// retroDE_ps2 — ps2_sh3_tex_upload (Ch352)
//
// Uploads the reconstructed 512x512 PSMT8 SH3 texture (65536 32-bit words) into FPGA-private LPDDR4B via the
// PS2 HPS-bridge write-probe, verifies the readback, arms the texture-cache fill, and retriggers the feeder.
// One command — mmap'd register pokes (NOT 65536 devmem process spawns). Same bridge protocol as ps2_feeder.c
// and docs/hardware/ps2_lpddr_tex_test.sh (Ch322), just scaled to the full 256 KiB texture.
//
// Build on the board:   gcc -O2 -o ps2_sh3_tex_upload ps2_sh3_tex_upload.c
// Run (after fit+boot): sudo ./ps2_sh3_tex_upload sh3_real_tex_lpddr.mem
//   (copy sh3_real_tex_lpddr.mem from sim/data/top_psmct32_raster_demo/ to the board alongside the binary.)
//
// Sequence: (1) write WRADDR=0x200000, stream 65536 words to WRDATA; (2) read each back via the read-probe and
// confirm sum32/xor32 match the file; (3) arm cache fill (0x054), poll fill_done, check beats/bytes/rd_errs;
// (4) pulse the feeder retrigger (0x0E8) so the scene re-renders with the now-warm cache.

#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <errno.h>
#include <fcntl.h>
#include <unistd.h>
#include <sys/mman.h>

#define OFF_LPDDR_STATUS   0x02C   // R: [3] rd_pending (read-probe in flight)
#define OFF_LPDDR_RDADDR   0x03C   // W: set read byte addr + trigger ; R: 32-bit word
#define OFF_LPDDR_WRADDR   0x04C   // W: set LPDDR byte addr (auto-increments +4 per WRDATA write)
#define OFF_LPDDR_WRDATA   0x050   // W: data word -> single 32-bit LPDDR write + addr += 4
#define OFF_TEX_FILL_CTRL  0x054   // W[0]: arm cache fill ; R: [0]fill_done [1]wr_busy
#define OFF_LPDDR_WR_ERRS  0x06C   // R: write-probe non-OKAY (BRESP) responses (expect 0)
#define WR_PENDING_BIT     0x4     // 0x054 bit2 — Ch352 STABLE write-done flag (poll instead of transient wr_busy)
#define RD_PENDING_BIT     0x8     // 0x02C bit3
#define OFF_TEX_FILL_BEATS 0x058   // R: beats filled  (expect TEX_BYTES/32 = 8192)
#define OFF_TEX_FILL_BYTES 0x05C   // R: bytes filled  (expect 262144)
#define OFF_TEX_RD_ERRS    0x068   // R: fill non-OKAY read responses (expect 0)
#define OFF_TEX_FILL_CRC   0x070   // R: sum32 of EVERY word the cache wrote into tex_mem (must == file sum32)
#define OFF_FEEDER_GO      0x0E8   // W[0]: trigger/retrigger the feeder
// Ch353 — GS_SH3_LPDDR_FB (--lpddr-fb) additions: preclear the LPDDR framebuffer + drive Codex's host-start sequence.
#define OFF_LPDDR_CTRL     0x018   // W: [0]arm [1]canary [2]video_src [3]scanout_lb (arm/base/canary latch on write)
#define OFF_LPDDR_FB_BASE  0x01C   // W: FB base byte addr (writer fb_base); set BEFORE the 0x018 arm-commit
#define OFF_LPDDR_STATUS   0x02C   // R: [0]idle (writer drained, stable post-render) [3]rd_pending
#define OFF_LPDDR_BYTES    0x030   // R: FB writer beats written (snapped at idle) — non-zero after a real render
#define OFF_FEEDER_STATUS  0x0D8   // R: [0] feeder control FSM in C_READY (setup + CLUT done)
#define IDLE_BIT           0x1     // 0x02C bit0 — writer idle (TRANSIENT: pulses between feeder render batches)
#define FRAME_DRAINED_BIT  0x40    // 0x02C bit6 — Ch353 STABLE ordered drain ack (poll THIS, not idle)
#define READY_BIT          0x1     // 0x0D8 bit0

