Enable 4-way parallel output on the ESP8266.

5 files changed, 52 insertions, 28 deletions

diff --git a/FastLED.cpp b/FastLED.cpp
index 0bf2e0f7..d3ad5732 100644
--- a/FastLED.cpp
+++ b/FastLED.cpp
@@ -16,6 +16,9 @@ CLEDController *CLEDController::m_pHead = NULL;
 CLEDController *CLEDController::m_pTail = NULL;
 static uint32_t lastshow = 0;
 
+uint32_t _frame_cnt=0;
+uint32_t _retry_cnt=0;
+
 // uint32_t CRGB::Squant = ((uint32_t)((__TIME__[4]-'0') * 28))<<16 | ((__TIME__[6]-'0')*50)<<8 | ((__TIME__[7]-'0')*28);
 
 CFastLED::CFastLED() {
@@ -237,7 +240,7 @@ namespace __cxxabiv1
 	#ifndef ESP8266
 	extern "C" void __cxa_pure_virtual (void) {}
 	#endif
-	
+
 	/* guard variables */
 
 	/* The ABI requires a 64-bit type.  */
diff --git a/platforms/esp/8266/clockless_block_esp8266.h b/platforms/esp/8266/clockless_block_esp8266.h
index 0738323f..47973fdd 100644
--- a/platforms/esp/8266/clockless_block_esp8266.h
+++ b/platforms/esp/8266/clockless_block_esp8266.h
@@ -2,15 +2,20 @@
 #define __INC_CLOCKLESS_BLOCK_ESP8266_H
 
 #define FASTLED_HAS_BLOCKLESS 1
-#define PORTA_FIRST_PIN 0
 
 #define PORT_MASK (((1<<LANES)-1) & 0x0000FFFFL)
 #define MIN(X,Y) (((X)<(Y)) ? (X):(Y))
-#define USED_LANES (MIN(LANES,16))
-#define LAST_PIN (FIRST_PIN + USED_LANES - 1)
+#define USED_LANES (MIN(LANES,4))
+#define REAL_FIRST_PIN 12
+#define LAST_PIN (12 + USED_LANES - 1)
 
 FASTLED_NAMESPACE_BEGIN
 
+#ifdef FASTLED_DEBUG_COUNT_FRAME_RETRIES
+extern uint32_t _frame_cnt;
+extern uint32_t _retry_cnt;
+#endif
+
 template <uint8_t LANES, int FIRST_PIN, int T1, int T2, int T3, EOrder RGB_ORDER = GRB, int XTRA0 = 0, bool FLIP = false, int WAIT_TIME = 5>
 class InlineBlockClocklessController : public CPixelLEDController<RGB_ORDER, LANES, PORT_MASK> {
 	typedef typename FastPin<FIRST_PIN>::port_ptr_t data_ptr_t;
@@ -28,6 +33,9 @@ public:
 		int cnt=FASTLED_INTERRUPT_RETRY_COUNT;
 		while(!showRGBInternal(pixels) && cnt--) {
       os_intr_unlock();
+			#ifdef FASTLED_DEBUG_COUNT_FRAME_RETRIES
+			_retry_cnt++;
+			#endif
       delayMicroseconds(WAIT_TIME * 10);
       os_intr_lock();
     }
@@ -41,24 +49,14 @@ public:
 	}
 
   template<int PIN> static void initPin() {
-    if(PIN >= FIRST_PIN && PIN <= LAST_PIN) {
-      FastPin<PIN>::setOutput();
+    if(PIN >= REAL_FIRST_PIN && PIN <= LAST_PIN) {
+			_ESPPIN<PIN, 1<<(PIN & 0xFF)>::setOutput();
+      // FastPin<PIN>::setOutput();
     }
   }
 
   virtual void init() {
-    // initPin<0>();
-    // initPin<1>();
-    // initPin<2>();
-    // initPin<3>();
-    // initPin<4>();
-    // initPin<5>();
-    // initPin<6>();
-    // initPin<7>();
-    // initPin<8>();
-    // initPin<9>();
-    // initPin<10>();
-    // initPin<11>();
+		// Only supportd on pins 12-15
     initPin<12>();
     initPin<13>();
     initPin<14>();
@@ -87,29 +85,31 @@ public:
 		register uint8_t scale = pixels.template getscale<PX>(pixels);
 
 		for(register uint32_t i = 0; i < USED_LANES; i++) {
-			while((__clock_cycles() - last_mark) < (T1+T2+T3));
+			while((__clock_cycles() - last_mark) < (T1+T2+T3+6));
 			last_mark = __clock_cycles();
-			*FastPin<FIRST_PIN>::sport() = PORT_MASK << FIRST_PIN;
+			*FastPin<FIRST_PIN>::sport() = PORT_MASK << REAL_FIRST_PIN;
 
+			uint32_t nword = ((uint32_t)(~b2.bytes[7-i]) & PORT_MASK) << REAL_FIRST_PIN;
 			while((__clock_cycles() - last_mark) < T1);
-			*FastPin<FIRST_PIN>::cport() = ((uint32_t)(~b2.bytes[7-i]) & PORT_MASK) << FIRST_PIN;
+			*FastPin<FIRST_PIN>::cport() = nword;
 
 			while((__clock_cycles() - last_mark) < (T1+T2));
-			*FastPin<FIRST_PIN>::cport() = PORT_MASK << FIRST_PIN;
+			*FastPin<FIRST_PIN>::cport() = PORT_MASK << REAL_FIRST_PIN;
 
 			b.bytes[i] = pixels.template loadAndScale<PX>(pixels,i,d,scale);
 		}
 
 		for(register uint32_t i = USED_LANES; i < 8; i++) {
-			while((__clock_cycles() - last_mark) < (T1+T2+T3));
-			last_mark = __clock_cycles() + (T1+T2+T3) + ESP_ADJUST;
-			*FastPin<FIRST_PIN>::sport() = PORT_MASK << FIRST_PIN;
+			while((__clock_cycles() - last_mark) < (T1+T2+T3+6));
+			last_mark = __clock_cycles();
+			*FastPin<FIRST_PIN>::sport() = PORT_MASK << REAL_FIRST_PIN;
 
