Merge pull request #2073 from luyahan:fix-rvv

PiperOrigin-RevId: 623416155
google · Apr 10, 2024 · 701c328 · 701c328
2 parents b15ba1e + fcf7101
commit 701c328
Show file tree

Hide file tree

Showing 14 changed files with 31 additions and 29 deletions.
diff --git a/hwy/abort.cc b/hwy/abort.cc
@@ -60,7 +60,7 @@ HWY_DLLEXPORT HWY_NORETURN void HWY_FORMAT(3, 4)
   fflush(stderr);
 
 // Now terminate the program:
-#if HWY_ARCH_RVV
+#if HWY_ARCH_RISCV
   exit(1);  // trap/abort just freeze Spike.
 #elif HWY_IS_DEBUG_BUILD && !HWY_COMPILER_MSVC
   // Facilitates breaking into a debugger, but don't use this in non-debug

diff --git a/hwy/aligned_allocator.cc b/hwy/aligned_allocator.cc
@@ -27,7 +27,7 @@
 namespace hwy {
 namespace {
 
-#if HWY_ARCH_RVV && defined(__riscv_v_intrinsic) && __riscv_v_intrinsic >= 11000
+#if HWY_ARCH_RISCV && defined(__riscv_v_intrinsic) && __riscv_v_intrinsic >= 11000
 // Not actually an upper bound on the size, but this value prevents crossing a
 // 4K boundary (relevant on Andes).
 constexpr size_t kAlignment = HWY_MAX(HWY_ALIGNMENT, 4096);

diff --git a/hwy/base.h b/hwy/base.h
@@ -370,7 +370,7 @@ HWY_API void ZeroBytes(void* to, size_t num_bytes) {
 
 #if HWY_ARCH_X86
 static constexpr HWY_MAYBE_UNUSED size_t kMaxVectorSize = 64;  // AVX-512
-#elif HWY_ARCH_RVV && defined(__riscv_v_intrinsic) && \
+#elif HWY_ARCH_RISCV && defined(__riscv_v_intrinsic) && \
     __riscv_v_intrinsic >= 11000
 // Not actually an upper bound on the size.
 static constexpr HWY_MAYBE_UNUSED size_t kMaxVectorSize = 4096;
@@ -386,7 +386,7 @@ static constexpr HWY_MAYBE_UNUSED size_t kMaxVectorSize = 16;
 // exceed the stack size.
 #if HWY_ARCH_X86
 #define HWY_ALIGN_MAX alignas(64)
-#elif HWY_ARCH_RVV && defined(__riscv_v_intrinsic) && \
+#elif HWY_ARCH_RISCV && defined(__riscv_v_intrinsic) && \
     __riscv_v_intrinsic >= 11000
 #define HWY_ALIGN_MAX alignas(8)  // only elements need be aligned
 #else
@@ -1046,7 +1046,7 @@ HWY_API HWY_BITCASTSCALAR_CONSTEXPR To BitCastScalar(const From& val) {
 
 // RVV with f16 extension supports _Float16 and f16 vector ops. If set, implies
 // HWY_HAVE_FLOAT16.
-#if HWY_ARCH_RVV && defined(__riscv_zvfh) && HWY_COMPILER_CLANG >= 1600
+#if HWY_ARCH_RISCV && defined(__riscv_zvfh) && HWY_COMPILER_CLANG >= 1600
 #define HWY_RVV_HAVE_F16_VEC 1
 #else
 #define HWY_RVV_HAVE_F16_VEC 0

diff --git a/hwy/contrib/sort/vqsort.cc b/hwy/contrib/sort/vqsort.cc
@@ -23,7 +23,7 @@
 // unavailable on Android and non-Linux RVV, we assume that those systems lack
 // getrandom. Note that the only supported sources of entropy are getrandom or
 // Windows, thus VQSORT_SECURE_SEED=0 when this is 0 and we are not on Windows.
-#if defined(ANDROID) || defined(__ANDROID__) || (HWY_ARCH_RVV && !HWY_OS_LINUX)
+#if defined(ANDROID) || defined(__ANDROID__) || (HWY_ARCH_RISCV && !HWY_OS_LINUX)
 #define VQSORT_GETRANDOM 0
 #endif
 

diff --git a/hwy/detect_compiler_arch.h b/hwy/detect_compiler_arch.h
@@ -233,10 +233,12 @@
 #endif
 
