feat(dnn): add elemwise multi type support i16xf32 and u8xf32

GitOrigin-RevId: 2fe469bb4e

dnn/scripts/opr_param_defs.py

@@ -497,7 +497,16 @@ pdef('ElemwiseMultiType').add_enum(
     Doc('QCOND_LEQ_MOV = 50', 'quantized cond_leq_mov'),
     Doc('QH_SWISH = 51', 'quantized h_swish'),
     Doc('QFUSE_ADD_H_SWISH = 52', 'quantized h_swish(x+y)'),
-    Doc('QH_SWISH_GRAD = 53', 'quantized h_swish_grad')
+    Doc('QH_SWISH_GRAD = 53', 'quantized h_swish_grad'),
+    Doc('FUSE_MUL_ADD3_INT16xF32xF32xF32 = 54',
+        'compute ``a * b + c`` requiring that ``a`` be int16 and ``b`` and '
+        '``c``  float32, and the result is float32.'),
+    Doc('MUL_INT16xF32xF32 = 55',
+        'compute ``a * b `` requiring that ``a`` be int16 and ``b`` float32, '
+        'and the result is float32.'),
+    Doc('FUSE_MUL_ADD3_UINT8xF32xF32xF32 = 56',
+        'compute ``a * b + c`` requiring that ``a`` be uint8 and ``b`` and '
+        '``c``  float32, and the result is float32.')
 )
 
 pdef('PowC', 'power with constant exponent').add_fields('float32', 'exp', 0)

