Improve my own performance hack

dn/kit-browser.cpp

@@ -319,7 +319,7 @@ profile_from_icc(dawn::Cmm &cmm, const shared_ptr<const vector<uint8_t>> &icc)
 		if (auto profile = cmm.get_profile(*icc))
 			return profile;
 	}
-	return cmm.get_profile_sRGB(false);
+	return cmm.get_profile_sRGB();
 }
 
 static int
@@ -590,7 +590,7 @@ make_thumb(shared_ptr<dawn::Cmm> cmm, const ThumbJob &job)
 	if (job.pending) {
 		if (const ThumbnailTierPixels *pixels = job.pending->find(tier)) {
 			result.ram = *pixels->pixels;
-			shared_ptr<dawn::Profile> p3 = cmm->get_profile_display_p3(false);
+			shared_ptr<dawn::Profile> p3 = cmm->get_profile_display_p3();
 			if (p3 && screen && !result.ram.empty() &&
 				cmm->transform_bgra16(
 					reinterpret_cast<uint8_t *>(result.ram.data()),
@@ -639,8 +639,7 @@ make_thumb(shared_ptr<dawn::Cmm> cmm, const ThumbJob &job)
 	dawn::OpenContext ctx;
 	ctx.uri = job.path;
 	ctx.cmm = cmm;
-	ctx.screen_profile =
-		job.cacheable ? cmm->get_profile_display_p3(false) : screen;
+	ctx.screen_profile = job.cacheable ? cmm->get_profile_display_p3() : screen;
 	ctx.first_frame_only = true;
 	ctx.screen_dpi = 96;
 
@@ -699,6 +698,11 @@ static shared_ptr<dawn::Cmm>
 worker_cmm()
 {
 	thread_local auto cmm = make_shared<dawn::Cmm>();
+
+	// Cmm memoizes these only for as long as somebody holds one, and
+	// every cache read and every display transform asks for them again.
+	thread_local auto srgb = cmm->get_profile_sRGB();
+	thread_local auto display_p3 = cmm->get_profile_display_p3();
 	return cmm;
 }
 
@@ -728,7 +732,7 @@ display_thumb(Browser *browser, FinishJob job)
 	update.ram_tier = job.tier;
 
 	auto cmm = worker_cmm();
-	shared_ptr<dawn::Profile> p3 = cmm->get_profile_display_p3(false);
+	shared_ptr<dawn::Profile> p3 = cmm->get_profile_display_p3();
 	shared_ptr<dawn::Profile> screen = profile_from_icc(*cmm, job.screen_icc);
 	vector<uint16_t> display = job.pixels ? *job.pixels : vector<uint16_t>{};
 	if (p3 && screen && !display.empty() &&

dn/kit-viewer.cpp

@@ -376,7 +376,7 @@ profile_from_icc(dawn::Cmm &cmm, const shared_ptr<const vector<uint8_t>> &icc)
 		if (auto profile = cmm.get_profile(*icc))
 			return profile;
 	}
-	return cmm.get_profile_sRGB(false);
+	return cmm.get_profile_sRGB();
 }
 
 static bool apply_action(Viewer &v, Action action);
@@ -797,7 +797,7 @@ sync_ui(Viewer &v, Page &ui)
 			shared_ptr<dawn::Profile> srgb;
 			if (!img.have_primaries) {
 				auto cmm = v.cmm_ ? v.cmm_ : dawn::Cmm::get_default();
-				srgb = cmm->get_profile_sRGB(false);
+				srgb = cmm->get_profile_sRGB();
 				img = profile_chromaticities(srgb.get());
 				image_dashed = true;
 			}

dn/kit.cpp

@@ -72,7 +72,7 @@ bake_rgb(dawn::Cmm *cmm, dawn::Profile *target, uint8_t r, uint8_t g, uint8_t b)
 		return colour;
 	uint16_t pixel[4] = {
 		uint16_t(b * 257), uint16_t(g * 257), uint16_t(r * 257), 65535};
-	auto srgb = cmm->get_profile_sRGB(false);
+	auto srgb = cmm->get_profile_sRGB();
 	if (!srgb)
 		return colour;
 	if (!cmm->transform_bgra16(

dn/thumbnail-cache.cpp

@@ -194,7 +194,7 @@ read_wide(const QString &path, const ThumbnailSource &source, int tier,
 	const string *tag = value(meta, kColorSpace);
 	const bool p3 = tag && *tag == "Display P3";
 	shared_ptr<dawn::Profile> src =
-		p3 ? cmm->get_profile_display_p3(true) : cmm->get_profile_sRGB(true);
+		p3 ? cmm->get_profile_display_p3() : cmm->get_profile_sRGB();
 	dawn::ImagePtr image = decode_webp(bytes, *cmm, src.get(), screen);
 	if (!image)
 		return hit;
@@ -237,7 +237,7 @@ read_png(const QString &path, const ThumbnailSource &source, int tier,
 	if (!valid_metadata(meta, source))
 		return {};
 
