Defer write-only USB transactions to better utilize bus

src/ch347jtag.cpp

@@ -43,20 +43,37 @@ enum CH347JtagSig {
 	SIG_TMS =      0b10,
 	SIG_TDI =   0b10000,
 };
+
 static void LIBUSB_CALL sync_cb(struct libusb_transfer *transfer) {
 	int *complete = (int *)transfer->user_data;
 	*complete = true;
 }
 
-int CH347Jtag::usb_xfer(unsigned wlen, unsigned rlen, unsigned *ract) {
+// defer should only be used with rlen == 0
-	wcomplete = 0;
+
+int CH347Jtag::usb_xfer(unsigned wlen, unsigned rlen, unsigned *ract, bool defer) {
 	if (_verbose) {
+		fprintf(stderr, "obuf - _obuf = %ld\n", obuf - _obuf);
 		fprintf(stderr, "obuf[%d] = {", wlen);
 		for (unsigned i = 0; i < wlen; ++i) {
 			fprintf(stderr, "%02x ", obuf[i]);
 		}
 		fprintf(stderr, "}\n\n");
 	}
+	
+	if (defer) {
+		obuf += wlen;
+		return 0;
+	}
+	// write out whole buffer
+	wlen += obuf - _obuf;
+	obuf = _obuf;
+
+	if (wlen == 0) {
+		return 0;
+	}
+	
+	wcomplete = 0;
 	libusb_fill_bulk_transfer(wtrans, dev_handle, CH347JTAG_WRITE_EP, obuf,
 		wlen, sync_cb, &wcomplete, CH347JTAG_TIMEOUT);
 	int r = libusb_submit_transfer(wtrans);
@@ -131,12 +148,14 @@ wait_rcompletion:
 }
 
 int CH347Jtag::setClk(const uint8_t &factor) {
+	// flush the obuf
+	usb_xfer(0, 0, 0, false); // is called from constructor, don't replace with virtual flush()
 	memset(obuf, 0, 16);
 	obuf[0] = CMD_CLK;
 	obuf[1] = 6;
 	obuf[4] = factor;
 	unsigned actual = 0;
-	int rv = usb_xfer(9, 4, &actual);
+	int rv = usb_xfer(9, 4, &actual, false);
 	if (rv || actual != 4)
 		return -1;
 	if (ibuf[0] != 0xd0 || ibuf[3] != 0)
@@ -145,7 +164,7 @@ int CH347Jtag::setClk(const uint8_t &factor) {
 }
 
 CH347Jtag::CH347Jtag(uint32_t clkHZ, int8_t verbose):
-	_verbose(verbose>1), dev_handle(NULL), usb_ctx(NULL)
+      _verbose(verbose>1), dev_handle(NULL), usb_ctx(NULL), obuf(_obuf)
 {
 	int actual_length = 0;
 	struct libusb_device_descriptor desc;
@@ -241,24 +260,25 @@ int CH347Jtag::_setClkFreq(uint32_t clkHZ)
 
 int CH347Jtag::writeTMS(const uint8_t *tms, uint32_t len, bool flush_buffer)
 {
-	(void) flush_buffer;
+	if (get_obuf_length() < (int)(len * 2 + 4)) { // check if there is enough room left
-
+		flush();
+	}
 	uint8_t *ptr = obuf;
 	for (uint32_t i = 0; i < len; ++i) {
 		if (ptr == obuf) {
 			*ptr++ = CMD_BITS_WO;
 			ptr += 2;  // leave place for length;
 		}
-		uint8_t x = /*SIG_TDI |*/ ((tms[i >> 3] & (1 << (i & 7))) ? SIG_TMS : 0);
+		uint8_t x = ((tms[i >> 3] & (1 << (i & 7))) ? SIG_TMS : 0);
 		*ptr++ = x;
 		*ptr++ = x | SIG_TCK;
-		unsigned wlen = ptr - obuf;
+		int wlen = ptr - obuf;
-		if (wlen > sizeof(obuf) - 3 || i == len - 1) {
+		if (wlen + 1 >= get_obuf_length() || i == len - 1) {
 			*ptr++ = x;  // clear TCK
-			++wlen;
+			wlen = ptr - obuf;
 			obuf[1] = wlen - 3;
 			obuf[2] = (wlen - 3) >> 8;
-			int ret = usb_xfer(wlen, 0, 0);
+			int ret = usb_xfer(wlen, 0, 0, !flush_buffer);
 			if (ret < 0) {
 				cerr << "writeTMS: usb bulk write failed: " <<
 					libusb_strerror(static_cast<libusb_error>(ret)) << endl;
@@ -275,9 +295,14 @@ int CH347Jtag::toggleClk(uint8_t tms, uint8_t tdi, uint32_t len)
 	uint8_t bits = 0;
 	if (tms) bits |= SIG_TMS;
 	if (tdi) bits |= SIG_TDI;
-	if (!bits && len > 7) {
+	if (!bits && len > 8) {
 		return writeTDI(0, 0, len, false);
 	}
+	
+	if (get_obuf_length() < (int)(len * 2 + 4)) {
+		flush();
+	}
+
 	uint8_t *ptr = obuf;
 	for (uint32_t i = 0; i < len; ++i) {
 		if (ptr == obuf) {
@@ -286,13 +311,13 @@ int CH347Jtag::toggleClk(uint8_t tms, uint8_t tdi, uint32_t len)
 		}
 		*ptr++ = bits;
 		*ptr++ = bits | SIG_TCK;
-		unsigned wlen = ptr - obuf;
+		int wlen = ptr - obuf;
-		if (wlen > sizeof(obuf) - 3 || i == len - 1) {
+		if (wlen + 1 >= get_obuf_length() || i == len - 1) {
 			*ptr++ = bits;  // clear TCK
-			++wlen;
+			wlen = ptr - obuf;
 			obuf[1] = wlen - 3;
 			obuf[2] = (wlen - 3) >> 8;
-			int ret = usb_xfer(wlen, 0, 0);
+			int ret = usb_xfer(wlen, 0, 0, true);
 			if (ret < 0) {
 				cerr << "writeCLK: usb bulk write failed: " <<
 					libusb_strerror(static_cast<libusb_error>(ret)) << endl;
@@ -306,17 +331,20 @@ int CH347Jtag::toggleClk(uint8_t tms, uint8_t tdi, uint32_t len)
 
