fix(dnn/arm_common): support more monotonous case in arm typecvt for performance

GitOrigin-RevId: 9e28a64d93799a04501910f5a05a988715ff4b04

dnn/src/arm_common/type_cvt/opr_impl.cpp

@@ -11,6 +11,7 @@
 #include "src/arm_common/type_cvt/opr_impl.h"
 
 #include <cstring>
+#include <deque>
 #include "midout.h"
 #include "src/arm_common/quantized_converter.h"
 #include "src/arm_common/simd_macro/marm_neon.h"
@@ -18,6 +19,7 @@
 #include "src/naive/handle.h"
 
 MIDOUT_DECL(megdnn_arm_typecvt_fix2float)
+MIDOUT_DECL(megdnn_arm_typecvt_quan2float)
 MIDOUT_DECL(megdnn_arm_typecvt_quantized)
 MIDOUT_DECL(megdnn_arm_typecvt_float)
 
@@ -326,8 +328,34 @@ struct FloatTypeCvter<float, __fp16> {
 };
 #endif
 
+template <typename TypeCvter>
+void do_typecvt(
+        const typename TypeCvter::stype* src, typename TypeCvter::dst_type* dst,
+        DType src_dtype, DType dst_dtype, size_t nr_elems) {
+    TypeCvter typecvt(src_dtype, dst_dtype);
+    size_t i = 0;
+    for (; i + TypeCvter::SIMD_WIDTH <= nr_elems; i += TypeCvter::SIMD_WIDTH) {
+        typecvt.cvt(src, dst);
+        src += TypeCvter::SIMD_WIDTH;
+        dst += TypeCvter::SIMD_WIDTH;
+    }
+#if MEGDNN_FIX_AARCH32_BUG
+// FIXME: as llvm may cause cannot select error if enable vectorize
+#pragma clang loop vectorize(disable)
+#endif
+    for (; i < nr_elems; i++) {
+        typecvt.cvt_remain(src, dst);
+        src++;
+        dst++;
+    }
+}
+
 template <typename ctype, typename dtype>
 struct Fix2FloatTypeCvter;
+
+template <typename ctype, typename dtype>
+struct Quan2FloatTypeCvter;
+
 template <>
 struct Fix2FloatTypeCvter<int16_t, float> {
     using stype = int16_t;
@@ -368,62 +396,184 @@ struct Fix2FloatTypeCvter<uint16_t, float> {
     void cvt_remain(const uint16_t* src, float* dst) { *dst = *src; }
 };
 