dnn/src/arm_common/elemwise_multi_type/kernels.cpp

@@ -0,0 +1,707 @@
+/**
+ * \file dnn/src/arm_common/elemwise_multi_type/kernels.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "kernels.h"
+#include "src/arm_common/simd_macro/marm_neon.h"
+
+namespace megdnn {
+namespace arm_common {
+
+#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
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_bcast111c_bcast111c(
+        size_t batch_size, size_t channel_stride, size_t channel_size,
+        const int16_t* src0, const float* src1, const float* src2, float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    float* __restrict dst_ptr = dst;
+    for (size_t batch = 0; batch < batch_size; ++batch) {
+        for (size_t s = 0; s < channel_stride; ++s) {
+            size_t i = 0;
+            for (; i + 15 < channel_size; i += 16, sptr0 += 16, dst_ptr += 16) {
+                auto vec0_01 = vld1q_s16(sptr0);
+                auto vec0_23 = vld1q_s16(sptr0 + 8);
+                auto vec1_0 = vld1q_f32(sptr1 + i);
+                auto vec1_1 = vld1q_f32(sptr1 + i + 4);
+                auto vec1_2 = vld1q_f32(sptr1 + i + 8);
+                auto vec1_3 = vld1q_f32(sptr1 + i + 12);
+                auto vec2_0 = vld1q_f32(sptr2 + i);
+                auto vec2_1 = vld1q_f32(sptr2 + i + 4);
+                auto vec2_2 = vld1q_f32(sptr2 + i + 8);
+                auto vec2_3 = vld1q_f32(sptr2 + i + 12);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+                auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+                auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+                auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+                auto dst_vec_2 = Vfmaq_f32(vec2_2, vec0_2, vec1_2);
+                auto dst_vec_3 = Vfmaq_f32(vec2_3, vec0_3, vec1_3);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+                vst1q_f32(dst_ptr + 8, dst_vec_2);
+                vst1q_f32(dst_ptr + 12, dst_vec_3);
+            }
+            for (; i + 7 < channel_size; i += 8, sptr0 += 8, dst_ptr += 8) {
+                auto vec0_01 = vld1q_s16(sptr0);
+                auto vec1_0 = vld1q_f32(sptr1 + i);
+                auto vec1_1 = vld1q_f32(sptr1 + i + 4);
+                auto vec2_0 = vld1q_f32(sptr2 + i);
+                auto vec2_1 = vld1q_f32(sptr2 + i + 4);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+                auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+            }
+            for (; i + 3 < channel_size; i += 4, sptr0 += 4, dst_ptr += 4) {
+                auto vec0_0 = vld1_s16(sptr0);
+                auto vec1_0 = vld1q_f32(sptr1 + i);
+                auto vec2_0 = vld1q_f32(sptr2 + i);
+
+                auto vec0_0_f32 = vcvtq_f32_s32(vmovl_s16(vec0_0));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0_f32, vec1_0);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+            }
+            for (; i < channel_size; ++i, ++sptr0, ++dst_ptr) {
+                *dst_ptr = (float)(*sptr0) * sptr1[i] + sptr2[i];
+            }
+        }
+    }
+}
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_bcast111c_bcast111c(
+        size_t batch_size, size_t channel_stride, size_t channel_size,
+        const uint8_t* src0, const float* src1, const float* src2, float* dst) {
+    const uint8_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    float* __restrict dst_ptr = dst;
+    for (size_t batch = 0; batch < batch_size; ++batch) {
+        for (size_t s = 0; s < channel_stride; ++s) {
+            size_t i = 0;
+            for (; i + 15 < channel_size; i += 16, sptr0 += 16, dst_ptr += 16) {
+                auto vec0_0123_u8 = vld1q_u8(sptr0);
+                auto vec1_0 = vld1q_f32(sptr1 + i);
+                auto vec1_1 = vld1q_f32(sptr1 + i + 4);
+                auto vec1_2 = vld1q_f32(sptr1 + i + 8);
+                auto vec1_3 = vld1q_f32(sptr1 + i + 12);
+                auto vec2_0 = vld1q_f32(sptr2 + i);
+                auto vec2_1 = vld1q_f32(sptr2 + i + 4);
+                auto vec2_2 = vld1q_f32(sptr2 + i + 8);
+                auto vec2_3 = vld1q_f32(sptr2 + i + 12);
+
+                auto vec0_01 =
+                        vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vec0_0123_u8)));
+                auto vec0_23 =
+                        vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(vec0_0123_u8)));
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+                auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+                auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+                auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+                auto dst_vec_2 = Vfmaq_f32(vec2_2, vec0_2, vec1_2);
+                auto dst_vec_3 = Vfmaq_f32(vec2_3, vec0_3, vec1_3);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+                vst1q_f32(dst_ptr + 8, dst_vec_2);
+                vst1q_f32(dst_ptr + 12, dst_vec_3);
+            }
+            for (; i + 7 < channel_size; i += 8, sptr0 += 8, dst_ptr += 8) {
+                auto vec0_01_u8 = vld1_u8(sptr0);
+                auto vec1_0 = vld1q_f32(sptr1 + i);
+                auto vec1_1 = vld1q_f32(sptr1 + i + 4);
+                auto vec2_0 = vld1q_f32(sptr2 + i);
+                auto vec2_1 = vld1q_f32(sptr2 + i + 4);
+
+                auto vec0_01 = vreinterpretq_s16_u16(vmovl_u8(vec0_01_u8));
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+                auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+            }
+            for (; i < channel_size; ++i, ++sptr0, ++dst_ptr) {
+                *dst_ptr = (float)(*sptr0) * sptr1[i] + sptr2[i];
+            }
+        }
+    }
+}
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_bcast101_bcast101(
+        size_t batch_size, size_t channel_size, size_t channel_stride,
+        const int16_t* src0, const float* src1, const float* src2, float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    float* __restrict dst_ptr = dst;
+    for (size_t batch = 0; batch < batch_size; ++batch) {
+        for (size_t chan = 0; chan < channel_size; ++chan) {
+            auto vec1 = vdupq_n_f32(sptr1[chan]);
+            auto vec2 = vdupq_n_f32(sptr2[chan]);
+            size_t i = 0;
+            for (; i + 15 < channel_stride; i += 16, sptr0 += 16, dst_ptr += 16) {
+                auto vec0_01 = vld1q_s16(sptr0);
+                auto vec0_23 = vld1q_s16(sptr0 + 8);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+                auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+                auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+                auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+                auto dst_vec_2 = Vfmaq_f32(vec2, vec0_2, vec1);
+                auto dst_vec_3 = Vfmaq_f32(vec2, vec0_3, vec1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+                vst1q_f32(dst_ptr + 8, dst_vec_2);
+                vst1q_f32(dst_ptr + 12, dst_vec_3);
+            }
+            for (; i + 7 < channel_stride; i += 8, sptr0 += 8, dst_ptr += 8) {
+                auto vec0_01 = vld1q_s16(sptr0);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+                auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+            }
+            for (; i + 3 < channel_stride; i += 4, sptr0 += 4, dst_ptr += 4) {
+                auto vec0_0 = vld1_s16(sptr0);
+                auto vec0_0_f32 = vcvtq_f32_s32(vmovl_s16(vec0_0));
+                auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0_f32, vec1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+            }
+            for (; i < channel_stride; ++i, ++sptr0, ++dst_ptr) {
+                *dst_ptr = (float)(*sptr0) * sptr1[chan] + sptr2[chan];
+            }
+        }
+    }
+}
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_bcast101_bcast101(
+        size_t batch_size, size_t channel_size, size_t channel_stride,
+        const uint8_t* src0, const float* src1, const float* src2, float* dst) {
+    const uint8_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    float* __restrict dst_ptr = dst;
+    for (size_t batch = 0; batch < batch_size; ++batch) {
+        for (size_t chan = 0; chan < channel_size; ++chan) {
+            auto vec1 = vdupq_n_f32(sptr1[chan]);
+            auto vec2 = vdupq_n_f32(sptr2[chan]);
+            size_t i = 0;
+            for (; i + 15 < channel_stride; i += 16, sptr0 += 16, dst_ptr += 16) {
+                auto vec0_0123_u8 = vld1q_u8(sptr0);
+
+                auto vec0_01 =
+                        vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vec0_0123_u8)));
+                auto vec0_23 =
+                        vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(vec0_0123_u8)));
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+                auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+                auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+                auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+                auto dst_vec_2 = Vfmaq_f32(vec2, vec0_2, vec1);
+                auto dst_vec_3 = Vfmaq_f32(vec2, vec0_3, vec1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+                vst1q_f32(dst_ptr + 8, dst_vec_2);
+                vst1q_f32(dst_ptr + 12, dst_vec_3);
+            }
+            for (; i + 7 < channel_stride; i += 8, sptr0 += 8, dst_ptr += 8) {
+                auto vec0_01_u8 = vld1_u8(sptr0);
+                auto vec0_01 = vreinterpretq_s16_u16(vmovl_u8(vec0_01_u8));
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+                auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+                auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+            }
+            for (; i < channel_stride; ++i, ++sptr0, ++dst_ptr) {
+                *dst_ptr = (float)(*sptr0) * sptr1[chan] + sptr2[chan];
+            }
+        }
+    }
+}
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_vec_vec(
+        size_t size, const int16_t* src0, const float* src1, const float* src2,
+        float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    float* __restrict dst_ptr = dst;
+    size_t i = 0;
+    for (; i + 15 < size;
+         i += 16, sptr0 += 16, sptr1 += 16, sptr2 += 16, dst_ptr += 16) {
+        auto vec0_01 = vld1q_s16(sptr0);
+        auto vec0_23 = vld1q_s16(sptr0 + 8);
+        auto vec1_0 = vld1q_f32(sptr1);
+        auto vec1_1 = vld1q_f32(sptr1 + 4);
+        auto vec1_2 = vld1q_f32(sptr1 + 8);
+        auto vec1_3 = vld1q_f32(sptr1 + 12);
+        auto vec2_0 = vld1q_f32(sptr2);
+        auto vec2_1 = vld1q_f32(sptr2 + 4);
+        auto vec2_2 = vld1q_f32(sptr2 + 8);
+        auto vec2_3 = vld1q_f32(sptr2 + 12);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+        auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+        auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+        auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+        auto dst_vec_2 = Vfmaq_f32(vec2_2, vec0_2, vec1_2);
+        auto dst_vec_3 = Vfmaq_f32(vec2_3, vec0_3, vec1_3);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+        vst1q_f32(dst_ptr + 8, dst_vec_2);
+        vst1q_f32(dst_ptr + 12, dst_vec_3);
+    }
+    for (; i + 7 < size; i += 8, sptr0 += 8, sptr1 += 8, sptr2 += 8, dst_ptr += 8) {
+        auto vec0_01 = vld1q_s16(sptr0);
+        auto vec1_0 = vld1q_f32(sptr1);
+        auto vec1_1 = vld1q_f32(sptr1 + 4);
+        auto vec2_0 = vld1q_f32(sptr2);
+        auto vec2_1 = vld1q_f32(sptr2 + 4);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+        auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+    }
+    for (; i + 3 < size; i += 4, sptr0 += 4, sptr1 += 4, sptr2 += 4, dst_ptr += 4) {
+        auto vec0_0 = vld1_s16(sptr0);
+        auto vec1_0 = vld1q_f32(sptr1);
+        auto vec2_0 = vld1q_f32(sptr2);
+
+        auto vec0_0_f32 = vcvtq_f32_s32(vmovl_s16(vec0_0));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0_f32, vec1_0);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+    }
+    for (; i < size; ++i, ++sptr0, ++sptr1, ++sptr2, ++dst_ptr) {
+        *dst_ptr = (float)(*sptr0) * (*sptr1) + (*sptr2);
+    }
+}
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_vec_vec(
+        size_t size, const uint8_t* src0, const float* src1, const float* src2,
+        float* dst) {
+    const uint8_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    float* __restrict dst_ptr = dst;
+    size_t i = 0;
+    for (; i + 15 < size;