 #ifdef __riscv
-#define HWY_ARCH_RVV 1
+#define HWY_ARCH_RISCV 1
 #else
-#define HWY_ARCH_RVV 0
+#define HWY_ARCH_RISCV 0
 #endif
+// DEPRECATED names; please use HWY_ARCH_RISCV instead.
+#define HWY_ARCH_RVV HWY_ARCH_RISCV 
 
 #if defined(__s390x__)
 #define HWY_ARCH_S390X 1
@@ -247,7 +249,7 @@
 // It is an error to detect multiple architectures at the same time, but OK to
 // detect none of the above.
 #if (HWY_ARCH_X86 + HWY_ARCH_PPC + HWY_ARCH_ARM + HWY_ARCH_ARM_OLD + \
-     HWY_ARCH_WASM + HWY_ARCH_RVV + HWY_ARCH_S390X) > 1
+     HWY_ARCH_WASM + HWY_ARCH_RISCV + HWY_ARCH_S390X) > 1
 #error "Must not detect more than one architecture"
 #endif
 

diff --git a/hwy/detect_targets.h b/hwy/detect_targets.h
@@ -505,7 +505,7 @@
 #endif
 
 // RVV requires intrinsics 0.11 or later, see #1156.
-#if HWY_ARCH_RVV && defined(__riscv_v_intrinsic) && __riscv_v_intrinsic >= 11000
+#if HWY_ARCH_RISCV && defined(__riscv_v_intrinsic) && __riscv_v_intrinsic >= 11000
 #define HWY_BASELINE_RVV HWY_RVV
 #else
 #define HWY_BASELINE_RVV 0
@@ -556,7 +556,7 @@
 // On Arm, PPC, S390X, and RISC-V: GCC and Clang 16+ do, and we require Linux
 // to detect CPU capabilities. Currently require opt-in for Clang on Arm
 // because it is experimental.
-#elif (HWY_ARCH_ARM || HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_RVV) && \
+#elif (HWY_ARCH_ARM || HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_RISCV) && \
     (HWY_COMPILER_GCC_ACTUAL ||                                           \
      (HWY_COMPILER_CLANG >= 1600 &&                                       \
       (!HWY_ARCH_ARM || defined(HWY_ENABLE_CLANG_ARM_DISPATCH)))) &&      \

diff --git a/hwy/foreach_target.h b/hwy/foreach_target.h
@@ -295,7 +295,7 @@
 #endif
 #endif
 
-// ------------------------------ HWY_ARCH_RVV
+// ------------------------------ HWY_ARCH_RISCV
 
 #if (HWY_TARGETS & HWY_RVV) && (HWY_STATIC_TARGET != HWY_RVV)
 #undef HWY_TARGET

diff --git a/hwy/ops/emu128-inl.h b/hwy/ops/emu128-inl.h
@@ -2081,7 +2081,7 @@ HWY_API VFromD<D> ConcatEven(D d, VFromD<D> hi, VFromD<D> lo) {
 
 // 2023-11-23: workaround for incorrect codegen (reduction_test fails for
 // SumsOf2 because PromoteOddTo, which uses ConcatOdd, returns zero).
-#if HWY_ARCH_RVV && HWY_TARGET == HWY_EMU128 && HWY_COMPILER_CLANG
+#if HWY_ARCH_RISCV && HWY_TARGET == HWY_EMU128 && HWY_COMPILER_CLANG
 #define HWY_EMU128_CONCAT_INLINE HWY_NOINLINE
 #else
 #define HWY_EMU128_CONCAT_INLINE HWY_API

diff --git a/hwy/targets.cc b/hwy/targets.cc
@@ -31,7 +31,7 @@
 #include <cpuid.h>
 #endif  // HWY_COMPILER_MSVC
 
