AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory #95393

arsenm · 2024-06-13T11:22:30Z

Make this legal for gfx940 and gfx12

arsenm · 2024-06-13T11:22:48Z

This stack of pull requests is managed by Graphite. Learn more about stacking.

Join @arsenm and the rest of your teammates on Graphite

llvmbot · 2024-06-13T11:23:21Z

@llvm/pr-subscribers-llvm-transforms
@llvm/pr-subscribers-llvm-globalisel

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

Make this legal for gfx940 and gfx12

Patch is 49.12 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/95393.diff

7 Files Affected:

(modified) llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp (+4-1)
(modified) llvm/lib/Target/AMDGPU/DSInstructions.td (+6)
(modified) llvm/lib/Target/AMDGPU/SIISelLowering.cpp (+13-1)
(modified) llvm/test/CodeGen/AMDGPU/GlobalISel/fp-atomics-gfx940.ll (+2-29)
(modified) llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-atomicrmw.ll (+2-25)
(modified) llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll (+8-198)
(modified) llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fadd.ll (+332-60)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index ee7fb20c23aa7..0234cd9088ae7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -301,6 +301,9 @@ static const LLT V10S16 = LLT::fixed_vector(10, 16);
 static const LLT V12S16 = LLT::fixed_vector(12, 16);
 static const LLT V16S16 = LLT::fixed_vector(16, 16);
 
+static const LLT V2F16 = LLT::fixed_vector(2, LLT::float16());
+static const LLT V2BF16 = V2F16; // FIXME
+
 static const LLT V2S32 = LLT::fixed_vector(2, 32);
 static const LLT V3S32 = LLT::fixed_vector(3, 32);
 static const LLT V4S32 = LLT::fixed_vector(4, 32);
@@ -1638,7 +1641,7 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
     if (ST.hasLdsAtomicAddF64())
       Atomic.legalFor({{S64, LocalPtr}});
     if (ST.hasAtomicDsPkAdd16Insts())
-      Atomic.legalFor({{V2S16, LocalPtr}});
+      Atomic.legalFor({{V2F16, LocalPtr}, {V2BF16, LocalPtr}});
   }
   if (ST.hasAtomicFaddInsts())
     Atomic.legalFor({{S32, GlobalPtr}});
diff --git a/llvm/lib/Target/AMDGPU/DSInstructions.td b/llvm/lib/Target/AMDGPU/DSInstructions.td
index 5150641c50ea2..e22f8b7d46171 100644
--- a/llvm/lib/Target/AMDGPU/DSInstructions.td
+++ b/llvm/lib/Target/AMDGPU/DSInstructions.td
@@ -1082,6 +1082,12 @@ defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_U32, DS_MAX_U32, i32, "atomic_load_umax
 defm : DSAtomicRetNoRetPat_mc<DS_MIN_RTN_F32, DS_MIN_F32, f32, "atomic_load_fmin">;
 defm : DSAtomicRetNoRetPat_mc<DS_MAX_RTN_F32, DS_MAX_F32, f32, "atomic_load_fmax">;
 
+
+let SubtargetPredicate = HasAtomicDsPkAdd16Insts in {
+defm : DSAtomicRetNoRetPat_mc<DS_PK_ADD_RTN_F16, DS_PK_ADD_F16, v2f16, "atomic_load_fadd">;
+defm : DSAtomicRetNoRetPat_mc<DS_PK_ADD_RTN_BF16, DS_PK_ADD_BF16, v2bf16, "atomic_load_fadd">;
+}
+
 let SubtargetPredicate = isGFX6GFX7GFX8GFX9GFX10 in {
 defm : DSAtomicCmpXChgSwapped_mc<DS_CMPST_RTN_B32, DS_CMPST_B32, i32, "atomic_cmp_swap">;
 }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 81098201e9c0f..6325e75aa8f96 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -15931,6 +15931,16 @@ static OptimizationRemark emitAtomicRMWLegalRemark(const AtomicRMWInst *RMW) {
          << " operation at memory scope " << MemScope;
 }
 
