diff --git a/clean/dx12/kernels/l_swin.hlsl b/clean/dx12/kernels/l_swin.hlsl index 2a86061e..bd7fb81f 100644 --- a/clean/dx12/kernels/l_swin.hlsl +++ b/clean/dx12/kernels/l_swin.hlsl @@ -7,6 +7,9 @@ // 13333: sw_gelu_fetch8_bn; a further -7 .. -9 / -13 .. -15 us). The groupshared table (SW_GELU 2) and patterns with a // transcendental-chain tile (10333) take them to 108 .. 112 VGPRs, over their bound of 96. l_swin_block_c32_nc keeps // HIP's transcendental chain (SW_GELU 0). +// Scheduling knobs of swin_block.hlsl (the same values, timed per kernel): _enc_lin_nc with the head loop in pairs +// (SW_HEAD_PAIRS), the softmax's staged form (SW_SMAX_STAGED) and norm1's loads before the first barrier (SW_NORM1_EARLY); +// _dec_coef12_nc with the staged softmax (SW_SMAX_ADDST, SW_SMAX_STAGED). Both stay within 96 VGPRs. // l_swin_block_c32_nc FP8 map in, FP8 out, mirror border (#117) // l_swin_block_c32_enc_nc the encoder block: FP16 halo map [H + 4][W + 4][32] in (SW_SRC 1), FP8 out // with its workgroups in reverse order (WG_REV_ORDER, wg_platform.hlsli; the same bytes): it @@ -20,11 +23,11 @@ // // @wg-kernel l_swin_block_c32_nc co=l_swin.co block=256 wave=32 args=p,p,p,p,u,u,u,u defines=SW_C=32,SW_HEADS=2,SW_MLP=128,SW_ADD=0,SW_SAFE=1,SW_ECL=0,SW_HPU=8,SW_GELU=0,SW_ENTRY_OVERRIDE=l_swin_block_c32_nc // @wg-kernel l_swin_block_c32_enc_nc co=l_swin.co block=256 wave=32 args=p,p,p,p,u,u,u,u defines=SW_C=32,SW_HEADS=2,SW_MLP=128,SW_ADD=0,SW_SAFE=1,SW_ECL=0,SW_HPU=8,SW_GELU=3,SW_GELU_PAT=13333,SW_SRC=1,SW_ENTRY_OVERRIDE=l_swin_block_c32_enc_nc,WG_REV_ORDER=1 -// @wg-kernel l_swin_block_c32_enc_lin_nc co=- block=256 wave=32 args=p,p,p,p,u,u,u,u,p,p defines=SW_C=32,SW_HEADS=2,SW_MLP=128,SW_ADD=0,SW_SAFE=1,SW_ECL=0,SW_HPU=8,SW_GELU=3,SW_GELU_PAT=13333,SW_SRC=3,SW_ENTRY_OVERRIDE=l_swin_block_c32_enc_lin_nc,WG_REV_ORDER=1 +// @wg-kernel l_swin_block_c32_enc_lin_nc co=- block=256 wave=32 args=p,p,p,p,u,u,u,u,p,p defines=SW_C=32,SW_HEADS=2,SW_MLP=128,SW_ADD=0,SW_SAFE=1,SW_ECL=0,SW_HPU=8,SW_GELU=3,SW_GELU_PAT=13333,SW_SRC=3,SW_ENTRY_OVERRIDE=l_swin_block_c32_enc_lin_nc,WG_REV_ORDER=1,SW_HEAD_PAIRS=1,SW_SMAX_STAGED=1,SW_NORM1_EARLY=1 // @wg-kernel l_swin_block_c32_dec_nc co=l_swin.co block=256 wave=32 args=p,p,p,p,u,u,u,u defines=SW_C=32,SW_HEADS=2,SW_MLP=128,SW_ADD=0,SW_SAFE=1,SW_ECL=0,SW_HPU=8,SW_GELU=1,SW_SRC=2,SW_OUT_F16=1,SW_ENTRY_OVERRIDE=l_swin_block_c32_dec_nc // l_swin_block_c32_dec_coef12_nc D3D12 only: the decoder block with l_coef_pack12 as its epilogue (SW_OUT_COEF12; // byte for byte the pair _dec_nc -> l_coef_pack12, which stays the fallback route); // the FP16 map is not written. Arguments: _dec_nc's (out unused), then A, C, latent, // dec0.coef, rw, rh -// @wg-kernel l_swin_block_c32_dec_coef12_nc co=- block=256 wave=32 args=p,p,p,p,u,u,u,u,p,p,p,p,u,u defines=SW_C=32,SW_HEADS=2,SW_MLP=128,SW_ADD=0,SW_SAFE=1,SW_ECL=0,SW_HPU=8,SW_GELU=3,SW_GELU_PAT=13333,SW_SRC=2,SW_OUT_F16=1,SW_OUT_COEF12=1,COEF_CVT_HW=1,SW_ENTRY_OVERRIDE=l_swin_block_c32_dec_coef12_nc +// @wg-kernel l_swin_block_c32_dec_coef12_nc co=- block=256 wave=32 args=p,p,p,p,u,u,u,u,p,p,p,p,u,u defines=SW_C=32,SW_HEADS=2,SW_MLP=128,SW_ADD=0,SW_SAFE=1,SW_ECL=0,SW_HPU=8,SW_GELU=3,SW_GELU_PAT=13333,SW_SRC=2,SW_OUT_F16=1,SW_OUT_COEF12=1,COEF_CVT_HW=1,SW_ENTRY_OVERRIDE=l_swin_block_c32_dec_coef12_nc,SW_SMAX_ADDST=1,SW_SMAX_STAGED=1 #include "swin_block.hlsl"