1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
|
// RUN: mlir-opt --split-input-file %s -convert-arith-to-amdgpu="chipset=gfx950" | FileCheck %s
// CHECK-LABEL: @conversion_f8_f32_fallback
// CHECK: %[[CST:.+]] = arith.constant dense<0.000000e+00> : vector<2x2xf32>
// CHECK-NEXT: %[[SCALE_EXT:.+]] = arith.extf %arg1 : vector<2x2xf8E8M0FNU> to vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_00:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_00:.+]] = vector.shape_cast %[[IN_SLICE_00]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_00:.+]] = vector.extract %[[SCALE_EXT]][0, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_00:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_00]][0], %[[SCALE_SCALAR_00]] : vector<1xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_00:.+]] = vector.extract_strided_slice %[[PACKED_00]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_00:.+]] = vector.shape_cast %[[OUT_SLICE_00]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_00]], %[[CST]] {offsets = [0, 0], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_01:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_01:.+]] = vector.shape_cast %[[IN_SLICE_01]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_01:.+]] = vector.extract %[[SCALE_EXT]][0, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_01:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_01]][0], %[[SCALE_SCALAR_01]] : vector<1xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_01:.+]] = vector.extract_strided_slice %[[PACKED_01]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_01:.+]] = vector.shape_cast %[[OUT_SLICE_01]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_01]], %[[ACC_A]] {offsets = [0, 1], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_10:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_10:.+]] = vector.shape_cast %[[IN_SLICE_10]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_10:.+]] = vector.extract %[[SCALE_EXT]][1, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_10:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_10]][0], %[[SCALE_SCALAR_10]] : vector<1xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_10:.+]] = vector.extract_strided_slice %[[PACKED_10]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_10:.+]] = vector.shape_cast %[[OUT_SLICE_10]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_10]], %[[ACC_B]] {offsets = [1, 0], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_11:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_11:.+]] = vector.shape_cast %[[IN_SLICE_11]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_11:.+]] = vector.extract %[[SCALE_EXT]][1, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_11:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_11]][0], %[[SCALE_SCALAR_11]] : vector<1xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_11:.+]] = vector.extract_strided_slice %[[PACKED_11]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_11:.+]] = vector.shape_cast %[[OUT_SLICE_11]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_11]], %[[ACC_A]] {offsets = [1, 1], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: return %[[ACC_B]] : vector<2x2xf32>
func.func @conversion_f8_f32_fallback(%in: vector<2x2xf8E5M2>, %scale: vector<2x2xf8E8M0FNU>) -> vector<2x2xf32> {
%ext = arith.scaling_extf %in, %scale : vector<2x2xf8E5M2>, vector<2x2xf8E8M0FNU> to vector<2x2xf32>
return %ext : vector<2x2xf32>
}
// -----
// CHECK-LABEL: @conversion_f4_f32_fallback
// CHECK: %[[CST:.+]] = arith.constant dense<0.000000e+00> : vector<2x2xf32>
