use new method name and 32->256

Author: isuruf

sumpy/e2p.py

@@ -127,7 +127,7 @@ def default_name(self):
     def get_kernel(self, max_ntargets_in_one_box):
         ncoeffs = len(self.expansion)
         loopy_args = self.get_loopy_args()
-        max_work_items = min(32, max(ncoeffs, max_ntargets_in_one_box))
+        max_work_items = min(256, max(ncoeffs, max_ntargets_in_one_box))
 
         loopy_knl = lp.make_kernel(
                 [
@@ -212,8 +212,8 @@ def get_optimized_kernel(self, max_ntargets_in_one_box):
         inner_knl, optimizations = self.get_cached_loopy_knl_and_optimizations()
         knl = self.get_kernel(max_ntargets_in_one_box=max_ntargets_in_one_box)
         knl = lp.tag_inames(knl, {"itgt_box": "g.0"})
-        knl = lp.split_iname(knl, "itgt_offset", 32, inner_tag="l.0")
-        knl = lp.split_iname(knl, "icoeff", 32, inner_tag="l.0")
+        knl = lp.split_iname(knl, "itgt_offset", 256, inner_tag="l.0")
+        knl = lp.split_iname(knl, "icoeff", 256, inner_tag="l.0")
         knl = lp.add_inames_to_insn(knl, "dummy",
             "id:fetch_init* or id:fetch_center or id:kernel_scaling")
         knl = lp.add_inames_to_insn(knl, "itgt_box", "id:kernel_scaling")
@@ -278,7 +278,7 @@ def default_name(self):
     def get_kernel(self, max_ntargets_in_one_box):
         ncoeffs = len(self.expansion)
         loopy_args = self.get_loopy_args()
-        max_work_items = min(32, max(ncoeffs, max_ntargets_in_one_box))
+        max_work_items = min(256, max(ncoeffs, max_ntargets_in_one_box))
 
         loopy_knl = lp.make_kernel(
                 [
@@ -385,8 +385,8 @@ def get_optimized_kernel(self, max_ntargets_in_one_box):
         knl = lp.tag_inames(knl, {"itgt_box": "g.0", "dummy": "l.0"})
         knl = lp.unprivatize_temporaries_with_inames(knl,
             "itgt_offset", "result_temp")
-        knl = lp.split_iname(knl, "itgt_offset", 32, inner_tag="l.0")
-        knl = lp.split_iname(knl, "icoeff", 32, inner_tag="l.0")
+        knl = lp.split_iname(knl, "itgt_offset", 256, inner_tag="l.0")
+        knl = lp.split_iname(knl, "icoeff", 256, inner_tag="l.0")
         knl = lp.privatize_temporaries_with_inames(knl,
             "itgt_offset_outer", "result_temp")
         knl = lp.duplicate_inames(knl, "itgt_offset_outer", "id:init_result")

sumpy/expansion/local.py

@@ -32,6 +32,11 @@
         VolumeTaylorExpansionMixin,
         LinearPDEConformingVolumeTaylorExpansion)
 from sumpy.tools import add_to_sac, mi_increment_axis
+from sumpy.kernel import Kernel
+
+import loopy as lp
+
+from typing import Sequence
 
 import logging
 logger = logging.getLogger(__name__)
@@ -405,7 +410,7 @@ def loopy_translate_from(self, src_expansion):
             f"A direct loopy kernel for translation from "
             f"{src_expansion} to {self} is not implemented.")
 
-    def loopy_evaluate(self, kernels):
+    def get_loopy_evaluator(self, kernels: Sequence[Kernel]) -> lp.TranslationUnit:
         from sumpy.expansion.loopy import (make_l2p_loopy_kernel_for_volume_taylor,
             make_e2p_loopy_kernel)
         try:

sumpy/expansion/loopy.py

@@ -515,11 +515,13 @@ def get_idx(v):
         if c != sync_split:
             tags["e2p_iorder3"] = "l.0"
 
+        nsplit = min(256, ncoeffs)
+
         optimizations += [
             lambda knl: lp.tag_inames(knl, tags),
             lambda knl: lp.set_temporary_address_space(knl, "e2p_coeffs_copy",
                 lp.AddressSpace.LOCAL),
-            lambda knl: lp.split_iname(knl, "e2p_icoeff", 32, inner_tag="l.0"),
+            lambda knl: lp.split_iname(knl, "e2p_icoeff", nsplit, inner_tag="l.0"),
         ]
 
     target_args = gather_loopy_arguments((expansion,) + tuple(kernels))