# gguf_out_dtype_gate.py -- the epilogue's bf16 store, witnessed.
#
# card      RTX 5090 (sm120, 170 SMs), box 10.1.103.200, sole tenant.
#           CORRECTNESS ONLY. The fleet standard is the 4090 and this card is
#           not it, so no timing is taken here and none may be read off this
#           run; the deviations and counts below are card-independent.
# build     I8_ONLY_BITS=4 I8_STREAM_K=1, image test-ef6b061, both trees.
#           I8_ONLY_CB is NOT set: narrowing to the GGUF codebook (22) fails to
#           compile on the pre-edit tree too -- duplicate `case` labels at 4720
#           -- so it is a pre-existing limit of that switch, not of this change.
# baseline  the same build of the commit before the edit, loaded beside the new
#           extension in one process, so every fp16 comparison is two builds on
#           one input rather than two runs.
# command   gguf_out_dtype_gate.py --baseline-so <pre-edit .so> \
#             --k 5120 6144 17408 --rows 1 2 6 8 64 --n 512
#
[card] NVIDIA GeForce RTX 5090  sm_count 170
[note] CORRECTNESS ONLY. No timing here; this card is not the fleet's 4090.
[type] Q4_K cb=22 N=512 baseline=yes
[shapes] [177, 170, 178, 169, 179, 173, 175, 168, 167]  tiles {177: (16, 64, 128), 170: (16, 128, 128), 178: (32, 64, 128), 169: (32, 128, 128), 179: (48, 64, 128), 173: (48, 128, 128), 175: (48, 128, 128), 168: (64, 128, 128), 167: (256, 128, 128)}

=== THE COMPILED ARMS, BEFORE AND AFTER ===
  PASS s177 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [128, 85, 12800, 12800, 5, 0]  after [128, 85, 12800, 12800, 5, 0]
  PASS s177 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [128, 107, 12800, 12800, 4, 0]  after [128, 107, 12800, 12800, 4, 0]
  PASS s170 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 85, 21504, 21504, 2, 0]  after [256, 85, 21504, 21504, 2, 0]
  PASS s170 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 107, 21504, 21504, 2, 0]  after [256, 107, 21504, 21504, 2, 0]
  PASS s178 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [128, 121, 16896, 16896, 4, 0]  after [128, 121, 16896, 16896, 4, 0]
  PASS s178 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [128, 143, 16896, 16896, 3, 0]  after [128, 143, 16896, 16896, 3, 0]
  PASS s169 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 121, 25600, 25600, 2, 0]  after [256, 121, 25600, 25600, 2, 0]
  PASS s169 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 145, 25600, 25600, 1, 0]  after [256, 145, 25600, 25600, 1, 0]
  PASS s179 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [128, 152, 20992, 20992, 3, 0]  after [128, 152, 20992, 20992, 3, 0]
  PASS s179 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [128, 171, 20992, 20992, 2, 0]  after [128, 171, 20992, 20992, 2, 0]
  PASS s173 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 152, 29696, 29696, 1, 0]  after [256, 152, 29696, 29696, 1, 0]
  PASS s173 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 173, 29696, 29696, 1, 0]  after [256, 173, 29696, 29696, 1, 0]
  PASS s175 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 128, 29696, 29696, 2, 0]  after [256, 128, 29696, 29696, 2, 0]
  PASS s175 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 128, 29696, 29696, 2, 0]  after [256, 128, 29696, 29696, 2, 0]
  PASS s168 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 184, 33792, 33792, 1, 0]  after [256, 184, 33792, 33792, 1, 0]
  PASS s168 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 207, 33792, 33792, 1, 0]  after [256, 207, 33792, 33792, 1, 0]
  PASS s167 dp regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 255, 82944, 82944, 1, 72]  after [256, 255, 82944, 82944, 1, 72]
  PASS s167 sk regs/smem/blocks unchanged -- threads/regs/smem/mainloop/blocks/local  before [256, 255, 82944, 82944, 1, 0]  after [256, 255, 82944, 82944, 1, 0]

=== K=5120 live 32-groups [7, 13, 100, 133] ===
  PASS M=1   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s177 dp bf16 store == bf16(v) for the same v
  INFO M=1   s177 dp double-rounded 31/512 (6.05%)
  PASS M=1   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s177 sk bf16 store == bf16(v) for the same v
  INFO M=1   s177 sk double-rounded 31/512 (6.05%)
  PASS M=1   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s170 dp bf16 store == bf16(v) for the same v
  INFO M=1   s170 dp double-rounded 31/512 (6.05%)
  PASS M=1   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s178 dp bf16 store == bf16(v) for the same v
  INFO M=1   s178 dp double-rounded 31/512 (6.05%)
  PASS M=1   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s178 sk bf16 store == bf16(v) for the same v