// CHECK-NEXT: %[[SCALE_EXT:.+]] = arith.extf %arg1 : vector<2x2xf8E8M0FNU> to vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_00:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_00:.+]] = vector.shape_cast %[[IN_SLICE_00]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_00:.+]] = vector.extract %[[SCALE_EXT]][0, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_00:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_00]][0], %[[SCALE_SCALAR_00]] : vector<1xf4E2M1FN> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_00:.+]] = vector.extract_strided_slice %[[PACKED_00]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_00:.+]] = vector.shape_cast %[[OUT_SLICE_00]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_00]], %[[CST]] {offsets = [0, 0], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_01:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_01:.+]] = vector.shape_cast %[[IN_SLICE_01]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_01:.+]] = vector.extract %[[SCALE_EXT]][0, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_01:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_01]][0], %[[SCALE_SCALAR_01]] : vector<1xf4E2M1FN> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_01:.+]] = vector.extract_strided_slice %[[PACKED_01]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_01:.+]] = vector.shape_cast %[[OUT_SLICE_01]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_01]], %[[ACC_A]] {offsets = [0, 1], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_10:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_10:.+]] = vector.shape_cast %[[IN_SLICE_10]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_10:.+]] = vector.extract %[[SCALE_EXT]][1, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_10:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_10]][0], %[[SCALE_SCALAR_10]] : vector<1xf4E2M1FN> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_10:.+]] = vector.extract_strided_slice %[[PACKED_10]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_10:.+]] = vector.shape_cast %[[OUT_SLICE_10]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_10]], %[[ACC_B]] {offsets = [1, 0], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_11:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_11:.+]] = vector.shape_cast %[[IN_SLICE_11]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_11:.+]] = vector.extract %[[SCALE_EXT]][1, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_11:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_11]][0], %[[SCALE_SCALAR_11]] : vector<1xf4E2M1FN> to vector<2xf32>
// CHECK-NEXT: %[[OUT_SLICE_11:.+]] = vector.extract_strided_slice %[[PACKED_11]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_VEC_11:.+]] = vector.shape_cast %[[OUT_SLICE_11]] : vector<1xf32> to vector<1x1xf32>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_11]], %[[ACC_A]] {offsets = [1, 1], strides = [1, 1]} : vector<1x1xf32> into vector<2x2xf32>
// CHECK-NEXT: return %[[ACC_B]] : vector<2x2xf32>
func.func @conversion_f4_f32_fallback(%in: vector<2x2xf4E2M1FN>, %scale: vector<2x2xf8E8M0FNU>) -> vector<2x2xf32> {
%ext = arith.scaling_extf %in, %scale : vector<2x2xf4E2M1FN>, vector<2x2xf8E8M0FNU> to vector<2x2xf32>
return %ext : vector<2x2xf32>
}
// -----
// CHECK-LABEL: @conversion_f8_f16_fallback
// CHECK: %[[CST:.+]] = arith.constant dense<0.000000e+00> : vector<2x2xf16>
