| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #pragma once | ||
| 2 | /* | ||
| 3 | * This file is part of the Falcon Player (FPP) and is Copyright (C) | ||
| 4 | * 2022 by the Falcon Player Developers. | ||
| 5 | * | ||
| 6 | * The Falcon Player (FPP) is free software, and is covered under | ||
| 7 | * multiple Open Source licenses. Please see the included 'LICENSES' | ||
| 8 | * file for descriptions of what files are covered by each license. | ||
| 9 | * | ||
| 10 | * This source file is covered under the CC-BY-ND as described in the | ||
| 11 | * included LICENSE.CC-BY-ND file. This file may be modified for | ||
| 12 | * personal use, but modified copies MAY NOT be redistributed in any form. | ||
| 13 | */ | ||
| 14 | |||
| 15 | #include <atomic> | ||
| 16 | #include <chrono> | ||
| 17 | #include <string> | ||
| 18 | #include "fpp-json-fwd.h" | ||
| 19 | #include <thread> | ||
| 20 | #include <vector> | ||
| 21 | |||
| 22 | #include "../../pru/BBShiftStringDefs.hp" | ||
| 23 | #include "../FalconV5Support/FalconV5Support.h" | ||
| 24 | #include "channeloutput/ChannelOutput.h" | ||
| 25 | #include "channeloutput/PixelString.h" | ||
| 26 | #include "util/BBBPruUtils.h" | ||
| 27 | |||
| 28 | // How deep the cape's shift register chain is, i.e. how many strings hang off | ||
| 29 | // each data pin. 8 is the long standing layout; 16 doubles the strings a | ||
| 30 | // single PRU can drive and is AM62x only - see the SHIFT16 firmware variant in | ||
| 31 | // BBShiftString.asm and the makefile. Buffers and maps are sized for the | ||
| 32 | // maximum; the live depth is m_stringsPerPin, from the cape's "stringsPerPin". | ||
| 33 | constexpr int MAX_STRINGS_PER_PIN = 16; | ||
| 34 | constexpr int MAX_PINS_PER_PRU = 8; | ||
| 35 | |||
| 36 | // when output restarts, there are some packets that need to be sent first | ||
| 37 | // but not when looped around | ||
| 38 | constexpr int FIRST_LOOPING_CONFIG_PACKET = 6; | ||
| 39 | constexpr int NUM_CONFIG_PACKETS = 96 + FIRST_LOOPING_CONFIG_PACKET; | ||
| 40 | |||
| 41 | // structure of the data at the start of the PRU ram | ||
| 42 | // that the pru program expects to see | ||
| 43 | typedef struct { | ||
| 44 | // in the DDR shared with the PRU | ||
| 45 | uint32_t address_dma; | ||
| 46 | uint32_t address_dma_packet; | ||
| 47 | |||
| 48 | // write data length to start, 0xFFFF to abort. will be cleared when started | ||
| 49 | volatile uint32_t command; | ||
| 50 | volatile uint32_t response; | ||
| 51 | |||
| 52 | uint32_t buffer[4]; // need a bit of a buffer | ||
| 53 | uint16_t commandTable[BBSS_COMMAND_TABLE_ENTRIES]; | ||
| 54 | } __attribute__((__packed__)) BBShiftStringData; | ||
| 55 | |||
| 56 | // GPIOCommand::channelOffset value for a command that applies in the FalconV5 | ||
| 57 | // packet phase rather than at a data byte position: once the first packet has | ||
| 58 | // gone out, and for the whole phase in a frame that opens a listen window. | ||
| 59 | // Larger than any real offset so it sorts last in the command table build. | ||
| 60 | constexpr int GPIO_CMD_AFTER_FIRST_PACKET = 0x7FFFFFFF; | ||
| 61 | |||
| 62 | class BBShiftStringOutput : public ChannelOutput { | ||
| 63 | public: | ||
| 64 | BBShiftStringOutput(unsigned int startChannel, unsigned int channelCount); | ||
| 65 | virtual ~BBShiftStringOutput(); | ||
| 66 | |||
| 67 | 308 | virtual std::string GetOutputType() const { | |
| 68 |
1/2✓ Branch 0 taken 308 times.