 int CH347Jtag::writeTDI(const uint8_t *tx, uint8_t *rx, uint32_t len, bool end)
 {
-	if (!len)
+	if (len == 0)
 		return 0;
-	unsigned bytes = (len - ((end)?1:0)) / 8;
+	unsigned bytes = (len - (end ? 1 : 0)) / 8;
 	unsigned bits = len - bytes * 8;
 	uint8_t *rptr = rx;
 	const uint8_t *tptr = tx;
 	const uint8_t *txend = tx + bytes;
-	uint8_t cmd = (rx) ? CMD_BYTES_WR : CMD_BYTES_WO;
+	uint8_t cmd = (rx != nullptr) ? CMD_BYTES_WR : CMD_BYTES_WO;
 	while (tptr < txend) {
-		unsigned avail = sizeof(obuf) - 3;
+		if (get_obuf_length() < 4) {
-		unsigned chunk = (txend - tptr < avail)? txend - tptr: avail;
+			flush();
+		}
+		int avail = get_obuf_length() - 3;
+		int chunk = (txend - tptr < avail)? txend - tptr: avail;
 		if (tx) {
 			memcpy(&obuf[3], tptr, chunk);
 		} else {
@@ -328,7 +356,7 @@ int CH347Jtag::writeTDI(const uint8_t *tx, uint8_t *rx, uint32_t len, bool end)
 		obuf[1] = chunk;
 		obuf[2] = chunk >> 8;
 		unsigned actual_length = 0;
-		int ret = usb_xfer(chunk + 3, (rx) ? chunk + 3 : 0, &actual_length);
+		int ret = usb_xfer(chunk + 3, (rx) ? chunk + 3 : 0, &actual_length, rx == 0 && get_obuf_length());
 		if (ret < 0) {
 			cerr << "writeTDI: usb bulk read failed: " <<
 				libusb_strerror(static_cast<libusb_error>(ret)) << endl;
@@ -348,9 +376,12 @@ int CH347Jtag::writeTDI(const uint8_t *tx, uint8_t *rx, uint32_t len, bool end)
 	if (bits == 0)
 		return EXIT_SUCCESS;
 	cmd = (rx) ? CMD_BITS_WR : CMD_BITS_WO;
+	if (get_obuf_length() < (int)(4 + bits * 2)) {
+		flush();
+	}
 	uint8_t *ptr = &obuf[3];
 	uint8_t x = 0;
-	const uint8_t *bptr = &tx[bytes];
+	const uint8_t *bptr = tx + bytes;
 	for (unsigned i = 0; i < bits; ++i) {
 		uint8_t txb = (tx) ? bptr[i >> 3] : 0;
 		uint8_t _tdi = (txb & (1 << (i & 7))) ? SIG_TDI : 0;
@@ -366,7 +397,7 @@ int CH347Jtag::writeTDI(const uint8_t *tx, uint8_t *rx, uint32_t len, bool end)
 	obuf[0] = cmd;
 	obuf[1] = wlen - 3;
 	obuf[2] = (wlen - 3) >> 8;
-	int ret = usb_xfer(wlen, (rx) ? (bits + 3) : 0, &actual_length);
+	int ret = usb_xfer(wlen, (rx) ? (bits + 3) : 0, &actual_length, rx == nullptr);
 
 	if (ret < 0) {
 		cerr << "writeTDI: usb bulk read failed: " <<

src/ch347jtag.hpp

@@ -22,11 +22,11 @@ class CH347Jtag : public JtagInterface {
 	/* clk */
 	int toggleClk(uint8_t tms, uint8_t tdo, uint32_t clk_len) override;
 
