Skip to content

Commit d727b6f

Browse files
arsenmPravin Jagtap
arsenm
and
Pravin Jagtap
authored
AMDGPU: MC support for v_cvt_scalef32_pk_fp4_{f|bf}16 on gfx950. (#117594)
These instructions have non-standard use of OPSEL bits to select dest write byte. The src2_modifiers operand is used without having its corresponding src2 operand by introducing dummy src2. Co-authored-by: Pravin Jagtap <[email protected]>
1 parent c767570 commit d727b6f

File tree

5 files changed

+123
-1
lines changed

5 files changed

+123
-1
lines changed

‎llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -8824,7 +8824,9 @@ void AMDGPUAsmParser::cvtVOP3P(MCInst &Inst, const OperandVector &Operands,
88248824

88258825
const bool IsPacked = (Desc.TSFlags & SIInstrFlags::IsPacked) != 0;
88268826

8827-
if (Opc == AMDGPU::V_CVT_SR_BF8_F32_vi ||
8827+
if (Opc == AMDGPU::V_CVT_SCALEF32_PK_FP4_F16_vi ||
8828+
Opc == AMDGPU::V_CVT_SCALEF32_PK_FP4_BF16_vi ||
8829+
Opc == AMDGPU::V_CVT_SR_BF8_F32_vi ||
88288830
Opc == AMDGPU::V_CVT_SR_FP8_F32_vi ||
88298831
Opc == AMDGPU::V_CVT_SR_BF8_F32_gfx12_e64_gfx12 ||
88308832
Opc == AMDGPU::V_CVT_SR_FP8_F32_gfx12_e64_gfx12) {

‎llvm/lib/Target/AMDGPU/VOP3Instructions.td‎

Lines changed: 26 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -899,6 +899,23 @@ def VOP3_CVT_SCALE_FP4FP8BF8_F32_Profile : VOP3_Profile<VOPProfile<[i32, f32, f3
899899
let HasOMod = 0;
900900
}
901901

902+
def VOP3_CVT_SCALE_FP4_F16BF16_Profile : VOP3_Profile<VOPProfile<[i32, v2f16, f32, f32]>,
903+
VOP3_OPSEL> {
904+
let InsVOP3OpSel = (ins FP32InputMods:$src0_modifiers, Src0RC64:$src0,
905+
FP32InputMods:$src1_modifiers, Src1RC64:$src1,
906+
FP32InputMods:$src2_modifiers, VGPR_32:$src2,
907+
op_sel0:$op_sel);
908+
let HasClamp = 0;
909+
let HasSrc2 = 0;
910+
let HasSrc2Mods = 1;
911+
let HasOpSel = 1;
912+
let AsmVOP3OpSel = !subst(", $src2_modifiers", "",
913+
getAsmVOP3OpSel<3, HasClamp, HasOMod,
914+
HasSrc0FloatMods, HasSrc1FloatMods,
915+
HasSrc2FloatMods>.ret);
916+
let HasExtVOP3DPP = 0;
917+
}
918+
902919
class VOP3_CVT_SCALE_PK_F16BF16F32_FP4FP8BF8_Profile<ValueType DstTy> : VOP3_Profile<VOPProfile<[DstTy, i32, f32, untyped]>,
903920
VOP3_OPSEL> {
904921
let InsVOP3OpSel = (ins FP32InputMods:$src0_modifiers, Src0RC64:$src0,
@@ -965,6 +982,13 @@ let SubtargetPredicate = HasFP4ConversionScaleInsts, mayRaiseFPException = 0 in
965982
defm V_CVT_SCALEF32_PK_FP4_F32 : VOP3Inst<"v_cvt_scalef32_pk_fp4_f32", VOP3_CVT_SCALE_FP4FP8BF8_F32_Profile>;
966983
defm V_CVT_SCALEF32_PK_F16_FP4 : VOP3Inst<"v_cvt_scalef32_pk_f16_fp4", VOP3_CVT_SCALE_PK_F16BF16F32_FP4FP8BF8_Profile<v2f16>>;
967984
defm V_CVT_SCALEF32_PK_BF16_FP4 : VOP3Inst<"v_cvt_scalef32_pk_bf16_fp4", VOP3_CVT_SCALE_PK_F16BF16F32_FP4FP8BF8_Profile<v2bf16>>;
985+
986+
// These instructions have non-standard use of op_sel. In particular they are
987+
// using op_sel bits 2 and 3 while only having two sources.
988+
let Constraints = "$vdst = $src2", DisableEncoding = "$src2" in {
989+
defm V_CVT_SCALEF32_PK_FP4_F16 : VOP3Inst<"v_cvt_scalef32_pk_fp4_f16", VOP3_CVT_SCALE_FP4_F16BF16_Profile>;
990+
defm V_CVT_SCALEF32_PK_FP4_BF16 : VOP3Inst<"v_cvt_scalef32_pk_fp4_bf16", VOP3_CVT_SCALE_FP4_F16BF16_Profile>;
991+
}
968992
}
969993