+static bool isHalf2OrBFloat2(Type *Ty) {
+  if (FixedVectorType *VT = dyn_cast<FixedVectorType>(Ty)) {
+    Type *EltTy = VT->getElementType();
+    return VT->getNumElements() == 2 &&
+           (EltTy->isHalfTy() || EltTy->isBFloatTy());
+  }
+
+  return false;
+}
+
 TargetLowering::AtomicExpansionKind
 SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
   unsigned AS = RMW->getPointerAddressSpace();
@@ -15989,7 +15999,9 @@ SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
                                                  : AtomicExpansionKind::CmpXChg;
       }
 
-      // TODO: Handle v2f16/v2bf16 cases for gfx940
+      if (Subtarget->hasAtomicDsPkAdd16Insts() && isHalf2OrBFloat2(Ty))
+        return AtomicExpansionKind::None;
+
       return AtomicExpansionKind::CmpXChg;
     }
 
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/fp-atomics-gfx940.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/fp-atomics-gfx940.ll
index 93c30a6a01e00..4e21ef8379342 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/fp-atomics-gfx940.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/fp-atomics-gfx940.ll
@@ -213,22 +213,8 @@ define <2 x half> @local_atomic_fadd_ret_v2f16_offset(ptr addrspace(3) %ptr, <2
 ; GFX940-LABEL: local_atomic_fadd_ret_v2f16_offset:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v2, v0 offset:65532
-; GFX940-NEXT:    s_mov_b64 s[0:1], 0
-; GFX940-NEXT:  .LBB15_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    ds_pk_add_rtn_f16 v0, v0, v1 offset:65532
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v3, v2
-; GFX940-NEXT:    v_pk_add_f16 v2, v3, v1
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    s_cbranch_execnz .LBB15_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
   %result = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst
@@ -239,21 +225,8 @@ define void @local_atomic_fadd_noret_v2f16_offset(ptr addrspace(3) %ptr, <2 x ha
 ; GFX940-LABEL: local_atomic_fadd_noret_v2f16_offset:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v2, v0 offset:65532
-; GFX940-NEXT:    s_mov_b64 s[0:1], 0
-; GFX940-NEXT:  .LBB16_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    ds_pk_add_f16 v0, v1 offset:65532
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_pk_add_f16 v3, v2, v1
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT:    v_mov_b32_e32 v2, v3
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    s_cbranch_execnz .LBB16_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
   %unused = atomicrmw fadd ptr addrspace(3) %gep, <2 x half> %val seq_cst
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-atomicrmw.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-atomicrmw.ll
index 5724cf471bae3..d4a7f3b2d387d 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-atomicrmw.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/irtranslator-atomicrmw.ll
@@ -52,36 +52,13 @@ define float @test_atomicrmw_fsub(ptr addrspace(3) %addr) {
 define <2 x half> @test_atomicrmw_fadd_vector(ptr addrspace(3) %addr) {
   ; CHECK-LABEL: name: test_atomicrmw_fadd_vector
   ; CHECK: bb.1 (%ir-block.0):
-  ; CHECK-NEXT:   successors: %bb.2(0x80000000)
   ; CHECK-NEXT:   liveins: $vgpr0
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0
   ; CHECK-NEXT:   [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3C00
   ; CHECK-NEXT:   [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR [[C]](s16), [[C]](s16)
-  ; CHECK-NEXT:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
-  ; CHECK-NEXT:   [[LOAD:%[0-9]+]]:_(<2 x s16>) = G_LOAD [[COPY]](p3) :: (load (<2 x s16>) from %ir.addr, addrspace 3)
-  ; CHECK-NEXT:   G_BR %bb.2
-  ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT: bb.2.atomicrmw.start:
-  ; CHECK-NEXT:   successors: %bb.3(0x40000000), %bb.2(0x40000000)
-  ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT:   [[PHI:%[0-9]+]]:_(s64) = G_PHI %19(s64), %bb.2, [[C1]](s64), %bb.1