// CHECK-NEXT: %[[SCALE_EXT:.+]] = arith.extf %arg1 : vector<2x2xf8E8M0FNU> to vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_00:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_00:.+]] = vector.shape_cast %[[IN_SLICE_00]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_00:.+]] = vector.extract %[[SCALE_EXT]][0, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_00:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_00]][0], %[[SCALE_SCALAR_00]] : vector<1xf8E5M2> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_00:.+]] = vector.extract_strided_slice %[[PACKED_00]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_00:.+]] = vector.shape_cast %[[OUT_SLICE_00]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_00]], %[[CST]] {offsets = [0, 0], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: %[[IN_SLICE_01:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_01:.+]] = vector.shape_cast %[[IN_SLICE_01]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_01:.+]] = vector.extract %[[SCALE_EXT]][0, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_01:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_01]][0], %[[SCALE_SCALAR_01]] : vector<1xf8E5M2> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_01:.+]] = vector.extract_strided_slice %[[PACKED_01]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_01:.+]] = vector.shape_cast %[[OUT_SLICE_01]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_01]], %[[ACC_A]] {offsets = [0, 1], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: %[[IN_SLICE_10:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_10:.+]] = vector.shape_cast %[[IN_SLICE_10]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_10:.+]] = vector.extract %[[SCALE_EXT]][1, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_10:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_10]][0], %[[SCALE_SCALAR_10]] : vector<1xf8E5M2> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_10:.+]] = vector.extract_strided_slice %[[PACKED_10]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_10:.+]] = vector.shape_cast %[[OUT_SLICE_10]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_10]], %[[ACC_B]] {offsets = [1, 0], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: %[[IN_SLICE_11:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf8E5M2> to vector<1x1xf8E5M2>
// CHECK-NEXT: %[[IN_VEC_11:.+]] = vector.shape_cast %[[IN_SLICE_11]] : vector<1x1xf8E5M2> to vector<1xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_11:.+]] = vector.extract %[[SCALE_EXT]][1, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_11:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_11]][0], %[[SCALE_SCALAR_11]] : vector<1xf8E5M2> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_11:.+]] = vector.extract_strided_slice %[[PACKED_11]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_11:.+]] = vector.shape_cast %[[OUT_SLICE_11]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_11]], %[[ACC_A]] {offsets = [1, 1], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: return %[[ACC_B]] : vector<2x2xf16>
func.func @conversion_f8_f16_fallback(%in: vector<2x2xf8E5M2>, %scale: vector<2x2xf8E8M0FNU>) -> vector<2x2xf16> {
%ext = arith.scaling_extf %in, %scale : vector<2x2xf8E5M2>, vector<2x2xf8E8M0FNU> to vector<2x2xf16>
return %ext : vector<2x2xf16>
}
// -----
// CHECK-LABEL: @conversion_f4_f16_fallback
// CHECK: %[[CST:.+]] = arith.constant dense<0.000000e+00> : vector<2x2xf16>
// CHECK-NEXT: %[[SCALE_EXT:.+]] = arith.extf %arg1 : vector<2x2xf8E8M0FNU> to vector<2x2xf32>
// CHECK-NEXT: %[[IN_SLICE_00:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_00:.+]] = vector.shape_cast %[[IN_SLICE_00]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_00:.+]] = vector.extract %[[SCALE_EXT]][0, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_00:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_00]][0], %[[SCALE_SCALAR_00]] : vector<1xf4E2M1FN> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_00:.+]] = vector.extract_strided_slice %[[PACKED_00]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_00:.+]] = vector.shape_cast %[[OUT_SLICE_00]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_00]], %[[CST]] {offsets = [0, 0], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: %[[IN_SLICE_01:.