AMDGPU: MC support for v_cvt_sr_{f16|bf16}_f32 instructions (llvm#117796)

arsenm · shiltian · pravinjagtap · commit 277ede4d40c9 · 2025-01-21T21:00:17.000+05:30
Co-authored-by: Shilei Tian &lt;shilei.tian@amd.com&gt;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -420,6 +420,12 @@ def FeatureF16BF16ToFP6BF6ConversionScaleInsts : SubtargetFeature<"f16bf16-to-fp
   "Has f16bf16 to fp6bf6 conversion scale instructions"
 >;
 
+def FeatureF32ToF16BF16ConversionSRInsts : SubtargetFeature<"f32-to-f16bf16-cvt-sr-insts",
+  "HasF32ToF16BF16ConversionSRInsts",
+  "true",
+  "Has f32 to f16bf16 conversion scale instructions"
+>;
+
 def FeatureAshrPkInsts : SubtargetFeature<"ashr-pk-insts",
   "HasAshrPkInsts",
   "true",
@@ -438,6 +444,7 @@ def FeatureGFX950Insts : SubtargetFeature<"gfx950-insts",
    FeatureFP4ConversionScaleInsts,
    FeatureFP6BF6ConversionScaleInsts,
    FeatureF16BF16ToFP6BF6ConversionScaleInsts,
+   FeatureF32ToF16BF16ConversionSRInsts,
    FeatureMinimum3Maximum3F32,
    FeatureMinimum3Maximum3PKF16
    ]
@@ -2509,6 +2516,9 @@ def HasFP6BF6ConversionScaleInsts : Predicate<"Subtarget->hasFP6BF6ConversionSca
 def HasF16BF16ToFP6BF6ConversionScaleInsts : Predicate<"Subtarget->hasF16BF16ToFP6BF6ConversionScaleInsts()">,
   AssemblerPredicate<(all_of FeatureF16BF16ToFP6BF6ConversionScaleInsts)>;
 
+def HasF32ToF16BF16ConversionSRInsts : Predicate<"Subtarget->hasF32ToF16BF16ConversionSRInsts()">,
+  AssemblerPredicate<(all_of FeatureF32ToF16BF16ConversionSRInsts)>;
+
 def HasGDS : Predicate<"Subtarget->hasGDS()">;
 
 def HasGWS : Predicate<"Subtarget->hasGWS()">;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -55,6 +55,7 @@ class AMDGPUSubtarget {
   bool HasFP4ConversionScaleInsts = false;
   bool HasFP6BF6ConversionScaleInsts = false;
   bool HasF16BF16ToFP6BF6ConversionScaleInsts = false;
+  bool HasF32ToF16BF16ConversionSRInsts = false;
   bool EnableRealTrue16Insts = false;
   bool HasBF16ConversionInsts = false;
   bool HasMadMixInsts = false;
@@ -190,6 +191,10 @@ class AMDGPUSubtarget {
 
   bool hasF16BF16ToFP6BF6ConversionScaleInsts() const { return HasF16BF16ToFP6BF6ConversionScaleInsts; }
 
+  bool hasF32ToF16BF16ConversionSRInsts() const {
+    return HasF32ToF16BF16ConversionSRInsts;
+  }
+
   bool hasMadMacF32Insts() const {
     return HasMadMacF32Insts || !isGCN();
   }
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -2657,6 +2657,8 @@ def VOP_V2I16_V2BF16_F32 : VOPProfile<[v2i16, v2bf16, f32, untyped]>;
 def VOP_I32_F32_F32_F32 : VOPProfile<[i32, f32, f32, f32]>;
 def VOP_I32_V2F16_F32_F32 : VOPProfile<[i32, v2f16, f32, f32]>;
 def VOP_I32_V2BF16_F32_F32: VOPProfile<[i32, v2bf16, f32, f32]>;
+def VOP_BF16_F32_I32 : VOPProfile<[bf16, f32, i32, untyped]>;
+def VOP_F16_F32_I32 : VOPProfile<[f16, f32, i32, untyped]>;
 
 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
 def VOP_I64_I32_I64 : VOPProfile <[i64, i32, i64, untyped]>;
diff --git a/llvm/lib/Target/AMDGPU/VOP3Instructions.td b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
@@ -1261,6 +1261,12 @@ let SubtargetPredicate = isGFX11Plus in {
   defm V_CVT_PK_U16_F32 : VOP3Inst<"v_cvt_pk_u16_f32", VOP3_Profile<VOP_V2I16_F32_F32>>;
 } // End SubtargetPredicate = isGFX11Plus
 