+			uint32_t nword = ((uint32_t)(~b2.bytes[7-i]) & PORT_MASK) << REAL_FIRST_PIN;
 			while((__clock_cycles() - last_mark) < T1);
-			*FastPin<FIRST_PIN>::cport() = ((uint32_t)(~b2.bytes[7-i]) & PORT_MASK) << FIRST_PIN;
+			*FastPin<FIRST_PIN>::cport() = nword;
 
 			while((__clock_cycles() - last_mark) < (T1+T2));
-			*FastPin<FIRST_PIN>::cport() = PORT_MASK << FIRST_PIN;
+			*FastPin<FIRST_PIN>::cport() = PORT_MASK << REAL_FIRST_PIN;
 		}
 	}
 
@@ -156,6 +156,9 @@ public:
 		};
 
     os_intr_unlock();
+		#ifdef FASTLED_DEBUG_COUNT_FRAME_RETRIES
+		_frame_cnt++;
+		#endif
     return __clock_cycles() - _start;
 	}
 };
diff --git a/platforms/esp/8266/clockless_esp8266.h b/platforms/esp/8266/clockless_esp8266.h
index 746e9613..20ae6412 100644
--- a/platforms/esp/8266/clockless_esp8266.h
+++ b/platforms/esp/8266/clockless_esp8266.h
@@ -2,6 +2,11 @@
 
 FASTLED_NAMESPACE_BEGIN
 
+#ifdef FASTLED_DEBUG_COUNT_FRAME_RETRIES
+extern uint32_t _frame_cnt;
+extern uint32_t _retry_cnt;
+#endif
+
 // Info on reading cycle counter from https://github.com/kbeckmann/nodemcu-firmware/blob/ws2812-dual/app/modules/ws2812.c
 __attribute__ ((always_inline)) inline static uint32_t __clock_cycles() {
   uint32_t cyc;
@@ -34,6 +39,9 @@ protected:
     // mWait.wait();
 		int cnt = FASTLED_INTERRUPT_RETRY_COUNT;
     while((showRGBInternal(pixels)==0) && cnt--) {
+      #ifdef FASTLED_DEBUG_COUNT_FRAME_RETRIES
+      _retry_cnt++;
+      #endif
       os_intr_unlock();
       delayMicroseconds(WAIT_TIME);
       os_intr_lock();
@@ -99,6 +107,9 @@ protected:
 		};
 
 		os_intr_unlock();
+    #ifdef FASTLED_DEBUG_COUNT_FRAME_RETRIES
+    _frame_cnt++;
+    #endif
 		return __clock_cycles() - start;
 	}
 };
diff --git a/platforms/esp/8266/fastled_esp8266.h b/platforms/esp/8266/fastled_esp8266.h
index 2c2e2f12..d209de97 100644
--- a/platforms/esp/8266/fastled_esp8266.h
+++ b/platforms/esp/8266/fastled_esp8266.h
@@ -4,4 +4,4 @@
 #include "fastled_delay.h"
 #include "fastpin_esp8266.h"
 #include "clockless_esp8266.h"
-// #include "clockless_block_esp8266.h"
+#include "clockless_block_esp8266.h"
diff --git a/platforms/esp/8266/fastpin_esp8266.h b/platforms/esp/8266/fastpin_esp8266.h
index 5d4534f8..b8095b85 100644
--- a/platforms/esp/8266/fastpin_esp8266.h
+++ b/platforms/esp/8266/fastpin_esp8266.h
@@ -56,6 +56,8 @@ _DEFPIN_ESP8266(4,4); _DEFPIN_ESP8266(5,5);
 
 _DEFPIN_ESP8266(12,12); _DEFPIN_ESP8266(13,13); _DEFPIN_ESP8266(14,14); _DEFPIN_ESP8266(15,15);
 _DEFPIN_ESP8266(16,16);
+
+#define PORTA_FIRST_PIN 12
 #elif defined(FASTLED_ESP8266_D1_PIN_ORDER)
 #define MAX_PIN 15
 _DEFPIN_ESP8266(0,3);
@@ -74,6 +76,9 @@ _DEFPIN_ESP8266(12,12);
 _DEFPIN_ESP8266(13,14);
 _DEFPIN_ESP8266(14,4);
 _DEFPIN_ESP8266(15,5);
+
+#define PORTA_FIRST_PIN 12
+
 #else // if defined(FASTLED_ESP8266_NODEMCU_PIN_ORDER)
 #define MAX_PIN 10
 
@@ -82,6 +87,8 @@ _DEFPIN_ESP8266(0,16); _DEFPIN_ESP8266(1,5); _DEFPIN_ESP8266(2,4); _DEFPIN_ESP82
 _DEFPIN_ESP8266(4,2); _DEFPIN_ESP8266(5,14); _DEFPIN_ESP8266(6,12); _DEFPIN_ESP8266(7,13);
 _DEFPIN_ESP8266(8,15); _DEFPIN_ESP8266(9,3); _DEFPIN_ESP8266(10,1);
 
+#define PORTA_FIRST_PIN 6
+
 // The rest of the pins - these are generally not available
 // _DEFPIN_ESP8266(11,6);
 // _DEFPIN_ESP8266(12,7); _DEFPIN_ESP8266(13,8); _DEFPIN_ESP8266(14,9); _DEFPIN_ESP8266(15,10);