brintos

brintos / llvm-project-archived public Read only

0
0
Text · 8.4 KiB · 4dd258b Raw
217 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn -mcpu=hawaii < %s | FileCheck --check-prefix=GFX7 %s3; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck --check-prefix=GFX8 %s4; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck --check-prefix=GFX9 %s5; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck --check-prefix=GFX10 %s6; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 < %s | FileCheck --check-prefix=GFX11 %s7; RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -amdgpu-enable-delay-alu=0 < %s | FileCheck --check-prefix=GFX12 %s8 9define bfloat @raw_ptr_buffer_load_bf16(ptr addrspace(8) inreg %rsrc) {10; GFX7-LABEL: raw_ptr_buffer_load_bf16:11; GFX7:       ; %bb.0:12; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)13; GFX7-NEXT:    buffer_load_ushort v0, off, s[16:19], 014; GFX7-NEXT:    s_waitcnt vmcnt(0)15; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v016; GFX7-NEXT:    s_setpc_b64 s[30:31]17;18; GFX8-LABEL: raw_ptr_buffer_load_bf16:19; GFX8:       ; %bb.0:20; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)21; GFX8-NEXT:    buffer_load_ushort v0, off, s[16:19], 022; GFX8-NEXT:    s_waitcnt vmcnt(0)23; GFX8-NEXT:    s_setpc_b64 s[30:31]24;25; GFX9-LABEL: raw_ptr_buffer_load_bf16:26; GFX9:       ; %bb.0:27; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)28; GFX9-NEXT:    buffer_load_ushort v0, off, s[16:19], 029; GFX9-NEXT:    s_waitcnt vmcnt(0)30; GFX9-NEXT:    s_setpc_b64 s[30:31]31;32; GFX10-LABEL: raw_ptr_buffer_load_bf16:33; GFX10:       ; %bb.0:34; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)35; GFX10-NEXT:    buffer_load_ushort v0, off, s[16:19], 036; GFX10-NEXT:    s_waitcnt vmcnt(0)37; GFX10-NEXT:    s_setpc_b64 s[30:31]38;39; GFX11-LABEL: raw_ptr_buffer_load_bf16:40; GFX11:       ; %bb.0:41; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)42; GFX11-NEXT:    buffer_load_u16 v0, off, s[0:3], 043; GFX11-NEXT:    s_waitcnt vmcnt(0)44; GFX11-NEXT:    s_setpc_b64 s[30:31]45;46; GFX12-LABEL: raw_ptr_buffer_load_bf16:47; GFX12:       ; %bb.0:48; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x049; GFX12-NEXT:    s_wait_kmcnt 0x050; GFX12-NEXT:    buffer_load_u16 v0, off, s[0:3], null51; GFX12-NEXT:    s_wait_loadcnt 0x052; GFX12-NEXT:    s_set_pc_i64 s[30:31]53  %val = call bfloat @llvm.amdgcn.raw.ptr.buffer.load.v2bf16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)54  ret bfloat %val55}56 57define <2 x bfloat> @raw_ptr_buffer_load_v2bf16(ptr addrspace(8) inreg %rsrc) {58; GFX7-LABEL: raw_ptr_buffer_load_v2bf16:59; GFX7:       ; %bb.0:60; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)61; GFX7-NEXT:    buffer_load_dword v1, off, s[16:19], 062; GFX7-NEXT:    s_waitcnt vmcnt(0)63; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v164; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v165; GFX7-NEXT:    s_setpc_b64 s[30:31]66;67; GFX8-LABEL: raw_ptr_buffer_load_v2bf16:68; GFX8:       ; %bb.0:69; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)70; GFX8-NEXT:    buffer_load_dword v0, off, s[16:19], 071; GFX8-NEXT:    s_waitcnt vmcnt(0)72; GFX8-NEXT:    s_setpc_b64 s[30:31]73;74; GFX9-LABEL: raw_ptr_buffer_load_v2bf16:75; GFX9:       ; %bb.0:76; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)77; GFX9-NEXT:    buffer_load_dword v0, off, s[16:19], 078; GFX9-NEXT:    s_waitcnt vmcnt(0)79; GFX9-NEXT:    s_setpc_b64 s[30:31]80;81; GFX10-LABEL: raw_ptr_buffer_load_v2bf16:82; GFX10:       ; %bb.0:83; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)84; GFX10-NEXT:    buffer_load_dword v0, off, s[16:19], 085; GFX10-NEXT:    s_waitcnt vmcnt(0)86; GFX10-NEXT:    s_setpc_b64 s[30:31]87;88; GFX11-LABEL: raw_ptr_buffer_load_v2bf16:89; GFX11:       ; %bb.0:90; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)91; GFX11-NEXT:    buffer_load_b32 v0, off, s[0:3], 092; GFX11-NEXT:    s_waitcnt vmcnt(0)93; GFX11-NEXT:    s_setpc_b64 s[30:31]94;95; GFX12-LABEL: raw_ptr_buffer_load_v2bf16:96; GFX12:       ; %bb.0:97; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x098; GFX12-NEXT:    s_wait_kmcnt 0x099; GFX12-NEXT:    buffer_load_b32 v0, off, s[0:3], null100; GFX12-NEXT:    s_wait_loadcnt 0x0101; GFX12-NEXT:    s_set_pc_i64 s[30:31]102  %val = call <2 x bfloat> @llvm.amdgcn.raw.ptr.buffer.load.v2bf16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)103  ret <2 x bfloat> %val104}105 106define <4 x bfloat> @raw_ptr_buffer_load_v4bf16(ptr addrspace(8) inreg %rsrc) {107; GFX7-LABEL: raw_ptr_buffer_load_v4bf16:108; GFX7:       ; %bb.0:109; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)110; GFX7-NEXT:    buffer_load_dwordx2 v[2:3], off, s[16:19], 0111; GFX7-NEXT:    s_waitcnt vmcnt(0)112; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v2113; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2114; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v3115; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3116; GFX7-NEXT:    s_setpc_b64 s[30:31]117;118; GFX8-LABEL: raw_ptr_buffer_load_v4bf16:119; GFX8:       ; %bb.0:120; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)121; GFX8-NEXT:    buffer_load_dwordx2 v[0:1], off, s[16:19], 0122; GFX8-NEXT:    s_waitcnt vmcnt(0)123; GFX8-NEXT:    s_setpc_b64 s[30:31]124;125; GFX9-LABEL: raw_ptr_buffer_load_v4bf16:126; GFX9:       ; %bb.0:127; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)128; GFX9-NEXT:    buffer_load_dwordx2 v[0:1], off, s[16:19], 0129; GFX9-NEXT:    s_waitcnt vmcnt(0)130; GFX9-NEXT:    s_setpc_b64 s[30:31]131;132; GFX10-LABEL: raw_ptr_buffer_load_v4bf16:133; GFX10:       ; %bb.0:134; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)135; GFX10-NEXT:    buffer_load_dwordx2 v[0:1], off, s[16:19], 0136; GFX10-NEXT:    s_waitcnt vmcnt(0)137; GFX10-NEXT:    s_setpc_b64 s[30:31]138;139; GFX11-LABEL: raw_ptr_buffer_load_v4bf16:140; GFX11:       ; %bb.0:141; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)142; GFX11-NEXT:    buffer_load_b64 v[0:1], off, s[0:3], 0143; GFX11-NEXT:    s_waitcnt vmcnt(0)144; GFX11-NEXT:    s_setpc_b64 s[30:31]145;146; GFX12-LABEL: raw_ptr_buffer_load_v4bf16:147; GFX12:       ; %bb.0:148; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0149; GFX12-NEXT:    s_wait_kmcnt 0x0150; GFX12-NEXT:    buffer_load_b64 v[0:1], off, s[0:3], null151; GFX12-NEXT:    s_wait_loadcnt 0x0152; GFX12-NEXT:    s_set_pc_i64 s[30:31]153  %val = call <4 x bfloat> @llvm.amdgcn.raw.ptr.buffer.load.v4bf16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)154  ret <4 x bfloat> %val155}156 157; FIXME158; define <6 x bfloat> @raw_ptr_buffer_load_v6bf16(ptr addrspace(8) inreg %rsrc) {159;   %val = call <6 x bfloat> @llvm.amdgcn.raw.ptr.buffer.load.v6bf16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)160;   ret <6 x bfloat> %val161; }162 163define <8 x bfloat> @raw_ptr_buffer_load_v8bf16(ptr addrspace(8) inreg %rsrc) {164; GFX7-LABEL: raw_ptr_buffer_load_v8bf16:165; GFX7:       ; %bb.0:166; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)167; GFX7-NEXT:    buffer_load_dwordx4 v[4:7], off, s[16:19], 0168; GFX7-NEXT:    s_waitcnt vmcnt(0)169; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v4170; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v4171; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v5172; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5173; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v6174; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v6175; GFX7-NEXT:    v_lshlrev_b32_e32 v6, 16, v7176; GFX7-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7177; GFX7-NEXT:    s_setpc_b64 s[30:31]178;179; GFX8-LABEL: raw_ptr_buffer_load_v8bf16:180; GFX8:       ; %bb.0:181; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)182; GFX8-NEXT:    buffer_load_dwordx4 v[0:3], off, s[16:19], 0183; GFX8-NEXT:    s_waitcnt vmcnt(0)184; GFX8-NEXT:    s_setpc_b64 s[30:31]185;186; GFX9-LABEL: raw_ptr_buffer_load_v8bf16:187; GFX9:       ; %bb.0:188; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)189; GFX9-NEXT:    buffer_load_dwordx4 v[0:3], off, s[16:19], 0190; GFX9-NEXT:    s_waitcnt vmcnt(0)191; GFX9-NEXT:    s_setpc_b64 s[30:31]192;193; GFX10-LABEL: raw_ptr_buffer_load_v8bf16:194; GFX10:       ; %bb.0:195; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)196; GFX10-NEXT:    buffer_load_dwordx4 v[0:3], off, s[16:19], 0197; GFX10-NEXT:    s_waitcnt vmcnt(0)198; GFX10-NEXT:    s_setpc_b64 s[30:31]199;200; GFX11-LABEL: raw_ptr_buffer_load_v8bf16:201; GFX11:       ; %bb.0:202; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)203; GFX11-NEXT:    buffer_load_b128 v[0:3], off, s[0:3], 0204; GFX11-NEXT:    s_waitcnt vmcnt(0)205; GFX11-NEXT:    s_setpc_b64 s[30:31]206;207; GFX12-LABEL: raw_ptr_buffer_load_v8bf16:208; GFX12:       ; %bb.0:209; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0210; GFX12-NEXT:    s_wait_kmcnt 0x0211; GFX12-NEXT:    buffer_load_b128 v[0:3], off, s[0:3], null212; GFX12-NEXT:    s_wait_loadcnt 0x0213; GFX12-NEXT:    s_set_pc_i64 s[30:31]214  %val = call <8 x bfloat> @llvm.amdgcn.raw.ptr.buffer.load.v8bf16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0)215  ret <8 x bfloat> %val216}217