-template <typename TypeCvter>
-void do_typecvt(
-        const typename TypeCvter::stype* src, typename TypeCvter::dst_type* dst,
-        DType src_dtype, DType dst_dtype, size_t nr_elems) {
-    TypeCvter typecvt(src_dtype, dst_dtype);
-    size_t i = 0;
-    for (; i + TypeCvter::SIMD_WIDTH <= nr_elems; i += TypeCvter::SIMD_WIDTH) {
-        typecvt.cvt(src, dst);
-        src += TypeCvter::SIMD_WIDTH;
-        dst += TypeCvter::SIMD_WIDTH;
+template <>
+struct Fix2FloatTypeCvter<int8_t, float> {
+    using stype = int8_t;
+    using dst_type = float;
+    static constexpr size_t SIMD_WIDTH = 16;
+
+    Fix2FloatTypeCvter(DType src_dtype, DType dst_dtype) {
+        MEGDNN_MARK_USED_VAR(src_dtype);
+        MEGDNN_MARK_USED_VAR(dst_dtype);
     }
-#if MEGDNN_FIX_AARCH32_BUG
-// FIXME: as llvm may cause cannot select error if enable vectorize
-#pragma clang loop vectorize(disable)
+
+    void cvt(const int8_t* src, float* dst) {
+        int8x16_t vitem = vld1q_s8(src);
+        int16x8_t vtrans_high = vmovl_s8(vget_high_s8(vitem));
+        int16x8_t vtrans_low = vmovl_s8(vget_low_s8(vitem));
+        auto vres_high = QConverter::convert<float32x4x2_t, int16x8_t>(vtrans_high);
+        auto vres_low = QConverter::convert<float32x4x2_t, int16x8_t>(vtrans_low);
+        vst1q_f32_x2(dst, vres_low);
+        vst1q_f32_x2(dst + 8, vres_high);
+    }
+
+    void cvt_remain(const int8_t* src, float* dst) { *dst = *src; }
+};
+
+template <>
+struct Fix2FloatTypeCvter<uint8_t, float> {
+    using stype = uint8_t;
+    using dst_type = float;
+    static constexpr size_t SIMD_WIDTH = 16;
+
+    Fix2FloatTypeCvter(DType src_dtype, DType dst_dtype) {
+        MEGDNN_MARK_USED_VAR(src_dtype);
+        MEGDNN_MARK_USED_VAR(dst_dtype);
+    }
+
+    void cvt(const uint8_t* src, float* dst) {
+        uint8x16_t vitem = vld1q_u8(src);
+        uint16x8_t vtrans_high = vmovl_u8(vget_high_u8(vitem));
+        uint16x8_t vtrans_low = vmovl_u8(vget_low_u8(vitem));
+        auto vres_high = QConverter::convert<float32x4x2_t, uint16x8_t>(vtrans_high);
+        auto vres_low = QConverter::convert<float32x4x2_t, uint16x8_t>(vtrans_low);
+        vst1q_f32_x2(dst, vres_low);
+        vst1q_f32_x2(dst + 8, vres_high);
+    }
+
+    void cvt_remain(const uint8_t* src, float* dst) { *dst = *src; }
+};
+
+template <>
+struct Quan2FloatTypeCvter<int8_t, float> {
+    using stype = int8_t;
+    using dst_type = float;
+    static constexpr size_t SIMD_WIDTH = 16;
+    float _scale = 0.0f;
+    float32x4_t _vscale;
+
+    Quan2FloatTypeCvter(DType src_dtype, DType dst_dtype) {
+        _scale = src_dtype.param<dtype::QuantizedS8>().scale;
+        _vscale = vdupq_n_f32(_scale);
+        MEGDNN_MARK_USED_VAR(dst_dtype);
+    }
+
+    void cvt(const int8_t* src, float* dst) {
+        int8x16_t vitem = vld1q_s8(src);
+        int16x8_t vtrans_high = vmovl_s8(vget_high_s8(vitem));
+        int16x8_t vtrans_low = vmovl_s8(vget_low_s8(vitem));
+        auto vres_high = QConverter::convert<float32x4x2_t, int16x8_t>(vtrans_high);
+        auto vres_low = QConverter::convert<float32x4x2_t, int16x8_t>(vtrans_low);
+        vst1q_f32(dst, vmulq_f32(vres_low.val[0], _vscale));
+        vst1q_f32(dst + 4, vmulq_f32(vres_low.val[1], _vscale));
+        vst1q_f32(dst + 8, vmulq_f32(vres_high.val[0], _vscale));
+        vst1q_f32(dst + 12, vmulq_f32(vres_high.val[1], _vscale));
+    }
+
+    void cvt_remain(const int8_t* src, float* dst) { *dst = *src * _scale; }
+};
+
+#if defined(__ARM_FEATURE_FMA)
+#define Vfmaq_f32(d, n, m) vfmaq_f32(d, n, m)
+#else
+#define Vfmaq_f32(d, n, m) vmlaq_f32(d, n, m)
 #endif
-    for (; i < nr_elems; i++) {
-        typecvt.cvt_remain(src, dst);
-        src++;
-        dst++;
+
+template <>
+struct Quan2FloatTypeCvter<uint8_t, float> {
+    using stype = uint8_t;
+    using dst_type = float;
+    static constexpr size_t SIMD_WIDTH = 16;
+    float _scale = 0.0f;
+    float32x4_t _vscale;
+    uint8_t _zp = 0;
+    float32x4_t _vbias;
+
+    Quan2FloatTypeCvter(DType src_dtype, DType dst_dtype) {
+        _scale = src_dtype.param<dtype::Quantized8Asymm>().scale;
+        _vscale = vdupq_n_f32(_scale);
+        _zp = src_dtype.param<dtype::Quantized8Asymm>().zero_point;
+        float bias = -_zp * 1.0f * _scale;
+        _vbias = vdupq_n_f32(bias);
+        MEGDNN_MARK_USED_VAR(dst_dtype);
     }
-}
+
+    void cvt(const uint8_t* src, float* dst) {
+        uint8x16_t vitem = vld1q_u8(src);
+        uint16x8_t vtrans_high = vmovl_u8(vget_high_u8(vitem));
+        uint16x8_t vtrans_low = vmovl_u8(vget_low_u8(vitem));
+        auto vres_high = QConverter::convert<float32x4x2_t, uint16x8_t>(vtrans_high);
+        auto vres_low = QConverter::convert<float32x4x2_t, uint16x8_t>(vtrans_low);
+        vst1q_f32(dst, Vfmaq_f32(_vbias, vres_low.val[0], _vscale));
+        vst1q_f32(dst + 4, Vfmaq_f32(_vbias, vres_low.val[1], _vscale));
+        vst1q_f32(dst + 8, Vfmaq_f32(_vbias, vres_high.val[0], _vscale));
+        vst1q_f32(dst + 12, Vfmaq_f32(_vbias, vres_high.val[1], _vscale));
+    }
+
+    void cvt_remain(const uint8_t* src, float* dst) { *dst = (*src - _zp) * _scale; }
+};
+
+#undef Vfmaq_f32
+
+template <typename stype, typename dtype>
+struct TypeCvtTask {
+    const stype* src;
+    dtype* dst;
+    size_t dim;
+    size_t nr_elems;
+
+    explicit TypeCvtTask(const stype* s, dtype* d, size_t n, size_t tot)
+            : src(s), dst(d), dim(n), nr_elems(tot) {}
+};
 
