Debian Codegraph @

IGC/ocloc_tests/SPIRV-extensions/SPV_INTEL_mini_floats/FP4/ClampStochasticRoundFP16ToInt4.ll

raw
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_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 @FP16_to_Int4_scalar(half addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) {
29 ; CHECK-LABEL: .kernel "FP16_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.hftoint4.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 [[INPUT0]] v_type=G type=ud num_elts=32
35 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32
36 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
37 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
38 %inputAddr = getelementptr half, half addrspace(1)* %input, i64 %gid
39 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
40 %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid
41
42 %inputVal = load half, half addrspace(1)* %inputAddr, align 2
43 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
44 %conv = call i4 @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDhi(half %inputVal, i32 %seedVal)
45 %v1 = insertelement <2 x i4> undef, i4 %conv, i32 0
46 %v2 = insertelement <2 x i4> %v1, i4 0, i32 1
47
48 store <2 x i4> %v2, <2 x i4> addrspace(1)* %outputAddr, align 1
49 ret void
50 }
51 declare dso_local spir_func i4 @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDhi(half, i32)
52
53 define spir_kernel void @FP16_to_Int4_vector2(<2 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) {
54 ; CHECK-LABEL: .kernel "FP16_to_Int4_vector2"
55 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
56 ; CHECK-NOT: lfsr.b8v4
57 ; CHECK-DAG: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0
58 ; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud
59 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32
60 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32
61 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
62 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
63 %inputAddr = getelementptr <2 x half>, <2 x half> addrspace(1)* %input, i64 %gid
64 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
65 %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid
66
67 %inputVal = load <2 x half>, <2 x half> addrspace(1)* %inputAddr, align 4
68 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
69 %conv = call <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv2_Dhi(<2 x half> %inputVal, i32 %seedVal)
70
71 store <2 x i4> %conv, <2 x i4> addrspace(1)* %outputAddr, align 1
72 ret void
73 }
74 declare dso_local spir_func <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv2_Dhi(<2 x half>, i32)
75
76 define spir_kernel void @FP16_to_Int4_vector3(<3 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) {
77 ; CHECK-LABEL: .kernel "FP16_to_Int4_vector3"
78 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
79 ; CHECK-NOT: lfsr.b8v4
80 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0
81 ; CHECK-NOT: dnscl.hftoint4
82 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32
83 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32
84 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
85 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
86 %inputAddr = getelementptr <3 x half>, <3 x half> addrspace(1)* %input, i64 %gid
87 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
88 %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid
89
90 %inputVal = load <3 x half>, <3 x half> addrspace(1)* %inputAddr, align 8
91 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
92 %conv = call <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv3_Dhi(<3 x half> %inputVal, i32 %seedVal)
93 %conv4 = shufflevector <3 x i4> %conv, <3 x i4> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 undef>
94
95 store <4 x i4> %conv4, <4 x i4> addrspace(1)* %outputAddr, align 2
96 ret void
97 }
98 declare dso_local spir_func <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv3_Dhi(<3 x half>, i32)
99
100 define spir_kernel void @FP16_to_Int4_vector4(<4 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) {
101 ; CHECK-LABEL: .kernel "FP16_to_Int4_vector4"
102 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
103 ; CHECK-NOT: lfsr.b8v4
104 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].128 [[RAND]].0
105 ; CHECK-NOT: dnscl.hftoint4
106 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=64
107 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
108 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
109 %inputAddr = getelementptr <4 x half>, <4 x half> addrspace(1)* %input, i64 %gid
110 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
111 %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid
112
113 %inputVal = load <4 x half>, <4 x half> addrspace(1)* %inputAddr, align 8
114 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
