Repository navigation
Expand file tree
/
Copy pathtest_x86.py
More file actions
424 lines (351 loc) · 13.3 KB
/
Copy pathtest_x86.py
File metadata and controls
424 lines (351 loc) · 13.3 KB
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
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
from __future__ import annotations
import itertools
import numpy as np
import pytest
from exo import proc
from exo.platforms.x86 import *
from exo.stdlib.scheduling import *
old_split = repeat(divide_loop)
old_unroll = repeat(unroll_loop)
def old_fission_after(proc, stmt_pattern, n_lifts=1):
def find_stmts(p):
return [ c.after() for c in p.find_all(stmt_pattern) ]
return loop_hack(autofission, find_stmts)(proc, n_lifts)
@pytest.mark.isa('AVX2')
def test_avx2_memcpy(compiler):
"""
Compute dst = src
"""
@proc
def memcpy_avx2(n: size, dst: R[n] @ DRAM,
src: R[n] @ DRAM): # pragma: no cover
for i in seq(0, (n + 7) / 8):
if n - 8 * i >= 8:
tmp: f32[8] @ AVX2
mm256_loadu_ps(tmp, src[8 * i:8 * i + 8])
mm256_storeu_ps(dst[8 * i:8 * i + 8], tmp)
else:
for j in seq(0, n - 8 * i):
dst[8 * i + j] = src[8 * i + j]
# TODO: -march=skylake here is a hack. Such flags should be somehow handled
# automatically. Maybe this should be inferred by the use of AVX2, but
# "skylake" isn't right anyway. We might need a first-class notion of
# a Target, which has certain memories available. Then we can say that
# e.g. Skylake-X has AVX2, AVX512, etc.
fn = compiler.compile(memcpy_avx2,
skip_on_fail=True,
CMAKE_C_FLAGS="-march=skylake")
for n in (7, 8, 9, 31, 32, 33, 127, 128, 129):
inp = np.array([float(i) for i in range(n)], dtype=np.float32)
out = np.array([float(0) for _ in range(n)], dtype=np.float32)
fn(None, n, out, inp)
assert np.array_equal(inp, out)
@pytest.mark.isa('AVX2')
def test_avx2_simple_math(compiler):
"""
Compute x = x * y^2
"""
@proc
def simple_math_avx2(n: size, x: R[n] @ DRAM,
y: R[n] @ DRAM): # pragma: no cover
assert n % 8 == 0
for i in seq(0, n / 8):
xVec: f32[8] @ AVX2
yVec: f32[8] @ AVX2
mm256_loadu_ps(xVec, x[8 * i:8 * i + 8])
mm256_loadu_ps(yVec, y[8 * i:8 * i + 8])
mm256_mul_ps(xVec, xVec, yVec)
mm256_mul_ps(xVec, xVec, yVec)
mm256_storeu_ps(x[8 * i:8 * i + 8], xVec)
fn = compiler.compile(simple_math_avx2,
skip_on_fail=True,
CMAKE_C_FLAGS="-march=skylake")
for n in (8, 16, 24, 32, 64, 128):
x = np.array([float(i) for i in range(n)], dtype=np.float32)
y = np.array([float(3 * i) for i in range(n)], dtype=np.float32)
expected = x * y * y
fn(None, n, x, y)
np.testing.assert_almost_equal(x, expected)
@pytest.fixture
def simple_math_avx2_sched():
"""
Compute x = x * y^2
"""
@proc
def simple_math_avx2_sched(n: size, x: R[n] @ DRAM,
y: R[n] @ DRAM): # pragma: no cover
for i in seq(0, n):
