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
8 lines
402 B
Plaintext
8 lines
402 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)
|
|
// weight images through plain global loads here (buffer loads were 1-3% slower in this layer)
|
|
// d4r-build-flags: -DPWIN_NO_BUFFER_IMG -DPWIN_VEC_FROM_WEIGHTS
|
|
#define PWIN_F32ACC
|
|
#include "pwin_layer.h"
|
|
PWIN_ENCODER(dltss_pwin_enc2_layer, 2, 64, 96)
|