+]] = vector.extract_strided_slice %arg0 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_01:.+]] = vector.shape_cast %[[IN_SLICE_01]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_01:.+]] = vector.extract %[[SCALE_EXT]][0, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_01:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_01]][0], %[[SCALE_SCALAR_01]] : vector<1xf4E2M1FN> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_01:.+]] = vector.extract_strided_slice %[[PACKED_01]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_01:.+]] = vector.shape_cast %[[OUT_SLICE_01]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_01]], %[[ACC_A]] {offsets = [0, 1], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: %[[IN_SLICE_10:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 0], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_10:.+]] = vector.shape_cast %[[IN_SLICE_10]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_10:.+]] = vector.extract %[[SCALE_EXT]][1, 0] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_10:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_10]][0], %[[SCALE_SCALAR_10]] : vector<1xf4E2M1FN> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_10:.+]] = vector.extract_strided_slice %[[PACKED_10]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_10:.+]] = vector.shape_cast %[[OUT_SLICE_10]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_A:.+]] = vector.insert_strided_slice %[[OUT_VEC_10]], %[[ACC_B]] {offsets = [1, 0], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: %[[IN_SLICE_11:.+]] = vector.extract_strided_slice %arg0 {offsets = [1, 1], sizes = [1, 1], strides = [1, 1]} : vector<2x2xf4E2M1FN> to vector<1x1xf4E2M1FN>
// CHECK-NEXT: %[[IN_VEC_11:.+]] = vector.shape_cast %[[IN_SLICE_11]] : vector<1x1xf4E2M1FN> to vector<1xf4E2M1FN>
// CHECK-NEXT: %[[SCALE_SCALAR_11:.+]] = vector.extract %[[SCALE_EXT]][1, 1] : f32 from vector<2x2xf32>
// CHECK-NEXT: %[[PACKED_11:.+]] = amdgpu.scaled_ext_packed %[[IN_VEC_11]][0], %[[SCALE_SCALAR_11]] : vector<1xf4E2M1FN> to vector<2xf16>
// CHECK-NEXT: %[[OUT_SLICE_11:.+]] = vector.extract_strided_slice %[[PACKED_11]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf16> to vector<1xf16>
// CHECK-NEXT: %[[OUT_VEC_11:.+]] = vector.shape_cast %[[OUT_SLICE_11]] : vector<1xf16> to vector<1x1xf16>
// CHECK-NEXT: %[[ACC_B:.+]] = vector.insert_strided_slice %[[OUT_VEC_11]], %[[ACC_A]] {offsets = [1, 1], strides = [1, 1]} : vector<1x1xf16> into vector<2x2xf16>
// CHECK-NEXT: return %[[ACC_B]] : vector<2x2xf16>
func.func @conversion_f4_f16_fallback(%in: vector<2x2xf4E2M1FN>, %scale: vector<2x2xf8E8M0FNU>) -> vector<2x2xf16> {
%ext = arith.scaling_extf %in, %scale : vector<2x2xf4E2M1FN>, vector<2x2xf8E8M0FNU> to vector<2x2xf16>
return %ext : vector<2x2xf16>
}
// -----
// CHECK-LABEL: @conversion_broadcast
// CHECK-DAG: %[[CST:.+]] = arith.constant dense<0.000000e+00> : vector<8x2x4xf32>
// CHECK-DAG: %[[BCAST:.+]] = vector.broadcast %arg1
// CHECK-DAG: %[[IN_CAST:.+]] = vector.shape_cast %arg0
// CHECK-DAG: %[[SCALE_CAST:.+]] = vector.shape_cast %[[BCAST]]
// CHECK-DAG: %[[SCALE_EXT:.+]] = arith.extf %[[SCALE_CAST]]
// CHECK-DAG: vector.extract_strided_slice %[[IN_CAST]] {offsets = [0, 0, 0], sizes = [1, 1, 4], strides = [1, 1, 1]}
// CHECK-NEXT: %[[IN_SLICE_CAST:.+]] = vector.shape_cast
// CHECK-NEXT: vector.extract %[[SCALE_EXT]][0, 0, 0]
// CHECK-NEXT: %[[LOWHALF:.+]] = amdgpu.scaled_ext_packed %[[IN_SLICE_CAST]][0]
// CHECK-NEXT: vector.insert_strided_slice %[[LOWHALF]], %{{.+}} {offsets = [0], strides = [1]}
// CHECK-NEXT: %[[HIGHHALF:.+]] = amdgpu.scaled_ext_packed %[[IN_SLICE_CAST]][1]
