/**
 * user-sample - CIC-IEC 2026 Stage 3 Evaluation Program
 *
 * (ULTIMATE PERFORMANCE DMA + PING-PONG BUFFER ARCHITECTURE v2)
 *
 * Optimizations over v1:
 *   1. Pre-fire DMA Ch0 load at loop tail → fully overlaps with IP compute
 *   2. Inlined wait loops   → zero function-call overhead in hot path
 *   3. Inlined DMA start    → skips redundant STATUS_BUSY_BIT check
 *   4. XOR bank toggle      → avoids g%2 recomputation per iteration
 */

#include <dma.h>
#include <confreg_time.h>

/* ─── UART 16550 direct access (bypasses slow printf/picolibc) ──────── */
#define UART_THR (*(volatile unsigned char *)0xbf000000)
#define UART_LSR (*(volatile unsigned char *)0xbf000005)
#define LSR_THRE  0x20

static inline void uart_tx(unsigned char c) {
    while (!(UART_LSR & LSR_THRE)) {}
    UART_THR = c;
}

static void uart_puts(const char *s) {
    while (*s) uart_tx(*s++);
}

static void uart_puthex(unsigned int val) {
    int i;
    for (i = 28; i >= 0; i -= 4) {
        unsigned char nibble = (val >> i) & 0xF;
        uart_tx(nibble < 10 ? '0' + nibble : 'a' + nibble - 10);
    }
}

static void uart_putdec(unsigned int val) {
    char buf[10];
    int i = 0;

    if (val == 0) {
        uart_tx('0');
        return;
    }

    while (val != 0) {
        buf[i++] = (char)('0' + (val % 10));
        val /= 10;
    }
    while (i > 0) {
        uart_tx((unsigned char)buf[--i]);
    }
}

/* ─── Configurable group count (default 5000) ──────────────────────── */
#ifndef MATMUL_GROUP_NUM
#define MATMUL_GROUP_NUM 5000
#endif

/* ─── BSP globals ─────────────────────────────────────────────────── */
unsigned long UART_BASE              = 0xbf000000;
unsigned long CONFREG_TIMER_BASE     = 0xbf20f100;
unsigned long CONFREG_CLOCKS_PER_SEC = 50000000L;
unsigned long CORE_CLOCKS_PER_SEC    = 33000000L;

/* ─── Memory layout ────────────────────────────────────────────────── */
#define EXTRAM_BASE_PHYS  0x1c400000
#define EXTRAM_BASE_UNCACHED (0xA0000000 | (EXTRAM_BASE_PHYS & 0x1FFFFFFF))

#define SRC_BASE_WORDS    0
#define DST_BASE_WORDS    (SRC_BASE_WORDS + MATMUL_GROUP_NUM * 32)
#define RESULT_WORDS      (MATMUL_GROUP_NUM * 48)

/* ExtRAM Physical Addresses for DMA */
#define EXTRAM_PHYS_AB(g) (EXTRAM_BASE_PHYS + (SRC_BASE_WORDS * 4) + (g) * 128)
#define EXTRAM_PHYS_C(g)  (EXTRAM_BASE_PHYS + (DST_BASE_WORDS * 4) + (g) * 192)

/* ─── MATMUL IP registers (Ping-Pong Memory Map) ──────────────────── */
#define MATMUL_BASE_PHYS     0x1f500000
#define MATMUL_BASE_UNCACHED 0xbf500000

#define MATMUL_CTRL       (MATMUL_BASE_UNCACHED + 0x000)
#define MATMUL_STATUS     (MATMUL_BASE_UNCACHED + 0x004)
#define MATMUL_BASE_PHYS_FOR_DMA MATMUL_BASE_PHYS

#define CTRL_START_B0     0x1
#define CTRL_START_B1     0x2
#define STATUS_DONE       0x2
#define STATUS_ERROR      0x4

#define MATMUL_CRC (MATMUL_BASE_UNCACHED + 0x010)

/* ─── Pre-computed DMA burst control word ──────────────────────────── */
#define DMA_CTRL_VAL (CTRL_BURST_SIZE(2) | CTRL_BURST_LEN(63) | CTRL_START_BIT)
#define MATMUL_EP_ID 0

/* ═══════════════════════════════════════════════════════════════════ */
int main(void)
{
    if (MATMUL_GROUP_NUM <= 0) return 0;

    /* ── Hot-path local volatile pointers (one-time addr calc) ──────── */
    volatile unsigned int *mat_crc    = (volatile unsigned int *)MATMUL_CRC;
    volatile dma_seq_regs_t *seq = DMA_SEQ;

    unsigned int crc;
    unsigned int cycle_start, cycle_end, cycles, time_us;

    /* ── Start marker (must be first output) ──────────────────── */
    uart_puts("MATMUL_START\r\n");
    cycle_start = get_clock_count();

    seq->STATUS     = SEQ_DONE_BIT;
    seq->SRC_BASE   = EXTRAM_PHYS_AB(0);
    seq->DST_BASE   = EXTRAM_PHYS_C(0);
    seq->GROUPS     = MATMUL_GROUP_NUM;
    seq->MAT_BASE   = MATMUL_BASE_PHYS_FOR_DMA;
    seq->SRC_STRIDE = 128;
    seq->DST_STRIDE = 192;
    seq->CTRL       = SEQ_START_BIT;
    while (!(seq->STATUS & SEQ_DONE_BIT)) {}
    seq->STATUS = SEQ_DONE_BIT;
    cycle_end = get_clock_count();
    cycles = cycle_end - cycle_start;
    time_us = cycles / (CORE_CLOCKS_PER_SEC / 1000000UL);

    /* ── Read hardware CRC32 ──────────────────────────────────── */
    crc = *mat_crc;

    uart_puts("MATMUL_CRC32=");
    uart_puthex(crc);
    uart_puts("\r\nMATMUL_CYCLES=");
    uart_putdec(cycles);
    uart_puts("\r\nMATMUL_TIME_US=");
    uart_putdec(time_us);
    uart_puts("\r\nMATMUL_DONE\r\n");

    return 0;
}