;=========================== begin_copyright_notice ============================ ; ; Copyright (C) 2025 Intel Corporation ; ; SPDX-License-Identifier: MIT ; ;============================ end_copyright_notice ============================= ; REQUIRES: llvm-spirv, cri-supported ; RUN: llvm-as %s -o %t.bc ; RUN: llvm-spirv %t.bc -o %t.spv --spirv-ext=+SPV_INTEL_float4,+SPV_KHR_bfloat16,+SPV_INTEL_int4,+SPV_INTEL_fp_conversions ; RUN: ocloc compile -spirv_input -file %t.spv -device cri -options "-igc_opts 'ForceOCLSIMDWidth=32,DumpVISAASMToConsole=1,AddVISADumpDeclarationsToEnd=1'" | FileCheck %s ; Test if conversion opcodes are present in spirv disassembly ; RUN: llvm-spirv --to-text %t.spv -o %t.spt ; RUN: cat %t.spt | FileCheck %s -check-prefix=CHECK-SPV ; CHECK-SPV: StochasticRoundFToFINTEL target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024" target triple = "spir64-unknown-unknown" declare spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32) ; ; Variants without pointers ; define spir_kernel void @BF16_to_E2M1_scalar(bfloat addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_scalar" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK-DAG: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 ; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud ; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr bfloat, bfloat addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid %inputVal = load bfloat, bfloat addrspace(1)* %inputAddr, align 2 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call i4 @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDF16bi(bfloat %inputVal, i32 %seedVal) %v1 = insertelement <2 x i4> undef, i4 %conv, i32 0 %v2 = insertelement <2 x i4> %v1, i4 0, i32 1 store <2 x i4> %v2, <2 x i4> addrspace(1)* %outputAddr, align 1 ret void } declare dso_local spir_func i4 @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDF16bi(bfloat, i32) define spir_kernel void @BF16_to_E2M1_vector2(<2 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_vector2" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK-DAG: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 ; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud ; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <2 x bfloat>, <2 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <2 x bfloat>, <2 x bfloat> addrspace(1)* %inputAddr, align 4 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <2 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv2_DF16bi(<2 x bfloat> %inputVal, i32 %seedVal) store <2 x i4> %conv, <2 x i4> addrspace(1)* %outputAddr, align 1 ret void } declare dso_local spir_func <2 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv2_DF16bi(<2 x bfloat>, i32) define spir_kernel void @BF16_to_E2M1_vector3(<3 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_vector3" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <3 x bfloat>, <3 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <3 x bfloat>, <3 x bfloat> addrspace(1)* %inputAddr, align 8 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <3 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv3_DF16bi(<3 x bfloat> %inputVal, i32 %seedVal) %conv4 = shufflevector <3 x i4> %conv, <3 x i4> undef, <4 x i32> store <4 x i4> %conv4, <4 x i4> addrspace(1)* %outputAddr, align 2 ret void } declare dso_local spir_func <3 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv3_DF16bi(<3 x bfloat>, i32) define spir_kernel void @BF16_to_E2M1_vector4(<4 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_vector4" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].128 [[RAND]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=64 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <4 x bfloat>, <4 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <4 x bfloat>, <4 x bfloat> addrspace(1)* %inputAddr, align 8 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <4 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv4_DF16bi(<4 x bfloat> %inputVal, i32 %seedVal) store <4 x i4> %conv, <4 x i4> addrspace(1)* %outputAddr, align 2 ret void } declare dso_local spir_func <4 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv4_DF16bi(<4 x bfloat>, i32) define spir_kernel void @BF16_to_E2M1_vector8(<8 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <8 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_vector8" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 ; CHECK: dnscl.bftoe2m1.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=128 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <8 x bfloat>, <8 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <8 x i4>, <8 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <8 x bfloat>, <8 x bfloat> addrspace(1)* %inputAddr, align 16 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <8 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv8_DF16bi(<8 x bfloat> %inputVal, i32 %seedVal) store <8 x i4> %conv, <8 x i4> addrspace(1)* %outputAddr, align 4 ret void } declare dso_local spir_func <8 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv8_DF16bi(<8 x bfloat>, i32) define spir_kernel void @BF16_to_E2M1_vector16(<16 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <16 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_vector16" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND2:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND3:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 ; CHECK: dnscl.bftoe2m1.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST2:[A-z0-9]*]].0 [[INPUT]].512 [[INPUT]].768 [[RAND2]].0 ; CHECK: dnscl.bftoe2m1.mode2.srnd (M1, 32) [[DST3:[A-z0-9]*]].0 [[INPUT]].640 [[INPUT]].896 [[RAND3]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[RAND2]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[RAND3]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=256 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST2]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST3]] v_type=G type=ud num_elts=32 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <16 x bfloat>, <16 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <16 x i4>, <16 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <16 x bfloat>, <16 x bfloat> addrspace(1)* %inputAddr, align 32 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <16 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv16_DF16bi(<16 x bfloat> %inputVal, i32 %seedVal) store <16 x i4> %conv, <16 x i4> addrspace(1)* %outputAddr, align 8 ret void } declare dso_local spir_func <16 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv16_DF16bi(<16 x bfloat>, i32) ; ; Variants with pointers ; define spir_kernel void @BF16_to_E2M1_ptr_scalar(bfloat addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_ptr_scalar" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK-DAG: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 ; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %seedMem = alloca i32, align 4 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr bfloat, bfloat addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid %inputVal = load bfloat, bfloat addrspace(1)* %inputAddr, align 2 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call i4 @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDF16biPi(bfloat %inputVal, i32 %seedVal, i32* %seedMem) %v1 = insertelement <2 x i4> undef, i4 %conv, i32 0 %v2 = insertelement <2 x i4> %v1, i4 0, i32 1 store <2 x i4> %v2, <2 x i4> addrspace(1)* %outputAddr, align 1 %seedMemVal = load i32, i32* %seedMem, align 4 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 ret void } declare dso_local spir_func i4 @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDF16biPi(bfloat, i32, i32*) define spir_kernel void @BF16_to_E2M1_ptr_vector2(<2 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_ptr_vector2" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK-DAG: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 ; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %seedMem = alloca i32, align 4 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <2 x bfloat>, <2 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <2 x bfloat>, <2 x bfloat> addrspace(1)* %inputAddr, align 4 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <2 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv2_DF16biPi(<2 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) store <2 x i4> %conv, <2 x i4> addrspace(1)* %outputAddr, align 1 %seedMemVal = load i32, i32* %seedMem, align 4 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 ret void } declare dso_local spir_func <2 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv2_DF16biPi(<2 x bfloat>, i32, i32*) define spir_kernel void @BF16_to_E2M1_ptr_vector3(<3 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_ptr_vector3" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %seedMem = alloca i32, align 4 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <3 x bfloat>, <3 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <3 x bfloat>, <3 x bfloat> addrspace(1)* %inputAddr, align 8 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <3 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv3_DF16biPi(<3 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) %conv4 = shufflevector <3 x i4> %conv, <3 x i4> undef, <4 x i32> store <4 x i4> %conv4, <4 x i4> addrspace(1)* %outputAddr, align 2 %seedMemVal = load i32, i32* %seedMem, align 4 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 ret void } declare dso_local spir_func <3 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv3_DF16biPi(<3 x bfloat>, i32, i32*) define spir_kernel void @BF16_to_E2M1_ptr_vector4(<4 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_ptr_vector4" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].128 [[RAND]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=64 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 %seedMem = alloca i32, align 4 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <4 x bfloat>, <4 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <4 x bfloat>, <4 x bfloat> addrspace(1)* %inputAddr, align 8 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <4 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv4_DF16biPi(<4 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) store <4 x i4> %conv, <4 x i4> addrspace(1)* %outputAddr, align 2 %seedMemVal = load i32, i32* %seedMem, align 4 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 ret void } declare dso_local spir_func <4 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv4_DF16biPi(<4 x bfloat>, i32, i32*) define spir_kernel void @BF16_to_E2M1_ptr_vector8(<8 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <8 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_ptr_vector8" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 ; CHECK: dnscl.bftoe2m1.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=128 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 %seedMem = alloca i32, align 4 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <8 x bfloat>, <8 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <8 x i4>, <8 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <8 x bfloat>, <8 x bfloat> addrspace(1)* %inputAddr, align 16 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <8 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv8_DF16biPi(<8 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) store <8 x i4> %conv, <8 x i4> addrspace(1)* %outputAddr, align 4 %seedMemVal = load i32, i32* %seedMem, align 4 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 ret void } declare dso_local spir_func <8 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv8_DF16biPi(<8 x bfloat>, i32, i32*) define spir_kernel void @BF16_to_E2M1_ptr_vector16(<16 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <16 x i4> addrspace(1)* %output) { ; CHECK-LABEL: .kernel "BF16_to_E2M1_ptr_vector16" ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND2:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK: lfsr.b8v4 (M1, 32) [[RAND3:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> ; CHECK-NOT: lfsr.b8v4 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 ; CHECK: dnscl.bftoe2m1.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 ; CHECK: dnscl.bftoe2m1.mode0.srnd (M1, 32) [[DST2:[A-z0-9]*]].0 [[INPUT]].512 [[INPUT]].768 [[RAND2]].0 ; CHECK: dnscl.bftoe2m1.mode2.srnd (M1, 32) [[DST3:[A-z0-9]*]].0 [[INPUT]].640 [[INPUT]].896 [[RAND3]].0 ; CHECK-NOT: dnscl.bftoe2m1 ; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[RAND2]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[RAND3]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=256 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST2]] v_type=G type=ud num_elts=32 ; CHECK-DAG: // .decl [[DST3]] v_type=G type=ud num_elts=32 %seedMem = alloca i32, align 4 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) %inputAddr = getelementptr <16 x bfloat>, <16 x bfloat> addrspace(1)* %input, i64 %gid %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid %outputAddr = getelementptr <16 x i4>, <16 x i4> addrspace(1)* %output, i64 %gid %inputVal = load <16 x bfloat>, <16 x bfloat> addrspace(1)* %inputAddr, align 32 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 %conv = call <16 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv16_DF16biPi(<16 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) store <16 x i4> %conv, <16 x i4> addrspace(1)* %outputAddr, align 8 %seedMemVal = load i32, i32* %seedMem, align 4 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 ret void } declare dso_local spir_func <16 x i4> @_Z46__builtin_spirv_StochasticRoundBF16ToE2M1INTELDv16_DF16biPi(<16 x bfloat>, i32, i32*)