#define N_WORDS   65536            // 512*512 PSMT8 / 4
#define TEX_BYTES 262144
#define N_BEATS   8192             // TEX_BYTES / 32
#define FB_WORDS  85504            // Ch353 — 256*334 PSMCT32 framebuffer (0x000000..0x0537FF)
#define FB_BYTES  342016

typedef struct { volatile uint8_t *base; int dry; } bridge_t;
static void     wr32(bridge_t *b, int off, uint32_t v){ if(!b->dry) *(volatile uint32_t*)(b->base+off)=v; }
static uint32_t rd32(bridge_t *b, int off){ return b->dry?0:*(volatile uint32_t*)(b->base+off); }

int main(int argc, char **argv){
    unsigned long base = 0x40000000UL;     // PS2 HPS-bridge base (override --base or PS2_BRIDGE_BASE)
    unsigned long lpddr_base = 0x00200000; // EMIF byte base where the texture is staged (= TEX_LPDDR_BASE RTL)
    const char *texfile = "sh3_real_tex_lpddr.mem";
    int dry=0, do_fill=1, do_retrig=1, lpddr_fb=0, fb_rows=334;   // fb_rows: 334 (Ch353 single) / 338 (Ch354 multi)
    char *env = getenv("PS2_BRIDGE_BASE"); if (env) base = strtoul(env,NULL,0);
    for (int i=1;i<argc;i++){
        if      (!strcmp(argv[i],"--base") && i+1<argc) base = strtoul(argv[++i],NULL,0);
        else if (!strcmp(argv[i],"--lpddr-base") && i+1<argc) lpddr_base = strtoul(argv[++i],NULL,0);
        else if (!strcmp(argv[i],"--dry-run")) dry=1;
        else if (!strcmp(argv[i],"--no-fill")) do_fill=0;
        else if (!strcmp(argv[i],"--no-retrigger")) do_retrig=0;
        else if (!strcmp(argv[i],"--lpddr-fb")) lpddr_fb=1;   // Ch353 — preclear FB + Codex host-start sequence
        else if (!strcmp(argv[i],"--fb-rows") && i+1<argc) fb_rows = (int)strtoul(argv[++i],NULL,0); // Ch354 — 338 for multi
        else if (argv[i][0] != '-') texfile = argv[i];
        else { fprintf(stderr,"usage: %s [tex.mem] [--base 0x40000000] [--lpddr-base 0x200000] [--dry-run] [--no-fill] [--no-retrigger] [--lpddr-fb]\n", argv[0]); return 2; }
    }

    // ---- load the texture hex (.mem: one 32-bit word per line) ----
    static uint32_t tex[N_WORDS];
    FILE *f = fopen(texfile,"r");
    if (!f){ fprintf(stderr,"error: cannot open '%s': %s\n", texfile, strerror(errno)); return 1; }
    int n=0; char line[64];
    while (n<N_WORDS && fgets(line,sizeof line,f)){
        char *s=line; while(*s==' '||*s=='\t') s++;
        if (*s=='/'||*s=='\n'||*s==0) continue;       // skip blank / // banner lines
        tex[n++] = (uint32_t)strtoul(s,NULL,16);
    }
    fclose(f);
    if (n != N_WORDS){ fprintf(stderr,"error: %s has %d words, expected %d\n", texfile, n, N_WORDS); return 1; }

    // expected checksums (source of truth = the file)
    uint32_t sum=0, xr=0; for (int i=0;i<N_WORDS;i++){ sum+=tex[i]; xr^=tex[i]; }
    printf("[ps2_sh3_tex_upload] %d words from %s  sum32=0x%08x xor32=0x%08x -> LPDDR 0x%lx (bridge base 0x%lx%s)\n",
           n, texfile, sum, xr, lpddr_base, base, dry?", DRY-RUN":"");

    // ---- open the bridge ----
    bridge_t br = {0,dry}; int fd=-1; void *map=NULL;
    if (!dry){
        fd=open("/dev/mem", O_RDWR|O_SYNC);
        if (fd<0){ fprintf(stderr,"error: open /dev/mem (run as root?): %s\n", strerror(errno)); return 1; }
        map=mmap(NULL,0x1000,PROT_READ|PROT_WRITE,MAP_SHARED,fd,(off_t)base);
        if (map==MAP_FAILED){ fprintf(stderr,"error: mmap 0x%lx: %s\n", base, strerror(errno)); close(fd); return 1; }
        br.base=(volatile uint8_t*)map;
    }