115 %conv = call <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv4_Dhi(<4 x half> %inputVal, i32 %seedVal)
116
117 store <4 x i4> %conv, <4 x i4> addrspace(1)* %outputAddr, align 2
118 ret void
119 }
120 declare dso_local spir_func <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv4_Dhi(<4 x half>, i32)
121
122 define spir_kernel void @FP16_to_Int4_vector8(<8 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <8 x i4> addrspace(1)* %output) {
123 ; CHECK-LABEL: .kernel "FP16_to_Int4_vector8"
124 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
125 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
126 ; CHECK-NOT: lfsr.b8v4
127 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0
128 ; CHECK: dnscl.hftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0
129 ; CHECK-NOT: dnscl.hftoint4
130 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=128
131 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32
132 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32
133 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
134 %inputAddr = getelementptr <8 x half>, <8 x half> addrspace(1)* %input, i64 %gid
135 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
136 %outputAddr = getelementptr <8 x i4>, <8 x i4> addrspace(1)* %output, i64 %gid
137
138 %inputVal = load <8 x half>, <8 x half> addrspace(1)* %inputAddr, align 16
139 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
140 %conv = call <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv8_Dhi(<8 x half> %inputVal, i32 %seedVal)
141
142 store <8 x i4> %conv, <8 x i4> addrspace(1)* %outputAddr, align 4
143 ret void
144 }
145 declare dso_local spir_func <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv8_Dhi(<8 x half>, i32)
146
147 define spir_kernel void @FP16_to_Int4_vector16(<16 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <16 x i4> addrspace(1)* %output) {
148 ; CHECK-LABEL: .kernel "FP16_to_Int4_vector16"
149 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
150 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
151 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND2:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
152 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND3:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
153 ; CHECK-NOT: lfsr.b8v4
154 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0
155 ; CHECK: dnscl.hftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0
156 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST2:[A-z0-9]*]].0 [[INPUT]].512 [[INPUT]].768 [[RAND2]].0
157 ; CHECK: dnscl.hftoint4.mode2.srnd (M1, 32) [[DST3:[A-z0-9]*]].0 [[INPUT]].640 [[INPUT]].896 [[RAND3]].0
158 ; CHECK-NOT: dnscl.hftoint4
159 ; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32
160 ; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32
161 ; CHECK-DAG: // .decl [[RAND2]] v_type=G type=ud num_elts=32
162 ; CHECK-DAG: // .decl [[RAND3]] v_type=G type=ud num_elts=32
163 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=256
164 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32
165 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32
166 ; CHECK-DAG: // .decl [[DST2]] v_type=G type=ud num_elts=32
167 ; CHECK-DAG: // .decl [[DST3]] v_type=G type=ud num_elts=32
168 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
169 %inputAddr = getelementptr <16 x half>, <16 x half> addrspace(1)* %input, i64 %gid
170 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
171 %outputAddr = getelementptr <16 x i4>, <16 x i4> addrspace(1)* %output, i64 %gid
172
173 %inputVal = load <16 x half>, <16 x half> addrspace(1)* %inputAddr, align 32
174 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
175 %conv = call <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv16_Dhi(<16 x half> %inputVal, i32 %seedVal)
176
177 store <16 x i4> %conv, <16 x i4> addrspace(1)* %outputAddr, align 8
178 ret void
179 }
180 declare dso_local spir_func <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv16_Dhi(<16 x half>, i32)
181
182 ;
183 ; Variants with pointers
184 ;
185
186 define spir_kernel void @FP16_to_Int4_ptr_scalar(half addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) {
187 ; CHECK-LABEL: .kernel "FP16_to_Int4_ptr_scalar"
188 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
189 ; CHECK-NOT: lfsr.b8v4
190 ; CHECK-DAG: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0
191 ; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud
192 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32
193 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32
194 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
195 %seedMem = alloca i32, align 4
196
197 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
198 %inputAddr = getelementptr half, half addrspace(1)* %input, i64 %gid