✗ Branch 1 not taken.
|
616 | return "BBB Pixel Strings"; |
| 69 | } | ||
| 70 | |||
| 71 | virtual int Init(Json::Value config) override; | ||
| 72 | virtual int Close(void) override; | ||
| 73 | |||
| 74 | virtual void StartingOutput() override; | ||
| 75 | virtual void StoppingOutput() override; | ||
| 76 | |||
| 77 | virtual int SendData(unsigned char* channelData) override; | ||
| 78 | virtual void PrepData(unsigned char* channelData) override; | ||
| 79 | virtual void DumpConfig(void) override; | ||
| 80 | |||
| 81 | virtual void GetRequiredChannelRanges(const std::function<void(int, int)>& addRange) override; | ||
| 82 | |||
| 83 | virtual void OverlayTestData(unsigned char* channelData, int cycleNum, float percentOfCycle, int testType, const Json::Value& config) override; | ||
| 84 | 124 | virtual bool SupportsTesting() const { return true; } | |
| 85 | |||
| 86 | private: | ||
| 87 | void StopPRU(bool wait = true); | ||
| 88 | int StartPRU(); | ||
| 89 | |||
| 90 | std::string m_subType; | ||
| 91 | |||
| 92 | class FalconV5PacketInfo { | ||
| 93 | public: | ||
| 94 | 402 | FalconV5PacketInfo(int l, uint8_t* d, bool list) : | |
| 95 | 402 | len(l), data(d), listen(list) { | |
| 96 | 402 | } | |
| 97 | 200 | FalconV5PacketInfo() : | |
| 98 | 200 | len(0), data(nullptr), listen(false) { | |
| 99 | 200 | } | |
| 100 | 601 | ~FalconV5PacketInfo() { | |
| 101 | 601 | } | |
| 102 | 299 | uint32_t getCommandFlags() { | |
| 103 | 299 | uint32_t f = 0x20000; | |
| 104 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 291 times.
|
299 | if (len == 2) { |
| 105 | 8 | f |= 0x40000; | |
| 106 | } | ||
| 107 |
2/2✓ Branch 0 taken 240 times.
✓ Branch 1 taken 59 times.
|
299 | if (listen) { |
| 108 | 240 | f |= 0x80000; | |
| 109 | } | ||
| 110 | 299 | return f; | |
| 111 | } | ||
| 112 | int len; | ||
| 113 | uint8_t* data; | ||
| 114 | bool listen = false; | ||
| 115 | }; | ||
| 116 | |||
| 117 | class FrameData { | ||
| 118 | public: | ||
| 119 | 100 | FrameData() { | |
| 120 |
2/2✓ Branch 0 taken 800 times.
✓ Branch 1 taken 100 times.
|
900 | for (int y = 0; y < MAX_PINS_PER_PRU; ++y) { |
| 121 |
2/2✓ Branch 0 taken 12800 times.
✓ Branch 1 taken 800 times.
|
13600 | for (int x = 0; x < MAX_STRINGS_PER_PIN; ++x) { |
| 122 | 12800 | stringMap[y][x] = -1; | |
| 123 | } | ||
| 124 | } | ||
| 125 |
2/2✓ Branch 0 taken 10200 times.
✓ Branch 1 taken 100 times.
|
10300 | for (int y = 0; y < NUM_CONFIG_PACKETS; ++y) { |
| 126 | 10200 | v5_config_packets[y] = nullptr; | |
| 127 | } | ||
| 128 | 100 | } | |
| 129 | 100 | ~FrameData() { | |
| 130 |
2/2✓ Branch 0 taken 94 times.
✓ Branch 1 taken 6 times.
|
100 | if (formattedData) |
| 131 | 94 | free(formattedData); | |
| 132 |
2/2✓ Branch 0 taken 94 times.
✓ Branch 1 taken 6 times.
|
100 | if (heapData) { |
| 133 |
1/2✓ Branch 0 taken 94 times.
✗ Branch 1 not taken.
|
94 | if (curData) |
| 134 | 94 | free(curData); | |
| 135 |
1/2✓ Branch 0 taken 94 times.
✗ Branch 1 not taken.
|
94 | if (lastData) |
| 136 | 94 | free(lastData); | |
| 137 | } | ||
| 138 |
2/2✓ Branch 0 taken 10200 times.