+         i += 16, sptr0 += 16, sptr1 += 16, sptr2 += 16, dst_ptr += 16) {
+        auto vec0_0123 = vld1q_u8(sptr0);
+        auto vec1_0 = vld1q_f32(sptr1);
+        auto vec1_1 = vld1q_f32(sptr1 + 4);
+        auto vec1_2 = vld1q_f32(sptr1 + 8);
+        auto vec1_3 = vld1q_f32(sptr1 + 12);
+        auto vec2_0 = vld1q_f32(sptr2);
+        auto vec2_1 = vld1q_f32(sptr2 + 4);
+        auto vec2_2 = vld1q_f32(sptr2 + 8);
+        auto vec2_3 = vld1q_f32(sptr2 + 12);
+
+        auto vec0_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vec0_0123)));
+        auto vec0_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(vec0_0123)));
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+        auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+        auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+        auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+        auto dst_vec_2 = Vfmaq_f32(vec2_2, vec0_2, vec1_2);
+        auto dst_vec_3 = Vfmaq_f32(vec2_3, vec0_3, vec1_3);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+        vst1q_f32(dst_ptr + 8, dst_vec_2);
+        vst1q_f32(dst_ptr + 12, dst_vec_3);
+    }
+
+    for (; i + 7 < size; i += 8, sptr0 += 8, sptr1 += 8, sptr2 += 8, dst_ptr += 8) {
+        auto vec0_01_u8 = vld1_u8(sptr0);
+        auto vec1_0 = vld1q_f32(sptr1);
+        auto vec1_1 = vld1q_f32(sptr1 + 4);
+        auto vec2_0 = vld1q_f32(sptr2);
+        auto vec2_1 = vld1q_f32(sptr2 + 4);
+
+        auto vec0_01 = vreinterpretq_s16_u16(vmovl_u8(vec0_01_u8));
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2_0, vec0_0, vec1_0);
+        auto dst_vec_1 = Vfmaq_f32(vec2_1, vec0_1, vec1_1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+    }
+    for (; i < size; ++i, ++sptr0, ++sptr1, ++sptr2, ++dst_ptr) {
+        *dst_ptr = (float)(*sptr0) * (*sptr1) + (*sptr2);
+    }
+}
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_scaler_scaler(
+        size_t size, const int16_t* src0, const float* src1, const float* src2,
+        float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    auto vec1 = vdupq_n_f32(sptr1[0]);
+    auto vec2 = vdupq_n_f32(sptr2[0]);
+    float* __restrict dst_ptr = dst;
+    size_t i = 0;
+    for (; i + 15 < size; i += 16, sptr0 += 16, dst_ptr += 16) {
+        auto vec0_01 = vld1q_s16(sptr0);
+        auto vec0_23 = vld1q_s16(sptr0 + 8);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+        auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+        auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+        auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+        auto dst_vec_2 = Vfmaq_f32(vec2, vec0_2, vec1);
+        auto dst_vec_3 = Vfmaq_f32(vec2, vec0_3, vec1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+        vst1q_f32(dst_ptr + 8, dst_vec_2);
+        vst1q_f32(dst_ptr + 12, dst_vec_3);
+    }
+    for (; i + 7 < size; i += 8, sptr0 += 8, sptr1 += 8, sptr2 += 8, dst_ptr += 8) {
+        auto vec0_01 = vld1q_s16(sptr0);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+        auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+    }
+    for (; i + 3 < size; i += 4, sptr0 += 4, sptr1 += 4, sptr2 += 4, dst_ptr += 4) {
+        auto vec0_0 = vld1_s16(sptr0);
+
+        auto vec0_0_f32 = vcvtq_f32_s32(vmovl_s16(vec0_0));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0_f32, vec1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+    }
+    for (; i < size; ++i, ++sptr0, ++dst_ptr) {
+        *dst_ptr = (float)(*sptr0) * (*sptr1) + (*sptr2);
+    }
+}
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_scaler_scaler(
+        size_t size, const uint8_t* src0, const float* src1, const float* src2,
+        float* dst) {
+    const uint8_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    const float* __restrict sptr2 = src2;
+    auto vec1 = vdupq_n_f32(sptr1[0]);
+    auto vec2 = vdupq_n_f32(sptr2[0]);
+    float* __restrict dst_ptr = dst;
+    size_t i = 0;
+    for (; i + 15 < size; i += 16, sptr0 += 16, dst_ptr += 16) {
+        auto vec0_0123 = vld1q_u8(sptr0);
+
+        auto vec0_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vec0_0123)));
+        auto vec0_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(vec0_0123)));
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+        auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+        auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+        auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+        auto dst_vec_2 = Vfmaq_f32(vec2, vec0_2, vec1);
+        auto dst_vec_3 = Vfmaq_f32(vec2, vec0_3, vec1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+        vst1q_f32(dst_ptr + 8, dst_vec_2);
+        vst1q_f32(dst_ptr + 12, dst_vec_3);
+    }
+
+    for (; i + 7 < size; i += 8, sptr0 += 8, dst_ptr += 8) {
+        auto vec0_01_u8 = vld1_u8(sptr0);
+
+        auto vec0_01 = vreinterpretq_s16_u16(vmovl_u8(vec0_01_u8));
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+        auto dst_vec_0 = Vfmaq_f32(vec2, vec0_0, vec1);
+        auto dst_vec_1 = Vfmaq_f32(vec2, vec0_1, vec1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+    }
+    for (; i < size; ++i, ++sptr0, ++dst_ptr) {
+        *dst_ptr = (float)(*sptr0) * (*sptr1) + (*sptr2);
+    }
+}
+
+void neon_mul_int16xf32xf32_vec_bcast111c(
+        size_t batch_size, size_t channel_stride, size_t channel_size,
+        const int16_t* src0, const float* src1, float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    float* __restrict dst_ptr = dst;
+    for (size_t batch = 0; batch < batch_size; ++batch) {
+        for (size_t s = 0; s < channel_stride; ++s) {
+            size_t i = 0;
+            for (; i + 15 < channel_size; i += 16, sptr0 += 16, dst_ptr += 16) {
+                auto vec0_01 = vld1q_s16(sptr0);
+                auto vec0_23 = vld1q_s16(sptr0 + 8);
+                auto vec1_0 = vld1q_f32(sptr1 + i);
+                auto vec1_1 = vld1q_f32(sptr1 + i + 4);
+                auto vec1_2 = vld1q_f32(sptr1 + i + 8);
+                auto vec1_3 = vld1q_f32(sptr1 + i + 12);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+                auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+                auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+                auto dst_vec_0 = vmulq_f32(vec0_0, vec1_0);
+                auto dst_vec_1 = vmulq_f32(vec0_1, vec1_1);
+                auto dst_vec_2 = vmulq_f32(vec0_2, vec1_2);
+                auto dst_vec_3 = vmulq_f32(vec0_3, vec1_3);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+                vst1q_f32(dst_ptr + 8, dst_vec_2);
+                vst1q_f32(dst_ptr + 12, dst_vec_3);
+            }
+            for (; i + 7 < channel_size; i += 8, sptr0 += 8, dst_ptr += 8) {
+                auto vec0_01 = vld1q_s16(sptr0);
+                auto vec1_0 = vld1q_f32(sptr1 + i);
+                auto vec1_1 = vld1q_f32(sptr1 + i + 4);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+                auto dst_vec_0 = vmulq_f32(vec0_0, vec1_0);
+                auto dst_vec_1 = vmulq_f32(vec0_1, vec1_1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+            }
+            for (; i < channel_size; ++i, ++sptr0, ++dst_ptr) {
+                *dst_ptr = (float)(*sptr0) * sptr1[i];
+            }
+        }
+    }
+}
+
+void neon_mul_int16xf32xf32_vec_bcast101(
+        size_t batch_size, size_t channel_size, size_t channel_stride,
+        const int16_t* src0, const float* src1, float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    float* __restrict dst_ptr = dst;
+    for (size_t batch = 0; batch < batch_size; ++batch) {
+        for (size_t chan = 0; chan < channel_size; ++chan) {
+            auto vec1 = vdupq_n_f32(sptr1[chan]);
+            size_t i = 0;
+            for (; i + 15 < channel_stride; i += 16, sptr0 += 16, dst_ptr += 16) {
+                auto vec0_01 = vld1q_s16(sptr0);
+                auto vec0_23 = vld1q_s16(sptr0 + 8);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+                auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+                auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+                auto dst_vec_0 = vmulq_f32(vec0_0, vec1);
+                auto dst_vec_1 = vmulq_f32(vec0_1, vec1);
+                auto dst_vec_2 = vmulq_f32(vec0_2, vec1);
+                auto dst_vec_3 = vmulq_f32(vec0_3, vec1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+                vst1q_f32(dst_ptr + 8, dst_vec_2);
+                vst1q_f32(dst_ptr + 12, dst_vec_3);
+            }
+            for (; i + 7 < channel_stride; i += 8, sptr0 += 8, dst_ptr += 8) {
+                auto vec0_01 = vld1q_s16(sptr0);
+
+                auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+                auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+                auto dst_vec_0 = vmulq_f32(vec0_0, vec1);
+                auto dst_vec_1 = vmulq_f32(vec0_1, vec1);
+
+                vst1q_f32(dst_ptr, dst_vec_0);
+                vst1q_f32(dst_ptr + 4, dst_vec_1);
+            }
+            for (; i < channel_stride; ++i, ++sptr0, ++dst_ptr) {
+                *dst_ptr = (float)(*sptr0) * sptr1[chan];
+            }
+        }
+    }
+}
+
+void neon_mul_int16xf32xf32_vec_vec(
+        size_t size, const int16_t* src0, const float* src1, float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    float* __restrict dst_ptr = dst;
+    size_t i = 0;
+    for (; i + 15 < size; i += 16, sptr0 += 16, sptr1 += 16, dst_ptr += 16) {
+        auto vec0_01 = vld1q_s16(sptr0);
+        auto vec0_23 = vld1q_s16(sptr0 + 8);
+        auto vec1_0 = vld1q_f32(sptr1);
+        auto vec1_1 = vld1q_f32(sptr1 + 4);
+        auto vec1_2 = vld1q_f32(sptr1 + 8);
+        auto vec1_3 = vld1q_f32(sptr1 + 12);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+        auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+        auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+        auto dst_vec_0 = vmulq_f32(vec0_0, vec1_0);
+        auto dst_vec_1 = vmulq_f32(vec0_1, vec1_1);
+        auto dst_vec_2 = vmulq_f32(vec0_2, vec1_2);
+        auto dst_vec_3 = vmulq_f32(vec0_3, vec1_3);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+        vst1q_f32(dst_ptr + 8, dst_vec_2);
+        vst1q_f32(dst_ptr + 12, dst_vec_3);
+    }
+
+    for (; i + 7 < size; i += 8, sptr0 += 8, sptr1 += 8, dst_ptr += 8) {
+        auto vec0_01 = vld1q_s16(sptr0);
+        auto vec1_0 = vld1q_f32(sptr1);
+        auto vec1_1 = vld1q_f32(sptr1 + 4);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+        auto dst_vec_0 = vmulq_f32(vec0_0, vec1_0);
+        auto dst_vec_1 = vmulq_f32(vec0_1, vec1_1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+    }
+    for (; i < size; ++i, ++sptr0, ++sptr1, ++dst_ptr) {
+        *dst_ptr = (float)(*sptr0) * (*sptr1);
+    }
+}
+
+void neon_mul_int16xf32xf32_vec_scaler(
+        size_t size, const int16_t* src0, const float* src1, float* dst) {
+    const int16_t* __restrict sptr0 = src0;
+    const float* __restrict sptr1 = src1;
+    auto vec1 = vdupq_n_f32(sptr1[0]);
+    float* __restrict dst_ptr = dst;
+    size_t i = 0;
+    for (; i + 15 < size; i += 16, sptr0 += 16, dst_ptr += 16) {
+        auto vec0_01 = vld1q_s16(sptr0);
+        auto vec0_23 = vld1q_s16(sptr0 + 8);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+        auto vec0_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_23)));
+        auto vec0_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_23)));
+
+        auto dst_vec_0 = vmulq_f32(vec0_0, vec1);
+        auto dst_vec_1 = vmulq_f32(vec0_1, vec1);
+        auto dst_vec_2 = vmulq_f32(vec0_2, vec1);
+        auto dst_vec_3 = vmulq_f32(vec0_3, vec1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+        vst1q_f32(dst_ptr + 8, dst_vec_2);
+        vst1q_f32(dst_ptr + 12, dst_vec_3);
+    }
+
+    for (; i + 7 < size; i += 8, sptr0 += 8, dst_ptr += 8) {
+        auto vec0_01 = vld1q_s16(sptr0);
+
+        auto vec0_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vec0_01)));
+        auto vec0_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vec0_01)));
+
+        auto dst_vec_0 = vmulq_f32(vec0_0, vec1);
+        auto dst_vec_1 = vmulq_f32(vec0_1, vec1);
+
+        vst1q_f32(dst_ptr, dst_vec_0);
+        vst1q_f32(dst_ptr + 4, dst_vec_1);
+    }
+    for (; i < size; ++i, ++sptr0, ++dst_ptr) {
+        *dst_ptr = (float)(*sptr0) * (*sptr1);
+    }
+}
+
+}  // namespace arm_common
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/arm_common/elemwise_multi_type/kernels.h