  INFO M=1   s178 sk double-rounded 31/512 (6.05%)
  PASS M=1   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s169 dp bf16 store == bf16(v) for the same v
  INFO M=1   s169 dp double-rounded 31/512 (6.05%)
  PASS M=1   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s179 dp bf16 store == bf16(v) for the same v
  INFO M=1   s179 dp double-rounded 31/512 (6.05%)
  PASS M=1   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s179 sk bf16 store == bf16(v) for the same v
  INFO M=1   s179 sk double-rounded 31/512 (6.05%)
  PASS M=1   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s173 dp bf16 store == bf16(v) for the same v
  INFO M=1   s173 dp double-rounded 31/512 (6.05%)
  PASS M=1   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s175 dp bf16 store == bf16(v) for the same v
  INFO M=1   s175 dp double-rounded 31/512 (6.05%)
  PASS M=1   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s168 dp bf16 store == bf16(v) for the same v
  INFO M=1   s168 dp double-rounded 31/512 (6.05%)
  PASS M=1   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s167 dp bf16 store == bf16(v) for the same v
  INFO M=1   s167 dp double-rounded 31/512 (6.05%)
  PASS M=2   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s177 dp bf16 store == bf16(v) for the same v
  INFO M=2   s177 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s177 sk bf16 store == bf16(v) for the same v
  INFO M=2   s177 sk double-rounded 49/1024 (4.79%)
  PASS M=2   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s170 dp bf16 store == bf16(v) for the same v
  INFO M=2   s170 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s178 dp bf16 store == bf16(v) for the same v
  INFO M=2   s178 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s178 sk bf16 store == bf16(v) for the same v
  INFO M=2   s178 sk double-rounded 49/1024 (4.79%)
  PASS M=2   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s169 dp bf16 store == bf16(v) for the same v
  INFO M=2   s169 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s179 dp bf16 store == bf16(v) for the same v
  INFO M=2   s179 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s179 sk bf16 store == bf16(v) for the same v
  INFO M=2   s179 sk double-rounded 49/1024 (4.79%)
  PASS M=2   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s173 dp bf16 store == bf16(v) for the same v
  INFO M=2   s173 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s175 dp bf16 store == bf16(v) for the same v
  INFO M=2   s175 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s168 dp bf16 store == bf16(v) for the same v
  INFO M=2   s168 dp double-rounded 49/1024 (4.79%)
  PASS M=2   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s167 dp bf16 store == bf16(v) for the same v
  INFO M=2   s167 dp double-rounded 49/1024 (4.79%)
  PASS M=6   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s177 dp bf16 store == bf16(v) for the same v
  INFO M=6   s177 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s177 sk bf16 store == bf16(v) for the same v
  INFO M=6   s177 sk double-rounded 191/3072 (6.22%)
  PASS M=6   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s170 dp bf16 store == bf16(v) for the same v
  INFO M=6   s170 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s178 dp bf16 store == bf16(v) for the same v
  INFO M=6   s178 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s178 sk bf16 store == bf16(v) for the same v
  INFO M=6   s178 sk double-rounded 191/3072 (6.22%)
  PASS M=6   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s169 dp bf16 store == bf16(v) for the same v
  INFO M=6   s169 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s179 dp bf16 store == bf16(v) for the same v
  INFO M=6   s179 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s179 sk bf16 store == bf16(v) for the same v
  INFO M=6   s179 sk double-rounded 191/3072 (6.22%)
  PASS M=6   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s173 dp bf16 store == bf16(v) for the same v
  INFO M=6   s173 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s175 dp bf16 store == bf16(v) for the same v
  INFO M=6   s175 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s168 dp bf16 store == bf16(v) for the same v
  INFO M=6   s168 dp double-rounded 191/3072 (6.22%)
  PASS M=6   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s167 dp bf16 store == bf16(v) for the same v
  INFO M=6   s167 dp double-rounded 191/3072 (6.22%)
  PASS M=8   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s177 dp bf16 store == bf16(v) for the same v
  INFO M=8   s177 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s177 sk bf16 store == bf16(v) for the same v
