Hugo MeilandandCursor
560caa56e6
TARGET=CORTEXA72 previously reused the A57 8x4 path. Add a dedicated 6x8 NEON ukernel, contiguous MR=6 panel packers (stock gemm_*copy_6 is 4+2), and UNROLL_M=6-aware TRSM kernels so HPL/dtrsm does not corrupt the heap. DTRMM falls back to generic 2x2 until a matching kernel exists. Co-authored-by: Cursor <cursoragent@cursor.com>
29 lines
1.2 KiB
Plaintext
29 lines
1.2 KiB
Plaintext
include $(KERNELDIR)/KERNEL.CORTEXA57
|
|
|
|
# Cortex-A72 DGEMM uses a dedicated 6x8 microkernel (see
|
|
# dgemm_kernel_6x8_cortexa72.S). Stock TARGET=CORTEXA72 previously
|
|
# aliased CORTEXA57 (8x4). Keep SGEMM/CGEMM/ZGEMM from A57 for now.
|
|
#
|
|
# Stock generic/gemm_{t,n}copy_6.c packs as 4+2, not contiguous MR=6 —
|
|
# use the panel packers. NR=8 reuses the asm dgemm_{n,t}copy_8.S.
|
|
# No dtrmm_kernel_6x8.S yet; fall back to generic 2x2 for DTRMM.
|
|
# Stock generic/trsm_kernel_*.c assumes power-of-2 UNROLL_M; use the
|
|
# UNROLLM6-aware kernels (same idea as loongarch64 UNROLLN6).
|
|
DGEMMKERNEL = dgemm_kernel_6x8_cortexa72.S
|
|
DTRMMKERNEL = ../generic/trmmkernel_2x2.c
|
|
|
|
DTRSMKERNEL_LN = trsm_kernel_LN_UNROLLM6.c
|
|
DTRSMKERNEL_LT = trsm_kernel_LT_UNROLLM6.c
|
|
DTRSMKERNEL_RN = trsm_kernel_RN_UNROLLM6.c
|
|
DTRSMKERNEL_RT = trsm_kernel_RT_UNROLLM6.c
|
|
|
|
DGEMMINCOPY = gemm_ncopy_6_panel.c
|
|
DGEMMITCOPY = gemm_tcopy_6_panel.c
|
|
DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
|
|
DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
|
|
|
|
DGEMMONCOPY = dgemm_ncopy_8.S
|
|
DGEMMOTCOPY = dgemm_tcopy_8.S
|
|
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
|
|
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
|