RISC-V RVV optimization (UTF-16 inputs)

.github/workflows/risc.yml

@@ -1,4 +1,4 @@
-name: Ubuntu RISC-V rvv VLEN=128 (clang 17)
+name: Ubuntu RISC-V rvv (clang 17)
 
 on: [push, pull_request]
 
@@ -14,10 +14,15 @@ jobs:
     - name: Build
       run: |
         CXX=clang++-17 CXXFLAGS="--target=riscv64-linux-gnu -march=rv64gcv" \
-        cmake --toolchain=cmake/toolchains-ci/riscv64-linux-gnu.cmake -DCMAKE_BUILD_TYPE=Release -B build
+        cmake --toolchain=cmake/toolchains-ci/riscv64-linux-gnu.cmake -DCMAKE_BUILD_TYPE=Release -DFASTFLOAT_TEST=ON -B build
         cmake --build build/ -j$(nproc)
     - name: Test VLEN=128
       run: |
         export QEMU_LD_PREFIX="/usr/riscv64-linux-gnu"
         export QEMU_CPU="rv64,v=on,vlen=128,rvv_ta_all_1s=on,rvv_ma_all_1s=on"
         ctest --timeout 1800 --output-on-failure --test-dir build -j $(nproc)
+    - name: Test VLEN=256
+      run: |
+        export QEMU_LD_PREFIX="/usr/riscv64-linux-gnu"
+        export QEMU_CPU="rv64,v=on,vlen=256,rvv_ta_all_1s=on,rvv_ma_all_1s=on"
+        ctest --timeout 1800 --output-on-failure --test-dir build -j $(nproc)

CONTRIBUTORS

@@ -9,3 +9,4 @@ Jan Pharago
 Maya Warrier
 Taha Khokhar
 Anders Dalvander
+Fenghao Li

include/fast_float/ascii_number.h

@@ -18,6 +18,10 @@
 #include <arm_neon.h>
 #endif
 
+#ifdef FASTFLOAT_RVV
+#include <riscv_vector.h>
+#endif
+
 namespace fast_float {
 
 template <typename UC> fastfloat_really_inline constexpr bool has_simd_opt() {
@@ -126,6 +130,24 @@ fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) {
   FASTFLOAT_SIMD_RESTORE_WARNINGS
 }
 
+#elif defined(FASTFLOAT_RVV)
+
+fastfloat_really_inline uint64_t simd_read8_to_u64(vuint16m1_t const data) {
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  vuint8mf2_t const packed = __riscv_vnsrl_wx_u8mf2(data, 0, 8);
+  uint64_t value;
+  __riscv_vse8_v_u8mf2(reinterpret_cast<uint8_t *>(&value), packed, 8);
+  return value;
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+}
+
+fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) {
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  return simd_read8_to_u64(
+      __riscv_vle16_v_u16m1(reinterpret_cast<uint16_t const *>(chars), 8));
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
+}
+
 #endif // FASTFLOAT_SSE2
 
 // MSVC SFINAE is broken pre-VS2017
@@ -222,6 +244,22 @@ simd_parse_if_eight_digits_unrolled(char16_t const *chars,
   } else
     return false;
   FASTFLOAT_SIMD_RESTORE_WARNINGS
+#elif defined(FASTFLOAT_RVV)
+  FASTFLOAT_SIMD_DISABLE_WARNINGS
+  vuint16m1_t const data =
+      __riscv_vle16_v_u16m1(reinterpret_cast<uint16_t const *>(chars), 8);
+
+  // (x - '0') <= 9
+  // http://0x80.pl/articles/simd-parsing-int-sequences.html
+  vuint16m1_t const t0 = __riscv_vsub_vx_u16m1(data, '0', 8);
+  vbool16_t const nondigit = __riscv_vmsgtu_vx_u16m1_b16(t0, 9, 8);
+
+  if (__riscv_vfirst_m_b16(nondigit, 8) < 0) {
+    i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
+    return true;
+  } else
+    return false;
+  FASTFLOAT_SIMD_RESTORE_WARNINGS
 #else
   static_cast<void>(chars);
   static_cast<void>(i);

include/fast_float/float_common.h

@@ -172,7 +172,14 @@ using parse_options = parse_options_t<char>;
 #define FASTFLOAT_NEON 1
 #endif
 
-#if defined(FASTFLOAT_SSE2) || defined(FASTFLOAT_NEON)
+// The RISC-V V extension guarantees VLEN >= 128. The __riscv_-prefixed
+// intrinsics used here require version 0.11 or later of the intrinsics spec.
+#if defined(__riscv_v) && defined(__riscv_v_intrinsic) &&                      \
+    __riscv_v_intrinsic >= 11000
+#define FASTFLOAT_RVV 1
+#endif
+
+#if defined(FASTFLOAT_SSE2) || defined(FASTFLOAT_NEON) || defined(FASTFLOAT_RVV)
 #define FASTFLOAT_HAS_SIMD 1
 #endif