| 1 |
;=========================== begin_copyright_notice ============================ |
| 2 |
; |
| 3 |
; Copyright (C) 2025 Intel Corporation |
| 4 |
; |
| 5 |
; SPDX-License-Identifier: MIT |
| 6 |
; |
| 7 |
;============================ end_copyright_notice ============================= |
| 8 |
|
| 9 |
; REQUIRES: llvm-spirv, cri-supported |
| 10 |
; RUN: llvm-as %s -o %t.bc |
| 11 |
; RUN: llvm-spirv %t.bc -o %t.spv --spirv-ext=+SPV_INTEL_int4,+SPV_KHR_bfloat16,+SPV_INTEL_fp_conversions |
| 12 |
; RUN: ocloc compile -spirv_input -file %t.spv -device cri -options "-igc_opts 'ForceOCLSIMDWidth=32,DumpVISAASMToConsole=1,AddVISADumpDeclarationsToEnd=1'" | FileCheck %s |
| 13 |
|
| 14 |
; Test if conversion opcodes are present in spirv disassembly |
| 15 |
; RUN: llvm-spirv --to-text %t.spv -o %t.spt |
| 16 |
; RUN: cat %t.spt | FileCheck %s -check-prefix=CHECK-SPV |
| 17 |
; CHECK-SPV: StochasticRoundFToSINTEL |
| 18 |
|
| 19 |
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" |
| 20 |
target triple = "spir64-unknown-unknown" |
| 21 |
|
| 22 |
declare spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32) |
| 23 |
|
| 24 |
; |
| 25 |
; Variants without pointers |
| 26 |
; |
| 27 |
|
| 28 |
define spir_kernel void @BF16_to_Int4_scalar(bfloat addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { |
| 29 |
; CHECK-LABEL: .kernel "BF16_to_Int4_scalar" |
| 30 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 31 |
; CHECK-NOT: lfsr.b8v4 |
| 32 |
; CHECK-DAG: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 |
| 33 |
; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud |
| 34 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 35 |
; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 |
| 36 |
; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 |
| 37 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 38 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 39 |
%inputAddr = getelementptr bfloat, bfloat addrspace(1)* %input, i64 %gid |
| 40 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 41 |
%outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid |
| 42 |
|
| 43 |
%inputVal = load bfloat, bfloat addrspace(1)* %inputAddr, align 2 |
| 44 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 45 |
%conv = call i4 @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDF16bi(bfloat %inputVal, i32 %seedVal) |
| 46 |
%v1 = insertelement <2 x i4> undef, i4 %conv, i32 0 |
| 47 |
%v2 = insertelement <2 x i4> %v1, i4 0, i32 1 |
| 48 |
|
| 49 |
store <2 x i4> %v2, <2 x i4> addrspace(1)* %outputAddr, align 1 |
| 50 |
ret void |
| 51 |
} |
| 52 |
declare dso_local spir_func i4 @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDF16bi(bfloat, i32) |
| 53 |
|
| 54 |
define spir_kernel void @BF16_to_Int4_vector2(<2 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { |
| 55 |
; CHECK-LABEL: .kernel "BF16_to_Int4_vector2" |
| 56 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 57 |
; CHECK-NOT: lfsr.b8v4 |
| 58 |
; CHECK-DAG: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 |
| 59 |
; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud |
| 60 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 61 |
; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 |
| 62 |
; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 |
| 63 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 64 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 65 |
%inputAddr = getelementptr <2 x bfloat>, <2 x bfloat> addrspace(1)* %input, i64 %gid |
| 66 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 67 |
%outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid |
| 68 |
|
| 69 |
%inputVal = load <2 x bfloat>, <2 x bfloat> addrspace(1)* %inputAddr, align 4 |
| 70 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 71 |
%conv = call <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv2_DF16bi(<2 x bfloat> %inputVal, i32 %seedVal) |
| 72 |
|
| 73 |
store <2 x i4> %conv, <2 x i4> addrspace(1)* %outputAddr, align 1 |
| 74 |
ret void |
| 75 |
} |
| 76 |
declare dso_local spir_func <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv2_DF16bi(<2 x bfloat>, i32) |
| 77 |
|
| 78 |
define spir_kernel void @BF16_to_Int4_vector3(<3 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { |
| 79 |
; CHECK-LABEL: .kernel "BF16_to_Int4_vector3" |
| 80 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 81 |
; CHECK-NOT: lfsr.b8v4 |
| 82 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 |
| 83 |
; CHECK-NOT: dnscl.bftoint4 |
