/**
 * matmul_test - CIC-IEC 2026 Stage 3 Evaluation Program
 *
 * Uses the CDMA-to-matmul direct endpoint sequencer. The CDMA still owns the
 * external-memory traffic, while matmul A/B/C transfers use local wires instead
 * of the AXI crossbar.
 */

#include <stdint.h>
#include <dma.h>

/* UART 16550 direct access avoids the printf/newlib overhead on the hot path. */
#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((unsigned char)*s++);
    }
}

static void uart_puthex(uint32_t val)
{
    int i;

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

#ifndef MATMUL_GROUP_NUM
#define MATMUL_GROUP_NUM 5000
#endif

unsigned long UART_BASE = 0xbf000000;

#define EXTRAM_BASE_PHYS  0x1c400000u

#define SRC_BASE_WORDS    0u
#define DST_BASE_WORDS    (SRC_BASE_WORDS + MATMUL_GROUP_NUM * 32u)

#define EXTRAM_PHYS_AB(g) (EXTRAM_BASE_PHYS + (SRC_BASE_WORDS * 4u) + (uint32_t)(g) * 128u)
#define EXTRAM_PHYS_C(g)  (EXTRAM_BASE_PHYS + (DST_BASE_WORDS * 4u) + (uint32_t)(g) * 192u)

#define MATMUL_BASE_UNCACHED 0xbf500000u
#define MATMUL_CRC          (*(volatile uint32_t *)(MATMUL_BASE_UNCACHED + 0x010u))

int main(void)
{
    uint32_t crc;

    if (MATMUL_GROUP_NUM <= 0) {
        return 0;
    }

    uart_puts("MATMUL_START\n");

    DMA_SEQ->STATUS = SEQ_DONE_BIT;
    DMA_SEQ->SRC_BASE = EXTRAM_PHYS_AB(0);
    DMA_SEQ->DST_BASE = EXTRAM_PHYS_C(0);
    DMA_SEQ->GROUPS = MATMUL_GROUP_NUM;
    DMA_SEQ->MAT_BASE = 0x1f500000u;
    DMA_SEQ->SRC_STRIDE = 128u;
    DMA_SEQ->DST_STRIDE = 192u;
    DMA_SEQ->CTRL = SEQ_START_BIT;

    while ((DMA_SEQ->STATUS & SEQ_DONE_BIT) == 0u) {
    }
    DMA_SEQ->STATUS = SEQ_DONE_BIT;

    crc = MATMUL_CRC;

    uart_puts("MATMUL_CRC32=");
    uart_puthex(crc);
    uart_puts("\nMATMUL_DONE\n");

    return 0;
}