-#elif (HWY_ARCH_ARM || HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_RVV) && \
+#elif (HWY_ARCH_ARM || HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_RISCV) && \
     HWY_OS_LINUX
 // sys/auxv.h does not always include asm/hwcap.h, or define HWCAP*, hence we
 // still include this directly. See #1199.
@@ -498,7 +498,7 @@ int64_t DetectTargets() {
   return bits;
 }
 }  // namespace s390x
-#elif HWY_ARCH_RVV && HWY_HAVE_RUNTIME_DISPATCH
+#elif HWY_ARCH_RISCV && HWY_HAVE_RUNTIME_DISPATCH
 namespace rvv {
 
 #ifndef HWCAP_RVV
@@ -537,7 +537,7 @@ int64_t DetectTargets() {
   bits |= ppc::DetectTargets();
 #elif HWY_ARCH_S390X && HWY_HAVE_RUNTIME_DISPATCH
   bits |= s390x::DetectTargets();
-#elif HWY_ARCH_RVV && HWY_HAVE_RUNTIME_DISPATCH
+#elif HWY_ARCH_RISCV && HWY_HAVE_RUNTIME_DISPATCH
   bits |= rvv::DetectTargets();
 
 #else

diff --git a/hwy/targets.h b/hwy/targets.h
@@ -29,7 +29,7 @@
 #include "hwy/detect_targets.h"
 #include "hwy/highway_export.h"
 
-#if !HWY_ARCH_RVV && !defined(HWY_NO_LIBCXX)
+#if !HWY_ARCH_RISCV && !defined(HWY_NO_LIBCXX)
 #include <atomic>
 #endif
 
@@ -141,7 +141,7 @@ static inline HWY_MAYBE_UNUSED const char* TargetName(int64_t target) {
       return "WASM_EMU256";
 #endif
 
-#if HWY_ARCH_RVV
+#if HWY_ARCH_RISCV
     case HWY_RVV:
       return "RVV";
 #endif
@@ -237,7 +237,7 @@ static inline HWY_MAYBE_UNUSED const char* TargetName(int64_t target) {
       HWY_CHOOSE_NEON(func_name),            /* NEON */         \
       HWY_CHOOSE_NEON_WITHOUT_AES(func_name) /* NEON without AES */
 
-#elif HWY_ARCH_RVV
+#elif HWY_ARCH_RISCV
 // See HWY_ARCH_X86 above for details.
 #define HWY_MAX_DYNAMIC_TARGETS 9
 #define HWY_HIGHEST_TARGET_BIT HWY_HIGHEST_TARGET_BIT_RVV
@@ -324,7 +324,7 @@ struct ChosenTarget {
 
  private:
   // TODO(janwas): remove RVV once <atomic> is available
-#if HWY_ARCH_RVV || defined(HWY_NO_LIBCXX)
+#if HWY_ARCH_RISCV || defined(HWY_NO_LIBCXX)
   int64_t LoadMask() const { return mask_; }
   void StoreMask(int64_t mask) { mask_ = mask; }
 
@@ -334,7 +334,7 @@ struct ChosenTarget {
   void StoreMask(int64_t mask) { mask_.store(mask); }
 
   std::atomic<int64_t> mask_{1};  // Initialized to 1 so GetIndex() returns 0.
-#endif  // HWY_ARCH_RVV
+#endif  // HWY_ARCH_RISCV
 };
 
 // For internal use (e.g. by FunctionCache and DisableTargets).

diff --git a/hwy/tests/convert_test.cc b/hwy/tests/convert_test.cc
@@ -1009,7 +1009,7 @@ class TestNonFiniteF2IConvertTo {
     // TODO(janwas): workaround for QEMU 7.2 crash on vfwcvt_rtz_x_f_v:
     // target/riscv/translate.c:213 in void decode_save_opc(DisasContext *):
     // ctx->insn_start != NULL.
-#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RVV && HWY_TARGET == HWY_EMU128)
+#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RISCV && HWY_TARGET == HWY_EMU128)
     if (sizeof(TTo) > sizeof(TF)) {
       return;
     }
@@ -1154,7 +1154,7 @@ struct TestF2IPromoteTo {
     // TODO(janwas): workaround for QEMU 7.2 crash on vfwcvt_rtz_x_f_v:
     // target/riscv/translate.c:213 in void decode_save_opc(DisasContext *):
     // ctx->insn_start != NULL.
-#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RVV && HWY_TARGET == HWY_EMU128)
+#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RISCV && HWY_TARGET == HWY_EMU128)
     return;
 #endif
 