// CHECK-NEXT: vector.insert_strided_slice %[[HIGHHALF]], %{{.+}} {offsets = [2], strides = [1]}
// CHECK-NEXT: vector.shape_cast
// CHECK-NEXT: vector.insert_strided_slice %{{.+}} {offsets = [0, 0, 0], strides = [1, 1, 1]}
// CHECK-NEXT: vector.extract_strided_slice %[[IN_CAST]] {offsets = [0, 1, 0], sizes = [1, 1, 4], strides = [1, 1, 1]}
// CHECK-NEXT: vector.shape_cast
// CHECK-NEXT: vector.extract %[[SCALE_EXT]][0, 1, 0]
// CHECK-NEXT: amdgpu.scaled_ext_packed
// CHECK-NEXT: vector.insert_strided_slice %{{.+}}, %{{.+}} {offsets = [0], strides = [1]}
// CHECK-NEXT: amdgpu.scaled_ext_packed
// CHECK-NEXT: vector.insert_strided_slice %{{.+}}, %{{.+}} {offsets = [2], strides = [1]}
// CHECK-NEXT: vector.shape_cast
// CHECK-NEXT: vector.insert_strided_slice %{{.+}}, %{{.+}} {offsets = [0, 1, 0], strides = [1, 1, 1]}
func.func @conversion_broadcast(%in: vector<8x8xf8E5M2>, %scale: vector<8x2xf8E8M0FNU>) -> vector<8x8xf32> {
%bc = vector.broadcast %scale : vector<8x2xf8E8M0FNU> to vector<4x8x2xf8E8M0FNU>
%cast1 = vector.shape_cast %in : vector<8x8xf8E5M2> to vector<8x2x4xf8E5M2>
%cast2 = vector.shape_cast %bc : vector<4x8x2xf8E8M0FNU> to vector<8x2x4xf8E8M0FNU>
%ext = arith.scaling_extf %cast1, %cast2 : vector<8x2x4xf8E5M2>, vector<8x2x4xf8E8M0FNU> to vector<8x2x4xf32>
%cast3 = vector.shape_cast %ext : vector<8x2x4xf32> to vector<8x8xf32>
return %cast3 : vector<8x8xf32>
}
// -----
// CHECK-LABEL: @conversion_broadcast_odd
// CHECK-NEXT: %[[CST_PARTIAL:.+]] = arith.constant dense<0.000000e+00> : vector<3xf32>
// CHECK-NEXT: %[[CST_FINAL:.+]] = arith.constant dense<0.000000e+00> : vector<6xf32>
// CHECK-NEXT: %[[SCALE_BC:.+]] = vector.broadcast %arg1 : vector<2xf8E8M0FNU> to vector<3x2xf8E8M0FNU>
// CHECK-NEXT: %[[SCALE_FLAT:.+]] = vector.shape_cast %[[SCALE_BC]] : vector<3x2xf8E8M0FNU> to vector<6xf8E8M0FNU>
// CHECK-NEXT: %[[SCALE_EXT:.+]] = arith.extf %[[SCALE_FLAT]] : vector<6xf8E8M0FNU> to vector<6xf32>
// CHECK-NEXT: %[[IN_SLICE_0:.+]] = vector.extract_strided_slice %arg0 {offsets = [0], sizes = [3], strides = [1]} : vector<6xf8E5M2> to vector<3xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_0:.+]] = vector.extract %[[SCALE_EXT]][0] : f32 from vector<6xf32>
// CHECK-NEXT: %[[PACKED_0A:.+]] = amdgpu.scaled_ext_packed %[[IN_SLICE_0]][0], %[[SCALE_SCALAR_0]] : vector<3xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[PARTIAL_ACC_0:.+]] = vector.insert_strided_slice %[[PACKED_0A]], %[[CST_PARTIAL]] {offsets = [0], strides = [1]} : vector<2xf32> into vector<3xf32>
// CHECK-NEXT: %[[PACKED_0B_RAW:.+]] = amdgpu.scaled_ext_packed %[[IN_SLICE_0]][1], %[[SCALE_SCALAR_0]] : vector<3xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[PACKED_0B:.+]] = vector.extract_strided_slice %[[PACKED_0B_RAW]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_SLICE_0:.+]] = vector.insert_strided_slice %[[PACKED_0B]], %[[PARTIAL_ACC_0]] {offsets = [2], strides = [1]} : vector<1xf32> into vector<3xf32>
// CHECK-NEXT: %[[FINAL_ACC_A:.+]] = vector.insert_strided_slice %[[OUT_SLICE_0]], %[[CST_FINAL]] {offsets = [0], strides = [1]} : vector<3xf32> into vector<6xf32>
// CHECK-NEXT: %[[IN_SLICE_1:.+]] = vector.extract_strided_slice %arg0 {offsets = [3], sizes = [3], strides = [1]} : vector<6xf8E5M2> to vector<3xf8E5M2>
// CHECK-NEXT: %[[SCALE_SCALAR_1:.+]] = vector.extract %[[SCALE_EXT]][3] : f32 from vector<6xf32>
// CHECK-NEXT: %[[PACKED_1A:.+]] = amdgpu.scaled_ext_packed %[[IN_SLICE_1]][0], %[[SCALE_SCALAR_1]] : vector<3xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[PARTIAL_ACC_1:.+]] = vector.insert_strided_slice %[[PACKED_1A]], %[[CST_PARTIAL]] {offsets = [0], strides = [1]} : vector<2xf32> into vector<3xf32>
