Improve dft performance on arm64

kfrlib · Jan 13, 2025 · 8306017 · 8306017
1 parent 6aea976
commit 8306017
Show file tree

Hide file tree

Showing 3 changed files with 44 additions and 11 deletions.
diff --git a/src/dft/bitrev.hpp b/src/dft/bitrev.hpp
@@ -49,15 +49,22 @@ constexpr inline static size_t bitrev_table_log2N = ilog2(arraysize(data::bitrev
 template <size_t Bits>
 CMT_GNU_CONSTEXPR inline u32 bitrev_using_table(u32 x)
 {
+#ifdef CMT_ARCH_NEON
+    return __builtin_bitreverse32(x) >> (32 - Bits);
+#else
     if constexpr (Bits > bitrev_table_log2N)
         return bitreverse<Bits>(x);
 
     return data::bitrev_table[x] >> (bitrev_table_log2N - Bits);
+#endif
 }
 
 template <bool use_table>
 CMT_GNU_CONSTEXPR inline u32 bitrev_using_table(u32 x, size_t bits, cbool_t<use_table>)
 {
+#ifdef CMT_ARCH_NEON
+    return __builtin_bitreverse32(x) >> (32 - bits);
+#else
     if constexpr (use_table)
     {
         return data::bitrev_table[x] >> (bitrev_table_log2N - bits);
@@ -66,10 +73,17 @@ CMT_GNU_CONSTEXPR inline u32 bitrev_using_table(u32 x, size_t bits, cbool_t<use_
     {
         return bitreverse<32>(x) >> (32 - bits);
     }
+#endif
 }
 
 CMT_GNU_CONSTEXPR inline u32 dig4rev_using_table(u32 x, size_t bits)
 {
+#ifdef CMT_ARCH_NEON
+    x = __builtin_bitreverse32(x);
+    x = (((x & 0xaaaaaaaa) >> 1) | ((x & 0x55555555) << 1));
+    x = x >> (32 - bits);
+    return x;
+#else
     if (bits > bitrev_table_log2N)
     {
         if (bits <= 16)
@@ -82,6 +96,7 @@ CMT_GNU_CONSTEXPR inline u32 dig4rev_using_table(u32 x, size_t bits)
     x = (((x & 0xaaaaaaaa) >> 1) | ((x & 0x55555555) << 1));
     x = x >> (bitrev_table_log2N - bits);
     return x;
+#endif
 }
 
 template <size_t log2n, size_t bitrev, typename T>

diff --git a/src/dft/fft-impl.hpp b/src/dft/fft-impl.hpp
@@ -52,22 +52,30 @@ template <typename T>
 inline std::bitset<DFT_MAX_STAGES> fft_algorithm_selection;
 
 template <>
-inline std::bitset<DFT_MAX_STAGES> fft_algorithm_selection<float>{ (1ull << 15) - 1 };
+inline std::bitset<DFT_MAX_STAGES> fft_algorithm_selection<float>{
+#ifdef CMT_ARCH_NEON
+    0
+#else
+    (1ull << 15) - 1
+#endif
+};
 
 template <>
 inline std::bitset<DFT_MAX_STAGES> fft_algorithm_selection<double>{ 0 };
 
 template <typename T>
-constexpr bool inline use_autosort(size_t log2n)
+inline bool use_autosort(size_t log2n)
 {
     return fft_algorithm_selection<T>[log2n];
 }
 
+#ifndef CMT_ARCH_NEON
+#define KFR_AUTOSORT_FOR_2048
 #define KFR_AUTOSORT_FOR_128D
 #define KFR_AUTOSORT_FOR_256D
 #define KFR_AUTOSORT_FOR_512
 #define KFR_AUTOSORT_FOR_1024
-#define KFR_AUTOSORT_FOR_2048
+#endif
 
 #ifdef CMT_ARCH_AVX
 template <>
@@ -855,7 +863,11 @@ template <typename T>
 struct fft_config
 {
     constexpr static inline const bool recursion = true;
-    constexpr static inline const bool prefetch  = true;
+#ifdef CMT_ARCH_NEON
+    constexpr static inline const bool prefetch = false;
+#else
+    constexpr static inline const bool prefetch = true;
+#endif
     constexpr static inline const size_t process_width =
         const_max(static_cast<size_t>(1), vector_capacity<T> / 16);
 };
@@ -1606,7 +1618,7 @@ struct fft_specialization<T, 10> : fft_final_stage_impl<T, false, 1024>
     {
         fft_final_stage_impl<T, false, 1024>::template do_execute<inverse>(out, in, nullptr);
         if (this->need_reorder)
-            fft_reorder(out, 10, cfalse);
+            fft_reorder(out, csize_t<10>{}, cbool_t<always_br2>{});
     }
 };
 #endif
@@ -1649,8 +1661,6 @@ struct fft_specialization<T, 11> : dft_stage<T>
         radix8_autosort_pass_last(256, csize<width>, no, no, no, cbool<inverse>, out, out, tw);
     }
 };
-
-#else
 #endif
 
 template <bool is_even, bool first, typename T, bool autosort>
@@ -1768,7 +1778,13 @@ KFR_INTRINSIC void init_fft(dft_plan<T>* self, size_t size, dft_order)
 {
     const size_t log2n = ilog2(size);
     cswitch(
-        csizes_t<0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11>(), log2n,
+        csizes_t<0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
+#ifdef KFR_AUTOSORT_FOR_2048
+                 ,
+                 11
+#endif
+                 >(),
+        log2n,
         [&](auto log2n)
         {
             (void)log2n;

diff --git a/tests/dft_test.cpp b/tests/dft_test.cpp
@@ -33,11 +33,13 @@ constexpr ctypes_t<float, double> dft_float_types{};
 constexpr ctypes_t<float> dft_float_types{};
 #endif
 
-#if defined(CMT_ARCH_X86) && !defined(KFR_NO_PERF_TESTS)
+#if !defined(KFR_NO_PERF_TESTS)
 
 static void full_barrier()
 {
-#ifdef CMT_COMPILER_GNU
+#if defined(CMT_ARCH_NEON)
+    asm volatile("dmb ish" ::: "memory");
+#elif defined(CMT_COMPILER_GNU)
     asm volatile("mfence" ::: "memory");
 #else
     _ReadWriteBarrier();
@@ -235,7 +237,7 @@ TEST(fft_accuracy)
 
             if (is_even(size))
             {
-                index_t csize            = dft_plan_real<float_type>::complex_size_for(size, dft_pack_format::CCs);
+                index_t csize = dft_plan_real<float_type>::complex_size_for(size, dft_pack_format::CCs);
                 univector<float_type> in = truncate(gen_random_range<float_type>(gen, -1.0, +1.0), size);
 
                 univector<complex<float_type>> out    = truncate(dimensions<1>(scalar(qnan)), csize);