    // ---- Ch353 GS_SH3_LPDDR_FB (Codex host sequence, steps 1-2): wait for the feeder to reach C_READY (setup +
    // CLUT complete; with FEEDER_AUTOSTART=0 this happens WITHOUT a boot render), then PRECLEAR the LPDDR
    // framebuffer 0x000000..0x0537FF (85504 words) via the same proven write-probe. This is the "no hardware clear
    // engine" path (Codex): ~30% more traffic than the 65536-word texture upload, poll write_pending + check BRESP. ----
    if (lpddr_fb && !dry){
        int g=0; while(!(rd32(&br,OFF_FEEDER_STATUS)&READY_BIT) && g<20000000) g++;
        printf("[ps2_sh3_tex_upload] (lpddr-fb) feeder ready=%d (setup+CLUT done, no boot render)\n",
               (rd32(&br,OFF_FEEDER_STATUS)&READY_BIT)?1:0);
        unsigned fb_words = 256u*(unsigned)fb_rows;        // 256x334=85504 (Ch353) ; 256x338=86528=0x54800 (Ch354)
        wr32(&br, OFF_LPDDR_WRADDR, 0x00000000u);          // FB base 0
        for (unsigned i=0;i<fb_words;i++){
            wr32(&br, OFF_LPDDR_WRDATA, 0u);
            { int h=0; while ((rd32(&br,OFF_TEX_FILL_CTRL)&WR_PENDING_BIT) && h<2000000) h++; }
        }
        uint32_t perr = rd32(&br, OFF_LPDDR_WR_ERRS);
        printf("[ps2_sh3_tex_upload] (lpddr-fb) precleared FB %u words (%u rows, %u KiB) @0..0x%x  wr_bresp_errs=%u (exp 0)\n",
               fb_words, (unsigned)fb_rows, fb_words*4/1024, fb_words*4, perr);
        if (perr) fprintf(stderr,"WARN: %u preclear BRESP errors.\n", perr);
    }

    // ---- (1) upload: set WRADDR then stream WRDATA. CRITICAL: poll wr_busy (0x054 bit1) clear after each word
    // so the write-probe actually COMMITS before the next write — otherwise the fast mmap writes outrun the
    // CDC/AXI commit and get DROPPED (the bug: most words read back as 0). The Ch322 devmem script got away with
    // no poll only because devmem process-spawns are slow. ----
    wr32(&br, OFF_LPDDR_WRADDR, (uint32_t)lpddr_base);
    for (int i=0;i<N_WORDS;i++){
        wr32(&br, OFF_LPDDR_WRDATA, tex[i]);
        // wait for the STABLE write_pending (0x054 bit2) to clear — the probe committed this word
        if (!dry){ int g=0; while ((rd32(&br, OFF_TEX_FILL_CTRL) & WR_PENDING_BIT) && g<2000000) g++; }
    }
    if (!dry){
        uint32_t werr = rd32(&br, OFF_LPDDR_WR_ERRS);
        printf("[ps2_sh3_tex_upload] uploaded %d words (%d KiB).  wr_bresp_errs=%u (exp 0)\n", N_WORDS, TEX_BYTES/1024, werr);
        if (werr) fprintf(stderr,"WARN: %u write BRESP errors.\n", werr);
    } else printf("[ps2_sh3_tex_upload] uploaded %d words (%d KiB).\n", N_WORDS, TEX_BYTES/1024);

    // ---- (2) readback verify via the read-probe (guardrail #2). Poll rd_pending (0x02C bit3) clear before
    // reading the latched word — the LPDDR read has latency; reading immediately returns 0/stale. ----
    int mism=0; uint32_t rsum=0, rxr=0;
    if (!dry){
        for (int i=0;i<N_WORDS;i++){
            wr32(&br, OFF_LPDDR_RDADDR, (uint32_t)(lpddr_base + (unsigned)i*4));  // set addr + trigger read
            { int g=0; while ((rd32(&br, OFF_LPDDR_STATUS) & RD_PENDING_BIT) && g<1000000) g++; }
            uint32_t v = rd32(&br, OFF_LPDDR_RDADDR);                              // latched word
            rsum+=v; rxr^=v;
            if (v != tex[i] && mism<8) fprintf(stderr,"  readback mismatch @word %d: got 0x%08x exp 0x%08x\n", i, v, tex[i]);
            if (v != tex[i]) mism++;
        }
        printf("[ps2_sh3_tex_upload] readback sum32=0x%08x xor32=0x%08x  mismatches=%d\n", rsum, rxr, mism);
        if (mism){ fprintf(stderr,"FAIL: %d readback mismatches — bad upload, NOT filling cache.\n", mism); munmap(map,0x1000); close(fd); return 1; }
    }