✓ Branch 1 taken 100 times.
|
10300 | for (int y = 0; y < NUM_CONFIG_PACKETS; ++y) { |
| 139 |
4/6✓ Branch 0 taken 401 times.
✓ Branch 1 taken 9799 times.
✓ Branch 2 taken 401 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 401 times.
✗ Branch 5 not taken.
|
10200 | if (v5_config_packets[y] && v5_config_packets[y] != &dynamicPacketInfo2 && v5_config_packets[y] != &dynamicPacketInfo1) { |
| 140 |
1/2✓ Branch 0 taken 401 times.
✗ Branch 1 not taken.
|
401 | delete v5_config_packets[y]; |
| 141 | } | ||
| 142 | } | ||
| 143 | 100 | } | |
| 144 | std::array<std::array<int, MAX_STRINGS_PER_PIN>, MAX_PINS_PER_PRU> stringMap; | ||
| 145 | BBBPru* pru = nullptr; | ||
| 146 | BBShiftStringData* pruData = nullptr; | ||
| 147 | |||
| 148 | uint8_t* formattedData = nullptr; | ||
| 149 | |||
| 150 | uint8_t* lastData = nullptr; | ||
| 151 | uint8_t* curData = nullptr; | ||
| 152 | uint32_t frameSize = 0; | ||
| 153 | int maxStringLen = 0; | ||
| 154 | int outputStringLen = 0; | ||
| 155 | |||
| 156 | // AM62x only: the PRU is fed through a ring in the PRU shared RAM by | ||
| 157 | // a pump thread instead of reading DDR (see pru/SMEMRing.hp); the | ||
| 158 | // frame buffers live in normal cached memory. Unused on AM335x. | ||
| 159 | BBBPruSMEMRing ring; | ||
| 160 | bool heapData = false; | ||
| 161 | |||
| 162 | // frame handoff to the pump thread; SendData fills pendingFrame then | ||
| 163 | // bumps pendingSeq, the pump snapshots it when the PRU is ready | ||
| 164 | struct PumpFrame { | ||
| 165 | const uint8_t* frameData = nullptr; | ||
| 166 | uint32_t frameBytes = 0; | ||
| 167 | const uint8_t* packetData = nullptr; | ||
| 168 | uint32_t packetBytes = 0; | ||
| 169 | uint32_t command = 0; | ||
| 170 | } pendingFrame; | ||
| 171 | std::atomic<uint32_t> pendingSeq{ 0 }; | ||
| 172 | // last values read back out of the firmware's counters, so only new | ||
| 173 | // events get reported | ||
| 174 | uint32_t renderedFrames = 0; | ||
| 175 | uint32_t resyncCount = 0; | ||
| 176 | // when the firmware first stopped making progress on a pending frame; | ||
| 177 | // zero while it is keeping up | ||
| 178 | std::chrono::steady_clock::time_point stalledSince{}; | ||
| 179 | // pump-internal streaming state | ||
| 180 | // atomic: the pump thread writes this while the output thread reads it | ||
| 181 | // in SendData's back-pressure gate. It is published with release only | ||
| 182 | // once the pump's copy of pendingFrame is known good - the gate reads | ||
| 183 | // it as permission to restage, so an early store would hand the output | ||
| 184 | // thread a frame the pump is still reading. | ||
| 185 | std::atomic<uint32_t> pumpedSeq{ 0 }; | ||
| 186 | PumpFrame activeFrame; | ||
| 187 | uint32_t activeOff = 0; | ||
| 188 | bool pumpActive = false; | ||
| 189 | |||
| 190 | FalconV5PacketInfo* v5_config_packets[NUM_CONFIG_PACKETS]; | ||
| 191 | FalconV5PacketInfo dynamicPacketInfo1; | ||
| 192 | FalconV5PacketInfo dynamicPacketInfo2; | ||
| 193 | FalconV5PacketInfo* dynamicPacketInfo = nullptr; | ||
| 194 | int curV5ConfigPacket = 0; | ||
| 195 | } m_pru0, m_pru1; | ||
| 196 | |||
| 197 | std::vector<PixelString*> m_strings; | ||
| 198 | // per string, per virtual string: nonzero when chMap[i+3] == chMap[i]+3 | ||
| 199 | // for the whole map, letting prepData skip the map indirection | ||
| 200 | std::vector<std::vector<uint8_t>> m_vsAffine; | ||
