feat: switch to signed int striped simd, small qol updates

CHANGELOG.md

@@ -1,6 +1,13 @@
 
 # Next
 
+# v1.3.0 (07-21-2025)
+
+- Switch to using signed ints instead of unsigned for semi-global
+- Fix alignment score check used to decided whether or not to do the reverse alignment for semi-global and local alignment
+- Allow specifying simd width via a flag at comple time
+- Fix / updates the bench_aligner for current stable Mojo
+
 # v1.2.1 (06-25-2025)
 
 - Fixed bug in logging module related to update from Mojo 24.3 to 24.4 replacing `write_args`.

benchmarking/ish_bench_aligner.mojo

@@ -188,7 +188,7 @@ struct ByteFastaRecord:
             seq_bytes = scoring_matrix.convert_ascii_to_encoding(seq_bytes)
         var rev = List[UInt8](capacity=len(seq_bytes))
         for s in reversed(seq_bytes):
-            rev.append(s[])
+            rev.append(s)
         self.name = name^
         self.seq = seq_bytes^
         self.rev = rev^
@@ -329,26 +329,26 @@ struct BenchmarkResults(ToDelimited):
         var gcups = 0.0
 
         for result in results:
-            total_query_seqs += result[].total_query_seqs
-            total_target_seqs += result[].total_target_seqs
-            runtime_secs += result[].runtime_secs
-            cells_updated += result[].cells_updated
-            gcups += result[].gcups
+            total_query_seqs += result.total_query_seqs
+            total_target_seqs += result.total_target_seqs
+            runtime_secs += result.runtime_secs
+            cells_updated += result.cells_updated
+            gcups += result.gcups
 
-            if query_len != result[].query_len:
+            if query_len != result.query_len:
                 # TODO: may want to change this if we want to process multiple queries in one go.
                 raise "Mismatching query len"
-            if matrix != result[].matrix:
+            if matrix != result.matrix:
                 raise "Mismatching matrix"
-            if gap_open != result[].gap_open:
+            if gap_open != result.gap_open:
                 raise "Mismatching gap open"
-            if gap_extend != result[].gap_extend:
+            if gap_extend != result.gap_extend:
                 raise "Mismatching gap extend"
-            if u8_width != result[].u8_width:
+            if u8_width != result.u8_width:
                 raise "Mismatching u8 width"
-            if u16_width != result[].u16_width:
+            if u16_width != result.u16_width:
                 raise "Mismatching u16 width"
-            if score_size != result[].score_size:
+            if score_size != result.score_size:
                 raise "Mismatching score size"
 
         return Self(
@@ -518,7 +518,7 @@ fn bench_striped_local(
     )
     for q in queries:
         profiles.append(
-            Profiles[SIMD_U8_WIDTH, SIMD_U16_WIDTH](q[], matrix, score_size)
+            Profiles[SIMD_U8_WIDTH, SIMD_U16_WIDTH](q, matrix, score_size)
         )
     var prep_end = perf_counter()
     print("Setup Time:", prep_end - prep_start, file=stderr)
@@ -644,7 +644,7 @@ fn bench_striped_semi_global(
     for q in queries:
         profiles.append(
             SemiGlobalProfiles[SIMD_U8_WIDTH, SIMD_U16_WIDTH](
-                q[], matrix, score_size
+                q, matrix, score_size
             )
         )
     var prep_end = perf_counter()

ishlib/gpu/searcher_device.mojo

@@ -157,7 +157,9 @@ struct SearcherDevice[func_type: AnyTrivialRegType, //, func: func_type]:
         max_devices: Int,
     ) raises -> List[Self]:
         var ret = List[Self]()
-        for i in range(0, min(DeviceContext.number_of_devices(), max_devices)):
+        for i in range(
+            0, min(DeviceContext.number_of_devices(), min(1, max_devices))
+        ):
             var device = DeviceContext(i)
             # Triple check it's a gpu
             if device.api() == "cuda" or device.api() == "hip":

ishlib/matcher/__init__.mojo

@@ -16,10 +16,16 @@ fn simd_width_selector[dtype: DType]() -> Int:
     constrained[sizeof[Scalar[dtype]]() <= 16, "dytpe size too large."]()
     alias target = env_get_string["ISH_SIMD_TARGET", "none"]().lower()
     alias SSE_U8_SIZE = 16
+    alias AVX_256_U8_SIZE = 32
+    alias AVX_512_U8_SIZE = 64
 
     @parameter
-    if target == "baseline":
+    if target == "baseline" or target == "sse":
         return SSE_U8_SIZE // sizeof[Scalar[dtype]]()
+    elif target == "avx256":
+        return AVX_256_U8_SIZE // sizeof[Scalar[dtype]]()
+    elif target == "avx512":
+        return AVX_512_U8_SIZE // sizeof[Scalar[dtype]]()
     else:
         return simdwidthof[dtype]()
 

ishlib/matcher/alignment/__init__.mojo

@@ -59,6 +59,7 @@ struct AlignedMemory[dtype: DType, width: Int, alignment: Int](
         if zero_mem:
             memset_zero(self.ptr, self.length)
 
+    @always_inline
     fn __getitem__[
         I: Indexer
     ](ref self, offset: I) -> ref [self] SIMD[dtype, width]:
@@ -78,9 +79,11 @@ struct AlignedMemory[dtype: DType, width: Int, alignment: Int](
     fn __del__(owned self):
         self.ptr.free()
 
+    @always_inline
     fn __len__(read self) -> Int:
         return self.length
 
+    @always_inline
     fn as_span(
         ref self,
     ) -> Span[SIMD[self.dtype, self.width], __origin_of(self)]:

ishlib/matcher/alignment/local_aln/striped.mojo

@@ -152,7 +152,7 @@ fn ssw_align[
     gap_extension_penalty: UInt8 = 1,
     return_only_alignment_end: Bool = False,
     mask_length: Int32 = 15,  # for second best score
-    score_cutoff: Int32 = 0,
+    score_cutoff: Float32 = 0.0,
 ) -> Optional[Alignment]:
     # Find the alignment scores and ending positions
     var bests: AlignmentResult
@@ -212,7 +212,7 @@ fn ssw_align[
         Logger.warn("Failed to provide a valid query profile")
         return None
 
-    if bests.best.score <= score_cutoff:
+    if Float32(bests.best.score) <= score_cutoff:
         Logger.debug("Worse than cutoff")
         return None
 
@@ -294,7 +294,9 @@ fn sw[
     p_vecs.init_columns(len(reference))
     var max_score = UInt8(0).cast[dt]()
     var end_query: Int32 = query_len - 1
-    var end_reference: Int32 = -1  # 0 based best alignment ending point; initialized as isn't aligned -1
+    var end_reference: Int32 = (
+        -1
+    )  # 0 based best alignment ending point; initialized as isn't aligned -1
     var segment_length = p_vecs.segment_length
 
     # Note: