/**
 * matmul_test - ExtRAM-based flow test for matmul_axi_slave IP
 *
 * Reads source matrices from ExtRAM, computes C = A x B via the IP,
 * writes results back to ExtRAM.  Designed to work with matmul_flow_tb.v.
 *
 * Data layout in ExtRAM (matches testbench & generate_matmul_testdata.py):
 *   SRC_BASE = 0                        (word offset)
 *   DST_BASE = SRC_BASE + N_GROUPS * 32
 *   Group i: A_i @ SRC_BASE + i*32 + 0  (16 words)
 *            B_i @ SRC_BASE + i*32 + 16 (16 words)
 *            C_i @ DST_BASE + i*48      (48 words, 3 per element)
 *
 * Register map (byte offsets from MATMUL_BASE = 0x1f500000):
 *   0x000  CTRL       [W] bit0=1 starts computation
 *   0x004  STATUS     [R] bit2=error, bit1=done, bit0=busy
 *   0x008  SRC_BASE   [R/W] unused (IP has no DMA yet)
 *   0x00c  DST_BASE   [R/W] unused
 *   0x010  GROUP_NUM  [R/W] unused
 *   0x020  A_BASE     [R/W] 16 words of matrix A (row-major)
 *   0x060  B_BASE     [R/W] 16 words of matrix B (row-major)
 *   0x0a0  C_BASE     [R]   48 words of result C (3 words per element)
 */

#include <stdio.h>
#include <stdint.h>

/* ─── BSP globals (required by common.mk / picolibc) ─────────────── */
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       0x1c400000   /* ExtRAM: bit22=1 in RAM controller */

#define SRC_BASE          0             /* word offset — source A/B matrices  */
#define N_GROUPS          10            /* must match testbench default       */
#define DST_BASE          (SRC_BASE + N_GROUPS * 32)   /* result region start */

/* ─── MATMUL IP register map ─────────────────────────────────────── */
#define MATMUL_BASE       0x1f500000

#define MATMUL_CTRL       (MATMUL_BASE + 0x000)
#define MATMUL_STATUS     (MATMUL_BASE + 0x004)
#define MATMUL_A_BASE     (MATMUL_BASE + 0x020)
#define MATMUL_B_BASE     (MATMUL_BASE + 0x060)
#define MATMUL_C_BASE     (MATMUL_BASE + 0x0a0)

/* CTRL bits */
#define CTRL_START        0x1

/* STATUS bits */
#define STATUS_BUSY       0x1
#define STATUS_DONE       0x2
#define STATUS_ERROR      0x4

/* ─── Convenience macros ─────────────────────────────────────────── */
#define REG_WR(addr, val) (*(volatile uint32_t *)(addr) = (uint32_t)(val))
#define REG_RD(addr)      (*(volatile uint32_t *)(addr))

/* ─── Read N words from IP register space ────────────────────────── */
static void read_ip(uint32_t base, volatile uint32_t *dst, int n)
{
    int i;
    for (i = 0; i < n; i++)
        dst[i] = REG_RD(base + i * 4);
}

/* ─── Write N words to IP register space ─────────────────────────── */
static void write_ip(uint32_t base, const volatile uint32_t *src, int n)
{
    int i;
    for (i = 0; i < n; i++)
        REG_WR(base + i * 4, src[i]);
}

/* ─── Wait until IP is done, return 0 on success, -1 on error ────── */
static int wait_done(void)
{
    while (1) {
        uint32_t status = REG_RD(MATMUL_STATUS);
        if (status & STATUS_ERROR) {
            printf("[ERROR] matmul IP error (status=0x%08lx)\n",
                   (unsigned long)status);
            return -1;
        }
        if (status & STATUS_DONE)
            return 0;
    }
}

/* ─── Main ────────────────────────────────────────────────────────── */
int main(void)
{
    volatile uint32_t *extram = (volatile uint32_t *)EXTRAM_BASE;
    uint32_t a_buf[16], b_buf[16], c_buf[48];
    int g, pass = 0, fail = 0;

    printf("\n===== MATMUL ExtRAM Flow Test =====\n");
    printf("ExtRAM base: 0x%08lx\n", (unsigned long)EXTRAM_BASE);
    printf("Groups: %d  SRC_BASE=%d  DST_BASE=%d\n", N_GROUPS, SRC_BASE, DST_BASE);
    printf("====================================\n");

    for (g = 0; g < N_GROUPS; g++) {
        int i;

        printf("\n--- Group %d ---\n", g);

        /* 1. Read A[16] and B[16] from ExtRAM source region */
        printf("  Reading A/B from ExtRAM...\n");
        for (i = 0; i < 16; i++) {
            a_buf[i] = extram[SRC_BASE + g * 32 + i];
            b_buf[i] = extram[SRC_BASE + g * 32 + 16 + i];
        }

        /* 2. Write A and B to IP registers */
        printf("  Loading A/B into IP...\n");
        write_ip(MATMUL_A_BASE, a_buf, 16);
        write_ip(MATMUL_B_BASE, b_buf, 16);

        /* 3. Clear stale done / error flags */
        REG_WR(MATMUL_STATUS, STATUS_DONE | STATUS_ERROR);

        /* 4. Start computation */
        printf("  Starting computation...\n");
        REG_WR(MATMUL_CTRL, CTRL_START);

        /* 5. Wait for done (poll STATUS) */
        if (wait_done() != 0) {
            printf("  Group %d: FAIL (IP error)\n", g);
            fail++;
            /* zero-fill result area so testbench clearly sees failure */
            for (i = 0; i < 48; i++) c_buf[i] = 0;
        } else {
            /* 6. Read result C[48] from IP */
            printf("  Reading C from IP...\n");
            read_ip(MATMUL_C_BASE, c_buf, 48);
            printf("  Group %d: OK\n", g);
            pass++;
        }

        /* 7. Write result C to ExtRAM result region */
        for (i = 0; i < 48; i++)
            extram[DST_BASE + g * 48 + i] = c_buf[i];
    }

    /* ── Final report ─────────────────────────────────────────── */
    printf("\n============================\n");
    printf("RESULTS: %d passed, %d failed\n", pass, fail);
    printf("============================\n");

    /* Signal completion to testbench via UART */
    printf("MATMUL_DONE\n");

    return 0;
}