199 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
200 %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid
201
202 %inputVal = load half, half addrspace(1)* %inputAddr, align 2
203 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
204 %conv = call i4 @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDhiPi(half %inputVal, i32 %seedVal, i32* %seedMem)
205 %v1 = insertelement <2 x i4> undef, i4 %conv, i32 0
206 %v2 = insertelement <2 x i4> %v1, i4 0, i32 1
207
208 store <2 x i4> %v2, <2 x i4> addrspace(1)* %outputAddr, align 1
209 %seedMemVal = load i32, i32* %seedMem, align 4
210 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4
211 ret void
212 }
213 declare dso_local spir_func i4 @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDhiPi(half, i32, i32*)
214
215 define spir_kernel void @FP16_to_Int4_ptr_vector2(<2 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <2 x i4> addrspace(1)* %output) {
216 ; CHECK-LABEL: .kernel "FP16_to_Int4_ptr_vector2"
217 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
218 ; CHECK-NOT: lfsr.b8v4
219 ; CHECK-DAG: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0
220 ; CHECK-DAG: mov (M1, 32) [[INPUT1]](0,0)<1> 0x0:ud
221 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32
222 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32
223 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
224 %seedMem = alloca i32, align 4
225
226 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
227 %inputAddr = getelementptr <2 x half>, <2 x half> addrspace(1)* %input, i64 %gid
228 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
229 %outputAddr = getelementptr <2 x i4>, <2 x i4> addrspace(1)* %output, i64 %gid
230
231 %inputVal = load <2 x half>, <2 x half> addrspace(1)* %inputAddr, align 4
232 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
233 %conv = call <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv2_DhiPi(<2 x half> %inputVal, i32 %seedVal, i32* %seedMem)
234
235 store <2 x i4> %conv, <2 x i4> addrspace(1)* %outputAddr, align 1
236 %seedMemVal = load i32, i32* %seedMem, align 4
237 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4
238 ret void
239 }
240 declare dso_local spir_func <2 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv2_DhiPi(<2 x half>, i32, i32*)
241
242 define spir_kernel void @FP16_to_Int4_ptr_vector3(<3 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) {
243 ; CHECK-LABEL: .kernel "FP16_to_Int4_ptr_vector3"
244 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
245 ; CHECK-NOT: lfsr.b8v4
246 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT0:[A-z0-9]*]].0 [[INPUT1:[A-z0-9]*]].0 [[RAND]].0
247 ; CHECK-NOT: dnscl.hftoint4
248 ; CHECK-DAG: // .decl [[INPUT0]] v_type=G type=ud num_elts=32
249 ; CHECK-DAG: // .decl [[INPUT1]] v_type=G type=ud num_elts=32
250 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
251 %seedMem = alloca i32, align 4
252
253 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
254 %inputAddr = getelementptr <3 x half>, <3 x half> addrspace(1)* %input, i64 %gid
255 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
256 %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid
257
258 %inputVal = load <3 x half>, <3 x half> addrspace(1)* %inputAddr, align 8
259 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
260 %conv = call <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv3_DhiPi(<3 x half> %inputVal, i32 %seedVal, i32* %seedMem)
261 %conv4 = shufflevector <3 x i4> %conv, <3 x i4> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 undef>
262
263 store <4 x i4> %conv4, <4 x i4> addrspace(1)* %outputAddr, align 2
264 %seedMemVal = load i32, i32* %seedMem, align 4
265 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4
266 ret void
267 }
268 declare dso_local spir_func <3 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv3_DhiPi(<3 x half>, i32, i32*)
269
270 define spir_kernel void @FP16_to_Int4_ptr_vector4(<4 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <4 x i4> addrspace(1)* %output) {
271 ; CHECK-LABEL: .kernel "FP16_to_Int4_ptr_vector4"
272 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
273 ; CHECK-NOT: lfsr.b8v4
274 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].128 [[RAND]].0
275 ; CHECK-NOT: dnscl.hftoint4
276 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=64
277 ; CHECK-DAG: // .decl [[DST]] v_type=G type=ud num_elts=32
278 %seedMem = alloca i32, align 4
279
280 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
281 %inputAddr = getelementptr <4 x half>, <4 x half> addrspace(1)* %input, i64 %gid
282 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
283 %outputAddr = getelementptr <4 x i4>, <4 x i4> addrspace(1)* %output, i64 %gid
284
285 %inputVal = load <4 x half>, <4 x half> addrspace(1)* %inputAddr, align 8
