diff --git a/pystencils/data_types.py b/pystencils/data_types.py
index 20953c7c810801c58407cdb02cc6d48fc8f73b5d..7300f01ea47f6693a2915b2cb063f907e46d1c32 100644
--- a/pystencils/data_types.py
+++ b/pystencils/data_types.py
@@ -697,7 +697,7 @@ class VectorType(Type):
if self.instruction_set is None:
return "%s[%d]" % (self.base_type, self.width)
else:
- if self.base_type == create_type("int64"):
+ if self.base_type == create_type("int64") or self.base_type == create_type("int32"):
return self.instruction_set['int']
elif self.base_type == create_type("float64"):
return self.instruction_set['double']
diff --git a/pystencils/include/aesni_rand.h b/pystencils/include/aesni_rand.h
index 8af30e1aba3794e078554dff2d6cb4ac9e4081b6..4206e37f634e01586223e1412ce7836e2499f65f 100644
--- a/pystencils/include/aesni_rand.h
+++ b/pystencils/include/aesni_rand.h
@@ -21,6 +21,7 @@
typedef std::uint32_t uint32;
typedef std::uint64_t uint64;
+#if defined(__AES__) || defined(_MSC_VER)
QUALIFIERS __m128i aesni_keygen_assist(__m128i temp1, __m128i temp2) {
__m128i temp3;
temp2 = _mm_shuffle_epi32(temp2 ,0xff);
@@ -302,6 +303,7 @@ QUALIFIERS void aesni_double2(uint32 ctr0, __m128i ctr1, uint32 ctr2, uint32 ctr
__m128d ignore;
aesni_double2(ctr0v, ctr1, ctr2v, ctr3v, key0, key1, key2, key3, rnd1, ignore, rnd2, ignore);
}
+#endif
#ifdef __AVX2__
diff --git a/pystencils/include/myintrin.h b/pystencils/include/myintrin.h
index 24b8c0967f6562365ac67dee3e896feb82c1fe81..a94c316c44de7aa420e4a6be807a510ee35687dd 100644
--- a/pystencils/include/myintrin.h
+++ b/pystencils/include/myintrin.h
@@ -1,6 +1,6 @@
#pragma once
-#ifdef __SSE2__
+#if defined(__SSE2__) || defined(_MSC_VER)
QUALIFIERS __m128 _my_cvtepu32_ps(const __m128i v)
{
#ifdef __AVX512VL__
@@ -28,7 +28,7 @@ QUALIFIERS void _MY_TRANSPOSE4_EPI32(__m128i & R0, __m128i & R1, __m128i & R2, _
}
#endif
-#ifdef __SSE4_1__
+#if defined(__SSE4_1__) || defined(_MSC_VER)
#if !defined(__AVX512VL__) && defined(__GNUC__) && __GNUC__ >= 5 && !defined(__clang__)
__attribute__((optimize("no-associative-math")))
#endif
diff --git a/pystencils/include/philox_rand.h b/pystencils/include/philox_rand.h
index 36ba7004f061ddd65378034023fddd425f2b7b69..b4c83669d0fda05aee1a7e018904c927a59c7ad1 100644
--- a/pystencils/include/philox_rand.h
+++ b/pystencils/include/philox_rand.h
@@ -1,6 +1,6 @@
#include <cstdint>
-#ifdef __SSE4_1__
+#if defined(__SSE4_1__) || defined(_MSC_VER)
#include <emmintrin.h> // SSE2
#endif
#ifdef __AVX2__
@@ -115,7 +115,7 @@ QUALIFIERS void philox_float4(uint32 ctr0, uint32 ctr1, uint32 ctr2, uint32 ctr3
}
#ifndef __CUDA_ARCH__
-#ifdef __SSE4_1__
+#if defined(__SSE4_1__) || defined(_MSC_VER)
QUALIFIERS void _philox4x32round(__m128i* ctr, __m128i* key)
{
__m128i lohi0a = _mm_mul_epu32(ctr[0], _mm_set1_epi32(PHILOX_M4x32_0));
diff --git a/pystencils_tests/test_random.py b/pystencils_tests/test_random.py
index a73d430d003ab772e719e7a8658b9d139a29901d..85b8d7ee66adb639bee818f49e1f7082ac227d80 100644
--- a/pystencils_tests/test_random.py
+++ b/pystencils_tests/test_random.py
@@ -5,11 +5,21 @@ import pytest
import pystencils as ps
from pystencils.rng import PhiloxFourFloats, PhiloxTwoDoubles, AESNIFourFloats, AESNITwoDoubles, random_symbol
from pystencils.backends.simd_instruction_sets import get_supported_instruction_sets
+from pystencils.cpu.cpujit import get_compiler_config
from pystencils.data_types import TypedSymbol
RNGs = {('philox', 'float'): PhiloxFourFloats, ('philox', 'double'): PhiloxTwoDoubles,