-	shared_ptr<dawn::Profile> srgb = cmm->get_profile_sRGB(false);
+	shared_ptr<dawn::Profile> srgb = cmm->get_profile_sRGB();
 	if (!cmm->transform_bgra16(image->data.data(), image->width, image->height,
 			srgb.get(), screen, true, true))
 		return {};

dn/window.cpp

@@ -697,7 +697,7 @@ Window::refresh_screen_profile(QScreen *target_screen)
 		}
 	}
 	if (!next)
-		next = this->cmm_->get_profile_sRGB(false);
+		next = this->cmm_->get_profile_sRGB();
 	this->screen_profile_fallback_ = source == "srgb";
 
 	const bool changed =

dnimaged/main.cpp

@@ -64,7 +64,7 @@ decode_job(const Job &job)
 	context.cmm = cmm;
 	context.first_frame_only = job.first_frame_only;
 	context.screen_profile = job.target_icc.empty()
-		? cmm->get_profile_sRGB(false)
+		? cmm->get_profile_sRGB()
 		: cmm->get_profile(span<const uint8_t>(
 			  reinterpret_cast<const uint8_t *>(job.target_icc.data()),
 			  job.target_icc.size()));

libdn/libdn.cpp

@@ -515,8 +515,7 @@ ensure_working_premul(
 		if (owned)
 			image.effective_profile = owned;
 		else if (!source) {
-			image.effective_profile =
-				cmm_or_default(ctx)->get_profile_sRGB(false);
+			image.effective_profile = cmm_or_default(ctx)->get_profile_sRGB();
 			image.profile_assumed = true;
 		}
 	}
@@ -544,8 +543,7 @@ ensure_working_premul_pages(
 		if (owned)
 			page.effective_profile = owned;
 		else if (!source) {
-			page.effective_profile =
-				cmm_or_default(ctx)->get_profile_sRGB(false);
+			page.effective_profile = cmm_or_default(ctx)->get_profile_sRGB();
 			page.profile_assumed = true;
 		}
 	}
@@ -649,7 +647,18 @@ profiles_equal(const Profile *a, const Profile *b)
 		return true;
 	if (!a || !b)
 		return false;
-	return a->to_bytes() == b->to_bytes();
+
+	// ICC.1 lays the header out identically in every version: 128 bytes,
+	// with the creation dateTimeNumber at 24.  It says when the profile
+	// was made, never what it does.
+	constexpr size_t header = 128, created = 24, created_size = 12;
+	vector<uint8_t> x = a->to_bytes(), y = b->to_bytes();
+	if (x.size() != y.size() || x.size() < header)
+		return x == y;
+
+	memset(x.data() + created, 0, created_size);
+	memset(y.data() + created, 0, created_size);
+	return x == y;
 }
 
 float
@@ -907,28 +916,27 @@ Cmm::get_profile(span<const uint8_t> bytes)
 }
 
 shared_ptr<Profile>
-Cmm::get_profile_sRGB(bool cache)
+Cmm::get_profile_sRGB()
 {
 	// We may use these a lot, no need to recreate each time.
-	if (this->cached_sRGB)
-		return this->cached_sRGB;
+	if (auto cached = this->cached_sRGB.lock())
+		return cached;
 
 	cmsHPROFILE p = cmsCreate_sRGBProfileTHR(cmsContext(context_));
 	if (!p)
 		return nullptr;
 
 	shared_ptr<Profile> result(new Profile(shared_from_this(), p));
-	if (cache)
-		this->cached_sRGB = result;
+	this->cached_sRGB = result;
 	return result;
 }
 
 shared_ptr<Profile>
-Cmm::get_profile_display_p3(bool cache)
+Cmm::get_profile_display_p3()
 {
 	// We may use these a lot, no need to recreate each time.
-	if (this->cached_display_p3)
-		return this->cached_display_p3;
+	if (auto cached = this->cached_display_p3.lock())
+		return cached;
 
 	constexpr size_t samples = 4096;
 	vector<cmsUInt16Number> transfer(samples);
@@ -958,8 +966,7 @@ Cmm::get_profile_display_p3(bool cache)
 	cmsSetProfileVersion(p, 4.3);
 
 	shared_ptr<Profile> result(new Profile(shared_from_this(), p));
-	if (cache)
-		this->cached_display_p3 = result;
+	this->cached_display_p3 = result;
 	return result;
 }
 