    // ---- (3) arm the cache fill + poll fill_done; check beats/bytes/rd_errs ----
    if (do_fill && !dry){
        wr32(&br, OFF_TEX_FILL_CTRL, 0x1);
        int done=0; for (int i=0;i<200000;i++){ if (rd32(&br,OFF_TEX_FILL_CTRL)&0x1){ done=1; break; } }
        uint32_t beats=rd32(&br,OFF_TEX_FILL_BEATS), bytes=rd32(&br,OFF_TEX_FILL_BYTES), errs=rd32(&br,OFF_TEX_RD_ERRS);
        uint32_t fcrc=rd32(&br,OFF_TEX_FILL_CRC);
        printf("[ps2_sh3_tex_upload] cache fill: done=%d beats=%u (exp %d) bytes=%u (exp %d) rd_errs=%u (exp 0)\n",
               done, beats, N_BEATS, bytes, TEX_BYTES, errs);
        // The cache's sum32 over the words it wrote into tex_mem must equal the file's sum32. If it matches,
        // tex_mem is byte-correct on silicon — so any residual texture corruption is NOT the cache contents.
        printf("[ps2_sh3_tex_upload] cache fill_crc=0x%08x (exp file sum32=0x%08x) -> tex_mem %s\n",
               fcrc, sum, (fcrc==sum) ? "INTEGRITY OK" : "CORRUPT");
        if (!done || beats!=N_BEATS || bytes!=TEX_BYTES || errs!=0)
            fprintf(stderr,"WARN: cache fill stats off — texels may be wrong; check EMIF cal + LPDDR base.\n");
        if (fcrc!=sum)
            fprintf(stderr,"WARN: cache fill_crc mismatch — tex_mem corrupt on board (NOT a divider/sampler issue).\n");
    }

    // ---- Ch353 GS_SH3_LPDDR_FB (Codex host sequence, steps 5-8): base 0 + canary off + ARM the writer, GO, await
    // the ordered drain, then enable the scanout. The render epoch (RTL) emits exactly ONE EOF on the post-GO
    // feeder_ready rise -> ONE frame_drained, which HARD-gates the line-buffer scanout (no timeout; failure stays
    // black). Confirm the drain via LPDDR_STATUS[0] idle (stable post-render) + LPDDR_BYTES!=0 (a frame was written). ----
    if (lpddr_fb && !dry){
        wr32(&br, OFF_LPDDR_FB_BASE, 0x00000000u);          // step 5: FB base 0 ...
        wr32(&br, OFF_LPDDR_CTRL,    0x1);                  // ... canary off, ARM the writer (commit snapshots base/arm)
        wr32(&br, OFF_FEEDER_GO,     0x1);                  // step 6: GO -> render epoch -> one EOF -> one frame_drained
        printf("[ps2_sh3_tex_upload] (lpddr-fb) writer armed (base 0, canary off), feeder GO issued.\n");
        // step 7: await the ORDERED drain ack (0x02C[6] frame_drained) — STABLE, asserts only after the EOF marker's
        // last BRESP. Do NOT poll [0]idle: it pulses between the feeder's render batches and reads a mid-render count.
        { int g=0; while(!(rd32(&br,OFF_LPDDR_STATUS)&FRAME_DRAINED_BIT) && g<40000000) g++; }
        uint32_t fd = (rd32(&br,OFF_LPDDR_STATUS)&FRAME_DRAINED_BIT)?1:0, fbb = rd32(&br,OFF_LPDDR_BYTES);
        printf("[ps2_sh3_tex_upload] (lpddr-fb) drain: frame_drained=%u, FB beats written=%u (exp ~6500 for the full frame)\n", fd, fbb);
        if (!fd || fbb==0)
            fprintf(stderr,"WARN: drain not confirmed (frame_drained=%u beats=%u) — display stays BLACK (hard gate, no timeout).\n", fd, fbb);
        wr32(&br, OFF_LPDDR_CTRL, 0x1|0x4);                 // step 8: arm=1 + video_src=1 -> HDMI sources the LPDDR scanout
        printf("[ps2_sh3_tex_upload] (lpddr-fb) video_src=1 -> HDMI now sources the LPDDR line-buffer scanout (256x%u).\n", (unsigned)fb_rows);
    }
    // ---- (4) retrigger the feeder so the scene re-renders with the warm cache (crop profile path) ----
    else if (do_retrig && !dry){ wr32(&br, OFF_FEEDER_GO, 0x1); printf("[ps2_sh3_tex_upload] feeder retriggered.\n"); }

    if (!dry){ munmap(map,0x1000); close(fd); }
    printf("[ps2_sh3_tex_upload] DONE — check HDMI vs the crop reference (recon/sh3_real_ref.png).\n");
    return 0;
}