 template <typename TypeCvter>
 void do_typecvt(
         const typename TypeCvter::stype* src, typename TypeCvter::dst_type* dst,
         DType src_dtype, DType dst_dtype, const TensorLayout& src_layout) {
     TypeCvter typecvt(src_dtype, dst_dtype);
-    size_t calc_num = 1;
-    size_t nr_elems = src_layout.total_nr_elems();
-    size_t src_stride = nr_elems;
-
-    //! adjust calc_num nr_elems and src_stride according to src_collapse_layout
     auto src_collapse_layout = src_layout.collapse_contiguous();
-    if (src_collapse_layout.ndim == 2) {
-        calc_num = src_collapse_layout.shape[0];
-        nr_elems = src_collapse_layout.shape[1];
-        src_stride = src_collapse_layout.stride[0];
-    }
-
-    for (size_t c = 0; c < calc_num; ++c) {
-        size_t i = 0;
-        for (; i + TypeCvter::SIMD_WIDTH <= nr_elems; i += TypeCvter::SIMD_WIDTH) {
-            typecvt.cvt(src, dst);
-            src += TypeCvter::SIMD_WIDTH;
-            dst += TypeCvter::SIMD_WIDTH;
-        }
+
+    using TypeCvtTaskWithType =
+            TypeCvtTask<typename TypeCvter::stype, typename TypeCvter::dst_type>;
+    std::deque<TypeCvtTaskWithType> task_queue;
+    task_queue.emplace_back(src, dst, 0, src_collapse_layout.total_nr_elems());
+
+    while (!task_queue.empty()) {
+        auto&& task = task_queue.front();
+        const typename TypeCvter::stype* psrc = task.src;
+        typename TypeCvter::dst_type* pdst = task.dst;
+        size_t dim = task.dim;
+        size_t nr_elems = task.nr_elems;
+
+        //! calc according to stride information
+        if (src_collapse_layout.stride[dim] == 1) {
+            size_t i = 0;
+            for (; i + TypeCvter::SIMD_WIDTH < nr_elems; i += TypeCvter::SIMD_WIDTH) {
+                typecvt.cvt(psrc, pdst);
+                psrc += TypeCvter::SIMD_WIDTH;
+                pdst += TypeCvter::SIMD_WIDTH;
+            }
 #if MEGDNN_FIX_AARCH32_BUG
 // FIXME: as llvm may cause cannot select error if enable vectorize
 #pragma clang loop vectorize(disable)
 #endif
-        for (; i < nr_elems; i++) {
-            typecvt.cvt_remain(src, dst);
-            src++;
-            dst++;
+            for (; i < nr_elems; i++) {
+                typecvt.cvt_remain(psrc, pdst);
+                psrc++;
+                pdst++;
+            }
+        } else {
+            size_t calc_num = src_collapse_layout.shape[dim];
+            size_t src_stride = src_collapse_layout.stride[dim];
+            size_t dst_stride = nr_elems / calc_num;
+            for (size_t i = 0; i < calc_num; ++i) {
+                task_queue.emplace_back(psrc, pdst, dim + 1, dst_stride);
+                psrc += src_stride;
+                pdst += dst_stride;
+            }
         }
-        src += src_stride - nr_elems;
+
+        task_queue.pop_front();
     }
 }
 