@@ -0,0 +1,71 @@
+/**
+ * \file dnn/src/arm_common/elemwise_multi_type/kernels.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#pragma once
+
+#include "stddef.h"
+#include "stdint.h"
+
+namespace megdnn {
+namespace arm_common {
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_bcast111c_bcast111c(
+        size_t batch_size, size_t channel_stride, size_t channel_size,
+        const int16_t* src0, const float* src1, const float* src2, float* dst);
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_bcast111c_bcast111c(
+        size_t batch_size, size_t channel_stride, size_t channel_size,
+        const uint8_t* src0, const float* src1, const float* src2, float* dst);
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_bcast101_bcast101(
+        size_t batch_size, size_t channel_size, size_t channel_stride,
+        const int16_t* src0, const float* src1, const float* src2, float* dst);
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_bcast101_bcast101(
+        size_t batch_size, size_t channel_size, size_t channel_stride,
+        const uint8_t* src0, const float* src1, const float* src2, float* dst);
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_vec_vec(
+        size_t size, const int16_t* src0, const float* src1, const float* src2,
+        float* dst);
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_vec_vec(
+        size_t size, const uint8_t* src0, const float* src1, const float* src2,
+        float* dst);
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_b1x_b1x(
+        size_t size, size_t vec, const int16_t* src0, const float* src1,
+        const float* src2, float* dst);
+
+void neon_fuse_mul_add3_int16xf32xf32xf32_vec_scaler_scaler(
+        size_t size, const int16_t* src0, const float* src1, const float* src2,
+        float* dst);
+
+void neon_fuse_mul_add3_uint8xf32xf32xf32_vec_scaler_scaler(
+        size_t size, const uint8_t* src0, const float* src1, const float* src2,
+        float* dst);
+
+void neon_mul_int16xf32xf32_vec_bcast111c(
+        size_t batch_size, size_t channel_stride, size_t channel_size,
+        const int16_t* src0, const float* src1, float* dst);
+
+void neon_mul_int16xf32xf32_vec_bcast101(
+        size_t batch_size, size_t channel_size, size_t channel_stride,
+        const int16_t* src0, const float* src1, float* dst);
+
+void neon_mul_int16xf32xf32_vec_vec(
+        size_t size, const int16_t* src0, const float* src1, float* dst);
+
+void neon_mul_int16xf32xf32_vec_scaler(
+        size_t size, const int16_t* src0, const float* src1, float* dst);
+}  // namespace arm_common
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/arm_common/elemwise_multi_type/opr_impl.cpp

