Bench simd_masked_load

ogxd · Nov 9, 2024 · 0ad6fb3 · 0ad6fb3
1 parent 27eeaae
commit 0ad6fb3
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 6 deletions.
diff --git a/src/gxhash/platform/arm.rs b/src/gxhash/platform/arm.rs
@@ -55,7 +55,14 @@ pub unsafe fn get_partial_safe(data: *const State, len: usize) -> State {
 pub unsafe fn get_partial_unsafe_no_ub(data: *const State, len: usize) -> State {
     let indices = vld1q_s8([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15].as_ptr());
     let mask = vcgtq_s8(vdupq_n_s8(len as i8), indices);
-    let partial_vector = vandq_s8(load_unaligned(data), vreinterpretq_s8_u8(mask));
+    use std::arch::asm;
+    let mut result: State;
+    asm!(
+        "ld1 {0}, [{1}]",
+        out(vreg) result, in(reg) data,
+        options(pure, nomem, nostack)
+    );
+    let partial_vector = vandq_s8(result, vreinterpretq_s8_u8(mask));
     vaddq_s8(partial_vector, vdupq_n_s8(len as i8))
 }
 

diff --git a/src/gxhash/platform/x86.rs b/src/gxhash/platform/x86.rs
@@ -65,13 +65,18 @@ pub unsafe fn get_partial_safe(data: *const State, len: usize) -> State {
 #[inline(always)]
 pub unsafe fn get_partial_unsafe_no_ub(data: *const State, len: usize) -> State {
     // Using inline assembly to load out-of-bounds
-    use std::arch::asm;
+    // use std::arch::asm;
+    // let indices = _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+    // let mask = _mm_cmpgt_epi8(_mm_set1_epi8(len as i8), indices);
+    // let mut result: State;
+    // asm!("movdqu [{}], {}", in(reg) data, out(xmm_reg) result, options(pure, nomem, nostack));
+    // let partial_vector = _mm_and_si128(result, mask);
+    // _mm_add_epi8(partial_vector, _mm_set1_epi8(len as i8))
+
+    // Using simd_masked_load
     let indices = _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
     let mask = _mm_cmpgt_epi8(_mm_set1_epi8(len as i8), indices);
-    let mut result: State;
-    asm!("movdqu [{}], {}", in(reg) data, out(xmm_reg) result, options(pure, nomem, nostack));
-    let partial_vector = _mm_and_si128(result, mask);
-    _mm_add_epi8(partial_vector, _mm_set1_epi8(len as i8))
+    State::from(std::intrinsics::simd::simd_masked_load(core::simd::i8x16::from(mask), data as *const i8, core::simd::i8x16::from(_mm_set1_epi8(len as i8))))
 
     // Using std::simd
     // use std::simd::*;