@@ -432,15 +582,56 @@ void do_typecvt(
 void TypeCvtImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_out dst) {
     DType src_dtype = src.layout.dtype;
     DType dst_dtype = dst.layout.dtype;
-    size_t nr_elems = src.layout.total_nr_elems();
     bool execed = false;
     auto src_collapse_layout = src.layout.collapse_contiguous();
-    bool has_int16_special_impl =
-            (src.layout.dtype.enumv() == DTypeEnum::Int16 ||
-             src.layout.dtype.enumv() == DTypeEnum::Uint16) &&
-            (src.layout.is_contiguous() || src_collapse_layout.ndim == 2) &&
-            dst.layout.is_contiguous();
-    if (has_int16_special_impl) {
+
+    if (src.layout.is_contiguous()) {
+        using namespace dtype;
+        size_t nr_elems = src.layout.total_nr_elems();
+#define DISPATCH_QUANTIZED(_stype_enumv, _stype, _dtype_enumv, _dtype, _midout_iv) \
+    if (src_dtype.enumv() == DTypeTrait<_stype_enumv>::enumv &&                    \
+        dst_dtype.enumv() == DTypeTrait<_dtype_enumv>::enumv) {                    \
+        MIDOUT_BEGIN(megdnn_arm_typecvt_quantized, midout_iv(_midout_iv)) {        \
+            using _TypeCvter = QuantizedTypeCvter<_stype, _dtype>;                 \
+            MEGDNN_DISPATCH_CPU_KERN_OPR(do_typecvt<_TypeCvter>(                   \
+                    src.compatible_ptr<_stype>(), dst.compatible_ptr<_dtype>(),    \
+                    src_dtype, dst_dtype, nr_elems));                              \
+            execed = true;                                                         \
+        }                                                                          \
+        MIDOUT_END();                                                              \
+    }
+        DISPATCH_QUANTIZED(QuantizedS32, int32_t, Quantized8Asymm, uint8_t, 0);
+        DISPATCH_QUANTIZED(QuantizedS32, int32_t, QuantizedS8, int8_t, 1);
+        DISPATCH_QUANTIZED(QuantizedS8, int8_t, QuantizedS32, int32_t, 2);
+        DISPATCH_QUANTIZED(QuantizedS8, int8_t, QuantizedS8, int8_t, 3);
+        DISPATCH_QUANTIZED(Quantized8Asymm, uint8_t, Quantized8Asymm, uint8_t, 4);
+        DISPATCH_QUANTIZED(QuantizedS32, int32_t, QuantizedS32, int32_t, 5);
+        DISPATCH_QUANTIZED(float, float, QuantizedS8, int8_t, 6);
+        DISPATCH_QUANTIZED(float, float, Quantized8Asymm, uint8_t, 7);
+#undef DISPATCH_QUANTIZED
+
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+#define DISPATCH_FLOAT(_stype_enumv, _stype, _dtype_enumv, _dtype, _midout_iv)      \
+    if (src_dtype.enumv() == DTypeTrait<_stype_enumv>::enumv &&                     \
+        dst_dtype.enumv() == DTypeTrait<_dtype_enumv>::enumv) {                     \
+        MIDOUT_BEGIN(megdnn_arm_typecvt_float, midout_iv(_midout_iv)) {             \
+            using _TypeCvter = FloatTypeCvter<_stype, _dtype>;                      \
+            MEGDNN_DISPATCH_CPU_KERN_OPR(do_typecvt<_TypeCvter>(                    \
+                    reinterpret_cast<_stype*>(src.raw_ptr()),                       \
+                    reinterpret_cast<_dtype*>(dst.raw_ptr()), src_dtype, dst_dtype, \
+                    nr_elems));                                                     \
+            execed = true;                                                          \
+        }                                                                           \
+        MIDOUT_END();                                                               \
+    }
+        DISPATCH_FLOAT(dt_float16, __fp16, float, float, 0);
+        DISPATCH_FLOAT(float, float, dt_float16, __fp16, 1);
+#undef DISPATCH_FLOAT
+#endif
+    }
+
+    size_t last_stride = src_collapse_layout.stride[src_collapse_layout.ndim - 1];
+    if (!execed && last_stride == 1 && dst.layout.is_contiguous()) {
         using namespace dtype;
 #define DISPATCH_FIX2FLOAT(_stype_enumv, _stype, _dtype_enumv, _dtype, _midout_iv) \