@@ -11,6 +11,7 @@
  */
 
 #include "./opr_impl.h"
+#include "kernels.h"
 #include "src/common/elemwise_multi_type/kern_defs.cuh"
 #include "src/naive/handle.h"
 
@@ -851,6 +852,154 @@ void ElemwiseMultiTypeImpl::on_quantized_mode(
 #undef DISPATCH_QUANTIZED_MODE
 }
 
+void ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16xf32xf32xf32(
+        const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+    auto &src0 = param[0], &src1 = param[1], &src2 = param[2];
+    BroadcastChannelInfo binfo;
+
+    if (is_vector(src0.layout) &&
+        is_NHWC_broadcasted_channel_like(src1.layout, binfo) &&
+        src1.layout.eq_layout(src2.layout)) {
+        // VEC_BCAST111C_BCAST111C
+        MEGDNN_DISPATCH_CPU_KERN_OPR(
+                neon_fuse_mul_add3_int16xf32xf32xf32_vec_bcast111c_bcast111c(
+                        binfo.x, binfo.y, binfo.z,
+                        static_cast<dt_int16*>(src0.raw_ptr()),
+                        static_cast<dt_float32*>(src1.raw_ptr()),
+                        static_cast<dt_float32*>(src2.raw_ptr()),
+                        dst.ptr<dt_float32>()));
+        return;
+    } else if (
+            is_vector(src0.layout) && is_broadcasted_channel_like(src1.layout, binfo) &&
+            src1.layout.eq_layout(src2.layout)) {
+        // VEC_BCAST101_BCAST101
+        MEGDNN_DISPATCH_CPU_KERN_OPR(
+                neon_fuse_mul_add3_int16xf32xf32xf32_vec_bcast101_bcast101(
+                        binfo.x, binfo.y, binfo.z,
+                        static_cast<dt_int16*>(src0.raw_ptr()),
+                        static_cast<dt_float32*>(src1.raw_ptr()),
+                        static_cast<dt_float32*>(src2.raw_ptr()),
+                        dst.ptr<dt_float32>()));
+        return;
+    } else if (
+            is_vector(src0.layout) && is_vector(src1.layout) &&
+            is_vector(src2.layout)) {
+        // VEC_VEC_VEC
+        auto size = param.size;
+        MEGDNN_DISPATCH_CPU_KERN_OPR(neon_fuse_mul_add3_int16xf32xf32xf32_vec_vec_vec(
+                size, static_cast<dt_int16*>(src0.raw_ptr()),
+                static_cast<dt_float32*>(src1.raw_ptr()),
+                static_cast<dt_float32*>(src2.raw_ptr()), dst.ptr<dt_float32>()));
+        return;
+    } else if (
+            is_vector(src0.layout) && is_broadcasted_scalar(src1.layout) &&
+            is_broadcasted_scalar(src2.layout)) {
+        // VEC_SCALAR_SCALAR
+        auto size = param.size;
+        MEGDNN_DISPATCH_CPU_KERN_OPR(
+                neon_fuse_mul_add3_int16xf32xf32xf32_vec_scaler_scaler(
+                        size, static_cast<dt_int16*>(src0.raw_ptr()),
+                        static_cast<dt_float32*>(src1.raw_ptr()),
+                        static_cast<dt_float32*>(src2.raw_ptr()),
+                        dst.ptr<dt_float32>()));
+        return;
+    }
+    naive::ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16xf32xf32xf32(param, dst);
+}
+
+void ElemwiseMultiTypeImpl::on_fuse_mul_add3_uint8xf32xf32xf32(
+        const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+    auto &src0 = param[0], &src1 = param[1], &src2 = param[2];
+    BroadcastChannelInfo binfo;
+
+    if (is_vector(src0.layout) &&
+        is_NHWC_broadcasted_channel_like(src1.layout, binfo) &&
+        src1.layout.eq_layout(src2.layout)) {
+        // VEC_BCAST111C_BCAST111C
+        MEGDNN_DISPATCH_CPU_KERN_OPR(
+                neon_fuse_mul_add3_uint8xf32xf32xf32_vec_bcast111c_bcast111c(
+                        binfo.x, binfo.y, binfo.z,
+                        static_cast<dt_uint8*>(src0.raw_ptr()),
+                        static_cast<dt_float32*>(src1.raw_ptr()),
+                        static_cast<dt_float32*>(src2.raw_ptr()),
+                        dst.ptr<dt_float32>()));
+        return;
+    } else if (
+            is_vector(src0.layout) && is_broadcasted_channel_like(src1.layout, binfo) &&
+            src1.layout.eq_layout(src2.layout)) {
+        // VEC_BCAST101_BCAST101
+        MEGDNN_DISPATCH_CPU_KERN_OPR(
+                neon_fuse_mul_add3_uint8xf32xf32xf32_vec_bcast101_bcast101(
+                        binfo.x, binfo.y, binfo.z,
+                        static_cast<dt_uint8*>(src0.raw_ptr()),
+                        static_cast<dt_float32*>(src1.raw_ptr()),
+                        static_cast<dt_float32*>(src2.raw_ptr()),
+                        dst.ptr<dt_float32>()));
+        return;
+    } else if (
+            is_vector(src0.layout) && is_vector(src1.layout) &&
+            is_vector(src2.layout)) {
+        // VEC_VEC_VEC
+        auto size = param.size;
+        MEGDNN_DISPATCH_CPU_KERN_OPR(neon_fuse_mul_add3_uint8xf32xf32xf32_vec_vec_vec(
+                size, static_cast<dt_uint8*>(src0.raw_ptr()),
+                static_cast<dt_float32*>(src1.raw_ptr()),
+                static_cast<dt_float32*>(src2.raw_ptr()), dst.ptr<dt_float32>()));
+        return;
+    } else if (
+            is_vector(src0.layout) && is_broadcasted_scalar(src1.layout) &&
+            is_broadcasted_scalar(src2.layout)) {
+        // VEC_SCALAR_SCALAR
+        auto size = param.size;
+        MEGDNN_DISPATCH_CPU_KERN_OPR(
+                neon_fuse_mul_add3_uint8xf32xf32xf32_vec_scaler_scaler(
+                        size, static_cast<dt_uint8*>(src0.raw_ptr()),
+                        static_cast<dt_float32*>(src1.raw_ptr()),
+                        static_cast<dt_float32*>(src2.raw_ptr()),
+                        dst.ptr<dt_float32>()));
+        return;
+    }
+
+    naive::ElemwiseMultiTypeImpl::on_fuse_mul_add3_uint8xf32xf32xf32(param, dst);
+}
+
+void ElemwiseMultiTypeImpl::on_mul_int16xf32xf32(
+        const ElemwiseOpParamN<2>& param, const TensorND& dst) {
+    auto &src0 = param[0], &src1 = param[1];
+    BroadcastChannelInfo binfo;
+
+    if (is_vector(src0.layout) &&
+        is_NHWC_broadcasted_channel_like(src1.layout, binfo)) {
+        // VEC_BCAST111C
+        MEGDNN_DISPATCH_CPU_KERN_OPR(neon_mul_int16xf32xf32_vec_bcast111c(
+                binfo.x, binfo.y, binfo.z, static_cast<dt_int16*>(src0.raw_ptr()),
+                static_cast<dt_float32*>(src1.raw_ptr()), dst.ptr<dt_float32>()));
+        return;
+    } else if (
+            is_vector(src0.layout) && is_broadcasted_channel_like(src1.layout, binfo)) {
+        // VEC_BCAST101
+        MEGDNN_DISPATCH_CPU_KERN_OPR(neon_mul_int16xf32xf32_vec_bcast101(
+                binfo.x, binfo.y, binfo.z, static_cast<dt_int16*>(src0.raw_ptr()),
+                static_cast<dt_float32*>(src1.raw_ptr()), dst.ptr<dt_float32>()));
+        return;
+    } else if (is_vector(src0.layout) && is_vector(src1.layout)) {
+        // VEC_VEC
+        auto size = param.size;
+        MEGDNN_DISPATCH_CPU_KERN_OPR(neon_mul_int16xf32xf32_vec_vec(
+                size, static_cast<dt_int16*>(src0.raw_ptr()),
+                static_cast<dt_float32*>(src1.raw_ptr()), dst.ptr<dt_float32>()));
+        return;
+    } else if (is_vector(src0.layout) && is_broadcasted_scalar(src1.layout)) {
+        auto size = param.size;
+        MEGDNN_DISPATCH_CPU_KERN_OPR(neon_mul_int16xf32xf32_vec_scaler(
+                size, static_cast<dt_int16*>(src0.raw_ptr()),
+                static_cast<dt_float32*>(src1.raw_ptr()), dst.ptr<dt_float32>()));
+        return;
+    }
+
+    naive::ElemwiseMultiTypeImpl::on_mul_int16xf32xf32(param, dst);
+}
+
 }  // namespace arm_common
 }  // namespace megdnn
 

dnn/src/arm_common/elemwise_multi_type/opr_impl.h

@@ -48,6 +48,15 @@ protected:
             const ElemwiseOpParamN<3>& param, const TensorND& dst,
             Elemwise::Mode mode) override;
 