| 201 | std::map<std::string, std::string> m_usedPins; | ||
| 202 | |||
| 203 | // per-PRU pin sets; the platform defaults unless the cape overrides | ||
| 204 | // them by name (combo capes with nonstandard pinouts). The cape only | ||
| 205 | // carries header pin names - the r30 bit numbers are resolved here | ||
| 206 | // from the platform pin table and published to the firmware at runtime | ||
| 207 | std::vector<std::string> m_dataPins[2]; | ||
| 208 | std::vector<std::string> m_ctrlPins[2]; | ||
| 209 | bool m_pinNamesOverridden[2] = { false, false }; | ||
| 210 | int m_clockBit[2] = { 0, 0 }; | ||
| 211 | int m_latchBit[2] = { 0, 0 }; | ||
| 212 | // sharing the PRUSS with a panel driver on the other PRU: never clear | ||
| 213 | // the shared RAM or the other PRU's memory, no FalconV5 listeners | ||
| 214 | bool m_sharedPRUSS = false; | ||
| 215 | |||
| 216 | // shift register chain depth: strings per data pin, 8 or 16. Everything | ||
| 217 | // the PRU consumes scales with stringsPerPru(): one shift phase clocks | ||
| 218 | // that many bytes, so a frame is outputStringLen * stringsPerPru() bytes. | ||
| 219 | int m_stringsPerPin = 8; | ||
| 220 | 239965 | int stringsPerPru() const { return MAX_PINS_PER_PRU * m_stringsPerPin; } | |
| 221 | |||
| 222 | // whether anything at all was mapped onto a PRU. maxStringLen is set | ||
| 223 | // alongside every stringMap entry Init() fills in, so it is zero exactly | ||
| 224 | // when there is no frame to build or send. (stringMap is a std::array, | ||
| 225 | // so its empty() is a compile time false and cannot answer this.) | ||
| 226 |
2/4✓ Branch 0 taken 19913 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 19913 times.
✗ Branch 3 not taken.
|
19913 | bool hasStrings() const { return m_pru0.maxStringLen || m_pru1.maxStringLen; } |
| 227 | |||
| 228 | uint32_t m_curFrame = 0; | ||
| 229 | uint32_t m_licensedOutputs = 0; | ||
| 230 | |||
| 231 | // back-pressure gate accounting, touched only on the output thread. | ||
| 232 | // Members rather than function statics so a config reload starts a fresh | ||
| 233 | // window and so a removed output cannot leave its warning stuck in the UI. | ||
| 234 | uint32_t m_bpOffered = 0; | ||
| 235 | uint32_t m_bpDeclined = 0; | ||
| 236 | bool m_bpLastDeclined = false; // previous window declined any frames | ||
| 237 | bool m_bpWarned = false; | ||
| 238 | std::chrono::steady_clock::time_point m_bpWindowStart{}; | ||
| 239 | void setFrameRateWarning(bool on); | ||
| 240 | |||
| 241 | // predictive frame-rate budget: computed once in Init() from the longest | ||
| 242 | // string, compared against the sequence rate whenever output starts | ||
| 243 | int m_frameTimeUs = 0; | ||
| 244 | float m_lastBudgetRate = -1.0f; | ||
| 245 | std::string m_budgetWarnText; | ||
| 246 | void checkFrameRateBudget(float rate); | ||
| 247 | void clearBudgetWarning(); | ||
| 248 | |||
| 249 | int m_testCycle = -1; | ||
| 250 | int m_testType = 0; | ||
| 251 | float m_testPercent = 0.0f; | ||
| 252 | |||
| 253 | FalconV5Support* falconV5Support = nullptr; | ||
| 254 | std::list<std::string> m_autoCreatedModelNames; | ||
| 255 | bool supportsV5Listeners = false; | ||
| 256 | // any configured Falcon V5 (bidirectional) receivers after the | ||
| 257 | // capability checks; V4 (send-only) chains do not set this | ||
| 258 | bool m_hasBidirSR = false; | ||
| 259 | // the cape wires the enable line to PRU1 and we configured that pin; | ||
