257 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck --check-prefixes=GFX942,GFX942-SDAG %s3; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck --check-prefixes=GFX942,GFX942-GISEL %s4 5declare <4 x float> @llvm.amdgcn.mfma.f32.16x16x8.xf32(<2 x float>, <2 x float>, <4 x float>, i32, i32, i32)6declare <16 x float> @llvm.amdgcn.mfma.f32.32x32x4.xf32(<2 x float>, <2 x float>, <16 x float>, i32, i32, i32)7 8define amdgpu_kernel void @test_mfma_f32_16x16x8xf32(ptr addrspace(1) %arg) #0 {9; GFX942-SDAG-LABEL: test_mfma_f32_16x16x8xf32:10; GFX942-SDAG: ; %bb.0: ; %bb11; GFX942-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x2412; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 1.013; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, 2.014; GFX942-SDAG-NEXT: v_mov_b32_e32 v0, 0x4040000015; GFX942-SDAG-NEXT: v_mov_b32_e32 v1, 4.016; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)17; GFX942-SDAG-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x018; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 019; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)20; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a0, s021; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a1, s122; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a2, s223; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a3, s324; GFX942-SDAG-NEXT: s_nop 125; GFX942-SDAG-NEXT: v_mfma_f32_16x16x8_xf32 a[0:3], v[4:5], v[0:1], a[0:3] cbsz:1 abid:2 blgp:326; GFX942-SDAG-NEXT: s_nop 627; GFX942-SDAG-NEXT: global_store_dwordx4 v2, a[0:3], s[6:7]28; GFX942-SDAG-NEXT: s_endpgm29;30; GFX942-GISEL-LABEL: test_mfma_f32_16x16x8xf32:31; GFX942-GISEL: ; %bb.0: ; %bb32; GFX942-GISEL-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x2433; GFX942-GISEL-NEXT: s_mov_b32 s4, 1.034; GFX942-GISEL-NEXT: s_mov_b32 s5, 2.035; GFX942-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[4:5]36; GFX942-GISEL-NEXT: s_mov_b32 s4, 0x4040000037; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)38; GFX942-GISEL-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x039; GFX942-GISEL-NEXT: s_mov_b32 s5, 4.040; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[4:5]41; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)42; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a0, s043; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a1, s144; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a2, s245; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a3, s346; GFX942-GISEL-NEXT: s_nop 147; GFX942-GISEL-NEXT: v_mfma_f32_16x16x8_xf32 a[0:3], v[0:1], v[2:3], a[0:3] cbsz:1 abid:2 blgp:348; GFX942-GISEL-NEXT: v_mov_b32_e32 v0, 049; GFX942-GISEL-NEXT: s_nop 550; GFX942-GISEL-NEXT: global_store_dwordx4 v0, a[0:3], s[6:7]51; GFX942-GISEL-NEXT: s_endpgm52bb:53 %in.1 = load <4 x float>, ptr addrspace(1) %arg54 %mai.1 = tail call <4 x float> @llvm.amdgcn.mfma.f32.16x16x8.xf32(<2 x float> <float 1.0, float 2.0>, <2 x float> <float 3.0, float 4.0>, <4 x float> %in.1, i32 1, i32 2, i32 3)55 store <4 x float> %mai.1, ptr addrspace(1) %arg56 ret void57}58 59define amdgpu_kernel void @test_mfma_f32_16x16x8xf32_vgprcd(ptr addrspace(1) %arg) #1 {60; GFX942-SDAG-LABEL: test_mfma_f32_16x16x8xf32_vgprcd:61; GFX942-SDAG: ; %bb.0: ; %bb62; GFX942-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x2463; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 1.064; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, 2.065; GFX942-SDAG-NEXT: v_mov_b32_e32 v6, 