x[i] = x[i] * y[i] * y[i]
def sched_simple_math_avx2_sched(p=simple_math_avx2_sched):
p = old_split(p, 'i', 8, ['io', 'ii'], tail='cut_and_guard')
p = stage_assn(p, 'x[_] = _ #0', 'xyy')
p = autolift_alloc(p, 'xyy: _', keep_dims=True)
p = set_memory(p, 'xyy', AVX2)
p = old_fission_after(p, 'xyy[_] = _')
p = replace_all(p, mm256_storeu_ps)
p = bind_expr(p, 'x[_]', 'xVec')
p = autolift_alloc(p, 'xVec: _', keep_dims=True)
p = set_memory(p, 'xVec', AVX2)
p = old_fission_after(p, 'xVec[_] = _')
p = bind_expr(p, 'y[_]', 'yVec', cse=True)
p = autolift_alloc(p, 'yVec: _', keep_dims=True)
p = set_memory(p, 'yVec', AVX2)
p = old_fission_after(p, 'yVec[_] = _')
p = replace_all(p, mm256_loadu_ps)
p = bind_expr(p, 'xVec[_] * yVec[_]', 'xy')
p = autolift_alloc(p, 'xy: _', keep_dims=True)
p = set_memory(p, 'xy', AVX2)
p = old_fission_after(p, 'xy[_] = _')
p = replace_all(p, mm256_mul_ps)
return p
simple_math_avx2_sched = sched_simple_math_avx2_sched()
return simple_math_avx2_sched
def test_gen_avx2_simple_math_scheduling(golden, simple_math_avx2_sched):
assert str(simple_math_avx2_sched) == golden
@pytest.mark.isa('AVX2')
def test_exec_avx2_simple_math_scheduling(compiler, simple_math_avx2_sched):
fn = compiler.compile(simple_math_avx2_sched,
skip_on_fail=True,
CMAKE_C_FLAGS="-march=skylake")
for n in (8, 16, 24, 32, 64, 128):
x = np.array([float(i) for i in range(n)], dtype=np.float32)
y = np.array([float(3 * i) for i in range(n)], dtype=np.float32)
expected = x * y * y
fn(None, n, x, y)
np.testing.assert_almost_equal(x, expected)
def _sgemm_test_cases(fn, M, N, K):
for m, n, k in itertools.product(M, N, K):
A = np.random.rand(m, k).astype(np.float32)
B = np.random.rand(k, n).astype(np.float32)
C = A @ B
C_out = np.zeros_like(C)
fn(None, m, n, k, C_out, A, B)
np.testing.assert_almost_equal(C, C_out, decimal=3)
@pytest.fixture
def sgemm_6x16():
@proc
def sgemm_6x16(
K: size,
C: [f32][6, 16] @ DRAM,
A: [f32][6, K] @ DRAM,
B: [f32][K, 16] @ DRAM,
):
for i in seq(0, 6):
for j in seq(0, 16):
for k in seq(0, K):
C[i, j] += A[i, k] * B[k, j]
return sgemm_6x16
@pytest.fixture
def avx2_sgemm_6x16(sgemm_6x16):
avx = rename(sgemm_6x16, "rank_k_reduce_6x16_scheduled")
print(avx)
avx = stage_assn(avx, 'C[_] += _', 'C_reg')
avx = set_memory(avx, 'C_reg', AVX2)
avx = old_split(avx, 'j', 8, ['jo', 'ji'], perfect=True)
avx = reorder_loops(avx, 'ji k')
avx = reorder_loops(avx, 'jo k')
avx = reorder_loops(avx, 'i k')
avx = autolift_alloc(avx, 'C_reg:_', n_lifts=3, keep_dims=True)
avx = old_fission_after(avx, 'C_reg = _ #0', n_lifts=3)
avx = old_fission_after(avx, 'C_reg[_] += _ #0', n_lifts=3)
avx = autolift_alloc(avx, 'C_reg:_', n_lifts=1)
avx = old_fission_after(avx, 'for i in _:_#0', n_lifts=1)