-	int get_buffer_size() override { return 0;}
+	int get_buffer_size() override {return get_obuf_length();}
 
-	bool isFull() override { return false;}
+	bool isFull() override {return get_obuf_length() == 0;}
 
-	int flush() override {return 0;}
+	int flush() override {return usb_xfer(0, 0, 0, false);}
 
  private:
 	bool _verbose;
@@ -37,6 +37,8 @@ class CH347Jtag : public JtagInterface {
 	struct libusb_transfer *wtrans, *rtrans;
 	int rcomplete, wcomplete;
 	uint8_t ibuf[512];
-	uint8_t obuf[512];
+	uint8_t _obuf[512];
-	int usb_xfer(unsigned wlen, unsigned rlen, unsigned *actual);
+	uint8_t *obuf;
+	int get_obuf_length() const {return 512 - (obuf - _obuf);}
+	int usb_xfer(unsigned wlen, unsigned rlen, unsigned *actual, bool defer);
 };

src/gowin.cpp

@@ -20,7 +20,7 @@
 #include "rawParser.hpp"
 #include "spiFlash.hpp"
 
-#include <byteswap.h>
+//#include <byteswap.h>
 
 using namespace std;
 
@@ -140,7 +140,6 @@ Gowin::Gowin(Jtag *jtag, const string filename, const string &file_type, std::st
 			}
 		}
 	}
-	_jtag->setClkFreq(2500000);
 
 	/* erase and program flash differ for GW1N1 */
 	if (idcode == 0x0900281B)
@@ -224,12 +223,16 @@ void Gowin::reset()
 
 void Gowin::programFlash()
 {
+	_jtag->setClkFreq(2500000);
+
 	const uint8_t *data = _fs->getData();
 	int length = _fs->getLength();
 	
-	send_command(0x3a);
+	
-	send_command(0);
+	send_command(CONFIG_DISABLE);
+	send_command(NOOP);
 	_jtag->go_test_logic_reset();
+	_jtag->flush();
 	usleep(500*1000);
 	
 	eraseSRAM();
@@ -452,6 +455,13 @@ bool Gowin::pollFlag(uint32_t mask, uint32_t value)
 	return true;
 }
 
+inline uint32_t bswap_32(uint32_t x)
+{
+	return ((x << 24) & 0xff000000 ) |
+	       ((x <<  8) & 0x00ff0000 ) |
+	       ((x >>  8) & 0x0000ff00 ) |
+	       ((x >> 24) & 0x000000ff );
+}
 /* TN653 p. 17-21 */
 bool Gowin::writeFLASH(uint32_t page, const uint8_t *data, int length)
 {
@@ -486,11 +496,12 @@ bool Gowin::writeFLASH(uint32_t page, const uint8_t *data, int length)
 			sendClkUs((is_gw1n1) ? 32 : 16);
 		}
 		sendClkUs((is_gw1n1) ? 2400 : 6);
+		_jtag->flush();
 		usleep(200);
 	}
 	send_command(CONFIG_DISABLE);
 	send_command(NOOP);
-
+	_jtag->flush();
 	usleep(600*1000);
 	send_command(CONFIG_DISABLE);
 	send_command(NOOP);
@@ -499,6 +510,7 @@ bool Gowin::writeFLASH(uint32_t page, const uint8_t *data, int length)
 	send_command(NOOP);
 	if (_verbose)
 		displayReadReg(readStatusReg());
+	_jtag->flush();
 	sleep(1);
 
 	return true;
@@ -519,7 +531,7 @@ bool Gowin::writeSRAM(const uint8_t *data, int length)
 	_jtag->shiftDR(data, NULL, length);
 	send_command(CONFIG_DISABLE); // config disable
 	send_command(NOOP); // noop
-
+	_jtag->flush();
 	sleep(1);
 
 	if (readStatusReg() & STATUS_DONE_FINAL)
@@ -553,6 +565,7 @@ bool Gowin::eraseFLASH()
 	send_command(NOOP);
 	if (_verbose)
 		displayReadReg(readStatusReg());
+	_jtag->flush();
 	usleep(500*1000);
 	if (_verbose)
 		displayReadReg(readStatusReg());
@@ -562,9 +575,9 @@ bool Gowin::eraseFLASH()
 
 void Gowin::sendClkUs(unsigned us)
 {
-	//unsigned freq = _jtag->getClkFreq() / 1000000;
+	uint64_t clocks = _jtag->getClkFreq();
-	//freq = (freq) ? freq : 1;
+	clocks *= us;
-	unsigned clocks = us * 2; // at 2MHz 1us ~ 2 clocks
+	clocks /= 1000000;
 	_jtag->toggleClk(clocks);
 }