+    void on_fuse_mul_add3_int16xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) override;
+
+    void on_mul_int16xf32xf32(
+            const ElemwiseOpParamN<2>& param, const TensorND& dst) override;
+
+    void on_fuse_mul_add3_uint8xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) override;
+
 public:
     using fallback::ElemwiseMultiTypeImpl::ElemwiseMultiTypeImpl;
 };

dnn/src/common/elemwise_multi_type/opr_impl.cpp

@@ -155,6 +155,29 @@ const ModeTrait& ModeTrait::from_mode(Mode mode) {
             dst.name = name;
             dst.need_specify_out_dtype = true;
         };
+        auto init_fma3_int16xf32xf32xf32 = [&](ModeTrait& dst, const char* name) {
+            dst.arity = 3;
+            dst.check_inp[0] = make_check_dtype_func(dtype::Int16());
+            dst.check_inp[1] = make_check_dtype_func(dtype::Float32());
+            dst.check_inp[2] = make_check_dtype_func(dtype::Float32());
+            dst.check_out = make_out_dtype_func(dtype::Float32());
+            dst.name = name;
+        };
+        auto init_mul_int16xf32xf32 = [&](ModeTrait& dst, const char* name) {
+            dst.arity = 2;
+            dst.check_inp[0] = make_check_dtype_func(dtype::Int16());
+            dst.check_inp[1] = make_check_dtype_func(dtype::Float32());
+            dst.check_out = make_out_dtype_func(dtype::Float32());
+            dst.name = name;
+        };
+        auto init_fma3_uint8xf32xf32xf32 = [&](ModeTrait& dst, const char* name) {
+            dst.arity = 3;
+            dst.check_inp[0] = make_check_dtype_func(dtype::Uint8());
+            dst.check_inp[1] = make_check_dtype_func(dtype::Float32());
+            dst.check_inp[2] = make_check_dtype_func(dtype::Float32());
+            dst.check_out = make_out_dtype_func(dtype::Float32());
+            dst.name = name;
+        };
 
 #define SET(f, m)                                                         \
     MIDOUT_BEGIN(megdnn_common_elemwise_multi_type, midout_iv(Mode::m)) { \
@@ -169,6 +192,9 @@ const ModeTrait& ModeTrait::from_mode(Mode mode) {
         SET(init_fuse_add_rmulh_rshr_int32x32x32x8,
             FUSE_ADD_RMULH_ROUND_SHR_SATURATE_INT32x32x32x8);
         SET(init_rshrs_iXxi8xi16, ROUND_SHR_SATURATE_IXxI8xI16);
+        SET(init_fma3_int16xf32xf32xf32, FUSE_MUL_ADD3_INT16xF32xF32xF32);
+        SET(init_mul_int16xf32xf32, MUL_INT16xF32xF32);
+        SET(init_fma3_uint8xf32xf32xf32, FUSE_MUL_ADD3_UINT8xF32xF32xF32);
 
         //! quantized opr, with specified dtype.
         //! dispatch elemwise mode internally

dnn/src/common/elemwise_multi_type/opr_impl_helper.cpp

@@ -43,6 +43,17 @@ void ElemwiseMultiTypeImplHelper::exec(
         case Mode::ROUND_SHR_SATURATE_IXxI8xI16:
             on_round_shr_saturate_iXxi8xi16(make_elemwise_op_param<2>(src, dst), dst);
             break;
+        case Mode::FUSE_MUL_ADD3_INT16xF32xF32xF32:
+            on_fuse_mul_add3_int16xf32xf32xf32(
+                    make_elemwise_op_param<3>(src, dst), dst);
+            break;
+        case Mode::MUL_INT16xF32xF32:
+            on_mul_int16xf32xf32(make_elemwise_op_param<2>(src, dst), dst);
+            break;
+        case Mode::FUSE_MUL_ADD3_UINT8xF32xF32xF32:
+            on_fuse_mul_add3_uint8xf32xf32xf32(
+                    make_elemwise_op_param<3>(src, dst), dst);
+            break;
             ON_QUANTIZED_MODE(RELU, 1);
             ON_QUANTIZED_MODE(ABS, 1);
             ON_QUANTIZED_MODE(ACOS, 1);

dnn/src/common/elemwise_multi_type/opr_impl_helper.h

@@ -50,6 +50,27 @@ protected:
     virtual void on_round_shr_saturate_iXxi8xi16(
             const ElemwiseOpParamN<2>& param, const TensorND& dst) = 0;
 
+    virtual void on_fuse_mul_add3_int16xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+        MEGDNN_MARK_USED_VAR(param);
+        MEGDNN_MARK_USED_VAR(dst);
+        megdnn_throw("unsupported ElemwiseMultiType fma3 int16xf32xf32xf32.");
+    }
+
+    virtual void on_mul_int16xf32xf32(
+            const ElemwiseOpParamN<2>& param, const TensorND& dst) {
+        MEGDNN_MARK_USED_VAR(param);
+        MEGDNN_MARK_USED_VAR(dst);
+        megdnn_throw("unsupported ElemwiseMultiType fma3 int16xf32xf32.");
+    }
+
+    virtual void on_fuse_mul_add3_uint8xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+        MEGDNN_MARK_USED_VAR(param);
+        MEGDNN_MARK_USED_VAR(dst);
+        megdnn_throw("unsupported ElemwiseMultiType fma3 uint8xf32xf32xf32.");
+    }
+
     virtual void on_quantized_mode(
             const ElemwiseOpParamN<1>& param, const TensorND& dst,
             Elemwise::Mode mode) {

dnn/src/fallback/elemwise_multi_type/opr_impl.cpp

@@ -56,6 +56,216 @@ void ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16x32x32x32(
     naive::ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16x32x32x32(param, dst);
 }
 
