Merge branch 'sve' into 'master'

Remove support for non-power-of-2 SVE vector widths

See merge request pycodegen/pystencils!325

.gitlab-ci.yml

@@ -175,7 +175,7 @@ arm64v9:
   extends: .multiarch_template
   image: i10git.cs.fau.de:5005/pycodegen/pycodegen/arm64
   variables:
-    PYSTENCILS_SIMD: "sve256,sve512,sve"
+    PYSTENCILS_SIMD: "sve128,sve256,sve512,sve"
   before_script:
     - *multiarch_before_script
     - sed -i s/march=native/march=armv8-a+sve/g ~/.config/pystencils/config.json

pystencils/backends/arm_instruction_sets.py

@@ -151,9 +151,7 @@ def get_vector_instruction_set_arm(data_type='double', instruction_set='neon'):
         result['any'] = f'vaddlvq_u8(vreinterpretq_u8_u{bits[data_type]}({{0}})) > 0'
         result['all'] = f'vaddlvq_u8(vreinterpretq_u8_u{bits[data_type]}({{0}})) == 16*0xff'
 
-    if instruction_set == 'sve' or bitwidth & (bitwidth - 1) == 0:
-        # only power-of-2 vector sizes will evenly divide a cacheline
-        result['cachelineSize'] = 'cachelineSize()'
-        result['cachelineZero'] = 'cachelineZero((void*) {0})'
+    result['cachelineSize'] = 'cachelineSize()'
+    result['cachelineZero'] = 'cachelineZero((void*) {0})'
 
     return result

pystencils/backends/simd_instruction_sets.py

-import math
 import os
 import platform
 from ctypes import CDLL
@@ -86,15 +85,12 @@ def get_supported_instruction_sets():
     if flags.issuperset(required_sve_flags):
         if platform.system() == 'Linux':
             libc = CDLL('libc.so.6')
-            native_length = 8 * libc.prctl(51, 0, 0, 0, 0)  # PR_SVE_GET_VL
-            if native_length < 0:
+            length = 8 * libc.prctl(51, 0, 0, 0, 0)  # PR_SVE_GET_VL
+            if length < 0:
                 raise OSError("SVE length query failed")
-            pwr2_length = int(2**math.floor(math.log2(native_length)))
-            if pwr2_length % 256 == 0:
-                result.append(f"sve{pwr2_length//2}")
-            if native_length != pwr2_length:
-                result.append(f"sve{pwr2_length}")
-            result.append(f"sve{native_length}")
+            while length > 128:
+                result.append(f"sve{length}")
+                length //= 2
         result.append("sve")
     return result