@@ -1251,7 +1251,7 @@ struct TestF2IPromoteUpperLowerTo {
     // TODO(janwas): workaround for QEMU 7.2 crash on vfwcvt_rtz_x_f_v:
     // target/riscv/translate.c:213 in void decode_save_opc(DisasContext *):
     // ctx->insn_start != NULL.
-#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RVV && HWY_TARGET == HWY_EMU128)
+#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RISCV && HWY_TARGET == HWY_EMU128)
     return;
 #endif
 
@@ -1395,7 +1395,7 @@ class TestNonFiniteF2IPromoteUpperLowerTo {
     // TODO(janwas): workaround for QEMU 7.2 crash on vfwcvt_rtz_x_f_v:
     // target/riscv/translate.c:213 in void decode_save_opc(DisasContext *):
     // ctx->insn_start != NULL.
-#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RVV && HWY_TARGET == HWY_EMU128)
+#if HWY_TARGET == HWY_RVV || (HWY_ARCH_RISCV && HWY_TARGET == HWY_EMU128)
     return;
 #endif
 

diff --git a/hwy/tests/test_util-inl.h b/hwy/tests/test_util-inl.h
@@ -864,7 +864,7 @@ void ForUIF163264(const Func& func) {
 // For tests that involve loops, adjust the trip count so that emulated tests
 // finish quickly (but always at least 2 iterations to ensure some diversity).
 constexpr size_t AdjustedReps(size_t max_reps) {
-#if HWY_ARCH_RVV
+#if HWY_ARCH_RISCV
   return HWY_MAX(max_reps / 32, 2);
 #elif HWY_IS_DEBUG_BUILD
   return HWY_MAX(max_reps / 8, 2);
@@ -880,7 +880,7 @@ constexpr size_t AdjustedReps(size_t max_reps) {
 // Same as above, but the loop trip count will be 1 << max_pow2.
 constexpr size_t AdjustedLog2Reps(size_t max_pow2) {
   // If "negative" (unsigned wraparound), use original.
-#if HWY_ARCH_RVV
+#if HWY_ARCH_RISCV
   return HWY_MIN(max_pow2 - 4, max_pow2);
 #elif HWY_IS_DEBUG_BUILD
   return HWY_MIN(max_pow2 - 1, max_pow2);

diff --git a/hwy/timer-inl.h b/hwy/timer-inl.h
@@ -139,7 +139,7 @@ inline Ticks Start() {
       // "memory" avoids reordering. rdx = TSC >> 32.
       // "cc" = flags modified by SHL.
       : "rdx", "memory", "cc");
-#elif HWY_ARCH_RVV
+#elif HWY_ARCH_RISCV
   asm volatile("fence; rdtime %0" : "=r"(t));
 #elif defined(_WIN32) || defined(_WIN64)
   LARGE_INTEGER counter;

diff --git a/hwy/timer.cc b/hwy/timer.cc
@@ -136,7 +136,7 @@ HWY_DLLEXPORT bool HaveTimerStop(char* cpu100) {
 HWY_DLLEXPORT double InvariantTicksPerSecond() {
 #if HWY_ARCH_PPC && defined(__GLIBC__) && defined(__powerpc64__)
   return static_cast<double>(__ppc_get_timebase_freq());
-#elif HWY_ARCH_X86 || HWY_ARCH_RVV || (HWY_ARCH_ARM_A64 && !HWY_COMPILER_MSVC)
+#elif HWY_ARCH_X86 || HWY_ARCH_RISCV || (HWY_ARCH_ARM_A64 && !HWY_COMPILER_MSVC)
   // We assume the x86 TSC is invariant; it is on all recent Intel/AMD CPUs.
   static const double freq = MeasureNominalClockRate();
   return freq;