hsw.asm (6971B)
1// SPDX-License-Identifier: MIT 2/* 3 * Copyright © 2020 Intel Corporation 4 */ 5 6/* 7 * Kernel for PAVP buffer clear. 8 * 9 * 1. Clear all 64 GRF registers assigned to the kernel with designated value; 10 * 2. Write 32x16 block of all "0" to render target buffer which indirectly clears 11 * 512 bytes of Render Cache. 12 */ 13 14/* Store designated "clear GRF" value */ 15mov(1) f0.1<1>UW g1.2<0,1,0>UW { align1 1N }; 16 17/** 18 * Curbe Format 19 * 20 * DW 1.0 - Block Offset to write Render Cache 21 * DW 1.1 [15:0] - Clear Word 22 * DW 1.2 - Delay iterations 23 * DW 1.3 - Enable Instrumentation (only for debug) 24 * DW 1.4 - Rsvd (intended for context ID) 25 * DW 1.5 - [31:16]:SliceCount, [15:0]:SubSlicePerSliceCount 26 * DW 1.6 - Rsvd MBZ (intended for Enable Wait on Total Thread Count) 27 * DW 1.7 - Rsvd MBZ (inteded for Total Thread Count) 28 * 29 * Binding Table 30 * 31 * BTI 0: 2D Surface to help clear L3 (Render/Data Cache) 32 * BTI 1: Wait/Instrumentation Buffer 33 * Size : (SliceCount * SubSliceCount * 16 EUs/SubSlice) rows * (16 threads/EU) cols (Format R32_UINT) 34 * Expected to be initialized to 0 by driver/another kernel 35 * Layout: 36 * RowN: Histogram for EU-N: (SliceID*SubSlicePerSliceCount + SSID)*16 + EUID [assume max 16 EUs / SS] 37 * Col-k[DW-k]: Threads Executed on ThreadID-k for EU-N 38 */ 39add(1) g1.2<1>UD g1.2<0,1,0>UD 0x00000001UD { align1 1N }; /* Loop count to delay kernel: Init to (g1.2 + 1) */ 40cmp.z.f0.0(1) null<1>UD g1.3<0,1,0>UD 0x00000000UD { align1 1N }; 41(+f0.0) jmpi(1) 352D { align1 WE_all 1N }; 42 43/** 44 * State Register has info on where this thread is running 45 * IVB: sr0.0 :: [15:13]: MBZ, 12: HSID (Half-Slice ID), [11:8]EUID, [2:0] ThreadSlotID 46 * HSW: sr0.0 :: 15: MBZ, [14:13]: SliceID, 12: HSID (Half-Slice ID), [11:8]EUID, [2:0] ThreadSlotID 47 */ 48mov(8) g3<1>UD 0x00000000UD { align1 1Q }; 49shr(1) g3<1>D sr0<0,1,0>D 12D { align1 1N }; 50and(1) g3<1>D g3<0,1,0>D 1D { align1 1N }; /* g3 has HSID */ 51shr(1) g3.1<1>D sr0<0,1,0>D 13D { align1 1N }; 52and(1) g3.1<1>D g3.1<0,1,0>D 3D { align1 1N }; /* g3.1 has sliceID */ 53mul(1) g3.5<1>D g3.1<0,1,0>D g1.10<0,1,0>UW { align1 1N }; 54add(1) g3<1>D g3<0,1,0>D g3.5<0,1,0>D { align1 1N }; /* g3 = sliceID * SubSlicePerSliceCount + HSID */ 55shr(1) g3.2<1>D sr0<0,1,0>D 8D { align1 1N }; 56and(1) g3.2<1>D g3.2<0,1,0>D 15D { align1 1N }; /* g3.2 = EUID */ 57mul(1) g3.4<1>D g3<0,1,0>D 16D { align1 1N }; 58add(1) g3.2<1>D g3.2<0,1,0>D g3.4<0,1,0>D { align1 1N }; /* g3.2 now points to EU row number (Y-pixel = V address ) in instrumentation surf */ 59 60mov(8) g5<1>UD 0x00000000UD { align1 1Q }; 61and(1) g3.3<1>D sr0<0,1,0>D 7D { align1 1N }; 62mul(1) g3.3<1>D