hw/misc/milkymist-pfpu.c

   1 /*
   2  *  QEMU model of the Milkymist programmable FPU.
   3  *
   4  *  Copyright (c) 2010 Michael Walle <michael@walle.cc>
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
  18  *
  19  *
  20  * Specification available at:
  21  *   http://milkymist.walle.cc/socdoc/pfpu.pdf
  22  *
  23  */
  24
  25 #include "qemu/osdep.h"
  26 #include "hw/irq.h"
  27 #include "hw/sysbus.h"
  28 #include "migration/vmstate.h"
  29 #include "trace.h"
  30 #include "qemu/log.h"
  31 #include "qemu/module.h"
  32 #include "qemu/error-report.h"
  33 #include <math.h>
  34 #include "qom/object.h"
  35
  36 /* #define TRACE_EXEC */
  37
  38 #ifdef TRACE_EXEC
  39 #    define D_EXEC(x) x
  40 #else
  41 #    define D_EXEC(x)
  42 #endif
  43
  44 enum {
  45     R_CTL = 0,
  46     R_MESHBASE,
  47     R_HMESHLAST,
  48     R_VMESHLAST,
  49     R_CODEPAGE,
  50     R_VERTICES,
  51     R_COLLISIONS,
  52     R_STRAYWRITES,
  53     R_LASTDMA,
  54     R_PC,
  55     R_DREGBASE,
  56     R_CODEBASE,
  57     R_MAX
  58 };
  59
  60 enum {
  61     CTL_START_BUSY = (1<<0),
  62 };
  63
  64 enum {
  65     OP_NOP = 0,
  66     OP_FADD,
  67     OP_FSUB,
  68     OP_FMUL,
  69     OP_FABS,
  70     OP_F2I,
  71     OP_I2F,
  72     OP_VECTOUT,
  73     OP_SIN,
  74     OP_COS,
  75     OP_ABOVE,
  76     OP_EQUAL,
  77     OP_COPY,
  78     OP_IF,
  79     OP_TSIGN,
  80     OP_QUAKE,
  81 };
  82
  83 enum {
  84     GPR_X = 0,
  85     GPR_Y = 1,
  86     GPR_FLAGS = 2,
  87 };
  88
  89 enum {
  90     LATENCY_FADD = 5,
  91     LATENCY_FSUB = 5,
  92     LATENCY_FMUL = 7,
  93     LATENCY_FABS = 2,
  94     LATENCY_F2I = 2,
  95     LATENCY_I2F = 3,
  96     LATENCY_VECTOUT = 0,
  97     LATENCY_SIN = 4,
  98     LATENCY_COS = 4,
  99     LATENCY_ABOVE = 2,
 100     LATENCY_EQUAL = 2,
 101     LATENCY_COPY = 2,
 102     LATENCY_IF = 2,
 103     LATENCY_TSIGN = 2,
 104     LATENCY_QUAKE = 2,
 105     MAX_LATENCY = 7
 106 };
 107
 108 #define GPR_BEGIN       0x100
 109 #define GPR_END         0x17f
 110 #define MICROCODE_BEGIN 0x200
 111 #define MICROCODE_END   0x3ff
 112 #define MICROCODE_WORDS 2048
 113
 114 #define REINTERPRET_CAST(type, val) (*((type *)&(val)))
 115
 116 #ifdef TRACE_EXEC
 117 static const char *opcode_to_str[] = {
 118     "NOP", "FADD", "FSUB", "FMUL", "FABS", "F2I", "I2F", "VECTOUT",
 119     "SIN", "COS", "ABOVE", "EQUAL", "COPY", "IF", "TSIGN", "QUAKE",
 120 };
 121 #endif
 122
 123 #define TYPE_MILKYMIST_PFPU "milkymist-pfpu"
 124 OBJECT_DECLARE_SIMPLE_TYPE(MilkymistPFPUState, MILKYMIST_PFPU)
 125
 126 struct MilkymistPFPUState {
 127     SysBusDevice parent_obj;
 128
 129     MemoryRegion regs_region;
 130     Chardev *chr;
 131     qemu_irq irq;
 132
 133     uint32_t regs[R_MAX];
 134     uint32_t gp_regs[128];
 135     uint32_t microcode[MICROCODE_WORDS];
 136
 137     int output_queue_pos;
 138     uint32_t output_queue[MAX_LATENCY];
 139 };
 140
 141 static inline uint32_t
 142 get_dma_address(uint32_t base, uint32_t x, uint32_t y)
 143 {
 144     return base + 8 * (128 * y + x);
 145 }
 146
 147 static inline void
 148 output_queue_insert(MilkymistPFPUState *s, uint32_t val, int pos)
 149 {
 150     s->output_queue[(s->output_queue_pos + pos) % MAX_LATENCY] = val;
 151 }
 152
 153 static inline uint32_t
 154 output_queue_remove(MilkymistPFPUState *s)
