()) { if (j == p_fonts.size()) {
f_rid = p_prev_font; f_rid = p_prev_font;
@@ -6848,6 +6873,11 @@ void TextServerAdvanced::_shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_star
} }
} }
bool cache_valid = false;
RID cached_font_rid = RID();
int cached_font_size = 0;
float cached_offset = 0;
double adv_rem = 0.0; double adv_rem = 0.0;
for (unsigned int i = 0; i < glyph_count; i++) { for (unsigned int i = 0; i < glyph_count; i++) {
if ((i > 0) && (last_cluster_id != glyph_info[i].cluster)) { if ((i > 0) && (last_cluster_id != glyph_info[i].cluster)) {
@@ -6933,10 +6963,16 @@ void TextServerAdvanced::_shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_star
adv_rem = full_adv + gl.advance; adv_rem = full_adv + gl.advance;
} }
} }
if (!cache_valid || cached_font_rid != gl.font_rid || cached_font_size != gl.font_size) {
cache_valid = true;
cached_font_size = gl.font_size;
cached_font_rid = gl.font_rid;
cached_offset = _font_get_baseline_offset(gl.font_rid) * (double)(_font_get_ascent(gl.font_rid, gl.font_size) + _font_get_descent(gl.font_rid, gl.font_size));
}
if (p_sd->orientation == ORIENTATION_HORIZONTAL) { if (p_sd->orientation == ORIENTATION_HORIZONTAL) {
gl.y_off += _font_get_baseline_offset(gl.font_rid) * (double)(_font_get_ascent(gl.font_rid, gl.font_size) + _font_get_descent(gl.font_rid, gl.font_size)); gl.y_off += cached_offset;
} else { } else {
gl.x_off += _font_get_baseline_offset(gl.font_rid) * (double)(_font_get_ascent(gl.font_rid, gl.font_size) + _font_get_descent(gl.font_rid, gl.font_size)); gl.x_off += cached_offset;
} }
} }
if ((!last_run || i < last_non_zero_w) && !Math::is_zero_approx(gl.advance)) { if ((!last_run || i < last_non_zero_w) && !Math::is_zero_approx(gl.advance)) {
@@ -6978,6 +7014,7 @@ void TextServerAdvanced::_shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_star
failed_subrun_start = p_end + 1; failed_subrun_start = p_end + 1;
failed_subrun_end = p_start; failed_subrun_end = p_start;
} }
p_sd->glyphs.reserve(p_sd->glyphs.size() + w[i].count);
for (int j = 0; j < w[i].count; j++) { for (int j = 0; j < w[i].count; j++) {
if (p_sd->orientation == ORIENTATION_HORIZONTAL) { if (p_sd->orientation == ORIENTATION_HORIZONTAL) {
p_sd->ascent = MAX(p_sd->ascent, -w[i + j].y_off); p_sd->ascent = MAX(p_sd->ascent, -w[i + j].y_off);
@@ -7222,27 +7259,8 @@ bool TextServerAdvanced::_shaped_text_shape(const RID &p_shaped) {
} }
sd->glyphs.push_back(gl); sd->glyphs.push_back(gl);
} else { } else {
Array fonts; FontPriorityList fonts(this, span.fonts, span.language, script_code);
Array fonts_scr_only; _shape_run(sd, MAX(span.start - sd->start, script_run_start), MIN(span.end - sd->start, script_run_end), sd->script_iter->script_ranges[j].script, bidi_run_direction, fonts, k, 0, 0, 0, RID());
Array fonts_no_match;
int font_count = span.fonts.size();
if (font_count > 0) {
fonts.push_back(sd->spans[k].fonts[0]);
}
for (int l = 1; l < font_count; l++) {
if (_font_is_script_supported(span.fonts[l], script_code)) {
if (_font_is_language_supported(span.fonts[l], span.language)) {
fonts.push_back(sd->spans[k].fonts[l]);
} else {
fonts_scr_only.push_back(sd->spans[k].fonts[l]);
}
} else {
fonts_no_match.push_back(sd->spans[k].fonts[l]);
}
}
fonts.append_array(fonts_scr_only);
fonts.append_array(fonts_no_match);
_shape_run(sd, MAX(sd->spans[k].start - sd->start, script_run_start), MIN(sd->spans[k].end - sd->start, script_run_end), sd->script_iter->script_ranges[j].script, bidi_run_direction, fonts, k, 0, 0, 0, RID());