| 260 | // this also means PRU1 must be running, see Init() | ||
| 261 | bool m_usesEnablePin = false; | ||
| 262 | |||
| 263 | void prepData(FrameData& d, unsigned char* channelData); | ||
| 264 | void sendData(FrameData& d); | ||
| 265 | void bitFlipData(uint8_t* stringChannelData, uint8_t* bitSwapped, size_t len); | ||
| 266 | // The hot paths are templated on the chain depth and dispatched once per | ||
| 267 | // frame rather than reading m_stringsPerPin in the inner loops, so the 8 | ||
| 268 | // deep path keeps the compile time constants (and the codegen) it had. | ||
| 269 | template<int SPP> | ||
| 270 | void prepDataT(FrameData& d, unsigned char* channelData); | ||
| 271 | template<int SPP> | ||
| 272 | void bitFlipDataT(uint8_t* stringChannelData, uint8_t* bitSwapped, size_t len); | ||
| 273 | |||
| 274 | // AM62x shared-memory ring pump (see pru/SMEMRing.hp) | ||
| 275 | std::thread m_pumpThread; | ||
| 276 | std::atomic<bool> m_pumpRunning{ false }; | ||
| 277 | uint8_t* m_fv5PacketMem = nullptr; | ||
| 278 | void runPumpThread(); | ||
| 279 | bool pumpFrameData(FrameData& d); | ||
| 280 | |||
| 281 | void createOutputLengths(FrameData& d, const std::string& pfx); | ||
| 282 | // report ring resynchronizations the firmware has had to make | ||
| 283 | void reportRingResync(FrameData& d, int pru); | ||
| 284 | |||
| 285 | // Bit cell timing in ns, published to the firmware as cycle counts. These | ||
| 286 | // are the values the firmware used to have compiled in; keeping them here | ||
| 287 | // is what lets a different bit cell (a 400kHz part) be selected without a | ||
| 288 | // separate firmware image. | ||
| 289 | int m_t0Ns = 320; | ||
| 290 | int m_t1Ns = 750; | ||
| 291 | int m_lowNs = 1120; | ||
| 292 | |||
| 293 | // ordering so a Timing can key a map; the values are small and exact | ||
| 294 | struct TimingLess { | ||
| 295 | 510 | bool operator()(const PixelString::Timing& a, const PixelString::Timing& b) const { | |
| 296 |
1/2✓ Branch 0 taken 510 times.
✗ Branch 1 not taken.
|
510 | return std::tie(a.t0Ns, a.t1Ns, a.periodNs) < std::tie(b.t0Ns, b.t1Ns, b.periodNs); |
| 297 | } | ||
| 298 | }; | ||
| 299 | // pick the controller's bit cell from the protocols actually in use | ||
| 300 | void resolveTiming(const Json::Value& config); | ||
| 301 | |||
| 302 | void demoteInvertedReceiverChains(); | ||
| 303 | |||
| 304 | void setupFalconV5Support(const Json::Value& root, uint8_t* memLoc); | ||
| 305 | void encodeFalconV5Packet(std::vector<std::array<uint8_t, 64>>& packets, uint8_t* memLocPru0, uint8_t* memLocPru1); | ||
| 306 | |||
| 307 | // The packet phase is clocked out on every pin of the cape at once, so a | ||
| 308 | // send-only (V4) chain is on the bus for every packet frame the V5 chains | ||
| 309 | // need, not just its own config frames. Its slot must carry a valid V4 | ||
| 310 | // config packet in every one of them: an all-zero slot is a framed packet | ||
| 311 | // with a token the receiver does not know, and a V4 receiver is specified | ||
| 312 | // to get its config packet regularly after a data frame anyway. The | ||
| 313 | // encoded packet is cached per chain and stamped into every frame. | ||
| 314 | std::vector<std::pair<int, std::array<uint8_t, 64>>> m_sendOnlyConfigPackets; | ||
| 315 | void applySendOnlyConfigPackets(std::vector<std::array<uint8_t, 64>>& packets); | ||
| 316 | }; | ||
| 317 |