ayo
22026ede12
Preset K, byte-identical to the previous output in all captured buffers and the harness image, measured on an RX 7700 XT: - LDS row padding so WMMA operand loads hit distinct banks - WGP mode for enc1/enc2/dec1/dec2, ds_bpermute half-wave exchange for enc0/dec0/dec3, more waves per window in enc4/dec4/dec5 - cheaper operand plumbing in common/wmma_layout.h (packed f16 conversions, no dynamic vector indexing), raw buffer weight loads, accumulator-init vectors prepared once - enc1 as a single-buffer recompute layer, dec1 Q/K/V chains split - dec5 as four dependent phase kernels
9 lines
477 B
Plaintext
9 lines
477 B
Plaintext
// WGP mode: with the padded LDS rows (PWIN_PAD) this layer is 2-5% faster in WGP mode than in CU mode
|
|
// (CU mode was 1-6% faster before the padding)
|
|
// accumulator init vectors from the weights (the prepared f32 table is 2-4% slower here)
|
|
// Q, K, V chains two at a time (fewer live accumulators, spills 164 -> 68 B; -2 %)
|
|
// d4r-build-flags: -DPWIN_VEC_FROM_WEIGHTS -DPWIN_QKV_SPLIT
|
|
#define PWIN_F32ACC
|
|
#include "pwin_layer.h"
|
|
PWIN_DECODER(dltss_pwin_dec1_layer, 2, 64, 64)
|