970994
let SubtargetPredicate = HasFP6BF6ConversionScaleInsts, mayRaiseFPException = 0 in {
@@ -1930,6 +1954,8 @@ defm V_CVT_SCALEF32_PK_F32_FP4 : VOP3OpSel_Real_gfx9 <0x23f>;
19301954
defm V_CVT_SCALEF32_PK_FP4_F32 : VOP3OpSel_Real_gfx9 <0x23d>;
19311955
defm V_CVT_SCALEF32_PK_F16_FP4 : VOP3OpSel_Real_gfx9 <0x250>;
19321956
defm V_CVT_SCALEF32_PK_BF16_FP4 : VOP3OpSel_Real_gfx9 <0x251>;
1957+
defm V_CVT_SCALEF32_PK_FP4_F16 : VOP3OpSel_Real_gfx9_forced_opsel2 <0x24c>;
1958+
defm V_CVT_SCALEF32_PK_FP4_BF16: VOP3OpSel_Real_gfx9_forced_opsel2 <0x24d>;
19331959
}
19341960
let OtherPredicates = [HasFP6BF6ConversionScaleInsts] in {
19351961
defm V_CVT_SCALEF32_PK32_F32_FP6 : VOP3_Real_gfx9<0x256, "v_cvt_scalef32_pk32_f32_fp6">;

‎llvm/test/MC/AMDGPU/gfx950_asm_features.s‎

Lines changed: 40 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1025,3 +1025,43 @@ v_cvt_scalef32_pk_bf16_bf8 v1, v2, s3 op_sel:[1,0,0]
10251025
// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
10261026
// GFX950: v_cvt_scalef32_pk_bf16_bf8 v1, s2, 3 op_sel:[1,0,0] ; encoding: [0x01,0x08,0x6a,0xd2,0x02,0x06,0x01,0x00]
10271027
v_cvt_scalef32_pk_bf16_bf8 v1, s2, 3 op_sel:[1,0,0]
1028+
1029+
// NOT-GFX950: error: instruction not supported on this GPU
1030+
// GFX950: v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 ; encoding: [0x01,0x00,0x4c,0xd2,0x02,0x07,0x02,0x00]
1031+
v_cvt_scalef32_pk_fp4_f16 v1, v2, v3
1032+
1033+
// NOT-GFX950: error: instruction not supported on this GPU
1034+
// GFX950: v_cvt_scalef32_pk_fp4_f16 v1, s2, 3 ; encoding: [0x01,0x00,0x4c,0xd2,0x02,0x06,0x01,0x00]
1035+
v_cvt_scalef32_pk_fp4_f16 v1, s2, 3
1036+
1037+
// NOT-GFX950: error: instruction not supported on this GPU
1038+
// GFX950: v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 op_sel:[0,0,1,1] ; encoding: [0x01,0x60,0x4c,0xd2,0x02,0x07,0x02,0x00]
1039+
v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 op_sel:[0,0,1,1]
1040+
1041+
// NOT-GFX950: error: instruction not supported on this GPU
1042+
// GFX950: v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x01,0x40,0x4c,0xd2,0x02,0x07,0x02,0x00]
1043+
v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 op_sel:[0,0,0,1]
1044+
1045+
// NOT-GFX950: error: instruction not supported on this GPU
1046+
// GFX950: v_cvt_scalef32_pk_fp4_f16 v1, -|s2|, v3 ; encoding: [0x01,0x01,0x4c,0xd2,0x02,0x06,0x02,0x20]
1047+
v_cvt_scalef32_pk_fp4_f16 v1, -|s2|, v3
1048+
1049+
// NOT-GFX950: error: instruction not supported on this GPU
1050+
// GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 ; encoding: [0x01,0x00,0x4d,0xd2,0x02,0x07,0x02,0x00]
1051+
v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3
1052+
1053+
// NOT-GFX950: error: instruction not supported on this GPU
1054+
// GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, s2, 3 ; encoding: [0x01,0x00,0x4d,0xd2,0x02,0x06,0x01,0x00]
1055+
v_cvt_scalef32_pk_fp4_bf16 v1, s2, 3
1056+
1057+
// NOT-GFX950: error: instruction not supported on this GPU
1058+
// GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 op_sel:[0,0,1,1] ; encoding: [0x01,0x60,0x4d,0xd2,0x02,0x07,0x02,0x00]
1059+
v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 op_sel:[0,0,1,1]
1060+
1061+
// NOT-GFX950: error: instruction not supported on this GPU
1062+
// GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x01,0x40,0x4d,0xd2,0x02,0x07,0x02,0x00]
1063+
v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 op_sel:[0,0,0,1]
1064+
1065+
// NOT-GFX950: error: instruction not supported on this GPU
1066+
// GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, -|s2|, v3 ; encoding: [0x01,0x01,0x4d,0xd2,0x02,0x06,0x02,0x20]
1067+
v_cvt_scalef32_pk_fp4_bf16 v1, -|s2|, v3