| 84 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 85 |
; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 |
| 86 |
; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 |
| 87 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 88 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 89 |
%inputAddr = getelementptr <3 x bfloat>, <3 x bfloat> addrspace(1)* %input, i64 %gid |
| 90 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 91 |
%outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid |
| 92 |
|
| 93 |
%inputVal = load <3 x bfloat>, <3 x bfloat> addrspace(1)* %inputAddr, align 8 |
| 94 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 95 |
%conv = call <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv3_DF16bi(<3 x bfloat> %inputVal, i32 %seedVal) |
| 96 |
%conv4 = shufflevector <3 x i4> %conv, <3 x i4> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 undef> |
| 97 |
|
| 98 |
store <4 x i4> %conv4, <4 x i4> addrspace(1)* %outputAddr, align 2 |
| 99 |
ret void |
| 100 |
} |
| 101 |
declare dso_local spir_func <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv3_DF16bi(<3 x bfloat>, i32) |
| 102 |
|
| 103 |
define spir_kernel void @BF16_to_Int4_vector4(<4 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { |
| 104 |
; CHECK-LABEL: .kernel "BF16_to_Int4_vector4" |
| 105 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 106 |
; CHECK-NOT: lfsr.b8v4 |
| 107 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].128 [[RAND]].0 |
| 108 |
; CHECK-NOT: dnscl.bftoint4 |
| 109 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 110 |
; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=64 |
| 111 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 112 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 113 |
%inputAddr = getelementptr <4 x bfloat>, <4 x bfloat> addrspace(1)* %input, i64 %gid |
| 114 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 115 |
%outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid |
| 116 |
|
| 117 |
%inputVal = load <4 x bfloat>, <4 x bfloat> addrspace(1)* %inputAddr, align 8 |
| 118 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 119 |
%conv = call <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv4_DF16bi(<4 x bfloat> %inputVal, i32 %seedVal) |
| 120 |
|
| 121 |
store <4 x i4> %conv, <4 x i4> addrspace(1)* %outputAddr, align 2 |
| 122 |
ret void |
| 123 |
} |
| 124 |
declare dso_local spir_func <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv4_DF16bi(<4 x bfloat>, i32) |
| 125 |
|
| 126 |
define spir_kernel void @BF16_to_Int4_vector8(<8 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <8 x i4> addrspace(1)* %output) { |
| 127 |
; CHECK-LABEL: .kernel "BF16_to_Int4_vector8" |
| 128 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 129 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 130 |
; CHECK-NOT: lfsr.b8v4 |
| 131 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 |
| 132 |
; CHECK: dnscl.bftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 |
| 133 |
; CHECK-NOT: dnscl.bftoint4 |
| 134 |
; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32 |
| 135 |
; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32 |
| 136 |
; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=128 |
| 137 |
; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 |
| 138 |
; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 |
| 139 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 140 |
%inputAddr = getelementptr <8 x bfloat>, <8 x bfloat> addrspace(1)* %input, i64 %gid |
| 141 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 142 |
%outputAddr = getelementptr <8 x i4>, <8 x i4> addrspace(1)* %output, i64 %gid |
| 143 |
|
| 144 |
%inputVal = load <8 x bfloat>, <8 x bfloat> addrspace(1)* %inputAddr, align 16 |
| 145 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 146 |
%conv = call <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv8_DF16bi(<8 x bfloat> %inputVal, i32 %seedVal) |
| 147 |
|
| 148 |
store <8 x i4> %conv, <8 x i4> addrspace(1)* %outputAddr, align 4 |
| 149 |
ret void |
| 150 |
} |
| 151 |
declare dso_local spir_func <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv8_DF16bi(<8 x bfloat>, i32) |
| 152 |
|
| 153 |
define spir_kernel void @BF16_to_Int4_vector16(<16 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <16 x i4> addrspace(1)* %output) { |
| 154 |
; CHECK-LABEL: .kernel "BF16_to_Int4_vector16" |
