; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx950 < %s | FileCheck -check-prefix=GFX950 %s
; RUN: llc -global-isel=1 -global-isel-abort=2 -mtriple=amdgcn -mcpu=gfx950 < %s | FileCheck -check-prefix=GFX950 %s

declare i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f16(i32 %old, <2 x half> %src, i32 %seed, float %scale, i32 %dst_sel)
declare i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.bf16(i32 %old, <2 x bfloat> %src, i32 %seed, float %scale, i32 %dst_sel)
declare i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f32(i32 %old, <2 x float> %src, i32 %seed, float %scale, i32 %dst_sel)

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f16_dst_sel_0(ptr addrspace(1) %out, <2 x half> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f16_dst_sel_0:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f16 v6, v2, v3, v4
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f16(i32 %old, <2 x half> %src, i32 %seed, float %scale, i32 0)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f16_dst_sel_1(ptr addrspace(1) %out, <2 x half> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f16_dst_sel_1:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f16 v6, v2, v3, v4 op_sel:[0,0,1,0]
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f16(i32 %old, <2 x half> %src, i32 %seed, float %scale, i32 1)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f16_dst_sel_2(ptr addrspace(1) %out, <2 x half> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f16_dst_sel_2:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f16 v6, v2, v3, v4 op_sel:[0,0,0,1]
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f16(i32 %old, <2 x half> %src, i32 %seed, float %scale, i32 2)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f16_dst_sel_3(ptr addrspace(1) %out, <2 x half> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f16_dst_sel_3:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f16 v6, v2, v3, v4 op_sel:[0,0,1,1]
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f16(i32 %old, <2 x half> %src, i32 %seed, float %scale, i32 3)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_bf16_dst_sel_0(ptr addrspace(1) %out, <2 x bfloat> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_bf16_dst_sel_0:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_bf16 v6, v2, v3, v4
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.bf16(i32 %old, <2 x bfloat> %src, i32 %seed, float %scale, i32 0)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_bf16_dst_sel_1(ptr addrspace(1) %out, <2 x bfloat> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_bf16_dst_sel_1:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_bf16 v6, v2, v3, v4 op_sel:[0,0,1,0]
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.bf16(i32 %old, <2 x bfloat> %src, i32 %seed, float %scale, i32 1)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_bf16_dst_sel_2(ptr addrspace(1) %out, <2 x bfloat> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_bf16_dst_sel_2:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_bf16 v6, v2, v3, v4 op_sel:[0,0,0,1]
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.bf16(i32 %old, <2 x bfloat> %src, i32 %seed, float %scale, i32 2)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_bf16_dst_sel_3(ptr addrspace(1) %out, <2 x bfloat> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_bf16_dst_sel_3:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v5, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_bf16 v6, v2, v3, v4 op_sel:[0,0,1,1]
; GFX950-NEXT:    global_store_dword v[0:1], v6, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.bf16(i32 %old, <2 x bfloat> %src, i32 %seed, float %scale, i32 3)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f32_dst_sel_0(ptr addrspace(1) %out, <2 x float> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f32_dst_sel_0:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v6, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f32 v7, v[2:3], v4, v5
; GFX950-NEXT:    global_store_dword v[0:1], v7, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f32(i32 %old, <2 x float> %src, i32 %seed, float %scale, i32 0)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f32_dst_sel_1(ptr addrspace(1) %out, <2 x float> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f32_dst_sel_1:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v6, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f32 v7, v[2:3], v4, v5 op_sel:[0,0,1,0]
; GFX950-NEXT:    global_store_dword v[0:1], v7, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f32(i32 %old, <2 x float> %src, i32 %seed, float %scale, i32 1)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f32_dst_sel_2(ptr addrspace(1) %out, <2 x float> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f32_dst_sel_2:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v6, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f32 v7, v[2:3], v4, v5 op_sel:[0,0,0,1]
; GFX950-NEXT:    global_store_dword v[0:1], v7, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f32(i32 %old, <2 x float> %src, i32 %seed, float %scale, i32 2)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}

define amdgpu_ps void @test_scalef32_sr_pk_fp4_f32_dst_sel_3(ptr addrspace(1) %out, <2 x float> %src, i32 %seed, float %scale) {
; GFX950-LABEL: test_scalef32_sr_pk_fp4_f32_dst_sel_3:
; GFX950:       ; %bb.0:
; GFX950-NEXT:    global_load_dword v6, v[0:1], off
; GFX950-NEXT:    s_waitcnt vmcnt(0)
; GFX950-NEXT:    v_cvt_scalef32_sr_pk_fp4_f32 v7, v[2:3], v4, v5 op_sel:[0,0,1,1]
; GFX950-NEXT:    global_store_dword v[0:1], v7, off
; GFX950-NEXT:    s_endpgm
  %old = load i32, ptr addrspace(1) %out, align 4
  %cvt = tail call i32 @llvm.amdgcn.cvt.scalef32.sr.pk.fp4.f32(i32 %old, <2 x float> %src, i32 %seed, float %scale, i32 3)
  store i32 %cvt, ptr addrspace(1) %out, align 4
  ret void
}