-  ; CHECK-NEXT:   [[PHI1:%[0-9]+]]:_(<2 x s16>) = G_PHI [[LOAD]](<2 x s16>), %bb.1, %18(<2 x s16>), %bb.2
-  ; CHECK-NEXT:   [[FADD:%[0-9]+]]:_(<2 x s16>) = G_FADD [[PHI1]], [[BUILD_VECTOR]]
-  ; CHECK-NEXT:   [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[FADD]](<2 x s16>)
-  ; CHECK-NEXT:   [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[PHI1]](<2 x s16>)
-  ; CHECK-NEXT:   [[ATOMIC_CMPXCHG_WITH_SUCCESS:%[0-9]+]]:_(s32), [[ATOMIC_CMPXCHG_WITH_SUCCESS1:%[0-9]+]]:_(s1) = G_ATOMIC_CMPXCHG_WITH_SUCCESS [[COPY]](p3), [[BITCAST1]], [[BITCAST]] :: (load store seq_cst seq_cst (s32) on %ir.addr, addrspace 3)
-  ; CHECK-NEXT:   [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[ATOMIC_CMPXCHG_WITH_SUCCESS]](s32)
-  ; CHECK-NEXT:   [[INT:%[0-9]+]]:_(s64) = G_INTRINSIC intrinsic(@llvm.amdgcn.if.break), [[ATOMIC_CMPXCHG_WITH_SUCCESS1]](s1), [[PHI]](s64)
-  ; CHECK-NEXT:   [[INT1:%[0-9]+]]:_(s1) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.loop), [[INT]](s64)
-  ; CHECK-NEXT:   G_BRCOND [[INT1]](s1), %bb.3
-  ; CHECK-NEXT:   G_BR %bb.2
-  ; CHECK-NEXT: {{  $}}
-  ; CHECK-NEXT: bb.3.atomicrmw.end:
-  ; CHECK-NEXT:   [[PHI2:%[0-9]+]]:_(<2 x s16>) = G_PHI [[BITCAST2]](<2 x s16>), %bb.2
-  ; CHECK-NEXT:   [[PHI3:%[0-9]+]]:_(s64) = G_PHI [[INT]](s64), %bb.2
-  ; CHECK-NEXT:   G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.end.cf), [[PHI3]](s64)
-  ; CHECK-NEXT:   $vgpr0 = COPY [[PHI2]](<2 x s16>)
+  ; CHECK-NEXT:   [[ATOMICRMW_FADD:%[0-9]+]]:_(<2 x s16>) = G_ATOMICRMW_FADD [[COPY]](p3), [[BUILD_VECTOR]] :: (load store seq_cst (<2 x s16>) on %ir.addr, addrspace 3)
+  ; CHECK-NEXT:   $vgpr0 = COPY [[ATOMICRMW_FADD]](<2 x s16>)
   ; CHECK-NEXT:   SI_RETURN implicit $vgpr0
   %oldval = atomicrmw fadd ptr addrspace(3) %addr, <2 x half> <half 1.0, half 1.0> seq_cst
   ret <2 x half> %oldval
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll b/llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll
index 4373b76070e32..e34a33ebdbb04 100644
--- a/llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll
+++ b/llvm/test/CodeGen/AMDGPU/local-atomics-fp.ll
@@ -3030,22 +3030,8 @@ define <2 x half> @lds_atomic_fadd_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %
 ; GFX940-LABEL: lds_atomic_fadd_ret_v2f16:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v2, v0
-; GFX940-NEXT:    s_mov_b64 s[0:1], 0
-; GFX940-NEXT:  .LBB14_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v3, v2
-; GFX940-NEXT:    v_pk_add_f16 v2, v3, v1
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2
+; GFX940-NEXT:    ds_pk_add_rtn_f16 v0, v0, v1
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
-; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    s_cbranch_execnz .LBB14_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX12-LABEL: lds_atomic_fadd_ret_v2f16:
@@ -3055,26 +3041,10 @@ define <2 x half> @lds_atomic_fadd_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %
 ; GFX12-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    ds_load_b32 v2, v0
-; GFX12-NEXT:    s_mov_b32 s0, 0
-; GFX12-NEXT:  .LBB14_1: ; %atomicrmw.start
-; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT:    s_wait_dscnt 0x0
-; GFX12-NEXT:    v_mov_b32_e32 v3, v2
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX12-NEXT:    v_pk_add_f16 v2, v3, v1
 ; GFX12-NEXT:    s_wait_storecnt 0x0
-; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v3
+; GFX12-NEXT:    ds_pk_add_rtn_f16 v0, v0, v1
 ; GFX12-NEXT:    s_wait_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_SE
-; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
-; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT:    s_cbranch_execnz .LBB14_1
-; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX12-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-LABEL: lds_atomic_fadd_ret_v2f16:
@@ -3231,21 +3201,8 @@ define void @lds_atomic_fadd_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val)