 155 {
 156     return s->output_queue[s->output_queue_pos];
 157 }
 158
 159 static inline void
 160 output_queue_advance(MilkymistPFPUState *s)
 161 {
 162     s->output_queue[s->output_queue_pos] = 0;
 163     s->output_queue_pos = (s->output_queue_pos + 1) % MAX_LATENCY;
 164 }
 165
 166 static int pfpu_decode_insn(MilkymistPFPUState *s)
 167 {
 168     uint32_t pc = s->regs[R_PC];
 169     uint32_t insn = s->microcode[pc];
 170     uint32_t reg_a = (insn >> 18) & 0x7f;
 171     uint32_t reg_b = (insn >> 11) & 0x7f;
 172     uint32_t op = (insn >> 7) & 0xf;
 173     uint32_t reg_d = insn & 0x7f;
 174     uint32_t r = 0;
 175     int latency = 0;
 176
 177     switch (op) {
 178     case OP_NOP:
 179         break;
 180     case OP_FADD:
 181     {
 182         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 183         float b = REINTERPRET_CAST(float, s->gp_regs[reg_b]);
 184         float t = a + b;
 185         r = REINTERPRET_CAST(uint32_t, t);
 186         latency = LATENCY_FADD;
 187         D_EXEC(qemu_log("ADD a=%f b=%f t=%f, r=%08x\n", a, b, t, r));
 188     } break;
 189     case OP_FSUB:
 190     {
 191         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 192         float b = REINTERPRET_CAST(float, s->gp_regs[reg_b]);
 193         float t = a - b;
 194         r = REINTERPRET_CAST(uint32_t, t);
 195         latency = LATENCY_FSUB;
 196         D_EXEC(qemu_log("SUB a=%f b=%f t=%f, r=%08x\n", a, b, t, r));
 197     } break;
 198     case OP_FMUL:
 199     {
 200         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 201         float b = REINTERPRET_CAST(float, s->gp_regs[reg_b]);
 202         float t = a * b;
 203         r = REINTERPRET_CAST(uint32_t, t);
 204         latency = LATENCY_FMUL;
 205         D_EXEC(qemu_log("MUL a=%f b=%f t=%f, r=%08x\n", a, b, t, r));
 206     } break;
 207     case OP_FABS:
 208     {
 209         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 210         float t = fabsf(a);
 211         r = REINTERPRET_CAST(uint32_t, t);
 212         latency = LATENCY_FABS;
 213         D_EXEC(qemu_log("ABS a=%f t=%f, r=%08x\n", a, t, r));
 214     } break;
 215     case OP_F2I:
 216     {
 217         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 218         int32_t t = a;
 219         r = REINTERPRET_CAST(uint32_t, t);
 220         latency = LATENCY_F2I;
 221         D_EXEC(qemu_log("F2I a=%f t=%d, r=%08x\n", a, t, r));
 222     } break;
 223     case OP_I2F:
 224     {
 225         int32_t a = REINTERPRET_CAST(int32_t, s->gp_regs[reg_a]);
 226         float t = a;
 227         r = REINTERPRET_CAST(uint32_t, t);
 228         latency = LATENCY_I2F;
 229         D_EXEC(qemu_log("I2F a=%08x t=%f, r=%08x\n", a, t, r));
 230     } break;
 231     case OP_VECTOUT:
 232     {
 233         uint32_t a = cpu_to_be32(s->gp_regs[reg_a]);
 234         uint32_t b = cpu_to_be32(s->gp_regs[reg_b]);
 235         hwaddr dma_ptr =
 236             get_dma_address(s->regs[R_MESHBASE],
 237                     s->gp_regs[GPR_X], s->gp_regs[GPR_Y]);
 238         cpu_physical_memory_write(dma_ptr, &a, 4);
 239         cpu_physical_memory_write(dma_ptr + 4, &b, 4);
 240         s->regs[R_LASTDMA] = dma_ptr + 4;
 241         D_EXEC(qemu_log("VECTOUT a=%08x b=%08x dma=%08x\n", a, b, dma_ptr));
 242         trace_milkymist_pfpu_vectout(a, b, dma_ptr);