+void ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16xf32xf32xf32(
+        const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+    BroadcastChannelInfo binfo0, binfo1;
+
+    if (is_vector(param[0].layout) &&
+        is_NHWC_broadcasted_channel_like(param[1].layout, binfo0) &&
+        is_NHWC_broadcasted_channel_like(param[2].layout, binfo1) && binfo0 == binfo1) {
+        auto x = binfo0.x, y = binfo0.y, z = binfo0.z;
+        auto src0 = param[0];
+        auto src1 = param[1];
+        auto src2 = param[2];
+        auto work = [=]() {
+            const dt_int16* __restrict__ a = static_cast<dt_int16*>(src0.raw_ptr());
+            const dt_float32* __restrict__ b = static_cast<dt_float32*>(src1.raw_ptr());
+            const dt_float32* __restrict__ c = static_cast<dt_float32*>(src2.raw_ptr());
+            dt_float32* __restrict__ d = dst.ptr<dt_float32>();
+            for (size_t i = 0; i < x; ++i) {
+                for (size_t j = 0; j < y; ++j) {
+                    auto off = i * (y * z) + j * z;
+                    size_t k = 0;
+                    for (; k + 4 <= z; k += 4) {
+                        d[off + k + 0] = a[off + k + 0] * b[k + 0] + c[k + 0];
+                        d[off + k + 1] = a[off + k + 1] * b[k + 1] + c[k + 1];
+                        d[off + k + 2] = a[off + k + 2] * b[k + 2] + c[k + 2];
+                        d[off + k + 3] = a[off + k + 3] * b[k + 3] + c[k + 3];
+                    }
+                    for (; k < z; ++k) {
+                        d[off + k] = a[off + k] * b[k] + c[k];
+                    }
+                }
+            }
+        };
+
+        MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+        return;
+    } else if (
+            is_vector(param[0].layout) &&
+            is_broadcasted_channel_like(param[1].layout, binfo0) &&
+            is_broadcasted_channel_like(param[2].layout, binfo1) && binfo0 == binfo1) {
+        auto x = binfo0.x, y = binfo0.y, z = binfo0.z;
+        auto src0 = param[0];
+        auto src1 = param[1];
+        auto src2 = param[2];
+        auto work = [=]() {
+            const dt_int16* __restrict__ a = static_cast<dt_int16*>(src0.raw_ptr());
+            const dt_float32* __restrict__ b = static_cast<dt_float32*>(src1.raw_ptr());
+            const dt_float32* __restrict__ c = static_cast<dt_float32*>(src2.raw_ptr());
+            dt_float32* __restrict__ d = dst.ptr<dt_float32>();
+            for (size_t j = 0; j < y; ++j) {
+                auto bv = b[j], cv = c[j];
+                for (size_t i = 0; i < x; ++i) {
+                    auto off = i * (y * z) + j * z, offt = off + z;
+                    for (; off + 4 <= offt; off += 4) {
+                        d[off + 0] = a[off + 0] * bv + cv;
+                        d[off + 1] = a[off + 1] * bv + cv;
+                        d[off + 2] = a[off + 2] * bv + cv;
+                        d[off + 3] = a[off + 3] * bv + cv;
+                    }
+                    for (; off < offt; ++off) {
+                        d[off] = a[off] * bv + cv;
+                    }
+                }
+            }
+        };
+
+        MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+        return;
+    }
+
+    naive::ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16xf32xf32xf32(param, dst);
+}
+
+void ElemwiseMultiTypeImpl::on_mul_int16xf32xf32(
+        const ElemwiseOpParamN<2>& param, const TensorND& dst) {
+    BroadcastChannelInfo binfo;
+
+    if (is_vector(param[0].layout) &&
+        is_NHWC_broadcasted_channel_like(param[1].layout, binfo)) {
+        auto x = binfo.x, y = binfo.y, z = binfo.z;
+        auto src0 = param[0];
+        auto src1 = param[1];
+        auto work = [=]() {
+            const dt_int16* __restrict__ a = static_cast<dt_int16*>(src0.raw_ptr());
+            const dt_float32* __restrict__ b = static_cast<dt_float32*>(src1.raw_ptr());
+            dt_float32* __restrict__ d = dst.ptr<dt_float32>();
+            for (size_t i = 0; i < x; ++i) {
+                for (size_t j = 0; j < y; ++j) {
+                    auto off = i * (y * z) + j * z;
+                    size_t k = 0;
+                    for (; k + 4 <= z; k += 4) {
+                        d[off + k + 0] = a[off + k + 0] * b[k + 0];
+                        d[off + k + 1] = a[off + k + 1] * b[k + 1];
+                        d[off + k + 2] = a[off + k + 2] * b[k + 2];
+                        d[off + k + 3] = a[off + k + 3] * b[k + 3];
+                    }
+                    for (; k < z; ++k) {
+                        d[off + k] = a[off + k] * b[k];
+                    }
+                }
+            }
+        };
+
+        MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+        return;
+    } else if (
+            is_vector(param[0].layout) &&
+            is_broadcasted_channel_like(param[1].layout, binfo)) {
+        auto x = binfo.x, y = binfo.y, z = binfo.z;
+        auto src0 = param[0];
+        auto src1 = param[1];
+        auto work = [=]() {
+            const dt_int16* __restrict__ a = static_cast<dt_int16*>(src0.raw_ptr());
+            const dt_float32* __restrict__ b = static_cast<dt_float32*>(src1.raw_ptr());
+            dt_float32* __restrict__ d = dst.ptr<dt_float32>();
+            for (size_t j = 0; j < y; ++j) {
+                auto bv = b[j];
+                for (size_t i = 0; i < x; ++i) {
+                    auto off = i * (y * z) + j * z, offt = off + z;
+                    for (; off + 4 <= offt; off += 4) {
+                        d[off + 0] = a[off + 0] * bv;
+                        d[off + 1] = a[off + 1] * bv;
+                        d[off + 2] = a[off + 2] * bv;
+                        d[off + 3] = a[off + 3] * bv;
+                    }
+                    for (; off < offt; ++off) {
+                        d[off] = a[off] * bv;
+                    }
+                }
+            }
+        };
+
+        MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+        return;
+    }
+
+    naive::ElemwiseMultiTypeImpl::on_mul_int16xf32xf32(param, dst);
+}
+
+void ElemwiseMultiTypeImpl::on_fuse_mul_add3_uint8xf32xf32xf32(
+        const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+    BroadcastChannelInfo binfo0, binfo1;
+
+    if (is_vector(param[0].layout) &&
+        is_NHWC_broadcasted_channel_like(param[1].layout, binfo0) &&
+        is_NHWC_broadcasted_channel_like(param[2].layout, binfo1) && binfo0 == binfo1) {
+        auto x = binfo0.x, y = binfo0.y, z = binfo0.z;
+        auto src0 = param[0];
+        auto src1 = param[1];
+        auto src2 = param[2];
+        auto work = [=]() {
+            const dt_uint8* __restrict__ a = static_cast<dt_uint8*>(src0.raw_ptr());
+            const dt_float32* __restrict__ b = static_cast<dt_float32*>(src1.raw_ptr());
+            const dt_float32* __restrict__ c = static_cast<dt_float32*>(src2.raw_ptr());
+            dt_float32* __restrict__ d = dst.ptr<dt_float32>();
+            for (size_t i = 0; i < x; ++i) {
+                for (size_t j = 0; j < y; ++j) {
+                    auto off = i * (y * z) + j * z;
+                    size_t k = 0;
+                    for (; k + 4 <= z; k += 4) {
+                        d[off + k + 0] = a[off + k + 0] * b[k + 0] + c[k + 0];
+                        d[off + k + 1] = a[off + k + 1] * b[k + 1] + c[k + 1];
+                        d[off + k + 2] = a[off + k + 2] * b[k + 2] + c[k + 2];
+                        d[off + k + 3] = a[off + k + 3] * b[k + 3] + c[k + 3];
+                    }
+                    for (; k < z; ++k) {
+                        d[off + k] = a[off + k] * b[k] + c[k];
+                    }
+                }
+            }
+        };
+
+        MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+        return;
+    } else if (
+            is_vector(param[0].layout) &&
+            is_broadcasted_channel_like(param[1].layout, binfo0) &&
+            is_broadcasted_channel_like(param[2].layout, binfo1) && binfo0 == binfo1) {
+        auto x = binfo0.x, y = binfo0.y, z = binfo0.z;
+        auto src0 = param[0];
+        auto src1 = param[1];
+        auto src2 = param[2];
+        auto work = [=]() {
+            const dt_uint8* __restrict__ a = static_cast<dt_uint8*>(src0.raw_ptr());
+            const dt_float32* __restrict__ b = static_cast<dt_float32*>(src1.raw_ptr());
+            const dt_float32* __restrict__ c = static_cast<dt_float32*>(src2.raw_ptr());
+            dt_float32* __restrict__ d = dst.ptr<dt_float32>();
+            for (size_t j = 0; j < y; ++j) {
+                auto bv = b[j], cv = c[j];
+                for (size_t i = 0; i < x; ++i) {
+                    auto off = i * (y * z) + j * z, offt = off + z;
+                    for (; off + 4 <= offt; off += 4) {
+                        d[off + 0] = a[off + 0] * bv + cv;
+                        d[off + 1] = a[off + 1] * bv + cv;
+                        d[off + 2] = a[off + 2] * bv + cv;
+                        d[off + 3] = a[off + 3] * bv + cv;
+                    }
+                    for (; off < offt; ++off) {
+                        d[off] = a[off] * bv + cv;
+                    }
+                }
+            }
+        };
+
+        MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+        return;
+    }
+
+    naive::ElemwiseMultiTypeImpl::on_fuse_mul_add3_uint8xf32xf32xf32(param, dst);
+}
+
 template <typename ctype>
 void ElemwiseMultiTypeImpl::dispatch_fma3_iXxf32xf32xi8_bcast_1x(
         const ElemwiseOpParamN<3>& param, const Broadcast1xInfo& binfo,