 ; GFX940-LABEL: lds_atomic_fadd_noret_v2f16:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v2, v0
-; GFX940-NEXT:    s_mov_b64 s[0:1], 0
-; GFX940-NEXT:  .LBB15_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    ds_pk_add_f16 v0, v1
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_pk_add_f16 v3, v2, v1
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
-; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
-; GFX940-NEXT:    v_mov_b32_e32 v2, v3
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
-; GFX940-NEXT:    s_cbranch_execnz .LBB15_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX12-LABEL: lds_atomic_fadd_noret_v2f16:
@@ -3255,24 +3212,10 @@ define void @lds_atomic_fadd_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val)
 ; GFX12-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    ds_load_b32 v2, v0
-; GFX12-NEXT:    s_mov_b32 s0, 0
-; GFX12-NEXT:  .LBB15_1: ; %atomicrmw.start
-; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT:    s_wait_dscnt 0x0
-; GFX12-NEXT:    v_pk_add_f16 v3, v2, v1
 ; GFX12-NEXT:    s_wait_storecnt 0x0
-; GFX12-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v2
+; GFX12-NEXT:    ds_pk_add_f16 v0, v1
 ; GFX12-NEXT:    s_wait_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_SE
-; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v2
-; GFX12-NEXT:    v_mov_b32_e32 v2, v3
-; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
-; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
-; GFX12-NEXT:    s_cbranch_execnz .LBB15_1
-; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
 ; GFX12-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-LABEL: lds_atomic_fadd_noret_v2f16:
@@ -3483,41 +3426,8 @@ define <2 x bfloat> @lds_atomic_fadd_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bflo
 ; GFX940-LABEL: lds_atomic_fadd_ret_v2bf16:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v2, v0
-; GFX940-NEXT:    s_mov_b64 s[2:3], 0
-; GFX940-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
-; GFX940-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX940-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; GFX940-NEXT:    s_mov_b32 s5, 0x7060302
-; GFX940-NEXT:  .LBB16_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v4, v2
-; GFX940-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
-; GFX940-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
-; GFX940-NEXT:    v_add_f32_e32 v2, v2, v3
-; GFX940-NEXT:    v_add_f32_e32 v5, v5, v1
-; GFX940-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX940-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX940-NEXT:    v_or_b32_e32 v7, 0x400000, v2
-; GFX940-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX940-NEXT:    v_add3_u32 v6, v6, v2, s4
-; GFX940-NEXT:    v_add3_u32 v8, v8, v5, s4
-; GFX940-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT:    v_cmp_u_f32_e64 s[0:1], v2, v2
-; GFX940-NEXT:    s_nop 0
-; GFX940-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX940-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[0:1]
-; GFX940-NEXT:    v_perm_b32 v2, v5, v2, s5
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2
+; GFX940-NEXT:    ds_pk_add_rtn_bf16 v0, v0, v1
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
-; GFX940-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT:    s_cbranch_execnz .LBB16_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[2:3]
-; GFX940-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX12-LABEL: lds_atomic_fadd_ret_v2bf16:
@@ -3527,45 +3437,10 @@ define <2 x bfloat> @lds_atomic_fadd_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bflo
 ; GFX12-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    ds_load_b32 v2, v0
-; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
-; GFX12-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; GFX12-NEXT:    s_mov_b32 s1, 0
-; GFX12-NEXT:  .LBB16_1: ; %atomicrmw.start
-; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX12-NEXT:    s_wait_dscnt 0x0
-; GFX12-NEXT:    v_mov_b32_e32 v4, v2
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
-; GFX12-NEXT:    v_add_f32_e32 v5, v5, v1
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-NEXT:    v_bfe_u32 v7, v5, 16, 1
-; GFX12-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX12-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
-; GFX12-NEXT:    v_add_f32_e32 v2, v2, v3
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX12-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX12-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX12-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
-; GFX12-NEXT:    v_add3_u32 v6, v6, v2, 0x7fff
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_cndmask_b32_e64 v2, v6, v8, s0
-; GFX12-NEXT:    v_perm_b32 v2, v5, v2, 0x7060302
 ; GFX12-NEXT:    s_wait_storecnt 0x0
-; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v4
+; GFX12-NEXT:    ds_pk_add_rtn_bf16 v0, v0, v1
 ; GFX12-NEXT:    s_wait_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_SE
-; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
-; GFX12-NEXT:    s_or_b32 s1, vcc_lo, s1
-; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT:    s_cbranch_execnz .LBB16_1
-; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s1
-; GFX12-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX12-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-LABEL: lds_atomic_fadd_ret_v2bf16:
@@ -3769,40 +3644,8 @@ define void @lds_atomic_fadd_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %v
 ; GFX940-LABEL: lds_atomic_fadd_noret_v2bf16:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    ds_read_b32 v3, v0
-; GFX940-NEXT:    s_mov_b64 s[2:3], 0
-; GFX940-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
-; GFX940-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX940-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; GFX940-NEXT:    s_mov_b32 s5, 0x7060302
-; GFX940-NEXT:  .LBB17_1: ; %atomicrmw.start
-; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
-; GFX940-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
-; GFX940-NEXT:    v_add_f32_e32 v4, v4, v2
-; GFX940-NEXT:    v_add_f32_e32 v5, v5, v1
-; GFX940-NEXT:    v_bfe_u32 v6, v4, 16, 1
-; GFX940-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX940-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX940-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX940-NEXT:    v_add3_u32 v6, v6, v4, s4
-; GFX940-NEXT:    v_add3_u32 v8, v8, v5, s4
-; GFX940-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
-; GFX940-NEXT:    v_cmp_u_f32_e64 s[0:1], v4, v4
-; GFX940-NEXT:    s_nop 0
-; GFX940-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
-; GFX940-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[0:1]
-; GFX940-NEXT:    v_perm_b32 v4, v5, v4, s5
-; GFX940-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
+; GFX940-NEXT:    ds_pk_add_bf16 v0, v1
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
-; GFX940-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
-; GFX940-NEXT:    v_mov_b32_e32 v3, v4
-; GFX940-NEXT:    s_andn2_b64 exec, exec, s[2:3]
-; GFX940-NEXT:    s_cbranch_execnz .LBB17_1
-; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX940-NEXT:    s_or_b64 exec, exec, s[2:3]
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX12-LABEL: lds_atomic_fadd_noret_v2bf16:
@@ -3812,43 +3655,10 @@ define void @lds_atomic_fadd_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %v
 ; GFX12-NEXT:    s_wait_samplecnt 0x0
 ; GFX12-NEXT:    s_wait_bvhcnt 0x0
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    ds_load_b32 v3, v0
-; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
-; GFX12-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; GFX12-NEXT:    s_mov_b32 s1, 0
-; GFX12-NEXT:  .LBB17_1: ; %atomicrmw.start
-;...
[truncated]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