  INFO M=8   s177 sk double-rounded 245/4096 (5.98%)
  PASS M=8   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s170 dp bf16 store == bf16(v) for the same v
  INFO M=8   s170 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s178 dp bf16 store == bf16(v) for the same v
  INFO M=8   s178 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s178 sk bf16 store == bf16(v) for the same v
  INFO M=8   s178 sk double-rounded 245/4096 (5.98%)
  PASS M=8   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s169 dp bf16 store == bf16(v) for the same v
  INFO M=8   s169 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s179 dp bf16 store == bf16(v) for the same v
  INFO M=8   s179 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s179 sk bf16 store == bf16(v) for the same v
  INFO M=8   s179 sk double-rounded 245/4096 (5.98%)
  PASS M=8   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s173 dp bf16 store == bf16(v) for the same v
  INFO M=8   s173 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s175 dp bf16 store == bf16(v) for the same v
  INFO M=8   s175 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s168 dp bf16 store == bf16(v) for the same v
  INFO M=8   s168 dp double-rounded 245/4096 (5.98%)
  PASS M=8   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s167 dp bf16 store == bf16(v) for the same v
  INFO M=8   s167 dp double-rounded 245/4096 (5.98%)
  PASS M=64  s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s177 dp bf16 store == bf16(v) for the same v
  INFO M=64  s177 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s177 sk bf16 store == bf16(v) for the same v
  INFO M=64  s177 sk double-rounded 2054/32768 (6.27%)
  PASS M=64  s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s170 dp bf16 store == bf16(v) for the same v
  INFO M=64  s170 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s178 dp bf16 store == bf16(v) for the same v
  INFO M=64  s178 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s178 sk bf16 store == bf16(v) for the same v
  INFO M=64  s178 sk double-rounded 2054/32768 (6.27%)
  PASS M=64  s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s169 dp bf16 store == bf16(v) for the same v
  INFO M=64  s169 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s179 dp bf16 store == bf16(v) for the same v
  INFO M=64  s179 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s179 sk bf16 store == bf16(v) for the same v
  INFO M=64  s179 sk double-rounded 2054/32768 (6.27%)
  PASS M=64  s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s173 dp bf16 store == bf16(v) for the same v
  INFO M=64  s173 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s175 dp bf16 store == bf16(v) for the same v
  INFO M=64  s175 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s168 dp bf16 store == bf16(v) for the same v
  INFO M=64  s168 dp double-rounded 2054/32768 (6.27%)
  PASS M=64  s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s167 dp bf16 store == bf16(v) for the same v
  INFO M=64  s167 dp double-rounded 2054/32768 (6.27%)

=== K=6144 live 32-groups [124, 133, 154, 163] ===
  PASS M=1   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s177 dp bf16 store == bf16(v) for the same v
  INFO M=1   s177 dp double-rounded 20/512 (3.91%)
  PASS M=1   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s177 sk bf16 store == bf16(v) for the same v
  INFO M=1   s177 sk double-rounded 20/512 (3.91%)
  PASS M=1   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s170 dp bf16 store == bf16(v) for the same v
  INFO M=1   s170 dp double-rounded 20/512 (3.91%)
  PASS M=1   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s178 dp bf16 store == bf16(v) for the same v
  INFO M=1   s178 dp double-rounded 20/512 (3.91%)
  PASS M=1   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s178 sk bf16 store == bf16(v) for the same v
  INFO M=1   s178 sk double-rounded 20/512 (3.91%)
  PASS M=1   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s169 dp bf16 store == bf16(v) for the same v
  INFO M=1   s169 dp double-rounded 20/512 (3.91%)
  PASS M=1   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s179 dp bf16 store == bf16(v) for the same v
  INFO M=1   s179 dp double-rounded 20/512 (3.91%)
  PASS M=1   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s179 sk bf16 store == bf16(v) for the same v
  INFO M=1   s179 sk double-rounded 20/512 (3.91%)
  PASS M=1   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s173 dp bf16 store == bf16(v) for the same v
  INFO M=1   s173 dp double-rounded 20/512 (3.91%)
  PASS M=1   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s175 dp bf16 store == bf16(v) for the same v