| 155 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 156 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 157 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND2:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 158 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND3:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 159 |
; CHECK-NOT: lfsr.b8v4 |
| 160 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 |
| 161 |
; CHECK: dnscl.bftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 |
| 162 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST2:[A-z0-9]*]].0 [[INPUT]].512 [[INPUT]].768 [[RAND2]].0 |
| 163 |
; CHECK: dnscl.bftoint4.mode2.srnd (M1, 32) [[DST3:[A-z0-9]*]].0 [[INPUT]].640 [[INPUT]].896 [[RAND3]].0 |
| 164 |
; CHECK-NOT: dnscl.bftoint4 |
| 165 |
; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32 |
| 166 |
; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32 |
| 167 |
; CHECK-DAG: // .decl [[RAND2]] v_type=G type=ud num_elts=32 |
| 168 |
; CHECK-DAG: // .decl [[RAND3]] v_type=G type=ud num_elts=32 |
| 169 |
; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=256 |
| 170 |
; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 |
| 171 |
; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 |
| 172 |
; CHECK-DAG: // .decl [[DST2]] v_type=G type=ud num_elts=32 |
| 173 |
; CHECK-DAG: // .decl [[DST3]] v_type=G type=ud num_elts=32 |
| 174 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 175 |
%inputAddr = getelementptr <16 x bfloat>, <16 x bfloat> addrspace(1)* %input, i64 %gid |
| 176 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 177 |
%outputAddr = getelementptr <16 x i4>, <16 x i4> addrspace(1)* %output, i64 %gid |
| 178 |
|
| 179 |
%inputVal = load <16 x bfloat>, <16 x bfloat> addrspace(1)* %inputAddr, align 32 |
| 180 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 181 |
%conv = call <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv16_DF16bi(<16 x bfloat> %inputVal, i32 %seedVal) |
| 182 |
|
| 183 |
store <16 x i4> %conv, <16 x i4> addrspace(1)* %outputAddr, align 8 |
| 184 |
ret void |
| 185 |
} |
| 186 |
declare dso_local spir_func <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv16_DF16bi(<16 x bfloat>, i32) |
| 187 |
|
| 188 |
; |
| 189 |
; Variants with pointers |
| 190 |
; |
| 191 |
|
| 192 |
define spir_kernel void @BF16_to_Int4_ptr_scalar(bfloat addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { |
| 193 |
; CHECK-LABEL: .kernel "BF16_to_Int4_ptr_scalar" |
| 194 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 195 |
; CHECK-NOT: lfsr.b8v4 |
| 196 |
; CHECK-DAG: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 |
| 197 |
; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud |
| 198 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 199 |
; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 |
| 200 |
; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 |
| 201 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 202 |
%seedMem = alloca i32, align 4 |
| 203 |
|
| 204 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 205 |
%inputAddr = getelementptr bfloat, bfloat addrspace(1)* %input, i64 %gid |
| 206 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 207 |
%outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid |
| 208 |
|
| 209 |
%inputVal = load bfloat, bfloat addrspace(1)* %inputAddr, align 2 |
| 210 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 211 |
%conv = call i4 @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDF16biPi(bfloat %inputVal, i32 %seedVal, i32* %seedMem) |
| 212 |
%v1 = insertelement <2 x i4> undef, i4 %conv, i32 0 |
| 213 |
%v2 = insertelement <2 x i4> %v1, i4 0, i32 1 |
| 214 |
|
| 215 |
store <2 x i4> %v2, <2 x i4> addrspace(1)* %outputAddr, align 1 |
| 216 |
%seedMemVal = load i32, i32* %seedMem, align 4 |
| 217 |
store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 |
| 218 |
ret void |
| 219 |
} |
| 220 |
declare dso_local spir_func i4 @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDF16biPi(bfloat, i32, i32*) |
| 221 |
|
| 222 |
define spir_kernel void @BF16_to_Int4_ptr_vector2(<2 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) { |
| 223 |
; CHECK-LABEL: .kernel "BF16_to_Int4_ptr_vector2" |
| 224 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 225 |
; CHECK-NOT: lfsr.b8v4 |
| 226 |
; CHECK-DAG: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 |
| 227 |
; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud |
| 228 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 229 |
; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 |
| 230 |
; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 |