     if (src_dtype.enumv() == DTypeTrait<_stype_enumv>::enumv &&                    \
@@ -456,9 +647,26 @@ void TypeCvtImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_out dst) {
     }
         DISPATCH_FIX2FLOAT(Int16, int16_t, Float32, float, 0);
         DISPATCH_FIX2FLOAT(Uint16, uint16_t, Float32, float, 1);
+        DISPATCH_FIX2FLOAT(Int8, int8_t, Float32, float, 2);
+        DISPATCH_FIX2FLOAT(Uint8, uint8_t, Float32, float, 3);
 #undef DISPATCH_FIX2FLOAT
-    } else if (src.layout.is_contiguous()) {
-        using namespace dtype;
+
+#define DISPATCH_QUAN2FLOAT(_stype_enumv, _stype, _dtype_enumv, _dtype, _midout_iv) \
+    if (src_dtype.enumv() == DTypeTrait<_stype_enumv>::enumv &&                     \
+        dst_dtype.enumv() == DTypeTrait<_dtype_enumv>::enumv) {                     \
+        MIDOUT_BEGIN(megdnn_arm_typecvt_quan2float, midout_iv(_midout_iv)) {        \
+            using _TypeCvter = Quan2FloatTypeCvter<_stype, _dtype>;                 \
+            MEGDNN_DISPATCH_CPU_KERN_OPR(do_typecvt<_TypeCvter>(                    \
+                    src.compatible_ptr<_stype>(), dst.compatible_ptr<_dtype>(),     \
+                    src_dtype, dst_dtype, src.layout));                             \
+            execed = true;                                                          \
+        }                                                                           \
+        MIDOUT_END();                                                               \
+    }
+        DISPATCH_QUAN2FLOAT(QuantizedS8, int8_t, Float32, float, 0);
+        DISPATCH_QUAN2FLOAT(Quantized8Asymm, uint8_t, Float32, float, 1);
+#undef DISPATCH_QUAN2FLOAT
+
 #define DISPATCH_QUANTIZED(_stype_enumv, _stype, _dtype_enumv, _dtype, _midout_iv) \
     if (src_dtype.enumv() == DTypeTrait<_stype_enumv>::enumv &&                    \
         dst_dtype.enumv() == DTypeTrait<_dtype_enumv>::enumv) {                    \
@@ -466,12 +674,11 @@ void TypeCvtImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_out dst) {
             using _TypeCvter = QuantizedTypeCvter<_stype, _dtype>;                 \
             MEGDNN_DISPATCH_CPU_KERN_OPR(do_typecvt<_TypeCvter>(                   \
                     src.compatible_ptr<_stype>(), dst.compatible_ptr<_dtype>(),    \
-                    src_dtype, dst_dtype, nr_elems));                              \
+                    src_dtype, dst_dtype, src.layout));                            \
             execed = true;                                                         \
         }                                                                          \
         MIDOUT_END();                                                              \
     }
-
         DISPATCH_QUANTIZED(QuantizedS32, int32_t, Quantized8Asymm, uint8_t, 0);
         DISPATCH_QUANTIZED(QuantizedS32, int32_t, QuantizedS8, int8_t, 1);
         DISPATCH_QUANTIZED(QuantizedS8, int8_t, QuantizedS32, int32_t, 2);
@@ -491,7 +698,7 @@ void TypeCvtImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_out dst) {
             MEGDNN_DISPATCH_CPU_KERN_OPR(do_typecvt<_TypeCvter>(                    \
                     reinterpret_cast<_stype*>(src.raw_ptr()),                       \
                     reinterpret_cast<_dtype*>(dst.raw_ptr()), src_dtype, dst_dtype, \
-                    nr_elems));                                                     \
+                    src.layout));                                                   \
             execed = true;                                                          \
         }                                                                           \
         MIDOUT_END();                                                               \
@@ -501,6 +708,7 @@ void TypeCvtImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_out dst) {
 #undef DISPATCH_FLOAT
 #endif
     }
+
     if (!execed) {
         fallback::TypeCvtImpl::exec(src, dst);
     }