  INFO M=1   s175 dp double-rounded 20/512 (3.91%)
  PASS M=1   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s168 dp bf16 store == bf16(v) for the same v
  INFO M=1   s168 dp double-rounded 20/512 (3.91%)
  PASS M=1   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s167 dp bf16 store == bf16(v) for the same v
  INFO M=1   s167 dp double-rounded 20/512 (3.91%)
  PASS M=2   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s177 dp bf16 store == bf16(v) for the same v
  INFO M=2   s177 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s177 sk bf16 store == bf16(v) for the same v
  INFO M=2   s177 sk double-rounded 60/1024 (5.86%)
  PASS M=2   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s170 dp bf16 store == bf16(v) for the same v
  INFO M=2   s170 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s178 dp bf16 store == bf16(v) for the same v
  INFO M=2   s178 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s178 sk bf16 store == bf16(v) for the same v
  INFO M=2   s178 sk double-rounded 60/1024 (5.86%)
  PASS M=2   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s169 dp bf16 store == bf16(v) for the same v
  INFO M=2   s169 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s179 dp bf16 store == bf16(v) for the same v
  INFO M=2   s179 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s179 sk bf16 store == bf16(v) for the same v
  INFO M=2   s179 sk double-rounded 60/1024 (5.86%)
  PASS M=2   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s173 dp bf16 store == bf16(v) for the same v
  INFO M=2   s173 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s175 dp bf16 store == bf16(v) for the same v
  INFO M=2   s175 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s168 dp bf16 store == bf16(v) for the same v
  INFO M=2   s168 dp double-rounded 60/1024 (5.86%)
  PASS M=2   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s167 dp bf16 store == bf16(v) for the same v
  INFO M=2   s167 dp double-rounded 60/1024 (5.86%)
  PASS M=6   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s177 dp bf16 store == bf16(v) for the same v
  INFO M=6   s177 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s177 sk bf16 store == bf16(v) for the same v
  INFO M=6   s177 sk double-rounded 187/3072 (6.09%)
  PASS M=6   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s170 dp bf16 store == bf16(v) for the same v
  INFO M=6   s170 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s178 dp bf16 store == bf16(v) for the same v
  INFO M=6   s178 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s178 sk bf16 store == bf16(v) for the same v
  INFO M=6   s178 sk double-rounded 187/3072 (6.09%)
  PASS M=6   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s169 dp bf16 store == bf16(v) for the same v
  INFO M=6   s169 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s179 dp bf16 store == bf16(v) for the same v
  INFO M=6   s179 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s179 sk bf16 store == bf16(v) for the same v
  INFO M=6   s179 sk double-rounded 187/3072 (6.09%)
  PASS M=6   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s173 dp bf16 store == bf16(v) for the same v
  INFO M=6   s173 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s175 dp bf16 store == bf16(v) for the same v
  INFO M=6   s175 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s168 dp bf16 store == bf16(v) for the same v
  INFO M=6   s168 dp double-rounded 187/3072 (6.09%)
  PASS M=6   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s167 dp bf16 store == bf16(v) for the same v
  INFO M=6   s167 dp double-rounded 187/3072 (6.09%)
  PASS M=8   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s177 dp bf16 store == bf16(v) for the same v
  INFO M=8   s177 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s177 sk bf16 store == bf16(v) for the same v
  INFO M=8   s177 sk double-rounded 247/4096 (6.03%)
  PASS M=8   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s170 dp bf16 store == bf16(v) for the same v
  INFO M=8   s170 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s178 dp bf16 store == bf16(v) for the same v
  INFO M=8   s178 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s178 sk bf16 store == bf16(v) for the same v
  INFO M=8   s178 sk double-rounded 247/4096 (6.03%)
  PASS M=8   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s169 dp bf16 store == bf16(v) for the same v
  INFO M=8   s169 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s179 dp bf16 store == bf16(v) for the same v
  INFO M=8   s179 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s179 sk bf16 store == bf16(v) for the same v