@@ -1121,7 +1128,7 @@ Cmm::transform_bgra16(uint8_t *data, uint32_t width, uint32_t height,
 {
 	shared_ptr<Profile> src_fallback;
 	if (target && !source) {
-		src_fallback = get_profile_sRGB(false);
+		src_fallback = get_profile_sRGB();
 		source = src_fallback.get();
 	}
 	if (!source || !target)
@@ -1142,7 +1149,7 @@ Cmm::transform_bgra8_to_bgra16(const uint8_t *src, uint8_t *dst, uint32_t width,
 {
 	shared_ptr<Profile> src_fallback;
 	if (target && !source) {
-		src_fallback = get_profile_sRGB(false);
+		src_fallback = get_profile_sRGB();
 		source = src_fallback.get();
 	}
 	if (!src || !dst || !source || !target)
@@ -1218,7 +1225,7 @@ Cmm::finish_page(Image &page, Profile *target)
 		if (source)
 			page.effective_profile = source;
 		else {
-			page.effective_profile = get_profile_sRGB(false);
+			page.effective_profile = get_profile_sRGB();
 			page.profile_assumed = true;
 		}
 	}

libdn/libdn.h

@@ -254,6 +254,7 @@ public:
 };
 
 /// Serialized ICC equality. Null equals null; lcms has no compare API.
+/// The header's creation date/time is excluded.
 bool profiles_equal(const Profile *a, const Profile *b);
 
 class Cmm : public std::enable_shared_from_this<Cmm>
@@ -262,8 +263,10 @@ class Cmm : public std::enable_shared_from_this<Cmm>
 	void *context_ = nullptr;  ///< cmsContext
 	bool broken_premul_ = false;
 
-	std::shared_ptr<Profile> cached_sRGB;
-	std::shared_ptr<Profile> cached_display_p3;
+	// Weak, because a Profile owns its Cmm: we don't want a cycle.
+	// Deduplicating profiles for as long as somebody else wants them.
+	std::weak_ptr<Profile> cached_sRGB;
+	std::weak_ptr<Profile> cached_display_p3;
 
 public:
 	Cmm();
@@ -275,8 +278,8 @@ public:
 
 	std::shared_ptr<Profile> get_profile_data(const void *data, size_t len);
 	std::shared_ptr<Profile> get_profile(std::span<const uint8_t> bytes);
-	std::shared_ptr<Profile> get_profile_sRGB(bool cache);
-	std::shared_ptr<Profile> get_profile_display_p3(bool cache);
+	std::shared_ptr<Profile> get_profile_sRGB();
+	std::shared_ptr<Profile> get_profile_display_p3();
 	std::shared_ptr<Profile> get_profile_sRGB_gamma(double gamma);
 	std::shared_ptr<Profile> get_profile_parametric(
 		double gamma, double whitepoint[2], double primaries[6]);

libdn/load-jpeg.cpp

@@ -204,7 +204,7 @@ parse_exif_profile(Cmm &cmm, span<const uint8_t> exif)
 
 	// If sRGB is claimed, assume all parameters are standard.
 	if (params.colorspace == Exif_ColorSpace_sRGB)
-		return cmm.get_profile_sRGB(false);
+		return cmm.get_profile_sRGB();
 
 	// AdobeRGB Nikon JPEGs provide all of these.
 	if (params.colorspace != Exif_ColorSpace_Uncalibrated ||
@@ -555,7 +555,7 @@ load_jpeg_finalize(ImagePtr &image, bool cmyk, bool argb, int bits,
 	if (source)
 		image->effective_profile = source;
 	else if (!cmyk) {
-		image->effective_profile = cmm->get_profile_sRGB(false);
+		image->effective_profile = cmm->get_profile_sRGB();
 		image->profile_assumed = true;
 	}
 

tests/bench-open.cpp

@@ -47,7 +47,7 @@ bench_one(const char *path, bool cms, int repeats, dawn::ScaleScaler *scaler)
 	ctx.first_frame_only = true;
 	ctx.uri = path;
 	if (cms) {
-		auto srgb = cmm->get_profile_sRGB(false);
+		auto srgb = cmm->get_profile_sRGB();
 		if (!srgb) {
 			fprintf(stderr, "%s: get_profile_sRGB failed\n", path);
 			return 1;