| 231 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 232 |
%seedMem = alloca i32, align 4 |
| 233 |
|
| 234 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 235 |
%inputAddr = getelementptr <2 x bfloat>, <2 x bfloat> addrspace(1)* %input, i64 %gid |
| 236 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 237 |
%outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid |
| 238 |
|
| 239 |
%inputVal = load <2 x bfloat>, <2 x bfloat> addrspace(1)* %inputAddr, align 4 |
| 240 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 241 |
%conv = call <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv2_DF16biPi(<2 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) |
| 242 |
|
| 243 |
store <2 x i4> %conv, <2 x i4> addrspace(1)* %outputAddr, align 1 |
| 244 |
%seedMemVal = load i32, i32* %seedMem, align 4 |
| 245 |
store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 |
| 246 |
ret void |
| 247 |
} |
| 248 |
declare dso_local spir_func <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv2_DF16biPi(<2 x bfloat>, i32, i32*) |
| 249 |
|
| 250 |
define spir_kernel void @BF16_to_Int4_ptr_vector3(<3 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { |
| 251 |
; CHECK-LABEL: .kernel "BF16_to_Int4_ptr_vector3" |
| 252 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 253 |
; CHECK-NOT: lfsr.b8v4 |
| 254 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0 |
| 255 |
; CHECK-NOT: dnscl.bftoint4 |
| 256 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 257 |
; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32 |
| 258 |
; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32 |
| 259 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 260 |
%seedMem = alloca i32, align 4 |
| 261 |
|
| 262 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 263 |
%inputAddr = getelementptr <3 x bfloat>, <3 x bfloat> addrspace(1)* %input, i64 %gid |
| 264 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 265 |
%outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid |
| 266 |
|
| 267 |
%inputVal = load <3 x bfloat>, <3 x bfloat> addrspace(1)* %inputAddr, align 8 |
| 268 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 269 |
%conv = call <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv3_DF16biPi(<3 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) |
| 270 |
%conv4 = shufflevector <3 x i4> %conv, <3 x i4> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 undef> |
| 271 |
|
| 272 |
store <4 x i4> %conv4, <4 x i4> addrspace(1)* %outputAddr, align 2 |
| 273 |
%seedMemVal = load i32, i32* %seedMem, align 4 |
| 274 |
store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 |
| 275 |
ret void |
| 276 |
} |
| 277 |
declare dso_local spir_func <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv3_DF16biPi(<3 x bfloat>, i32, i32*) |
| 278 |
|
| 279 |
define spir_kernel void @BF16_to_Int4_ptr_vector4(<4 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) { |
| 280 |
; CHECK-LABEL: .kernel "BF16_to_Int4_ptr_vector4" |
| 281 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 282 |
; CHECK-NOT: lfsr.b8v4 |
| 283 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].128 [[RAND]].0 |
| 284 |
; CHECK-NOT: dnscl.bftoint4 |
| 285 |
; CHECK-DAG: // .decl [[RAND]] v_type=G type=ud num_elts=32 |
| 286 |
; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=64 |
| 287 |
; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32 |
| 288 |
%seedMem = alloca i32, align 4 |
| 289 |
|
| 290 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 291 |
%inputAddr = getelementptr <4 x bfloat>, <4 x bfloat> addrspace(1)* %input, i64 %gid |
| 292 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 293 |
%outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid |
| 294 |
|
| 295 |
%inputVal = load <4 x bfloat>, <4 x bfloat> addrspace(1)* %inputAddr, align 8 |
| 296 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 297 |
%conv = call <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv4_DF16biPi(<4 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) |
| 298 |
|
| 299 |
store <4 x i4> %conv, <4 x i4> addrspace(1)* %outputAddr, align 2 |
| 300 |
%seedMemVal = load i32, i32* %seedMem, align 4 |
| 301 |
store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 |
| 302 |
ret void |
| 303 |
} |
| 304 |
declare dso_local spir_func <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv4_DF16biPi(<4 x bfloat>, i32, i32*) |
| 305 |
|
| 306 |