286 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
287 %conv = call <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv4_DhiPi(<4 x half> %inputVal, i32 %seedVal, i32* %seedMem)
288
289 store <4 x i4> %conv, <4 x i4> addrspace(1)* %outputAddr, align 2
290 %seedMemVal = load i32, i32* %seedMem, align 4
291 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4
292 ret void
293 }
294 declare dso_local spir_func <4 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv4_DhiPi(<4 x half>, i32, i32*)
295
296 define spir_kernel void @FP16_to_Int4_ptr_vector8(<8 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <8 x i4> addrspace(1)* %output) {
297 ; CHECK-LABEL: .kernel "FP16_to_Int4_ptr_vector8"
298 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
299 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
300 ; CHECK-NOT: lfsr.b8v4
301 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0
302 ; CHECK: dnscl.hftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0
303 ; CHECK-NOT: dnscl.hftoint4
304 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=128
305 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32
306 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32
307 %seedMem = alloca i32, align 4
308
309 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
310 %inputAddr = getelementptr <8 x half>, <8 x half> addrspace(1)* %input, i64 %gid
311 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
312 %outputAddr = getelementptr <8 x i4>, <8 x i4> addrspace(1)* %output, i64 %gid
313
314 %inputVal = load <8 x half>, <8 x half> addrspace(1)* %inputAddr, align 16
315 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
316 %conv = call <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv8_DhiPi(<8 x half> %inputVal, i32 %seedVal, i32* %seedMem)
317
318 store <8 x i4> %conv, <8 x i4> addrspace(1)* %outputAddr, align 4
319 %seedMemVal = load i32, i32* %seedMem, align 4
320 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4
321 ret void
322 }
323 declare dso_local spir_func <8 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv8_DhiPi(<8 x half>, i32, i32*)
324
325 define spir_kernel void @FP16_to_Int4_ptr_vector16(<16 x half> addrspace(1)* %input, i32 addrspace(1)* %seed, <16 x i4> addrspace(1)* %output) {
326 ; CHECK-LABEL: .kernel "FP16_to_Int4_ptr_vector16"
327 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND0:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
328 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND1:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
329 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND2:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
330 ; CHECK: lfsr.b8v4 (M1, 32) [[RAND3:[A-z0-9]*]](0,0)<1> {{.*}}(0,0)<1;1,0> {{.*}}(0,0)<1;1,0>
331 ; CHECK-NOT: lfsr.b8v4
332 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST0:[A-z0-9]*]].0 [[INPUT:[A-z0-9]*]].0 [[INPUT]].256 [[RAND0]].0
333 ; CHECK: dnscl.hftoint4.mode2.srnd (M1, 32) [[DST1:[A-z0-9]*]].0 [[INPUT]].128 [[INPUT]].384 [[RAND1]].0
334 ; CHECK: dnscl.hftoint4.mode0.srnd (M1, 32) [[DST2:[A-z0-9]*]].0 [[INPUT]].512 [[INPUT]].768 [[RAND2]].0
335 ; CHECK: dnscl.hftoint4.mode2.srnd (M1, 32) [[DST3:[A-z0-9]*]].0 [[INPUT]].640 [[INPUT]].896 [[RAND3]].0
336 ; CHECK-NOT: dnscl.hftoint4
337 ; CHECK-DAG: // .decl [[RAND0]] v_type=G type=ud num_elts=32
338 ; CHECK-DAG: // .decl [[RAND1]] v_type=G type=ud num_elts=32
339 ; CHECK-DAG: // .decl [[RAND2]] v_type=G type=ud num_elts=32
340 ; CHECK-DAG: // .decl [[RAND3]] v_type=G type=ud num_elts=32
341 ; CHECK-DAG: // .decl [[INPUT]] v_type=G type=ud num_elts=256
342 ; CHECK-DAG: // .decl [[DST0]] v_type=G type=ud num_elts=32
343 ; CHECK-DAG: // .decl [[DST1]] v_type=G type=ud num_elts=32
344 ; CHECK-DAG: // .decl [[DST2]] v_type=G type=ud num_elts=32
345 ; CHECK-DAG: // .decl [[DST3]] v_type=G type=ud num_elts=32
346 %seedMem = alloca i32, align 4
347
348 %gid = call spir_func i64 @_Z33__spirv_BuiltInGlobalInvocationIdi(i32 0)
349 %inputAddr = getelementptr <16 x half>, <16 x half> addrspace(1)* %input, i64 %gid
350 %seedAddr = getelementptr i32, i32 addrspace(1)* %seed, i64 %gid
351 %outputAddr = getelementptr <16 x i4>, <16 x i4> addrspace(1)* %output, i64 %gid
352
353 %inputVal = load <16 x half>, <16 x half> addrspace(1)* %inputAddr, align 32
354 %seedVal = load i32, i32 addrspace(1)* %seedAddr, align 4
355 %conv = call <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv16_DhiPi(<16 x half> %inputVal, i32 %seedVal, i32* %seedMem)
356
357 store <16 x i4> %conv, <16 x i4> addrspace(1)* %outputAddr, align 8
358 %seedMemVal = load i32, i32* %seedMem, align 4
359 store i32 %seedMemVal, i32 addrspace(1)* %seedAddr, align 4
360 ret void
361 }
362 declare dso_local spir_func <16 x i4> @_Z51__builtin_spirv_ClampStochasticRoundFP16ToInt4INTELDv16_DhiPi(<16 x half>, i32, i32*)