avx = old_fission_after(avx, 'for i in _:_#1', n_lifts=1)
avx = simplify(avx)
avx2_sgemm_6x16 = avx
def sched_avx2_sgemm_6x16(p=avx2_sgemm_6x16):
p = bind_expr(p, 'A[i, k]', 'a_vec')
p = set_memory(p, 'a_vec', AVX2)
p = expand_dim(p, 'a_vec:_', '8', 'ji')
p = autolift_alloc(p, 'a_vec:_')
p = old_fission_after(p, 'a_vec[_] = _')
#
p = bind_expr(p, 'B[k, _]', 'b_vec')
p = set_memory(p, 'b_vec', AVX2)
p = expand_dim(p, 'b_vec:_', '8', 'ji')
p = autolift_alloc(p, 'b_vec:_')
p = old_fission_after(p, 'b_vec[_] = _')
#
p = replace_all(p, avx2_set0_ps)
p = replace_all(p, mm256_broadcast_ss)
p = replace_all(p, mm256_fmadd_ps)
p = replace_all(p, avx2_fmadd_memu_ps)
p = replace(p, 'for ji in _:_ #0', mm256_loadu_ps)
p = replace(p, 'for ji in _:_ #0', mm256_loadu_ps)
p = replace(p, 'for ji in _:_ #0', mm256_storeu_ps)
#
p = old_unroll(p, 'jo')
p = old_unroll(p, 'i')
#
p = simplify(p)
return p
avx2_sgemm_6x16 = sched_avx2_sgemm_6x16()
return avx2_sgemm_6x16
#@pytest.mark.skip(reason='apparently unifying the broadcast is '
# 'non-deterministic')
# Until this non-determinism can be removed, at least
# try to run the code here...
def test_gen_avx2_sgemm_kernel(avx2_sgemm_6x16):
pass
#assert str(avx2_sgemm_6x16) == golden
@pytest.fixture
def sgemm_full():
@proc
def sgemm_full(
N: size,
M: size,
K: size,
C: f32[N, M] @ DRAM,
A: f32[N, K] @ DRAM,
B: f32[K, M] @ DRAM,
):
assert K > 0
for i in seq(0, N):
for j in seq(0, M):
for k in seq(0, K):
C[i, j] += A[i, k] * B[k, j]
return sgemm_full
@pytest.fixture
def avx2_sgemm_full(sgemm_full, sgemm_6x16, avx2_sgemm_6x16):
cache_i = 16
cache_j = 4
cache_k = 2
def sched_avx_sgemm_full(p=sgemm_full):
print(p)
p = rename(p, 'avx_sgemm_full')
# initial i,j tiling
p = old_split(p, 'i', 6, ['io', 'ii'], tail='cut')
p = reorder_loops(p, 'ii j #0')
p = divide_loop(p, 'j #0', 16, ['jo', 'ji'], tail='cut')
p = reorder_loops(p, 'ji ii')
# breaking off the main loop
p = old_fission_after(p, 'for jo in _: _')
# introduce k-tiling for later
p = divide_loop(p, 'k #0', cache_k * 16, ['ko', 'ki'], tail='cut')
p = old_fission_after(p, 'for ko in _: _', n_lifts=2)
p = reorder_loops(p, 'ji ko')
p = reorder_loops(p, 'ii ko')
p = replace_all(p, sgemm_6x16)
# insert uses of micro-kernel now
p = call_eqv(p, 'sgemm_6x16(_, _, _, _)', avx2_sgemm_6x16)
p = call_eqv(p, 'sgemm_6x16(_, _, _, _)', avx2_sgemm_6x16)
# do outer tiling for cache-locality
p = divide_loop(p, 'io #0', cache_i, ['io', 'im'], tail='cut')
p = reorder_loops(p, 'im jo')
p = divide_loop(p, 'jo #0', cache_j, ['jo', 'jm'], tail='cut')
p = divide_loop(p, 'jo #1', cache_j, ['jo', 'jm'], tail='cut')
p = reorder_loops(p, 'jm im')
# move the ko loop up and out
p = old_fission_after(p, 'for ko in _: _', n_lifts=2)