dnn/test/arm_common/type_cvt.cpp

@@ -161,24 +161,251 @@ TEST_F(ARM_COMMON, TYPE_CVT_RECORD) {
             .execs({{1, 32, 24, 128}, {1, 32, 24, 128}});
 }
 
-TEST_F(ARM_COMMON, TYPE_CVT_16_F32) {
+TEST_F(ARM_COMMON, TYPE_CVT_NONCONTIGUOUS) {
+    UniformIntRNG rng32{INT32_MIN >> 1, INT32_MAX >> 1};
+    UniformIntRNG rng16{INT16_MIN >> 1, INT16_MAX >> 1};
+    UniformIntRNG rng8{INT8_MIN >> 1, INT8_MAX >> 1};
+
     Checker<TypeCvt> checker(handle());
-    UniformIntRNG rng{INT16_MIN >> 1, INT16_MAX >> 1};
 
+    size_t N = 1;
+    size_t C = 96;
+    size_t H = 64;
+    size_t W = 120;
+    TensorShape shape{N, C, H, W};
+    std::vector<ptrdiff_t> stride{
+            static_cast<long>(C * H * (W + 8)), static_cast<long>(H * (W + 8)),
+            static_cast<long>(W + 8), 1};
+    TensorLayout src, dst;
+
+    //! float32 -> float16
+    src = TensorLayout{shape, stride, dtype::Float32()};
+    dst = TensorLayout{shape, dtype::Float16()};
+    checker.execl({src, dst});
+
+    //! float16 -> float32
+    src = TensorLayout{shape, stride, dtype::Float16()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    //! float -> s8
+    src = TensorLayout{shape, stride, dtype::Float32()};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.245121f)};
+    checker.execl({src, dst});
+
+    //! float -> as8
+    src = TensorLayout{shape, stride, dtype::Float32()};
+    dst = TensorLayout{shape, dtype::Quantized8Asymm(0.1f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    checker.set_rng(0, &rng32);
+    //! s32 -> as8
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0000113264f)};
+    dst = TensorLayout{
+            shape, dtype::Quantized8Asymm(0.018909f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0003f)};
+    dst = TensorLayout{shape, dtype::Quantized8Asymm(0.1f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    //! s32 -> s8
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.000815917f)};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.245121f)};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0003f)};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.2f)};
+    checker.execl({src, dst});
+
+    checker.set_rng(0, &rng8);
+    //! s32 -> s32
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0004f)};
+    dst = TensorLayout{shape, dtype::QuantizedS32(0.0002f)};
+    checker.execl({src, dst});
+
+    //! s8 -> s8
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.3f)};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.2f)};
+    checker.execl({src, dst});
+
+    //! as8 -> as8
+    src = TensorLayout{
+            shape, stride, dtype::Quantized8Asymm(0.3f, static_cast<uint8_t>(8))};
+    dst = TensorLayout{shape, dtype::Quantized8Asymm(0.1f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    //! s8 -> s32
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.245121f)};
+    dst = TensorLayout{shape, dtype::QuantizedS32(0.000815917f)};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.2f)};
+    dst = TensorLayout{shape, dtype::QuantizedS32(0.0003f)};
+    checker.execl({src, dst});
+
+    //! s8 -> float
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.3f)};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    //! as8 -> float
+    src = TensorLayout{
+            shape, stride, dtype::Quantized8Asymm(0.3f, static_cast<uint8_t>(8))};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    //! int8/uint8 -> float
+    src = TensorLayout{shape, stride, dtype::Int8()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::Uint8()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    //! int16/uint16 -> float
+    checker.set_rng(0, &rng16);
     for (size_t size : {3, 7, 15, 33, 10000}) {
-        checker.set_rng(0, &rng);
         checker.set_dtype(0, dtype::Int16()).execs({{size}, {size}});
         checker.set_dtype(0, dtype::Uint16()).execs({{size}, {size}});
     }
-    TensorLayout src_int16{
-            {1, 96, 64, 120}, {128 * 64 * 96, 128 * 64, 128, 1}, dtype::Int16()};
-    TensorLayout dst_int16{{1, 96, 64, 120}, dtype::Float32()};
-    checker.execl({src_int16, dst_int16});
-
-    TensorLayout src_uint16{
-            {1, 96, 64, 120}, {128 * 64 * 96, 128 * 64, 128, 1}, dtype::Uint16()};
-    TensorLayout dst_uint16{{1, 96, 64, 120}, dtype::Float32()};
-    checker.execl({src_uint16, dst_uint16});
+
