diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm | 153 | ||||
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.h | 38 |
2 files changed, 191 insertions, 0 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm new file mode 100644 index 000000000000..d5325ef80ab0 --- /dev/null +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.asm @@ -0,0 +1,153 @@ +/* SPDX-License-Identifier: MIT */ +/* + * Copyright 2024 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + */ + +// This shader is to clean LDS, SGPRs and VGPRs. It is first 64 Dwords or 256 bytes of 192 Dwords cleaner shader. +//To turn this shader program on for complitaion change this to main and lower shader main to main_1 + +// MI300 : Clear SGPRs, VGPRs and LDS +// Uses two kernels launched separately: +// 1. Clean VGPRs, LDS, and lower SGPRs +// Launches one workgroup per CU, each workgroup with 4x wave64 per SIMD in the CU +// Waves are "wave64" and have 128 VGPRs each, which uses all 512 VGPRs per SIMD +// Waves in the workgroup share the 64KB of LDS +// Each wave clears SGPRs 0 - 95. Because there are 4 waves/SIMD, this is physical SGPRs 0-383 +// Each wave clears 128 VGPRs, so all 512 in the SIMD +// The first wave of the workgroup clears its 64KB of LDS +// The shader starts with "S_BARRIER" to ensure SPI has launched all waves of the workgroup +// before any wave in the workgroup could end. Without this, it is possible not all SGPRs get cleared. +// 2. Clean remaining SGPRs +// Launches a workgroup with 24 waves per workgroup, yielding 6 waves per SIMD in each CU +// Waves are allocating 96 SGPRs +// CP sets up SPI_RESOURCE_RESERVE_* registers to prevent these waves from allocating SGPRs 0-223. +// As such, these 6 waves per SIMD are allocated physical SGPRs 224-799 +// Barriers do not work for >16 waves per workgroup, so we cannot start with S_BARRIER +// Instead, the shader starts with an S_SETHALT 1. Once all waves are launched CP will send unhalt command +// The shader then clears all SGPRs allocated to it, cleaning out physical SGPRs 224-799 + +shader main + asic(MI300) + type(CS) + wave_size(64) +// Note: original source code from SQ team + +// (theorhetical fastest = ~512clks vgpr + 1536 lds + ~128 sgpr = 2176 clks) + + s_cmp_eq_u32 s0, 1 // Bit0 is set, sgpr0 is set then clear VGPRS and LDS as FW set COMPUTE_USER_DATA_3 + s_cbranch_scc0 label_0023 // Clean VGPRs and LDS if sgpr0 of wave is set, scc = (s3 == 1) + S_BARRIER + + s_movk_i32 m0, 0x0000 + s_mov_b32 s2, 0x00000078 // Loop 128/8=16 times (loop unrolled for performance) + // + // CLEAR VGPRs + // + s_set_gpr_idx_on s2, 0x8 // enable Dest VGPR indexing +label_0005: + v_mov_b32 v0, 0 + v_mov_b32 v1, 0 + v_mov_b32 v2, 0 + v_mov_b32 v3, 0 + v_mov_b32 v4, 0 + v_mov_b32 v5, 0 + v_mov_b32 v6, 0 + v_mov_b32 v7, 0 + s_sub_u32 s2, s2, 8 + s_set_gpr_idx_idx s2 + s_cbranch_scc0 label_0005 + s_set_gpr_idx_off + + // + // + + s_mov_b32 s2, 0x80000000 // Bit31 is first_wave + s_and_b32 s2, s2, s1 // sgpr0 has tg_size (first_wave) term as in ucode only COMPUTE_PGM_RSRC2.tg_size_en is set + s_cbranch_scc0 label_clean_sgpr_1 // Clean LDS if its first wave of ThreadGroup/WorkGroup + // CLEAR LDS + // + s_mov_b32 exec_lo, 0xffffffff + s_mov_b32 exec_hi, 0xffffffff + v_mbcnt_lo_u32_b32 v1, exec_hi, 0 // Set V1 to thread-ID (0..63) + v_mbcnt_hi_u32_b32 v1, exec_lo, v1 // Set V1 to thread-ID (0..63) + v_mul_u32_u24 v1, 0x00000008, v1 // * 8, so each thread is a double-dword address (8byte) + s_mov_b32 s2, 0x00000003f // 64 loop iteraions + s_mov_b32 m0, 0xffffffff + // Clear all of LDS space + // Each FirstWave of WorkGroup clears 64kbyte block + +label_001F: + ds_write2_b64 v1, v[2:3], v[2:3] offset1:32 + ds_write2_b64 v1, v[4:5], v[4:5] offset0:64 offset1:96 + v_add_co_u32 v1, vcc, 0x00000400, v1 + s_sub_u32 s2, s2, 1 + s_cbranch_scc0 label_001F + // + // CLEAR SGPRs + // +label_clean_sgpr_1: + s_mov_b32 m0, 0x0000005c // Loop 96/4=24 times (loop unrolled for performance) + s_nop 0 +label_sgpr_loop: + s_movreld_b32 s0, 0 + s_movreld_b32 s1, 0 + s_movreld_b32 s2, 0 + s_movreld_b32 s3, 0 + s_sub_u32 m0, m0, 4 + s_cbranch_scc0 label_sgpr_loop + + //clear vcc, flat scratch + s_mov_b32 flat_scratch_lo, 0 //clear flat scratch lo SGPR + s_mov_b32 flat_scratch_hi, 0 //clear flat scratch hi SGPR + s_mov_b64 vcc, 0 //clear vcc + s_mov_b64 ttmp0, 0 //Clear ttmp0 and ttmp1 + s_mov_b64 ttmp2, 0 //Clear ttmp2 and ttmp3 + s_mov_b64 ttmp4, 0 //Clear ttmp4 and ttmp5 + s_mov_b64 ttmp6, 0 //Clear ttmp6 and ttmp7 + s_mov_b64 ttmp8, 0 //Clear ttmp8 and ttmp9 + s_mov_b64 ttmp10, 0 //Clear ttmp10 and ttmp11 + s_mov_b64 ttmp12, 0 //Clear ttmp12 and ttmp13 + s_mov_b64 ttmp14, 0 //Clear ttmp14 and ttmp15 +s_endpgm + +label_0023: + + s_sethalt 1 + + s_mov_b32 m0, 0x0000005c // Loop 96/4=24 times (loop unrolled for performance) + s_nop 0 +label_sgpr_loop1: + + s_movreld_b32 s0, 0 + s_movreld_b32 s1, 0 + s_movreld_b32 s2, 0 + s_movreld_b32 s3, 0 + s_sub_u32 m0, m0, 4 + s_cbranch_scc0 label_sgpr_loop1 + + //clear vcc, flat scratch + s_mov_b32 flat_scratch_lo, 0 //clear flat scratch lo SGPR + s_mov_b32 flat_scratch_hi, 0 //clear flat scratch hi SGPR + s_mov_b64 vcc, 0xee //clear vcc + +s_endpgm +end + diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.h b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.h index 042944ac75df..69aa567c6c1d 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.h +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3_cleaner_shader.h @@ -23,4 +23,42 @@ /* Define the cleaner shader gfx_9_4_3 */ static const u32 gfx_9_4_3_cleaner_shader_hex[] = { + 0xbf068100, 0xbf84003b, + 0xbf8a0000, 0xb07c0000, + 0xbe8200ff, 0x00000078, + 0xbf110802, 0x7e000280, + 0x7e020280, 0x7e040280, + 0x7e060280, 0x7e080280, + 0x7e0a0280, 0x7e0c0280, + 0x7e0e0280, 0x80828802, + 0xbe803202, 0xbf84fff5, + 0xbf9c0000, 0xbe8200ff, + 0x80000000, 0x86020102, + 0xbf840011, 0xbefe00c1, + 0xbeff00c1, 0xd28c0001, + 0x0001007f, 0xd28d0001, + 0x0002027e, 0x10020288, + 0xbe8200bf, 0xbefc00c1, + 0xd89c2000, 0x00020201, + 0xd89c6040, 0x00040401, + 0x320202ff, 0x00000400, + 0x80828102, 0xbf84fff8, + 0xbefc00ff, 0x0000005c, + 0xbf800000, 0xbe802c80, + 0xbe812c80, 0xbe822c80, + 0xbe832c80, 0x80fc847c, + 0xbf84fffa, 0xbee60080, + 0xbee70080, 0xbeea0180, + 0xbeec0180, 0xbeee0180, + 0xbef00180, 0xbef20180, + 0xbef40180, 0xbef60180, + 0xbef80180, 0xbefa0180, + 0xbf810000, 0xbf8d0001, + 0xbefc00ff, 0x0000005c, + 0xbf800000, 0xbe802c80, + 0xbe812c80, 0xbe822c80, + 0xbe832c80, 0x80fc847c, + 0xbf84fffa, 0xbee60080, + 0xbee70080, 0xbeea01ff, + 0x000000ee, 0xbf810000, }; |