  INFO M=8   s179 sk double-rounded 247/4096 (6.03%)
  PASS M=8   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s173 dp bf16 store == bf16(v) for the same v
  INFO M=8   s173 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s175 dp bf16 store == bf16(v) for the same v
  INFO M=8   s175 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s168 dp bf16 store == bf16(v) for the same v
  INFO M=8   s168 dp double-rounded 247/4096 (6.03%)
  PASS M=8   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s167 dp bf16 store == bf16(v) for the same v
  INFO M=8   s167 dp double-rounded 247/4096 (6.03%)
  PASS M=64  s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s177 dp bf16 store == bf16(v) for the same v
  INFO M=64  s177 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s177 sk bf16 store == bf16(v) for the same v
  INFO M=64  s177 sk double-rounded 1924/32768 (5.87%)
  PASS M=64  s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s170 dp bf16 store == bf16(v) for the same v
  INFO M=64  s170 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s178 dp bf16 store == bf16(v) for the same v
  INFO M=64  s178 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s178 sk bf16 store == bf16(v) for the same v
  INFO M=64  s178 sk double-rounded 1924/32768 (5.87%)
  PASS M=64  s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s169 dp bf16 store == bf16(v) for the same v
  INFO M=64  s169 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s179 dp bf16 store == bf16(v) for the same v
  INFO M=64  s179 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s179 sk bf16 store == bf16(v) for the same v
  INFO M=64  s179 sk double-rounded 1924/32768 (5.87%)
  PASS M=64  s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s173 dp bf16 store == bf16(v) for the same v
  INFO M=64  s173 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s175 dp bf16 store == bf16(v) for the same v
  INFO M=64  s175 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s168 dp bf16 store == bf16(v) for the same v
  INFO M=64  s168 dp double-rounded 1924/32768 (5.87%)
  PASS M=64  s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s167 dp bf16 store == bf16(v) for the same v
  INFO M=64  s167 dp double-rounded 1924/32768 (5.87%)

=== K=17408 live 32-groups [79, 194, 246, 492] ===
  PASS M=1   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s177 dp bf16 store == bf16(v) for the same v
  INFO M=1   s177 dp double-rounded 29/512 (5.66%)
  PASS M=1   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s177 sk bf16 store == bf16(v) for the same v
  INFO M=1   s177 sk double-rounded 29/512 (5.66%)
  PASS M=1   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s170 dp bf16 store == bf16(v) for the same v
  INFO M=1   s170 dp double-rounded 29/512 (5.66%)
  PASS M=1   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s178 dp bf16 store == bf16(v) for the same v
  INFO M=1   s178 dp double-rounded 29/512 (5.66%)
  PASS M=1   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s178 sk bf16 store == bf16(v) for the same v
  INFO M=1   s178 sk double-rounded 29/512 (5.66%)
  PASS M=1   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s169 dp bf16 store == bf16(v) for the same v
  INFO M=1   s169 dp double-rounded 29/512 (5.66%)
  PASS M=1   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s179 dp bf16 store == bf16(v) for the same v
  INFO M=1   s179 dp double-rounded 29/512 (5.66%)
  PASS M=1   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=1   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s179 sk bf16 store == bf16(v) for the same v
  INFO M=1   s179 sk double-rounded 29/512 (5.66%)
  PASS M=1   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s173 dp bf16 store == bf16(v) for the same v
  INFO M=1   s173 dp double-rounded 29/512 (5.66%)
  PASS M=1   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s175 dp bf16 store == bf16(v) for the same v
  INFO M=1   s175 dp double-rounded 29/512 (5.66%)
  PASS M=1   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s168 dp bf16 store == bf16(v) for the same v
  INFO M=1   s168 dp double-rounded 29/512 (5.66%)
  PASS M=1   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=1   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=1   s167 dp bf16 store == bf16(v) for the same v
  INFO M=1   s167 dp double-rounded 29/512 (5.66%)
  PASS M=2   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s177 dp bf16 store == bf16(v) for the same v
  INFO M=2   s177 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s177 sk bf16 store == bf16(v) for the same v
  INFO M=2   s177 sk double-rounded 61/1024 (5.96%)
