Increased the amount of prefetch done (and small parameter changes)

sterrettm2 · sterrettm2 · commit d4ecb7ed92bd · 2023-10-04T11:07:17.000-07:00
diff --git a/src/avx512-32bit-qsort.hpp b/src/avx512-32bit-qsort.hpp
@@ -37,7 +37,7 @@ struct zmm_vector<int32_t> {
     using opmask_t = __mmask16;
     static const uint8_t numlanes = 16;
     static constexpr int network_sort_threshold = 512;
-    static constexpr int partition_unroll_factor = 4;
+    static constexpr int partition_unroll_factor = 8;
     
     using swizzle_ops = avx512_32bit_swizzle_ops;
 
@@ -158,7 +158,7 @@ struct zmm_vector<uint32_t> {
     using opmask_t = __mmask16;
     static const uint8_t numlanes = 16;
     static constexpr int network_sort_threshold = 512;
-    static constexpr int partition_unroll_factor = 4;
+    static constexpr int partition_unroll_factor = 8;
     
     using swizzle_ops = avx512_32bit_swizzle_ops;
 
diff --git a/src/avx512-common-qsort.h b/src/avx512-common-qsort.h
@@ -440,16 +440,16 @@ X86_SIMD_SORT_INLINE arrsize_t partition_avx512_unrolled(type_t *arr,
             X86_SIMD_SORT_UNROLL_LOOP(8)
             for (int ii = 0; ii < num_unroll; ++ii) {
                 curr_vec[ii] = vtype::loadu(arr + right + ii * vtype::numlanes);
+                _mm_prefetch(arr + right + ii * vtype::numlanes - num_unroll * vtype::numlanes, _MM_HINT_T0);
             }
-            _mm_prefetch(arr + right - num_unroll * vtype::numlanes, _MM_HINT_T0);
         }
         else {
             X86_SIMD_SORT_UNROLL_LOOP(8)
             for (int ii = 0; ii < num_unroll; ++ii) {
                 curr_vec[ii] = vtype::loadu(arr + left + ii * vtype::numlanes);
+                _mm_prefetch(arr + left + ii * vtype::numlanes + num_unroll * vtype::numlanes, _MM_HINT_T0);
             }
             left += num_unroll * vtype::numlanes;
-            _mm_prefetch(arr + left, _MM_HINT_T0);
         }
         // partition the current vector and save it on both sides of the array
         X86_SIMD_SORT_UNROLL_LOOP(8)

Original file line number	Diff line number	Diff line change
`@@ -440,16 +440,16 @@ X86_SIMD_SORT_INLINE arrsize_t partition_avx512_unrolled(type_t *arr,`
`440`	`440`	`X86_SIMD_SORT_UNROLL_LOOP(8)`
`441`	`441`	`for (int ii = 0; ii < num_unroll; ++ii) {`
`442`	`442`	`curr_vec[ii] = vtype::loadu(arr + right + ii * vtype::numlanes);`
	`443`	`+ _mm_prefetch(arr + right + ii * vtype::numlanes - num_unroll * vtype::numlanes, _MM_HINT_T0);`
`443`	`444`	`}`
`444`		`- _mm_prefetch(arr + right - num_unroll * vtype::numlanes, _MM_HINT_T0);`
`445`	`445`	`}`
`446`	`446`	`else {`
`447`	`447`	`X86_SIMD_SORT_UNROLL_LOOP(8)`
`448`	`448`	`for (int ii = 0; ii < num_unroll; ++ii) {`
`449`	`449`	`curr_vec[ii] = vtype::loadu(arr + left + ii * vtype::numlanes);`
	`450`	`+ _mm_prefetch(arr + left + ii * vtype::numlanes + num_unroll * vtype::numlanes, _MM_HINT_T0);`
`450`	`451`	`}`
`451`	`452`	`left += num_unroll * vtype::numlanes;`
`452`		`- _mm_prefetch(arr + left, _MM_HINT_T0);`
`453`	`453`	`}`
`454`	`454`	`// partition the current vector and save it on both sides of the array`
`455`	`455`	`X86_SIMD_SORT_UNROLL_LOOP(8)`