} }
} }
} }

View File

@@ -705,7 +705,73 @@ class TextServerAdvanced : public TextServerExtension {
int64_t _convert_pos(const ShapedTextDataAdvanced *p_sd, int64_t p_pos) const; int64_t _convert_pos(const ShapedTextDataAdvanced *p_sd, int64_t p_pos) const;
int64_t _convert_pos_inv(const ShapedTextDataAdvanced *p_sd, int64_t p_pos) const; int64_t _convert_pos_inv(const ShapedTextDataAdvanced *p_sd, int64_t p_pos) const;
bool _shape_substr(ShapedTextDataAdvanced *p_new_sd, const ShapedTextDataAdvanced *p_sd, int64_t p_start, int64_t p_length) const; bool _shape_substr(ShapedTextDataAdvanced *p_new_sd, const ShapedTextDataAdvanced *p_sd, int64_t p_start, int64_t p_length) const;
void _shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_start, int64_t p_end, hb_script_t p_script, hb_direction_t p_direction, TypedArray<RID> p_fonts, int64_t p_span, int64_t p_fb_index, int64_t p_prev_start, int64_t p_prev_end, RID p_prev_font);
struct FontPriorityList {
friend class TextServerAdvanced;
const int MAX_PRIORITY = 2;
int current_priority = 0;
uint32_t current_index = 0;
uint32_t font_count = 0;
const String *language;
const String *script_code;
LocalVector<Pair<RID, int>> unprocessed_fonts;
LocalVector<RID> fonts;
const TextServerAdvanced *text_server;
FontPriorityList(const TextServerAdvanced *p_text_server, const Array &p_fonts, const String &p_language, const String &p_script_code) {
text_server = p_text_server;
language = &p_language;
script_code = &p_script_code;
font_count = p_fonts.size();
unprocessed_fonts.reserve(font_count);
for (uint32_t i = 0; i < font_count; i++) {
unprocessed_fonts.push_back(Pair<RID, int>(p_fonts[i], -1));
}
fonts.reserve(font_count);
if (font_count > 0) {
fonts.push_back(p_fonts[0]);
unprocessed_fonts[0].second = 0;
current_index++;
}
}
_FORCE_INLINE_ uint32_t size() const {
return font_count;
}
_FORCE_INLINE_ int _get_priority(const RID &font) {
return text_server->_font_is_script_supported(font, *script_code) ? (text_server->_font_is_language_supported(font, *language) ? 0 : 1) : 2;
}
RID operator[](uint32_t index) {
if (index < fonts.size()) {
return fonts[index];
}
while (current_priority < MAX_PRIORITY || current_index < font_count) {
if (current_index >= font_count) {
current_priority++;
current_index = 0;
}
const RID &font = unprocessed_fonts[current_index].first;
int &priority = unprocessed_fonts[current_index].second;
if (priority < 0) {
priority = _get_priority(font);
}
if (priority == current_priority) {
fonts.push_back(font);
if (index < fonts.size()) {
return fonts[index];
}
}
current_index++;
}
return RID();
}
};
void _shape_run(ShapedTextDataAdvanced *p_sd, int64_t p_start, int64_t p_end, hb_script_t p_script, hb_direction_t p_direction, FontPriorityList &p_fonts, int64_t p_span, int64_t p_fb_index, int64_t p_prev_start, int64_t p_prev_end, RID p_prev_font);
Glyph _shape_single_glyph(ShapedTextDataAdvanced *p_sd, char32_t p_char, hb_script_t p_script, hb_direction_t p_direction, const RID &p_font, int64_t p_font_size); Glyph _shape_single_glyph(ShapedTextDataAdvanced *p_sd, char32_t p_char, hb_script_t p_script, hb_direction_t p_direction, const RID &p_font, int64_t p_font_size);
_FORCE_INLINE_ RID _find_sys_font_for_text(const RID &p_fdef, const String &p_script_code, const String &p_language, const String &p_text); _FORCE_INLINE_ RID _find_sys_font_for_text(const RID &p_fdef, const String &p_script_code, const String &p_language, const String &p_text);