// CHECK-NEXT: %[[PACKED_1B_RAW:.+]] = amdgpu.scaled_ext_packed %[[IN_SLICE_1]][1], %[[SCALE_SCALAR_1]] : vector<3xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[PACKED_1B:.+]] = vector.extract_strided_slice %[[PACKED_1B_RAW]] {offsets = [0], sizes = [1], strides = [1]} : vector<2xf32> to vector<1xf32>
// CHECK-NEXT: %[[OUT_SLICE_1:.+]] = vector.insert_strided_slice %[[PACKED_1B]], %[[PARTIAL_ACC_1]] {offsets = [2], strides = [1]} : vector<1xf32> into vector<3xf32>
// CHECK-NEXT: %[[RESULT:.+]] = vector.insert_strided_slice %[[OUT_SLICE_1]], %[[FINAL_ACC_A]] {offsets = [3], strides = [1]} : vector<3xf32> into vector<6xf32>
// CHECK-NEXT: return %[[RESULT]] : vector<6xf32>
func.func @conversion_broadcast_odd(%in: vector<6xf8E5M2>, %scale: vector<2xf8E8M0FNU>) -> vector<6xf32> {
%bc = vector.broadcast %scale : vector<2xf8E8M0FNU> to vector<3x2xf8E8M0FNU>
%cast = vector.shape_cast %bc : vector<3x2xf8E8M0FNU> to vector<6xf8E8M0FNU>
%ext = arith.scaling_extf %in, %cast : vector<6xf8E5M2>, vector<6xf8E8M0FNU> to vector<6xf32>
return %ext : vector<6xf32>
}
// -----
// CHECK-LABEL: @conversion_broadcast
// CHECK-DAG: %[[CST:.+]] = arith.constant dense<0.000000e+00> : vector<4xf32>
// CHECK-DAG: %[[SCALE_SPLAT:.+]] = vector.broadcast %arg1 : f8E8M0FNU to vector<4xf8E8M0FNU>
// CHECK-DAG: %[[SCALE_EXTF:.+]] = arith.extf %[[SCALE_SPLAT]] : vector<4xf8E8M0FNU> to vector<4xf32>
// CHECK-DAG: %[[SCALE_SCALAR:.+]] = vector.extract %[[SCALE_EXTF]][0] : f32 from vector<4xf32>
// CHECK: %[[OUT_CHUNK0:.+]] = amdgpu.scaled_ext_packed %arg0[0], %[[SCALE_SCALAR]] : vector<4xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[ACCUM_A:.+]] = vector.insert_strided_slice %[[OUT_CHUNK0]], %[[CST]] {offsets = [0], strides = [1]} : vector<2xf32> into vector<4xf32>
// CHECK-NEXT: %[[OUT_CHUNK1:.+]] = amdgpu.scaled_ext_packed %arg0[1], %[[SCALE_SCALAR]] : vector<4xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[FINAL_RESULT:.+]] = vector.insert_strided_slice %[[OUT_CHUNK1]], %[[ACCUM_A]] {offsets = [2], strides = [1]} : vector<2xf32> into vector<4xf32>
// CHECK-NEXT: return %[[FINAL_RESULT]] : vector<4xf32>
func.func @conversion_broadcast(%in: vector<4xf8E5M2>, %scale: f8E8M0FNU) -> vector<4xf32> {
%splat = vector.broadcast %scale : f8E8M0FNU to vector<4xf8E8M0FNU>
%ext = arith.scaling_extf %in, %splat : vector<4xf8E5M2>, vector<4xf8E8M0FNU> to vector<4xf32>
return %ext : vector<4xf32>
}
// -----
// CHECK-LABEL: @conversion_scalar
// CHECK: %[[SCALE_F32:.+]] = arith.extf %arg1 : f8E8M0FNU to f32
// CHECK-NEXT: %[[SPLAT_IN:.+]] = vector.broadcast %arg0 : f8E5M2 to vector<1xf8E5M2>
// CHECK-NEXT: %[[PACKED_EXT:.+]] = amdgpu.scaled_ext_packed %[[SPLAT_IN]][0], %[[SCALE_F32]] : vector<1xf8E5M2> to vector<2xf32>
// CHECK-NEXT: %[[RESULT:.+]] = vector.extract %[[PACKED_EXT]][0] : f32 from vector<2xf32>
// CHECK-NEXT: return %[[RESULT]] : f32
func.func @conversion_scalar(%in: f8E5M2, %scale: f8E8M0FNU) -> f32 {
%ext = arith.scaling_extf %in, %scale : f8E5M2, f8E8M0FNU to f32
return %ext : f32
}
// -----
// CHECK-LABEL: @long_fp4_broadcast
// CHECK-COUNT-4: amdgpu.scaled_ext_packed %{{.+}}[3]
// CHECK-NOT: amdgpu.scaled_ext_packed
// CHECK: return
func.func @long_fp4_broadcast(%in: vector<32xf4E2M1FN>, %scale: f32) -> vector<32xf32> {
%splat = vector.broadcast %scale : f32 to vector<32xf32>
%ext = arith.scaling_extf %in, %splat : vector<32xf4E2M1FN>, vector<32xf32> to vector<32xf32>
return %ext : vector<32xf32>
}
// -----
// CHECK-LABEL: @long_fp8_broadcast
// CHECK-COUNT-8: amdgpu.scaled_ext_packed %{{.+}}[1]
// CHECK-NOT: amdgpu.scaled_ext_packed
// CHECK: return
func.func @long_fp8_broadcast(%in: vector<32xf8E4M3FN>, %scale: f32) -> vector<32xf32> {
%splat = vector.broadcast %scale : f32 to vector<32xf32>
%ext = arith.scaling_extf %in, %splat : vector<32xf8E4M3FN>, vector<32xf32> to vector<32xf32>
return %ext : vector<32xf32>
}
|