‎llvm/test/MC/AMDGPU/gfx950_err.s‎

Lines changed: 24 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -293,3 +293,27 @@ v_cvt_scalef32_pk_bf16_bf8 v[20:25], v[10:25], v8 div:2
293293

294294
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: not a valid operand
295295
v_cvt_scalef32_pk_bf16_bf8 v[20:25], v[10:25], v8 clamp div:2
296+
297+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: invalid operand for instruction
298+
v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 clamp
299+
300+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: not a valid operand
301+
v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 mul:2
302+
303+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: not a valid operand
304+
v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 div:2
305+
306+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: not a valid operand
307+
v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 clamp div:2
308+
309+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: invalid operand for instruction
310+
v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 clamp
311+
312+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: not a valid operand
313+
v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 mul:2
314+
315+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: not a valid operand
316+
v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 div:2
317+
318+
// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: not a valid operand
319+
v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 clamp div:2

‎llvm/test/MC/Disassembler/AMDGPU/gfx950_dasm_vop3.txt‎

Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -713,3 +713,33 @@
713713

714714
# GFX950: v_cvt_scalef32_pk_bf16_bf8 v1, s2, 3 op_sel:[1,0,0] ; encoding: [0x01,0x08,0x6a,0xd2,0x02,0x06,0x01,0x00]
715715
0x01,0x08,0x6a,0xd2,0x02,0x06,0x01,0x00
716+
717+
# GFX950: v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 ; encoding: [0x01,0x00,0x4c,0xd2,0x02,0x07,0x02,0x00]
718+
0x01,0x00,0x4c,0xd2,0x02,0x07,0x02,0x00
719+
720+
# GFX950: v_cvt_scalef32_pk_fp4_f16 v1, s2, 3 ; encoding: [0x01,0x00,0x4c,0xd2,0x02,0x06,0x01,0x00]
721+
0x01,0x00,0x4c,0xd2,0x02,0x06,0x01,0x00
722+
723+
# GFX950: v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 op_sel:[0,0,1,1] ; encoding: [0x01,0x60,0x4c,0xd2,0x02,0x07,0x02,0x00]
724+
0x01,0x60,0x4c,0xd2,0x02,0x07,0x02,0x00
725+
726+
# GFX950: v_cvt_scalef32_pk_fp4_f16 v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x01,0x40,0x4c,0xd2,0x02,0x07,0x02,0x00]
727+
0x01,0x40,0x4c,0xd2,0x02,0x07,0x02,0x00
728+
729+
# GFX950: v_cvt_scalef32_pk_fp4_f16 v1, -|s2|, v3 ; encoding: [0x01,0x01,0x4c,0xd2,0x02,0x06,0x02,0x20]
730+
0x01,0x01,0x4c,0xd2,0x02,0x06,0x02,0x20
731+
732+
# GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 ; encoding: [0x01,0x00,0x4d,0xd2,0x02,0x07,0x02,0x00]
733+
0x01,0x00,0x4d,0xd2,0x02,0x07,0x02,0x00
734+
735+
# GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, s2, 3 ; encoding: [0x01,0x00,0x4d,0xd2,0x02,0x06,0x01,0x00]
736+
0x01,0x00,0x4d,0xd2,0x02,0x06,0x01,0x00
737+
738+
# GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 op_sel:[0,0,1,1] ; encoding: [0x01,0x60,0x4d,0xd2,0x02,0x07,0x02,0x00]
739+
0x01,0x60,0x4d,0xd2,0x02,0x07,0x02,0x00
740+
741+
# GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x01,0x40,0x4d,0xd2,0x02,0x07,0x02,0x00]
742+
0x01,0x40,0x4d,0xd2,0x02,0x07,0x02,0x00
743+
744+
# GFX950: v_cvt_scalef32_pk_fp4_bf16 v1, -|s2|, v3 ; encoding: [0x01,0x01,0x4d,0xd2,0x02,0x06,0x02,0x20]
745+
0x01,0x01,0x4d,0xd2,0x02,0x06,0x02,0x20

0 commit comments

Comments
 (0)