g3.3<0,1,0>D 4D { align1 1N }; 63 64mov(8) g4<1>UD g0<8,8,1>UD { align1 1Q }; /* Initialize message header with g0 */ 65mov(1) g4<1>UD g3.3<0,1,0>UD { align1 1N }; /* Block offset */ 66mov(1) g4.1<1>UD g3.2<0,1,0>UD { align1 1N }; /* Block offset */ 67mov(1) g4.2<1>UD 0x00000003UD { align1 1N }; /* Block size (1 row x 4 bytes) */ 68and(1) g4.3<1>UD g4.3<0,1,0>UW 0xffffffffUD { align1 1N }; 69 70/* Media block read to fetch current value at specified location in instrumentation buffer */ 71sendc(8) g5<1>UD g4<8,8,1>F 0x02190001 72 73 render MsgDesc: media block read MsgCtrl = 0x0 Surface = 1 mlen 1 rlen 1 { align1 1Q }; 74add(1) g5<1>D g5<0,1,0>D 1D { align1 1N }; 75 76/* Media block write for updated value at specified location in instrumentation buffer */ 77sendc(8) g5<1>UD g4<8,8,1>F 0x040a8001 78 render MsgDesc: media block write MsgCtrl = 0x0 Surface = 1 mlen 2 rlen 0 { align1 1Q }; 79 80/* Delay thread for specified parameter */ 81add.nz.f0.0(1) g1.2<1>UD g1.2<0,1,0>UD -1D { align1 1N }; 82(+f0.0) jmpi(1) -32D { align1 WE_all 1N }; 83 84/* Store designated "clear GRF" value */ 85mov(1) f0.1<1>UW g1.2<0,1,0>UW { align1 1N }; 86 87/* Initialize looping parameters */ 88mov(1) a0<1>D 0D { align1 1N }; /* Initialize a0.0:w=0 */ 89mov(1) a0.4<1>W 127W { align1 1N }; /* Loop count. Each loop contains 16 GRF's */ 90 91/* Write 32x16 all "0" block */ 92mov(8) g2<1>UD g0<8,8,1>UD { align1 1Q }; 93mov(8) g127<1>UD g0<8,8,1>UD { align1 1Q }; 94mov(2) g2<1>UD g1<2,2,1>UW { align1 1N }; 95mov(1) g2.2<1>UD 0x000f000fUD { align1 1N }; /* Block size (16x16) */ 96and(1) g2.3<1>UD g2.3<0,1,0>UW 0xffffffefUD { align1 1N }; 97mov(16) g3<1>UD 0x00000000UD { align1 1H }; 98mov(16) g4<1>UD 0x00000000UD { align1 1H }; 99mov(16) g5<1>UD 0x00000000UD { align1 1H }; 100mov(16) g6<1>UD 0x00000000UD { align1 1H }; 101mov(16) g7<1>UD 0x00000000UD { align1 1H }; 102mov(16) g8<1>UD 0x00000000UD { align1 1H }; 103mov(16) g9<1>UD 0x00000000UD { align1 1H }; 104mov(16) g10<1>UD 0x00000000UD { align1 1H }; 105sendc(8) null<1>UD g2<8,8,1>F 0x120a8000 106 render MsgDesc: media block write MsgCtrl = 0x0 Surface = 0 mlen 9 rlen 0 { align1 1Q }; 107add(1) g2<1>UD g1<0,1,0>UW 0x0010UW { align1 1N }; 108sendc(8) null<1>UD g2<8,8,1>F 0x120a8000 109 render MsgDesc: media block write MsgCtrl = 0x0 Surface = 0 mlen 9 rlen 0 { align1 1Q }; 110 111/* Now, clear all GRF registers */ 112add.nz.f0.0(1) a0.4<1>W a0.4<0,1,0>W -1W { align1 1N }; 113mov(16) g[a0]<1>UW f0.1<0,1,0>UW { align1 1H }; 114add(1) a0<1>D a0<0,1,0>D 32D { align1 1N }; 115(+f0.0) jmpi(1) -64D { align1 WE_all 1N }; 116 117/* Terminante the thread */ 118sendc(8) null<1>UD g127<8,8,1>F 0x82000010 119 thread_spawner MsgDesc: mlen 1 rlen 0 { align1 1Q EOT };