('aesni', 'float'): AESNIFourFloats, ('aesni', 'double'): AESNITwoDoubles}
+instruction_sets = get_supported_instruction_sets()
+if get_compiler_config()['os'] == 'windows':
+ # skip instruction sets supported by CPU but not the compiler
+ if '/arch:avx2' not in get_compiler_config()['flags'].lower() and \
+ '/arch:avx512' not in get_compiler_config()['flags'].lower():
+ instruction_sets.remove('avx')
+ if '/arch:avx512' not in get_compiler_config()['flags'].lower():
+ instruction_sets.remove('avx512')
+
@pytest.mark.parametrize('target,rng', (('cpu', 'philox'), ('cpu', 'aesni'), ('gpu', 'philox')))
@pytest.mark.parametrize('precision', ('float', 'double'))
@@ -89,20 +99,20 @@ def test_rng(target, rng, precision, dtype, t=124, offsets=(0, 0), keys=(0, 0),
def test_rng_offsets(kind, vectorized):
if vectorized:
test = test_rng_vectorized
- if not get_supported_instruction_sets():
+ if not instruction_sets:
pytest.skip("cannot detect CPU instruction set")
else:
test = test_rng
if kind == 'value':
- test(get_supported_instruction_sets()[0] if vectorized else 'cpu', 'philox', 'float', 'float', t=8,
+ test(instruction_sets[0] if vectorized else 'cpu', 'philox', 'float', 'float', t=8,
offsets=(6, 7), keys=(5, 309))
elif kind == 'symbol':
offsets = (TypedSymbol("x0", np.uint32), TypedSymbol("y0", np.uint32))
- test(get_supported_instruction_sets()[0] if vectorized else 'cpu', 'philox', 'float', 'float', t=8,
+ test(instruction_sets[0] if vectorized else 'cpu', 'philox', 'float', 'float', t=8,
offsets=offsets, offset_values=(6, 7), keys=(5, 309))
-@pytest.mark.parametrize('target', get_supported_instruction_sets())
+@pytest.mark.parametrize('target', instruction_sets)
@pytest.mark.parametrize('rng', ('philox', 'aesni'))
@pytest.mark.parametrize('precision,dtype', (('float', 'float'), ('double', 'double')))
def test_rng_vectorized(target, rng, precision, dtype, t=130, offsets=(1, 3), keys=(0, 0), offset_values=None):
@@ -139,11 +149,11 @@ def test_rng_vectorized(target, rng, precision, dtype, t=130, offsets=(1, 3), ke
def test_rng_symbol(vectorized):
"""Make sure that the RNG symbol generator generates symbols and that the resulting code compiles"""
if vectorized:
- if not get_supported_instruction_sets():
+ if not instruction_sets:
pytest.skip("cannot detect CPU instruction set")
else:
cpu_vectorize_info = {'assume_inner_stride_one': True, 'assume_aligned': True,
- 'instruction_set': get_supported_instruction_sets()[0]}
+ 'instruction_set': instruction_sets[0]}
else:
cpu_vectorize_info = None
@@ -155,7 +165,7 @@ def test_rng_symbol(vectorized):
ac.main_assignments[i] = ps.Assignment(ac.main_assignments[i].lhs, next(rng_symbol_gen))
symbols = [a.rhs for a in ac.main_assignments]
assert len(symbols) == f.shape[-1] and len(set(symbols)) == f.shape[-1]
- kernel = ps.create_kernel(ac, target=dh.default_target, cpu_vectorize_info=cpu_vectorize_info).compile()
+ ps.create_kernel(ac, target=dh.default_target, cpu_vectorize_info=cpu_vectorize_info).compile()
@pytest.mark.parametrize('vectorized', (False, True))
@@ -171,11 +181,11 @@ def test_staggered(vectorized):
if not vectorized:
return
- if not get_supported_instruction_sets():
+ if not instruction_sets:
pytest.skip("cannot detect CPU instruction set")
pytest.importorskip('islpy')
cpu_vectorize_info = {'assume_inner_stride_one': True, 'assume_aligned': False,
- 'instruction_set': get_supported_instruction_sets()[0]}
+ 'instruction_set': instruction_sets[0]}
dh.fill(j.name, 867)
dh.run_kernel(kernel, seed=5, time_step=309)