shiltian

LGTM with nits

arsenm · 2024-06-15T07:51:25Z

Merge activity

Jun 15, 3:51 AM EDT: @arsenm started a stack merge that includes this pull request via Graphite.
Jun 15, 3:53 AM EDT: Graphite rebased this pull request as part of a merge.
Jun 15, 3:55 AM EDT: @arsenm merged this pull request with Graphite.

Make this legal for gfx940 and gfx12

As far as AMDGPU target supports vectorization for atomic_rmw operation, allow construction of LLVM_AtomicRMWOp with 16 bit floating point values. See also: llvm#94845, llvm#95393, llvm#95394 Signed-off-by: Ilya Veselov <iveselov.nn@gmail.com>

As far as AMDGPU target supports vectorization for atomic_rmw operation, allow construction of LLVM_AtomicRMWOp with 16 bit floating point values. This patch enables building of LLVM_AtomicRMWOp with fixed vectors of 16 bit fp values as operands. See also: llvm#94845, llvm#95393, llvm#95394 Signed-off-by: Ilya Veselov <iveselov.nn@gmail.com>

As far as AMDGPU target supports vectorization for `atomic_rmw fadd` operation, enable building of `LLVM_AtomicRMWOp fadd` with fixed vectors of 16 bit fp values as operands. See also: llvm#94845, llvm#95393, llvm#95394 Signed-off-by: Ilya Veselov <iveselov.nn@gmail.com>

This was referenced Jun 13, 2024

AMDGPU: Handle legal v2f16/v2bf16 atomicrmw fadd for global/flat #95394

Merged

clang/AMDGPU: Emit atomicrmw from ds_fadd builtins #95395

Merged

AMDGPU: Remove ds atomic fadd intrinsics #95396

Merged

arsenm added the backend:AMDGPU label Jun 13, 2024 — with Graphite App

arsenm requested review from rampitec, Sisyph, giuseros, shiltian and kuhar June 13, 2024 11:25

arsenm marked this pull request as ready for review June 13, 2024 11:29

llvmbot added llvm:globalisel llvm:transforms labels Jun 13, 2024

arsenm mentioned this pull request Jun 13, 2024

[ROCDL] Add the global.atomic.fadd intrinsic in ROCDL #94486

Closed

shiltian reviewed Jun 13, 2024

View reviewed changes

llvm/lib/Target/AMDGPU/SIISelLowering.cpp Outdated Show resolved Hide resolved

shiltian reviewed Jun 13, 2024

View reviewed changes

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp Show resolved Hide resolved

shiltian approved these changes Jun 13, 2024

View reviewed changes

arsenm force-pushed the users/arsenm/amdgpu-legalize-vector-atomicrmw-fadd-local branch 2 times, most recently from ba80b77 to 76cbecc Compare June 14, 2024 19:42

AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory

122eec1

Make this legal for gfx940 and gfx12

arsenm force-pushed the users/arsenm/amdgpu-legalize-vector-atomicrmw-fadd-local branch from 76cbecc to 122eec1 Compare June 15, 2024 07:53

arsenm merged commit 0a9a5f9 into main Jun 15, 2024
3 of 5 checks passed

arsenm deleted the users/arsenm/amdgpu-legalize-vector-atomicrmw-fadd-local branch June 15, 2024 07:55

jpienaar mentioned this pull request Jun 17, 2024

[mlirc] Add missing extern C #95829

Merged

joviliast mentioned this pull request Sep 30, 2024

Allow fixed vector operand for LLVM_AtomicRMWOp #110553

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory #95393

AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory #95393

arsenm commented Jun 13, 2024

arsenm commented Jun 13, 2024 •

edited

Loading

llvmbot commented Jun 13, 2024 •

edited

Loading

shiltian left a comment

arsenm commented Jun 15, 2024 •

edited

Loading

AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory #95393

AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory #95393

Conversation

arsenm commented Jun 13, 2024

arsenm commented Jun 13, 2024 • edited Loading

llvmbot commented Jun 13, 2024 • edited Loading

shiltian left a comment

Choose a reason for hiding this comment

arsenm commented Jun 15, 2024 • edited Loading

Merge activity

arsenm commented Jun 13, 2024 •

edited

Loading

llvmbot commented Jun 13, 2024 •

edited

Loading

arsenm commented Jun 15, 2024 •

edited

Loading