p = reorder_loops(p, 'jm ko #0')
p = reorder_loops(p, 'im ko #0')
return p
avx_sgemm_full = sched_avx_sgemm_full()
return avx_sgemm_full
# just make sure the scheduling works
def test_gen_avx2_sgemm_full(avx2_sgemm_full):
pass
@pytest.mark.isa('AVX2')
def test_avx2_sgemm_full(compiler, avx2_sgemm_full):
fn = compiler.compile(avx2_sgemm_full,
skip_on_fail=True,
CMAKE_C_FLAGS="-march=skylake")
_sgemm_test_cases(fn,
M=range(10, 600, 200),
N=range(20, 400, 120),
K=range(1, 512, 160))
@pytest.mark.isa('AVX2')
def test_avx2_sgemm_6x16(compiler, avx2_sgemm_6x16):
@proc
def avx2_sgemm_6x16_wrapper(
M: size,
N: size,
K: size,
C: f32[6, 16] @ DRAM,
A: f32[6, K] @ DRAM,
B: f32[K, 16] @ DRAM,
):
avx2_sgemm_6x16(K, C, A, B)
fn = compiler.compile(avx2_sgemm_6x16_wrapper,
skip_on_fail=True,
CMAKE_C_FLAGS="-march=skylake")
_sgemm_test_cases(fn, M=[6], N=[16], K=range(1, 512))
@pytest.fixture
def spec_kernel(sgemm_full, sgemm_6x16, avx2_sgemm_6x16):
@proc
def sgemm_micro_kernel_staged(
M: size,
N: size,
K: size,
A: f32[M, K],
B: f32[K, 16 * ((N + 15) / 16)],
C: [f32][M, N],
):
assert M >= 1
assert N >= 1
assert K >= 1
assert stride(A, 1) == 1
assert stride(B, 1) == 1
assert stride(C, 1) == 1
C_reg: f32[M, ((N + 15) / 16), 16] @ AVX512
for i in seq(0, M):
for j in seq(0, N / 16):
mm512_loadu_ps(C_reg[i, j, :], C[i, 16 * j:16 * j + 16])
if N % 16 > 0:
mm512_maskz_loadu_ps(
N % 16,
C_reg[i, N / 16, :],
C[i, 16 * (N / 16):16 * (N / 16) + N % 16]
)
for k in seq(0, K):
for i in seq(0, M):
a_vec: f32[16] @ AVX512
mm512_set1_ps(a_vec, A[i, k:k + 1])
for j in seq(0, ((N + 15) / 16)):
b_vec: f32[16] @ AVX512
mm512_loadu_ps(b_vec, B[k, j * 16:j * 16 + 16])
mm512_fmadd_ps(a_vec, b_vec, C_reg[i, j, :])
for i in seq(0, M):
for j in seq(0, N / 16):
mm512_storeu_ps(C[i, 16 * j:16 * j + 16], C_reg[i, j, :])
if N % 16 > 0:
mm512_mask_storeu_ps(
N % 16,
C[i, 16 * (N / 16):16 * (N / 16) + N % 16],
C_reg[i, N / 16, :]
)
def sched_spec_kernel(p=sgemm_micro_kernel_staged):
p = p.partial_eval(6, 64)
p = simplify(p)
p = old_unroll(p, 'j')
p = old_unroll(p, 'i')
p = simplify(p)
return p
spec_kernel = sched_spec_kernel()
return spec_kernel
# just make sure the scheduling works
def test_gen_avx512_sgemm_full(spec_kernel):
pass
@pytest.mark.isa('AVX512f')
def test_avx512_sgemm_full(compiler, spec_kernel):
spec_kernel.c_code_str()
@proc
def sgemm_full(
N: size,
M: size,
K: size,
C: f32[N, M] @ DRAM,
A: f32[N, K] @ DRAM,
B: f32[K, M] @ DRAM,
):
for i in seq(0, N):
for j in seq(0, M):
for k in seq(0, K):
C[i, j] += A[i, k] * B[k, j]
fn = compiler.compile(sgemm_full,
skip_on_fail=True,
CMAKE_C_FLAGS="-march=skylake-avx512")
_sgemm_test_cases(fn,
M=range(10, 600, 200),
N=range(20, 400, 120),
K=range(1, 512, 160))