  PASS M=2   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s170 dp bf16 store == bf16(v) for the same v
  INFO M=2   s170 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s178 dp bf16 store == bf16(v) for the same v
  INFO M=2   s178 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s178 sk bf16 store == bf16(v) for the same v
  INFO M=2   s178 sk double-rounded 61/1024 (5.96%)
  PASS M=2   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s169 dp bf16 store == bf16(v) for the same v
  INFO M=2   s169 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s179 dp bf16 store == bf16(v) for the same v
  INFO M=2   s179 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=2   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s179 sk bf16 store == bf16(v) for the same v
  INFO M=2   s179 sk double-rounded 61/1024 (5.96%)
  PASS M=2   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s173 dp bf16 store == bf16(v) for the same v
  INFO M=2   s173 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s175 dp bf16 store == bf16(v) for the same v
  INFO M=2   s175 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s168 dp bf16 store == bf16(v) for the same v
  INFO M=2   s168 dp double-rounded 61/1024 (5.96%)
  PASS M=2   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=2   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=2   s167 dp bf16 store == bf16(v) for the same v
  INFO M=2   s167 dp double-rounded 61/1024 (5.96%)
  PASS M=6   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s177 dp bf16 store == bf16(v) for the same v
  INFO M=6   s177 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s177 sk bf16 store == bf16(v) for the same v
  INFO M=6   s177 sk double-rounded 198/3072 (6.45%)
  PASS M=6   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s170 dp bf16 store == bf16(v) for the same v
  INFO M=6   s170 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s178 dp bf16 store == bf16(v) for the same v
  INFO M=6   s178 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s178 sk bf16 store == bf16(v) for the same v
  INFO M=6   s178 sk double-rounded 198/3072 (6.45%)
  PASS M=6   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s169 dp bf16 store == bf16(v) for the same v
  INFO M=6   s169 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s179 dp bf16 store == bf16(v) for the same v
  INFO M=6   s179 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=6   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s179 sk bf16 store == bf16(v) for the same v
  INFO M=6   s179 sk double-rounded 198/3072 (6.45%)
  PASS M=6   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s173 dp bf16 store == bf16(v) for the same v
  INFO M=6   s173 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s175 dp bf16 store == bf16(v) for the same v
  INFO M=6   s175 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s168 dp bf16 store == bf16(v) for the same v
  INFO M=6   s168 dp double-rounded 198/3072 (6.45%)
  PASS M=6   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=6   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=6   s167 dp bf16 store == bf16(v) for the same v
  INFO M=6   s167 dp double-rounded 198/3072 (6.45%)
  PASS M=8   s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s177 dp bf16 store == bf16(v) for the same v
  INFO M=8   s177 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s177 sk bf16 store == bf16(v) for the same v
  INFO M=8   s177 sk double-rounded 261/4096 (6.37%)
  PASS M=8   s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s170 dp bf16 store == bf16(v) for the same v
  INFO M=8   s170 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s178 dp bf16 store == bf16(v) for the same v
  INFO M=8   s178 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s178 sk bf16 store == bf16(v) for the same v
  INFO M=8   s178 sk double-rounded 261/4096 (6.37%)
  PASS M=8   s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s169 dp bf16 store == bf16(v) for the same v
  INFO M=8   s169 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s179 dp bf16 store == bf16(v) for the same v
  INFO M=8   s179 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=8   s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s179 sk bf16 store == bf16(v) for the same v
  INFO M=8   s179 sk double-rounded 261/4096 (6.37%)
  PASS M=8   s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s173 dp bf16 store == bf16(v) for the same v
  INFO M=8   s173 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s175 dp bf16 store == bf16(v) for the same v
  INFO M=8   s175 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s168 dp bf16 store == bf16(v) for the same v