0x4040000066; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, 4.067; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)68; GFX942-SDAG-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x069; GFX942-SDAG-NEXT: v_mov_b32_e32 v8, 070; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)71; GFX942-SDAG-NEXT: v_mov_b64_e32 v[0:1], s[0:1]72; GFX942-SDAG-NEXT: v_mov_b64_e32 v[2:3], s[2:3]73; GFX942-SDAG-NEXT: s_nop 174; GFX942-SDAG-NEXT: v_mfma_f32_16x16x8_xf32 v[0:3], v[4:5], v[6:7], v[0:3] cbsz:1 abid:2 blgp:375; GFX942-SDAG-NEXT: s_nop 676; GFX942-SDAG-NEXT: global_store_dwordx4 v8, v[0:3], s[6:7]77; GFX942-SDAG-NEXT: s_endpgm78;79; GFX942-GISEL-LABEL: test_mfma_f32_16x16x8xf32_vgprcd:80; GFX942-GISEL: ; %bb.0: ; %bb81; GFX942-GISEL-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x2482; GFX942-GISEL-NEXT: s_mov_b32 s4, 1.083; GFX942-GISEL-NEXT: s_mov_b32 s5, 2.084; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[4:5]85; GFX942-GISEL-NEXT: s_mov_b32 s4, 0x4040000086; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)87; GFX942-GISEL-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x088; GFX942-GISEL-NEXT: s_mov_b32 s5, 4.089; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[4:5]90; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)91; GFX942-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[0:1]92; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[2:3]93; GFX942-GISEL-NEXT: s_nop 194; GFX942-GISEL-NEXT: v_mfma_f32_16x16x8_xf32 v[0:3], v[4:5], v[6:7], v[0:3] cbsz:1 abid:2 blgp:395; GFX942-GISEL-NEXT: v_mov_b32_e32 v4, 096; GFX942-GISEL-NEXT: s_nop 597; GFX942-GISEL-NEXT: global_store_dwordx4 v4, v[0:3], s[6:7]98; GFX942-GISEL-NEXT: s_endpgm99bb:100 %in.1 = load <4 x float>, ptr addrspace(1) %arg101 %mai.1 = tail call <4 x float> @llvm.amdgcn.mfma.f32.16x16x8.xf32(<2 x float> <float 1.0, float 2.0>, <2 x float> <float 3.0, float 4.0>, <4 x float> %in.1, i32 1, i32 2, i32 3)102 store <4 x float> %mai.1, ptr addrspace(1) %arg103 ret void104}105 106define amdgpu_kernel void @test_mfma_f32_32x32x4xf32(ptr addrspace(1) %arg) #0 {107; GFX942-SDAG-LABEL: test_mfma_f32_32x32x4xf32:108; GFX942-SDAG: ; %bb.0: ; %bb109; GFX942-SDAG-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x24110; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 1.0111; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 2.0112; GFX942-SDAG-NEXT: v_mov_b32_e32 v0, 0x40400000113; GFX942-SDAG-NEXT: v_mov_b32_e32 v1, 4.0114; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)115; GFX942-SDAG-NEXT: s_load_dwordx16 s[0:15], s[16:17], 0x0116; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)117; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a0, s0118; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a1, s1119; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a2, s2120; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a3, s3121; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a4, s4122; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a5, s5123; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a6, s6124; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a7, s7125; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a8, s8126; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a9, s9127; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a10, s10128; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a11, s11129; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a12, s12130; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a13, s13131; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a14, s14132; GFX942-SDAG-NEXT: v_accvgpr_write_b32 a15, s15133; GFX942-SDAG-NEXT: s_nop 1134; GFX942-SDAG-NEXT: v_mfma_f32_32x32x4_xf32 a[0:15], v[2:3], v[0:1], a[0:15] cbsz:1 abid:2 blgp:3135; GFX942-SDAG-NEXT: v_mov_b32_e32 v0, 0136; GFX942-SDAG-NEXT: s_nop 9137; GFX942-SDAG-NEXT: global_store_dwordx4 v0, a[12:15], s[16:17] offset:48138; GFX942-SDAG-NEXT: global_store_dwordx4 v0, a[8:11], s[16:17] offset:32139; GFX942-SDAG-NEXT: global_store_dwordx4 v0, a[4:7], s[16:17] offset:16140; GFX942-SDAG-NEXT: global_store_dwordx4 v0, a[0:3], s[16:17]141; GFX942-SDAG-NEXT: s_endpgm142;143; GFX942-GISEL-LABEL: test_mfma_f32_32x32x4xf32:144; GFX942-GISEL: ; %bb.0: ; %bb145; GFX942-GISEL-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x24146; GFX942-GISEL-NEXT: s_mov_b32 s18, 1.0147; GFX942-GISEL-NEXT: s_mov_b32 s19, 2.0148; GFX942-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[18:19]149; GFX942-GISEL-NEXT: s_mov_b32 s18, 0x40400000150; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)151; GFX942-GISEL-NEXT: s_load_dwordx16 s[0:15], s[16:17], 0x0152; GFX942-GISEL-NEXT: s_mov_b32 s19, 4.0153; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[18:19]154; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)155; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a0, s0156; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a1, s1157; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a2, s2158; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a3, s3159; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a4, s4160; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a5, s5161; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a6, s6162; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a7, s7163; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a8, s8164; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a9, s9165; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a10, s10166; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a11, s11167; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a12, s12168; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a13, s13169; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a14, s14170; GFX942-GISEL-NEXT: v_accvgpr_write_b32 a15, s15171; GFX942-GISEL-NEXT: s_nop 1172; GFX942-GISEL-NEXT: v_mfma_f32_32x32x4_xf32 a[0:15], v[0:1], v[2:3], a[0:15] cbsz:1 abid:2 blgp:3173; GFX942-GISEL-NEXT: v_mov_b32_e32 v0, 0174; GFX942-GISEL-NEXT: s_nop 9175; GFX942-GISEL-NEXT: global_store_dwordx4 v0, a[0:3], s[16:17]176; GFX942-GISEL-NEXT: global_store_dwordx4 v0, a[4:7], s[16:17] offset:16177; GFX942-GISEL-NEXT: global_store_dwordx4 v0, a[8:11], s[16:17] offset:32178; GFX942-GISEL-NEXT: global_store_dwordx4 v0, a[12:15], s[16:17] offset:48179; GFX942-GISEL-NEXT: s_endpgm180bb:181 %in.1 = load <16 x float>, ptr addrspace(1) %arg182 %mai.1 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x4.xf32(<2 x float> <float 1.0, float 2.0>, <2 x float> <float 3.0, float 4.0>, <16 x float> %in.1, i32 1, i32 2, i32 3)183 store <16 x float> %mai.1, ptr addrspace(1) %arg184 ret void185}186 187define amdgpu_kernel void @test_mfma_f32_32x32x4xf32_vgprcd(ptr addrspace(1) %arg) #1 {188; GFX942-SDAG-LABEL: test_mfma_f32_32x32x4xf32_vgprcd:189; GFX942-SDAG: ; %bb.0: ; %bb190; GFX942-SDAG-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x24191; GFX942-SDAG-NEXT: v_mov_b32_e32 v16, 1.0192; GFX942-SDAG-NEXT: v_mov_b32_e32 v17, 2.0193; GFX942-SDAG-NEXT: v_mov_b32_e32 v18, 0x40400000194; GFX942-SDAG-NEXT: v_mov_b32_e32 v19, 4.0195; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)196; GFX942-SDAG-NEXT: s_load_dwordx16 s[0:15], s[16:17], 0x0197; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0)198; GFX942-SDAG-NEXT: v_mov_b64_e32 v[0:1], s[0:1]199; GFX942-SDAG-NEXT: v_mov_b64_e32 v[2:3], s[2:3]200; GFX942-SDAG-NEXT: v_mov_b64_e32 v[4:5], s[4:5]201; GFX942-SDAG-NEXT: v_mov_b64_e32 v[6:7], s[6:7]202; GFX942-SDAG-NEXT: v_mov_b64_e32 v[8:9], s[8:9]203; GFX942-SDAG-NEXT: v_mov_b64_e32 v[10:11], s[10:11]204; GFX942-SDAG-NEXT: v_mov_b64_e32 v[12:13], s[12:13]205; GFX942-SDAG-NEXT: v_mov_b64_e32 v[14:15], s[14:15]206; GFX942-SDAG-NEXT: s_nop 1207; GFX942-SDAG-NEXT: v_mfma_f32_32x32x4_xf32 v[0:15], v[16:17], v[18:19], v[0:15] cbsz:1 abid:2 blgp:3208; GFX942-SDAG-NEXT: v_mov_b32_e32 v16, 0209; GFX942-SDAG-NEXT: s_nop 9210; GFX942-SDAG-NEXT: global_store_dwordx4 v16, v[12:15], s[16:17] offset:48211; GFX942-SDAG-NEXT: global_store_dwordx4 v16, v[8:11], s[16:17] offset:32212; GFX942-SDAG-NEXT: global_store_dwordx4 v16, v[4:7], s[16:17] offset:16213; GFX942-SDAG-NEXT: global_store_dwordx4 v16, v[0:3], s[16:17]214; GFX942-SDAG-NEXT: s_endpgm215;216; GFX942-GISEL-LABEL: test_mfma_f32_32x32x4xf32_vgprcd:217; GFX942-GISEL: ; %bb.0: ; %bb218; GFX942-GISEL-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x24219; GFX942-GISEL-NEXT: s_mov_b32 s18, 1.0220; GFX942-GISEL-NEXT: s_mov_b32 s19, 2.0221; GFX942-GISEL-NEXT: v_mov_b64_e32 v[16:17], s[18:19]222; GFX942-GISEL-NEXT: s_mov_b32 s18, 0x40400000223; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)224; GFX942-GISEL-NEXT: s_load_dwordx16 s[0:15], s[16:17], 0x0225; GFX942-GISEL-NEXT: s_mov_b32 s19, 4.0226; GFX942-GISEL-NEXT: v_mov_b64_e32 v[18:19], s[18:19]227; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0)228; GFX942-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[0:1]229; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[2:3]230; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[4:5]231; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[6:7]232; GFX942-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[8:9]233; GFX942-GISEL-NEXT: v_mov_b64_e32 v[10:11], s[10:11]234; GFX942-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[12:13]235; GFX942-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[14:15]236; GFX942-GISEL-NEXT: s_nop 1237; GFX942-GISEL-NEXT: v_mfma_f32_32x32x4_xf32 v[0:15], v[16:17], v[18:19], v[0:15] cbsz:1 abid:2 blgp:3238; GFX942-GISEL-NEXT: v_mov_b32_e32 v16, 0239; GFX942-GISEL-NEXT: s_nop 9240; GFX942-GISEL-NEXT: global_store_dwordx4 v16, v[0:3], s[16:17]241; GFX942-GISEL-NEXT: global_store_dwordx4 v16, v[4:7], s[16:17] offset:16242; GFX942-GISEL-NEXT: global_store_dwordx4 v16, v[8:11], s[16:17] offset:32243; GFX942-GISEL-NEXT: global_store_dwordx4 v16, v[12:15], s[16:17] offset:48244; GFX942-GISEL-NEXT: s_endpgm245bb:246 %in.1 = load <16 x float>, ptr addrspace(1) %arg247 %mai.1 = tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x4.xf32(<2 x float> <float 1.0, float 2.0>, <2 x float> <float 3.0, float 4.0>, <16 x float> %in.1, i32 1, i32 2, i32 3)248 store <16 x float> %mai.1, ptr addrspace(1) %arg249 ret void250}251 252attributes #0 = { "amdgpu-flat-work-group-size"="1,256" }253attributes #1 = { "amdgpu-flat-work-group-size"="512,512" "amdgpu-agpr-alloc"="0,0" }254 255;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:256; GFX942: {{.*}}257