 243     } break;
 244     case OP_SIN:
 245     {
 246         int32_t a = REINTERPRET_CAST(int32_t, s->gp_regs[reg_a]);
 247         float t = sinf(a * (1.0f / (M_PI * 4096.0f)));
 248         r = REINTERPRET_CAST(uint32_t, t);
 249         latency = LATENCY_SIN;
 250         D_EXEC(qemu_log("SIN a=%d t=%f, r=%08x\n", a, t, r));
 251     } break;
 252     case OP_COS:
 253     {
 254         int32_t a = REINTERPRET_CAST(int32_t, s->gp_regs[reg_a]);
 255         float t = cosf(a * (1.0f / (M_PI * 4096.0f)));
 256         r = REINTERPRET_CAST(uint32_t, t);
 257         latency = LATENCY_COS;
 258         D_EXEC(qemu_log("COS a=%d t=%f, r=%08x\n", a, t, r));
 259     } break;
 260     case OP_ABOVE:
 261     {
 262         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 263         float b = REINTERPRET_CAST(float, s->gp_regs[reg_b]);
 264         float t = (a > b) ? 1.0f : 0.0f;
 265         r = REINTERPRET_CAST(uint32_t, t);
 266         latency = LATENCY_ABOVE;
 267         D_EXEC(qemu_log("ABOVE a=%f b=%f t=%f, r=%08x\n", a, b, t, r));
 268     } break;
 269     case OP_EQUAL:
 270     {
 271         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 272         float b = REINTERPRET_CAST(float, s->gp_regs[reg_b]);
 273         float t = (a == b) ? 1.0f : 0.0f;
 274         r = REINTERPRET_CAST(uint32_t, t);
 275         latency = LATENCY_EQUAL;
 276         D_EXEC(qemu_log("EQUAL a=%f b=%f t=%f, r=%08x\n", a, b, t, r));
 277     } break;
 278     case OP_COPY:
 279     {
 280         r = s->gp_regs[reg_a];
 281         latency = LATENCY_COPY;
 282         D_EXEC(qemu_log("COPY"));
 283     } break;
 284     case OP_IF:
 285     {
 286         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 287         float b = REINTERPRET_CAST(float, s->gp_regs[reg_b]);
 288         uint32_t f = s->gp_regs[GPR_FLAGS];
 289         float t = (f != 0) ? a : b;
 290         r = REINTERPRET_CAST(uint32_t, t);
 291         latency = LATENCY_IF;
 292         D_EXEC(qemu_log("IF f=%u a=%f b=%f t=%f, r=%08x\n", f, a, b, t, r));
 293     } break;
 294     case OP_TSIGN:
 295     {
 296         float a = REINTERPRET_CAST(float, s->gp_regs[reg_a]);
 297         float b = REINTERPRET_CAST(float, s->gp_regs[reg_b]);
 298         float t = (b < 0) ? -a : a;
 299         r = REINTERPRET_CAST(uint32_t, t);
 300         latency = LATENCY_TSIGN;
 301         D_EXEC(qemu_log("TSIGN a=%f b=%f t=%f, r=%08x\n", a, b, t, r));
 302     } break;
 303     case OP_QUAKE:
 304     {
 305         uint32_t a = s->gp_regs[reg_a];
 306         r = 0x5f3759df - (a >> 1);
 307         latency = LATENCY_QUAKE;
 308         D_EXEC(qemu_log("QUAKE a=%d r=%08x\n", a, r));
 309     } break;
 310
 311     default:
 312         error_report("milkymist_pfpu: unknown opcode %d", op);
 313         break;
 314     }
 315
 316     if (!reg_d) {
 317         D_EXEC(qemu_log("%04d %8s R%03d, R%03d <L=%d, E=%04d>\n",
 318                     s->regs[R_PC], opcode_to_str[op], reg_a, reg_b, latency,
 319                     s->regs[R_PC] + latency));
 320     } else {
 321         D_EXEC(qemu_log("%04d %8s R%03d, R%03d <L=%d, E=%04d> -> R%03d\n",
 322                     s->regs[R_PC], opcode_to_str[op], reg_a, reg_b, latency,
 323                     s->regs[R_PC] + latency, reg_d));
 324     }
 325
 326     if (op == OP_VECTOUT) {