  INFO M=8   s168 dp double-rounded 261/4096 (6.37%)
  PASS M=8   s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=8   s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=8   s167 dp bf16 store == bf16(v) for the same v
  INFO M=8   s167 dp double-rounded 261/4096 (6.37%)
  PASS M=64  s177 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s177 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s177 dp bf16 store == bf16(v) for the same v
  INFO M=64  s177 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s177 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s177 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s177 sk bf16 store == bf16(v) for the same v
  INFO M=64  s177 sk double-rounded 2058/32768 (6.28%)
  PASS M=64  s170 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s170 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s170 dp bf16 store == bf16(v) for the same v
  INFO M=64  s170 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s178 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s178 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s178 dp bf16 store == bf16(v) for the same v
  INFO M=64  s178 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s178 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s178 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s178 sk bf16 store == bf16(v) for the same v
  INFO M=64  s178 sk double-rounded 2058/32768 (6.28%)
  PASS M=64  s169 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s169 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s169 dp bf16 store == bf16(v) for the same v
  INFO M=64  s169 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s179 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s179 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s179 dp bf16 store == bf16(v) for the same v
  INFO M=64  s179 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s179 sk fp16 store byte-identical to the pre-edit build
  PASS M=64  s179 sk fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s179 sk bf16 store == bf16(v) for the same v
  INFO M=64  s179 sk double-rounded 2058/32768 (6.28%)
  PASS M=64  s173 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s173 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s173 dp bf16 store == bf16(v) for the same v
  INFO M=64  s173 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s175 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s175 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s175 dp bf16 store == bf16(v) for the same v
  INFO M=64  s175 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s168 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s168 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s168 dp bf16 store == bf16(v) for the same v
  INFO M=64  s168 dp double-rounded 2058/32768 (6.28%)
  PASS M=64  s167 dp fp16 store byte-identical to the pre-edit build
  PASS M=64  s167 dp fp16 store == fp16(v) for the exactly-known v
  PASS M=64  s167 dp bf16 store == bf16(v) for the same v
  INFO M=64  s167 dp double-rounded 2058/32768 (6.28%)

=== SATURATION: v beyond fp16's 65504, on purpose ===
  w_scale=2^20  max|v| 4.3858e+07  fp16-chain inf 4008   bf16-store inf 0       saturation-rescued 4008
  w_scale=2^24  max|v| 7.0172e+08  fp16-chain inf 4088   bf16-store inf 0       saturation-rescued 4088
  w_scale=2^28  max|v| 1.1228e+10  fp16-chain inf 4095   bf16-store inf 0       saturation-rescued 4095
  PASS the saturation case is EXERCISED, not asserted -- 12191 elements are inf through the fp16 chain and finite through the bf16 store

=== DEVIATION FROM THE fp64 TRUTH ===

  ALL SERVED-SHAPE ELEMENTS (1492992)
    fp16 store + cast  max|abs|  1.4014e-01  mean|abs|  8.8492e-03  max|rel| 4.3753e-03  mean|rel| 1.4327e-03  inf 0
    bf16 store         max|abs|  1.2494e-01  mean|abs|  8.7111e-03  max|rel| 3.8911e-03  mean|rel| 1.4096e-03  inf 0

  THE DOUBLE-ROUNDED SUBSET (91380/1492992, 6.121% -- the elements where the two stores disagree at all)
    fp16 store + cast  max|abs|  1.4014e-01  mean|abs|  1.8747e-02  max|rel| 4.3753e-03  mean|rel| 3.0080e-03  inf 0
    bf16 store         max|abs|  1.2494e-01  mean|abs|  1.6491e-02  max|rel| 3.8758e-03  mean|rel| 2.6320e-03  inf 0

  THE SATURATION ROWS, DRIVEN (12191 elements rescued of 12288)
    fp16 store + cast  max|abs|         inf  mean|abs|         inf  max|rel|        inf  mean|rel|        inf  inf 12191
    bf16 store         max|abs|  2.0341e+07  mean|abs|  5.5683e+05  max|rel| 3.8911e-03  mean|rel| 1.4315e-03  inf 0
  PASS no element is FURTHER from the truth through the bf16 store -- nearer on 103571 elements, equal on the rest

GATE PASS