+class VOP3_CVT_SR_FP16_TiedInput_Profile<VOPProfile P> : VOP3_CVT_SCALE_F1632_FP8BF8_TiedInput_Profile<P> {
+  let InsVOP3OpSel = (ins FP32InputMods:$src0_modifiers, Src0RC64:$src0,
+                          Int32InputMods:$src1_modifiers, Src1RC64:$src1,
+                          VGPR_32:$vdst_in, op_sel0:$op_sel);
+}
+
 // FIXME: GlobalISel cannot distinguish f16 and bf16 and may start using bf16 patterns
 //        instead of less complex f16. Disable GlobalISel for these for now.
 def bf16_fpround : PatFrag <(ops node:$src0),  (fpround $src0), [{ return true; }]> {
@@ -1285,6 +1291,13 @@ let SubtargetPredicate = HasBF16ConversionInsts in {
                (V_CVT_PK_BF16_F32_e64 0, (f32 (V_CVT_F32_F64_e64 $src0_modifiers, $src0)), 0, (f32 (IMPLICIT_DEF)))>;
 }
 
+let SubtargetPredicate = HasF32ToF16BF16ConversionSRInsts in {
+  let Constraints = "$vdst = $vdst_in", DisableEncoding = "$vdst_in" in {
+    defm V_CVT_SR_F16_F32   : VOP3Inst<"v_cvt_sr_f16_f32", VOP3_CVT_SR_FP16_TiedInput_Profile<VOP_F16_F32_I32>>;
+    defm V_CVT_SR_BF16_F32  : VOP3Inst<"v_cvt_sr_bf16_f32", VOP3_CVT_SR_FP16_TiedInput_Profile<VOP_BF16_F32_I32>>;
+  }
+}
+
 let SubtargetPredicate = isGFX12Plus, ReadsModeReg = 0 in {
   defm V_MAXIMUMMINIMUM_F32 : VOP3Inst<"v_maximumminimum_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
   defm V_MINIMUMMAXIMUM_F32 : VOP3Inst<"v_minimummaximum_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
@@ -2144,6 +2157,11 @@ defm V_CVT_SCALEF32_PK32_BF6_F16  : VOP3_Real_gfx9<0x25a, "v_cvt_scalef32_pk32_b
 defm V_CVT_SCALEF32_PK32_BF6_BF16 : VOP3_Real_gfx9<0x25b, "v_cvt_scalef32_pk32_bf6_bf16">;
 }
 
+let OtherPredicates = [HasF32ToF16BF16ConversionSRInsts] in {
+defm V_CVT_SR_F16_F32 : VOP3OpSel_Real_gfx9 <0x2a6>;
+defm V_CVT_SR_BF16_F32: VOP3OpSel_Real_gfx9 <0x2a7>;
+}
+
 defm V_ASHR_PK_I8_I32 : VOP3OpSel_Real_gfx9 <0x265>;
 defm V_ASHR_PK_U8_I32 : VOP3OpSel_Real_gfx9 <0x266>;
 
diff --git a/llvm/test/MC/AMDGPU/gfx950_asm_features.s b/llvm/test/MC/AMDGPU/gfx950_asm_features.s
@@ -1375,3 +1375,35 @@ v_cvt_scalef32_sr_pk_fp4_f32 v0, |v[2:3]|, v4, v5
 // NOT-GFX950: error: instruction not supported on this GPU
 // GFX950: v_cvt_scalef32_sr_pk_fp4_f32 v0, v[2:3], v4, |v5| ; encoding: [0x00,0x04,0x3e,0xd2,0x02,0x09,0x16,0x04]
 v_cvt_scalef32_sr_pk_fp4_f32 v0, v[2:3], v4, |v5|
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_f16_f32 v0, v1, v2             ; encoding: [0x00,0x00,0xa6,0xd2,0x01,0x05,0x02,0x00]
+v_cvt_sr_f16_f32 v0, v1, v2
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_bf16_f32 v0, v1, v2            ; encoding: [0x00,0x00,0xa7,0xd2,0x01,0x05,0x02,0x00]
+v_cvt_sr_bf16_f32 v0, v1, v2
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_f16_f32 v0, v1, v2 op_sel:[0,0,1] ; encoding: [0x00,0x40,0xa6,0xd2,0x01,0x05,0x02,0x00]
+v_cvt_sr_f16_f32 v0, v1, v2 op_sel:[0,0,1]
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_bf16_f32 v0, v1, v2 op_sel:[0,0,1] ; encoding: [0x00,0x40,0xa7,0xd2,0x01,0x05,0x02,0x00]
+v_cvt_sr_bf16_f32 v0, v1, v2 op_sel:[0,0,1]
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_f16_f32 v0, -v1, v2            ; encoding: [0x00,0x00,0xa6,0xd2,0x01,0x05,0x02,0x20]
+v_cvt_sr_f16_f32 v0, -v1, v2
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_f16_f32 v0, |v1|, v2           ; encoding: [0x00,0x01,0xa6,0xd2,0x01,0x05,0x02,0x00]
+v_cvt_sr_f16_f32 v0, |v1|, v2
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_bf16_f32 v0, -v1, v2           ; encoding: [0x00,0x00,0xa7,0xd2,0x01,0x05,0x02,0x20]
+v_cvt_sr_bf16_f32 v0, -v1, v2
+
+// NOT-GFX950: :[[@LINE+2]]:{{[0-9]+}}: error:
+// GFX950: v_cvt_sr_bf16_f32 v0, |v1|, v2          ; encoding: [0x00,0x01,0xa7,0xd2,0x01,0x05,0x02,0x00]
+v_cvt_sr_bf16_f32 v0, |v1|, v2
diff --git a/llvm/test/MC/AMDGPU/gfx950_err.s b/llvm/test/MC/AMDGPU/gfx950_err.s
@@ -392,3 +392,9 @@ v_pk_minimum3_f16 v0, s1, s2, v3
 
 // GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: invalid operand (violates constant bus restrictions)
 v_pk_maximum3_f16 v0, s1, s2, v3
+
+// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: invalid operand for instruction
+v_cvt_sr_f16_f32 v1, v2, v3 clamp
+
+// GFX950: :[[@LINE+1]]:{{[0-9]+}}: error: invalid operand for instruction
+v_cvt_sr_bf16_f32 v1, v2, v3 clamp
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx950_dasm_vop3.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx950_dasm_vop3.txt
@@ -1014,3 +1014,27 @@
 
 # GFX950: v_cvt_scalef32_sr_pk_fp4_f32 v0, v[2:3], v4, |v5| ; encoding: [0x00,0x04,0x3e,0xd2,0x02,0x09,0x16,0x04]
 0x00,0x04,0x3e,0xd2,0x02,0x09,0x16,0x04
+
+# GFX950: v_cvt_sr_f16_f32 v0, v1, v2             ; encoding: [0x00,0x00,0xa6,0xd2,0x01,0x05,0x02,0x00]
+0x00,0x00,0xa6,0xd2,0x01,0x05,0x02,0x00
+
+# GFX950: v_cvt_sr_bf16_f32 v0, v1, v2            ; encoding: [0x00,0x00,0xa7,0xd2,0x01,0x05,0x02,0x00]
+0x00,0x00,0xa7,0xd2,0x01,0x05,0x02,0x00
+
+# GFX950: v_cvt_sr_f16_f32 v0, v1, v2 op_sel:[0,0,1] ; encoding: [0x00,0x40,0xa6,0xd2,0x01,0x05,0x02,0x00]
+0x00,0x40,0xa6,0xd2,0x01,0x05,0x02,0x00
+
+# GFX950: v_cvt_sr_bf16_f32 v0, v1, v2 op_sel:[0,0,1] ; encoding: [0x00,0x40,0xa7,0xd2,0x01,0x05,0x02,0x00]
+0x00,0x40,0xa7,0xd2,0x01,0x05,0x02,0x00
+
+# GFX950: v_cvt_sr_f16_f32 v0, -v1, v2            ; encoding: [0x00,0x00,0xa6,0xd2,0x01,0x05,0x02,0x20]
+0x00,0x00,0xa6,0xd2,0x01,0x05,0x02,0x20
+
+# GFX950: v_cvt_sr_f16_f32 v0, |v1|, v2           ; encoding: [0x00,0x01,0xa6,0xd2,0x01,0x05,0x02,0x00]
+0x00,0x01,0xa6,0xd2,0x01,0x05,0x02,0x00
+
+# GFX950: v_cvt_sr_bf16_f32 v0, -v1, v2           ; encoding: [0x00,0x00,0xa7,0xd2,0x01,0x05,0x02,0x20]
+0x00,0x00,0xa7,0xd2,0x01,0x05,0x02,0x20
+
+# GFX950: v_cvt_sr_bf16_f32 v0, |v1|, v2          ; encoding: [0x00,0x01,0xa7,0xd2,0x01,0x05,0x02,0x00]
+0x00,0x01,0xa7,0xd2,0x01,0x05,0x02,0x00
