Skip to content

i[gfx1250] Switch BF16 GEMM LDS loads to ds_read_b256, +0.4 TB/s #1077

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
amd-hhashemi wants to merge 1 commit into
base: main
Choose a base branch
from
Open

i[gfx1250] Switch BF16 GEMM LDS loads to ds_read_b256, +0.4 TB/s #1077

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 8 additions & 9 deletions kernels/gemm/gemm_bf16_gfx1250.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,7 +64,7 @@ def launch_gemm_bf16(
if K_WS < 2:
raise ValueError(f"tile_k={tile_k} needs at least 2 WMMA K-steps to hide the LDS reads")

LDS_PAD = 16 # keeps consecutive rows 4 dwords apart -> conflict-free b128 reads
LDS_PAD = 32 # 32-byte pad keeps rows 32-byte aligned for ds_read_b256
LDS_ROW = KB + LDS_PAD
STAGE_A = ((tile_m * LDS_ROW + 15) // 16) * 16
STAGE_B = ((tile_n * LDS_ROW + 15) // 16) * 16
Expand Down Expand Up @@ -157,21 +157,20 @@ def issue(s, kt):

wmb = wave_m * warp_tile_m
wnb = wave_n * warp_tile_n
lds_load_b128, _ = make_lds_copy_ops(128)
lds_load_b256, _ = make_lds_copy_ops(256)

def _frag(buf, b0):
"""One 16-element operand fragment: two 16-byte chunks 32 bytes apart."""
v0 = Vec(lds_load_b128(buf, b0))
v1 = Vec(lds_load_b128(buf, b0 + 32))
return v0.shuffle(v1, list(range(8)))
"""One 8-int32 fragment via single ds_read_b256."""
return Vec(lds_load_b256(buf, b0))

def load_a(buf, wm, ks):
row = wmb + wm * WMMA_M + lane16
return _frag(buf, fx.Int64(row * LDS_ROW + ks * KSB + kgrp * 16))
# kgrp*32: kgrp=0→bytes[0-31], kgrp=1→bytes[32-63] of K-step row
return _frag(buf, fx.Int64(row * LDS_ROW + ks * KSB + kgrp * 32))

def load_b(buf, wn, ks):
col = wnb + wn * WMMA_N + lane16
return _frag(buf, fx.Int64(STAGE_A + col * LDS_ROW + ks * KSB + kgrp * 16))
return _frag(buf, fx.Int64(STAGE_A + col * LDS_ROW + ks * KSB + kgrp * 32))

wmma_atom = fx.make_mma_atom(fx.rocdl.WMMA(WMMA_M, WMMA_N, WMMA_K, elem_cls, fx.Float32))
c_frags = [fx.make_rmem_tensor(8, fx.Float32) for _ in range_constexpr(n_acc)]
Expand All @@ -183,7 +182,7 @@ def _rmem(n, v):
t.store(v)
return t

DS_A = DS_B = 2 # ds_read_b128 per fragment
DS_A = DS_B = 1 # ds_read_b256 per fragment (halved from 2×b128)
KS_DS = wmma_m_rep * DS_A + wmma_n_rep * DS_B # ds_reads for one WMMA K-step

def _load_ks(buf, ks):
Expand Down
Loading