 327         return 0;
 328     }
 329
 330     /* store output for this cycle */
 331     if (reg_d) {
 332         uint32_t val = output_queue_remove(s);
 333         D_EXEC(qemu_log("R%03d <- 0x%08x\n", reg_d, val));
 334         s->gp_regs[reg_d] = val;
 335     }
 336
 337     output_queue_advance(s);
 338
 339     /* store op output */
 340     if (op != OP_NOP) {
 341         output_queue_insert(s, r, latency-1);
 342     }
 343
 344     /* advance PC */
 345     s->regs[R_PC]++;
 346
 347     return 1;
 348 };
 349
 350 static void pfpu_start(MilkymistPFPUState *s)
 351 {
 352     int x, y;
 353     int i;
 354
 355     for (y = 0; y <= s->regs[R_VMESHLAST]; y++) {
 356         for (x = 0; x <= s->regs[R_HMESHLAST]; x++) {
 357             D_EXEC(qemu_log("\nprocessing x=%d y=%d\n", x, y));
 358
 359             /* set current position */
 360             s->gp_regs[GPR_X] = x;
 361             s->gp_regs[GPR_Y] = y;
 362
 363             /* run microcode on this position */
 364             i = 0;
 365             while (pfpu_decode_insn(s)) {
 366                 /* decode at most MICROCODE_WORDS instructions */
 367                 if (++i >= MICROCODE_WORDS) {
 368                     error_report("milkymist_pfpu: too many instructions "
 369                             "executed in microcode. No VECTOUT?");
 370                     break;
 371                 }
 372             }
 373
 374             /* reset pc for next run */
 375             s->regs[R_PC] = 0;
 376         }
 377     }
 378
 379     s->regs[R_VERTICES] = x * y;
 380
 381     trace_milkymist_pfpu_pulse_irq();
 382     qemu_irq_pulse(s->irq);
 383 }
 384
 385 static inline int get_microcode_address(MilkymistPFPUState *s, uint32_t addr)
 386 {
 387     return (512 * s->regs[R_CODEPAGE]) + addr - MICROCODE_BEGIN;
 388 }
 389
 390 static uint64_t pfpu_read(void *opaque, hwaddr addr,
 391                           unsigned size)
 392 {
 393     MilkymistPFPUState *s = opaque;
 394     uint32_t r = 0;
 395
 396     addr >>= 2;
 397     switch (addr) {
 398     case R_CTL:
 399     case R_MESHBASE:
 400     case R_HMESHLAST:
 401     case R_VMESHLAST:
 402     case R_CODEPAGE:
 403     case R_VERTICES:
 404     case R_COLLISIONS:
 405     case R_STRAYWRITES:
 406     case R_LASTDMA:
 407     case R_PC:
 408     case R_DREGBASE:
 409     case R_CODEBASE:
 410         r = s->regs[addr];
 411         break;
 412     case GPR_BEGIN ... GPR_END:
 413         r = s->gp_regs[addr - GPR_BEGIN];
 414         break;
 415     case MICROCODE_BEGIN ...  MICROCODE_END:
 416         r = s->microcode[get_microcode_address(s, addr)];
 417         break;
 418
 419     default:
 420         error_report("milkymist_pfpu: read access to unknown register 0x"
 421                 TARGET_FMT_plx, addr << 2);
 422         break;
 423     }
 424
 425     trace_milkymist_pfpu_memory_read(addr << 2, r);
 426
 427     return r;
 428 }
 429
 430 static void pfpu_write(void *opaque, hwaddr addr, uint64_t value,
 431                        unsigned size)
 432 {
 433     MilkymistPFPUState *s = opaque;
 434
 435     trace_milkymist_pfpu_memory_write(addr, value);
 436
 437     addr >>= 2;
 438     switch (addr) {
 439     case R_CTL:
 440         if (value & CTL_START_BUSY) {
 441             pfpu_start(s);
 442         }
 443         break;