+    src = TensorLayout{shape, stride, dtype::Int16()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::Uint16()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    UniformIntRNG narrow_rng{-40000, 40000};
+    checker.set_rng(0, &narrow_rng);
+    //! s32 -> as8
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.000163794f)};
+    dst = TensorLayout{
+            shape, dtype::Quantized8Asymm(0.0479196f, static_cast<uint8_t>(144))};
+    checker.execl({src, dst});
+}
+
+TEST_F(ARM_COMMON, TYPE_CVT_MONOTONOUS) {
+    UniformIntRNG rng32{INT32_MIN >> 1, INT32_MAX >> 1};
+    UniformIntRNG rng16{INT16_MIN >> 1, INT16_MAX >> 1};
+    UniformIntRNG rng8{INT8_MIN >> 1, INT8_MAX >> 1};
+
+    Checker<TypeCvt> checker(handle());
+
+    size_t N = 1;
+    size_t C = 96;
+    size_t H = 64;
+    size_t W = 120;
+    TensorShape shape{N, C, H, W};
+    std::vector<ptrdiff_t> stride{
+            static_cast<long>((C + 8) * (H + 8) * (W + 8)),
+            static_cast<long>((H + 8) * (W + 8)), static_cast<long>(W + 8), 1};
+    TensorLayout src, dst;
+
+    //! float32 -> float16
+    src = TensorLayout{shape, stride, dtype::Float32()};
+    dst = TensorLayout{shape, dtype::Float16()};
+    checker.execl({src, dst});
+
+    //! float16 -> float32
+    src = TensorLayout{shape, stride, dtype::Float16()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    //! float -> s8
+    src = TensorLayout{shape, stride, dtype::Float32()};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.245121f)};
+    checker.execl({src, dst});
+
+    //! float -> as8
+    src = TensorLayout{shape, stride, dtype::Float32()};
+    dst = TensorLayout{shape, dtype::Quantized8Asymm(0.1f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    checker.set_rng(0, &rng32);
+    //! s32 -> as8
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0000113264f)};
+    dst = TensorLayout{
+            shape, dtype::Quantized8Asymm(0.018909f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0003f)};
+    dst = TensorLayout{shape, dtype::Quantized8Asymm(0.1f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    //! s32 -> s8
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.000815917f)};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.245121f)};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0003f)};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.2f)};
+    checker.execl({src, dst});
+
+    checker.set_rng(0, &rng8);
+    //! s32 -> s32
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.0004f)};
+    dst = TensorLayout{shape, dtype::QuantizedS32(0.0002f)};
+    checker.execl({src, dst});
+
+    //! s8 -> s8
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.3f)};
+    dst = TensorLayout{shape, dtype::QuantizedS8(0.2f)};
+    checker.execl({src, dst});
+
+    //! as8 -> as8
+    src = TensorLayout{
+            shape, stride, dtype::Quantized8Asymm(0.3f, static_cast<uint8_t>(8))};
+    dst = TensorLayout{shape, dtype::Quantized8Asymm(0.1f, static_cast<uint8_t>(3))};
+    checker.execl({src, dst});
+
+    //! s8 -> s32
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.245121f)};
+    dst = TensorLayout{shape, dtype::QuantizedS32(0.000815917f)};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.2f)};
+    dst = TensorLayout{shape, dtype::QuantizedS32(0.0003f)};
+    checker.execl({src, dst});
+
+    //! s8 -> float
+    src = TensorLayout{shape, stride, dtype::QuantizedS8(0.3f)};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    //! as8 -> float
+    src = TensorLayout{
+            shape, stride, dtype::Quantized8Asymm(0.3f, static_cast<uint8_t>(8))};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    //! int8/uint8 -> float
+    src = TensorLayout{shape, stride, dtype::Int8()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::Uint8()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::Int16()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    src = TensorLayout{shape, stride, dtype::Uint16()};
+    dst = TensorLayout{shape, dtype::Float32()};
+    checker.execl({src, dst});
+
+    UniformIntRNG narrow_rng{-40000, 40000};
+    checker.set_rng(0, &narrow_rng);
+    //! s32 -> as8
+    src = TensorLayout{shape, stride, dtype::QuantizedS32(0.000163794f)};
+    dst = TensorLayout{
+            shape, dtype::Quantized8Asymm(0.0479196f, static_cast<uint8_t>(144))};
+    checker.execl({src, dst});
 }
 
 #if MEGDNN_WITH_BENCHMARK