tests/test-channels.cpp

@@ -194,7 +194,7 @@ static void
 test_jpeg_cms_8_to_16()
 {
 	auto cmm = dawn::Cmm::get_default();
-	auto srgb = cmm->get_profile_sRGB(false);
+	auto srgb = cmm->get_profile_sRGB();
 	CHECK(srgb != nullptr);
 
 	const uint8_t src[4] = {0, 0, 255, 255};
@@ -254,7 +254,7 @@ static void
 test_cmyk_cms_opaque()
 {
 	auto cmm = dawn::Cmm::get_default();
-	auto srgb = cmm->get_profile_sRGB(false);
+	auto srgb = cmm->get_profile_sRGB();
 	CHECK(srgb != nullptr);
 
 	const fs::path icc = fs::path(DAWN_TEST_FIXTURES_DIR) / "cmyk-lab.icc";
@@ -281,7 +281,7 @@ static void
 test_cms_tiled()
 {
 	auto cmm = dawn::Cmm::get_default();
-	auto srgb = cmm->get_profile_sRGB(false);
+	auto srgb = cmm->get_profile_sRGB();
 	CHECK(srgb != nullptr);
 
 	// Above the serial threshold so row bands run on more than one worker.
@@ -451,7 +451,7 @@ test_chromaticities()
 		dawn::ColorModel::Unknown);
 
 	auto cmm = dawn::Cmm::get_default();
-	auto srgb = cmm->get_profile_sRGB(false);
+	auto srgb = cmm->get_profile_sRGB();
 	CHECK(srgb != nullptr);
 	dawn::Chromaticities s = dawn::profile_chromaticities(srgb.get());
 	CHECK(s.model == dawn::ColorModel::Rgb);
@@ -463,7 +463,7 @@ test_chromaticities()
 	CHECK(s.have_white);
 	near_xy("sRGB W", s.wx, s.wy, 0.3127, 0.3290, 0.002);
 
-	auto display_p3 = cmm->get_profile_display_p3(false);
+	auto display_p3 = cmm->get_profile_display_p3();
 	CHECK(display_p3 != nullptr);
 	CHECK(dawn::profile_transfer(display_p3.get()) == dawn::Transfer::Srgb);
 	dawn::Chromaticities d = dawn::profile_chromaticities(display_p3.get());
@@ -547,7 +547,7 @@ test_profile_transfer()
 	CHECK(dawn::profile_transfer(nullptr) == dawn::Transfer::Srgb);
 
 	auto cmm = dawn::Cmm::get_default();
-	auto srgb = cmm->get_profile_sRGB(false);
+	auto srgb = cmm->get_profile_sRGB();
 	CHECK(srgb != nullptr);
 	CHECK(dawn::profile_transfer(srgb.get()) == dawn::Transfer::Srgb);
 

tests/test-thumbnail-cache.cpp

@@ -112,7 +112,7 @@ test_cache_entries(const QTemporaryDir &inputs)
 		fprintf(stderr, "%s\n", qUtf8Printable(error));
 
 	auto cmm = make_shared<dawn::Cmm>();
-	auto p3 = cmm->get_profile_display_p3(false);
+	auto p3 = cmm->get_profile_display_p3();
 	dn::ThumbnailHit hit = dn::thumbnail_cache_lookup(source, 0, cmm, p3.get());
 	CHECK(!hit.pixels.empty());
 	CHECK(hit.width == 2 && hit.height == 1);

tools/dnimage.cpp

@@ -175,7 +175,7 @@ main()
 	const auto *pixels = reinterpret_cast<const uint8_t *>(bytes.data());
 
 	auto cmm = make_shared<dawn::Cmm>();
-	auto icc = cmm->get_profile_sRGB(false)->to_bytes();
+	auto icc = cmm->get_profile_sRGB()->to_bytes();
 	string error;
 	if (!write_png16(stdout, pixmap.width, pixmap.height, pixmap.stride, pixels,
 			icc, pixmap.orientation, &error)) {

tools/dnthumb.cpp

@@ -88,7 +88,7 @@ main(int argc, char **argv)
 	auto cmm = make_shared<dawn::Cmm>();
 	dawn::OpenContext context;
 	context.cmm = cmm;
-	context.screen_profile = cmm->get_profile_sRGB(false);
+	context.screen_profile = cmm->get_profile_sRGB();
 	context.first_frame_only = true;
 	dawn::Error open_error;
 	auto image = dawn::open_from_data(input, context, &open_error);