 444     case R_MESHBASE:
 445     case R_HMESHLAST:
 446     case R_VMESHLAST:
 447     case R_CODEPAGE:
 448     case R_VERTICES:
 449     case R_COLLISIONS:
 450     case R_STRAYWRITES:
 451     case R_LASTDMA:
 452     case R_PC:
 453     case R_DREGBASE:
 454     case R_CODEBASE:
 455         s->regs[addr] = value;
 456         break;
 457     case GPR_BEGIN ...  GPR_END:
 458         s->gp_regs[addr - GPR_BEGIN] = value;
 459         break;
 460     case MICROCODE_BEGIN ...  MICROCODE_END:
 461         s->microcode[get_microcode_address(s, addr)] = value;
 462         break;
 463
 464     default:
 465         error_report("milkymist_pfpu: write access to unknown register 0x"
 466                 TARGET_FMT_plx, addr << 2);
 467         break;
 468     }
 469 }
 470
 471 static const MemoryRegionOps pfpu_mmio_ops = {
 472     .read = pfpu_read,
 473     .write = pfpu_write,
 474     .valid = {
 475         .min_access_size = 4,
 476         .max_access_size = 4,
 477     },
 478     .endianness = DEVICE_NATIVE_ENDIAN,
 479 };
 480
 481 static void milkymist_pfpu_reset(DeviceState *d)
 482 {
 483     MilkymistPFPUState *s = MILKYMIST_PFPU(d);
 484     int i;
 485
 486     for (i = 0; i < R_MAX; i++) {
 487         s->regs[i] = 0;
 488     }
 489     for (i = 0; i < 128; i++) {
 490         s->gp_regs[i] = 0;
 491     }
 492     for (i = 0; i < MICROCODE_WORDS; i++) {
 493         s->microcode[i] = 0;
 494     }
 495     s->output_queue_pos = 0;
 496     for (i = 0; i < MAX_LATENCY; i++) {
 497         s->output_queue[i] = 0;
 498     }
 499 }
 500
 501 static void milkymist_pfpu_realize(DeviceState *dev, Error **errp)
 502 {
 503     MilkymistPFPUState *s = MILKYMIST_PFPU(dev);
 504     SysBusDevice *sbd = SYS_BUS_DEVICE(dev);
 505
 506     sysbus_init_irq(sbd, &s->irq);
 507
 508     memory_region_init_io(&s->regs_region, OBJECT(dev), &pfpu_mmio_ops, s,
 509             "milkymist-pfpu", MICROCODE_END * 4);
 510     sysbus_init_mmio(sbd, &s->regs_region);
 511 }
 512
 513 static const VMStateDescription vmstate_milkymist_pfpu = {
 514     .name = "milkymist-pfpu",
 515     .version_id = 1,
 516     .minimum_version_id = 1,
 517     .fields = (VMStateField[]) {
 518         VMSTATE_UINT32_ARRAY(regs, MilkymistPFPUState, R_MAX),
 519         VMSTATE_UINT32_ARRAY(gp_regs, MilkymistPFPUState, 128),
 520         VMSTATE_UINT32_ARRAY(microcode, MilkymistPFPUState, MICROCODE_WORDS),
 521         VMSTATE_INT32(output_queue_pos, MilkymistPFPUState),
 522         VMSTATE_UINT32_ARRAY(output_queue, MilkymistPFPUState, MAX_LATENCY),
 523         VMSTATE_END_OF_LIST()
 524     }
 525 };
 526
 527 static void milkymist_pfpu_class_init(ObjectClass *klass, void *data)
 528 {
 529     DeviceClass *dc = DEVICE_CLASS(klass);
 530
 531     dc->realize = milkymist_pfpu_realize;
 532     dc->reset = milkymist_pfpu_reset;
 533     dc->vmsd = &vmstate_milkymist_pfpu;
 534 }
 535
 536 static const TypeInfo milkymist_pfpu_info = {
 537     .name          = TYPE_MILKYMIST_PFPU,
 538     .parent        = TYPE_SYS_BUS_DEVICE,
 539     .instance_size = sizeof(MilkymistPFPUState),
 540     .class_init    = milkymist_pfpu_class_init,
 541 };
 542
 543 static void milkymist_pfpu_register_types(void)
 544 {
 545     type_register_static(&milkymist_pfpu_info);
 546 }
 547
 548 type_init(milkymist_pfpu_register_types)