code wiki / _hdl_build / _gpu_bm_qmd_v5_gate.nx
_gpu_bm_qmd_v5_gate.nx
buildroot/runtime/_hdl_build/_gpu_bm_qmd_v5_gate.nx
about
_gpu_bm_qmd_v5_gate.nx -- BARE-METAL SOVEREIGN-GPU rung BM-GPU-7a: REAL Blackwell QMDV05_00 descriptor.
Hardens BM-GPU-C2 (_gpu_bm_qmd_gate): that rung's LAUNCH METHODS were faithful (SET_OBJECT 0xCDC0 +
SEND_PCAS_A 0x2b4 + SEND_SIGNALING_PCAS2_B 0x2c0 + PCAS_ACTION=SCHEDULE) but its QMD-INTERNAL field
offsets were REPRESENTATIVE ("QMDV0x bit-packing wasn't locatable"). THIS rung writes the REAL
QMDV05_00 bit-packed layout, recovered (deep-research, 3-0 verified) verbatim from NVIDIA's open
header classes/compute/clcdc0qmd.h and independently mirrored in tinygrad's NVCEC0_QMDV05_00. So the
GPU LAUNCH DESCRIPTOR is now byte/bit-FAITHFUL to NVIDIA's published Blackwell spec, not a stand-in.
QMDV05_00 fields are MW(hi:lo) BIT ranges over a 3072-bit (384-byte) struct (NOT plain byte fields):
PROGRAM_ADDRESS_LOWER_SHIFTED4 MW(1055:1024), UPPER_SHIFTED4 MW(1076:1056) [stored = addr>>4]
CTA_THREAD_DIMENSION0/1/2 MW(1103:1088)/(1119:1104)/(1127:1120) [block dims]
CTA_RASTER_WIDTH/HEIGHT/DEPTH MW(1279:1248)/(1295:1280)/(1327:1312) [grid dims]
REGISTER_COUNT MW(1136:1128); SHARED_MEMORY_SIZE_SHIFTED7 MW(1162:1152)
CONSTANT_BUFFER_ADDR_LOWER_SHIFTED6(i) MW((1375+i*64):(1344+i*64)), UPPER_SHIFTED6(i) MW((1394+i*64):(1376+i*64))
CONSTANT_BUFFER_SIZE_SHIFTED4(i) MW((1407+i*64):(1395+i*64)); VALID(i) MW((1856+i*4):(1856+i*4))
QMD_GROUP_ID MW(149:144)
HONEST SCOPE (no-false-green): the QMD LAYOUT is now REAL (round-trip: every field written at its real
bit range reads back exactly; the SKED extracts program_address + the const-buffer VA from the REAL
offsets and the GEMM runs off that real cbuf pointer). The SASS kernel itself STILL STANDS IN -- real
Blackwell sm_120 SASS is not hand-emittable from open sources yet (encoding/control-bits unsolved; only
Mesa NAK can emit it) = BM-GPU-7b, blocked on the open ecosystem / real silicon. So: descriptor REAL,
machine-code modeled.
GREEN iff: A) all real QMDV05_00 fields round-trip (write->read at real bit offsets == value); B) the
SKED reading program_address + cbuf from the REAL offsets dispatches a GEMM == independent reference;
C) TAMPER: zeroing CONSTANT_BUFFER_VALID(0) OR a wrong bit range -> field/dispatch mismatch (the real
layout is load-bearing). Marker -> knowledge/status/gpu_baremetal.log (BMQMDV5GATE). license_tier: ORIGINAL
dependencies 1 imports · 0 importers
imports: nx_syscalls.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| 32 | const COMPUTE_CLASS: i64 = 0xCDC0 // BLACKWELL_COMPUTE_A (NVCDC0), confirmed |
| 33 | const SET_OBJECT: i64 = 0x0000 |
| 34 | const SEND_PCAS_A: i64 = 0x02b4 |
| 35 | const SEND_SIGNALING_PCAS2_B: i64 = 0x02c0 |
| 36 | const PCAS_ACTION_SCHEDULE: i64 = 0x2 |
| 37 | const DIM: i64 = 4 |
functions
| 39 | func p(s: *u8) -> i64 { var nn: i64=0; while s[nn]!=(0 as u8){nn=nn+1} sys_write(1,s,nn); return 0 } |
| 40 | func fp(fd: i64, s: *u8) -> i64 { var nn: i64=0; while s[nn]!=(0 as u8){nn=nn+1} sys_write(fd,s,nn); return 0 } |
| 41 | func n(v: i64) -> i64 { let bb: *u8=sys_mmap(28); var m: i64=v; if m<0{m=0-m;sys_write(1,"-" as *u8,1)}; let t: *u8=sys_mmap(28); var k: i64=0; if m==0{t[0]=48;k=1}; while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1}; var i: i64=0; while i<k{bb[i]=t[k-1-i];i=i+1}; sys_write(1,bb,k); return 0 } |
| 42 | func x(v: i64) -> i64 { p("0x" as *u8); let bb:*u8=sys_mmap(20); var k:i64=0; var m:i64=v; if m==0{bb[0]=48;k=1}; while m>0{ let d:i64=m&15; if d<10{bb[k]=(48+d) as u8}else{bb[k]=(87+d) as u8}; m=(m>>4); k=k+1 } var i:i64=0; let o:*u8=sys_mmap(20); while i<k{o[i]=bb[k-1-i];i=i+1} sys_write(1,o,k); return 0 } |
| 43 | func fx(fd: i64, v: i64) -> i64 { let bb:*u8=sys_mmap(20); var k:i64=0; var m:i64=v; if m==0{bb[0]=48;k=1}; while m>0{ let d:i64=m&15; if d<10{bb[k]=(48+d) as u8}else{bb[k]=(87+d) as u8}; m=(m>>4); k=k+1 } let o:*u8=sys_mmap(20); var i:i64=0; while i<k{o[i]=bb[k-1-i];i=i+1} fp(fd,"0x" as *u8); sys_write(fd,o,k); return 0 } |
| 44 | func rd32(b: *u8, o: i64) -> i64 { return (b[o] as i64)|((b[o+1] as i64)<<8)|((b[o+2] as i64)<<16)|((b[o+3] as i64)<<24) } |
| 45 | func wr32(b: *u8, o: i64, v: i64) -> i64 { b[o]=(v&0xff) as u8; b[o+1]=((v>>8)&0xff) as u8; b[o+2]=((v>>16)&0xff) as u8; b[o+3]=((v>>24)&0xff) as u8; return 0 } |
| 48 | func qmd_set_bit(q: *u8, b: i64, v: i64) -> i64 called by 1: qmd_set |
| 53 | func qmd_get_bit(q: *u8, b: i64) -> i64 { return ((q[b >> 3] as i64) >> (b & 7)) & 1 } called by 1: qmd_get |
| 55 | func qmd_set(q: *u8, hi: i64, lo: i64, val: i64) -> i64 |
| 60 | func qmd_get(q: *u8, hi: i64, lo: i64) -> i64 |
| 68 | func gpfifo_run_qmd(gmem: *u8, pb_off: i64, pb_dwords: i64, cs: *i64) -> i64 |
| 108 | func main() -> i64 |