define spir_kernel void @BF16_to_Int4_ptr_vector8(<8 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <8 x i4> addrspace(1)* %output) { |
| 307 |
; CHECK-LABEL: .kernel "BF16_to_Int4_ptr_vector8" |
| 308 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 309 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 310 |
; CHECK-NOT: lfsr.b8v4 |
| 311 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 |
| 312 |
; CHECK: dnscl.bftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 |
| 313 |
; CHECK-NOT: dnscl.bftoint4 |
| 314 |
; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32 |
| 315 |
; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32 |
| 316 |
; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=128 |
| 317 |
; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 |
| 318 |
; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 |
| 319 |
%seedMem = alloca i32, align 4 |
| 320 |
|
| 321 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 322 |
%inputAddr = getelementptr <8 x bfloat>, <8 x bfloat> addrspace(1)* %input, i64 %gid |
| 323 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 324 |
%outputAddr = getelementptr <8 x i4>, <8 x i4> addrspace(1)* %output, i64 %gid |
| 325 |
|
| 326 |
%inputVal = load <8 x bfloat>, <8 x bfloat> addrspace(1)* %inputAddr, align 16 |
| 327 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 328 |
%conv = call <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv8_DF16biPi(<8 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) |
| 329 |
|
| 330 |
store <8 x i4> %conv, <8 x i4> addrspace(1)* %outputAddr, align 4 |
| 331 |
%seedMemVal = load i32, i32* %seedMem, align 4 |
| 332 |
store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 |
| 333 |
ret void |
| 334 |
} |
| 335 |
declare dso_local spir_func <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv8_DF16biPi(<8 x bfloat>, i32, i32*) |
| 336 |
|
| 337 |
define spir_kernel void @BF16_to_Int4_ptr_vector16(<16 x bfloat> addrspace(1)* %input, i32 addrspace(1)* %seed, <16 x i4> addrspace(1)* %output) { |
| 338 |
; CHECK-LABEL: .kernel "BF16_to_Int4_ptr_vector16" |
| 339 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 340 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 341 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND2:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 342 |
; CHECK: lfsr.b8v4 (M1, 32) [[RAND3:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0> |
| 343 |
; CHECK-NOT: lfsr.b8v4 |
| 344 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0 |
| 345 |
; CHECK: dnscl.bftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0 |
| 346 |
; CHECK: dnscl.bftoint4.mode0.srnd (M1, 32) [[DST2:[A-z0-9]*]].0 [[INPUT]].512 [[INPUT]].768 [[RAND2]].0 |
| 347 |
; CHECK: dnscl.bftoint4.mode2.srnd (M1, 32) [[DST3:[A-z0-9]*]].0 [[INPUT]].640 [[INPUT]].896 [[RAND3]].0 |
| 348 |
; CHECK-NOT: dnscl.bftoint4 |
| 349 |
; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32 |
| 350 |
; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32 |
| 351 |
; CHECK-DAG: // .decl [[RAND2]] v_type=G type=ud num_elts=32 |
| 352 |
; CHECK-DAG: // .decl [[RAND3]] v_type=G type=ud num_elts=32 |
| 353 |
; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=256 |
| 354 |
; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32 |
| 355 |
; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32 |
| 356 |
; CHECK-DAG: // .decl [[DST2]] v_type=G type=ud num_elts=32 |
| 357 |
; CHECK-DAG: // .decl [[DST3]] v_type=G type=ud num_elts=32 |
| 358 |
%seedMem = alloca i32, align 4 |
| 359 |
|
| 360 |
%gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0) |
| 361 |
%inputAddr = getelementptr <16 x bfloat>, <16 x bfloat> addrspace(1)* %input, i64 %gid |
| 362 |
%seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid |
| 363 |
%outputAddr = getelementptr <16 x i4>, <16 x i4> addrspace(1)* %output, i64 %gid |
| 364 |
|
| 365 |
%inputVal = load <16 x bfloat>, <16 x bfloat> addrspace(1)* %inputAddr, align 32 |
| 366 |
%seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4 |
| 367 |
%conv = call <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv16_DF16biPi(<16 x bfloat> %inputVal, i32 %seedVal, i32* %seedMem) |
| 368 |
|
| 369 |
store <16 x i4> %conv, <16 x i4> addrspace(1)* %outputAddr, align 8 |
| 370 |
%seedMemVal = load i32, i32* %seedMem, align 4 |
| 371 |
store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4 |
| 372 |
ret void |
| 373 |
} |
| 374 |
declare dso_local spir_func <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundBF16ToInt4INTELDv16_DF16biPi(<16 x bfloat>, i32, i32*) |