dnn/src/fallback/elemwise_multi_type/opr_impl.h

@@ -43,6 +43,12 @@ protected:
             const ElemwiseOpParamN<6>& param, const TensorND& dst) override;
     void on_round_shr_saturate_iXxi8xi16(
             const ElemwiseOpParamN<2>& param, const TensorND& dst) override;
+    void on_fuse_mul_add3_int16xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) override;
+    void on_mul_int16xf32xf32(
+            const ElemwiseOpParamN<2>& param, const TensorND& dst) override;
+    void on_fuse_mul_add3_uint8xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) override;
 
 public:
     using naive::ElemwiseMultiTypeImpl::ElemwiseMultiTypeImpl;

dnn/src/naive/elemwise_multi_type/opr_impl.cpp

@@ -39,6 +39,66 @@ void ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16x32x32x32(
     MEGDNN_DISPATCH_CPU_KERN_OPR(work());
 }
 
+void ElemwiseMultiTypeImpl::on_fuse_mul_add3_int16xf32xf32xf32(
+        const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+    auto size = param.size;
+    auto src0 = param[0];
+    auto src1 = param[1];
+    auto src2 = param[2];
+    auto work = [src0, src1, src2, size, dst]() {
+        auto i0 = tensor_iter_valonly<dt_int16>(src0).begin();
+        auto i1 = tensor_iter_valonly<dt_float32>(src1).begin();
+        auto i2 = tensor_iter_valonly<dt_float32>(src2).begin();
+        auto dst_ptr = dst.ptr<dt_float32>();
+        for (size_t i = 0; i < size; ++i) {
+            dst_ptr[i] = (*i0) * (*i1) + (*i2);
+            ++i0;
+            ++i1;
+            ++i2;
+        }
+    };
+    MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+}
+
+void ElemwiseMultiTypeImpl::on_fuse_mul_add3_uint8xf32xf32xf32(
+        const ElemwiseOpParamN<3>& param, const TensorND& dst) {
+    auto size = param.size;
+    auto src0 = param[0];
+    auto src1 = param[1];
+    auto src2 = param[2];
+    auto work = [src0, src1, src2, size, dst]() {
+        auto i0 = tensor_iter_valonly<dt_uint8>(src0).begin();
+        auto i1 = tensor_iter_valonly<dt_float32>(src1).begin();
+        auto i2 = tensor_iter_valonly<dt_float32>(src2).begin();
+        auto dst_ptr = dst.ptr<dt_float32>();
+        for (size_t i = 0; i < size; ++i) {
+            dst_ptr[i] = (*i0) * (*i1) + (*i2);
+            ++i0;
+            ++i1;
+            ++i2;
+        }
+    };
+    MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+}
+
+void ElemwiseMultiTypeImpl::on_mul_int16xf32xf32(
+        const ElemwiseOpParamN<2>& param, const TensorND& dst) {
+    auto size = param.size;
+    auto src0 = param[0];
+    auto src1 = param[1];
+    auto work = [src0, src1, size, dst]() {
+        auto i0 = tensor_iter_valonly<dt_int16>(src0).begin();
+        auto i1 = tensor_iter_valonly<dt_float32>(src1).begin();
+        auto dst_ptr = dst.ptr<dt_float32>();
+        for (size_t i = 0; i < size; ++i) {
+            dst_ptr[i] = (*i0) * (*i1);
+            ++i0;
+            ++i1;
+        }
+    };
+    MEGDNN_DISPATCH_CPU_KERN_OPR(work());
+}
+
 void ElemwiseMultiTypeImpl::on_fuse_mul_add3_iXxf32xf32xi8(
         const ElemwiseOpParamN<3>& param, const TensorND& dst) {
     switch (param[0].layout.dtype.enumv()) {

dnn/src/naive/elemwise_multi_type/opr_impl.h

@@ -60,6 +60,12 @@ protected:
             const ElemwiseOpParamN<6>& param, const TensorND& dst) override;
     void on_round_shr_saturate_iXxi8xi16(
             const ElemwiseOpParamN<2>& param, const TensorND& dst) override;
+    void on_fuse_mul_add3_int16xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) override;
+    void on_mul_int16xf32xf32(
+            const ElemwiseOpParamN<2>& param, const TensorND& dst) override;
+    void on_fuse_mul_add3_uint8xf32xf32xf32(
+            const ElemwiseOpParamN<3>& param, const TensorND& dst) override;
 
     void on_quantized_mode(
             const ElemwiseOpParamN<1>& param, const TensorND& dst,

dnn/test/arm_common/elemwise_multi_type.cpp

@@ -456,4 +456,107 @@ TEST_F(ARM_COMMON, ELEMWISE_QUANTIZED_MODE_TERNARY_RECORD) {
     }
 }
 
+TEST_F(ARM_COMMON, ELEMWISE_FMA3_INT16xF32xF32xF32) {
+    Checker<ElemwiseMultiType> checker(handle());
+    checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_INT16xF32xF32xF32});
+    checker.set_dtype(0, dtype::Int16());
+    checker.set_dtype(1, dtype::Float32());
+    checker.set_dtype(2, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.set_rng(2, &rng);
+    checker.execs({{5, 7, 16}, {1, 1, 16}, {1, 1, 16}, {}})
+            .execs({{2, 700, 600}, {1, 1, 600}, {1, 1, 600}, {}})
+            .execs({{2, 700, 600}, {2, 700, 600}, {2, 700, 600}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {1, 128, 1, 1}, {}});
+}
+
+TEST_F(ARM_COMMON, ELEMWISE_FMA3_INT16xF32xF32xF32_RECORD) {
+    TaskRecordChecker<ElemwiseMultiType> checker(0);
+    checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_INT16xF32xF32xF32});
+    checker.set_dtype(0, dtype::Int16());
+    checker.set_dtype(1, dtype::Float32());
+    checker.set_dtype(2, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.set_rng(2, &rng);
+    checker.execs({{5, 7, 16}, {1, 1, 16}, {1, 1, 16}, {}})
+            .execs({{2, 700, 600}, {1, 1, 600}, {1, 1, 600}, {}})
+            .execs({{2, 700, 600}, {2, 700, 600}, {2, 700, 600}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 18}, {1, 1, 1, 18}, {1, 1, 1, 18}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {1, 1, 1, 1}, {}});
+}
+
+TEST_F(ARM_COMMON, ELEMWISE_MUL_INT16xF32xF32) {
+    Checker<ElemwiseMultiType> checker(handle());
+    checker.set_param({ElemwiseMultiType::Mode::MUL_INT16xF32xF32});
+    checker.set_dtype(0, dtype::Int16());
+    checker.set_dtype(1, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.execs({{5, 7, 16}, {1, 1, 16}, {}})
+            .execs({{2, 700, 600}, {1, 1, 600}, {}})
+            .execs({{2, 700, 600}, {2, 700, 600}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {}});
+}
+
+TEST_F(ARM_COMMON, ELEMWISE_ELEMWISE_MUL_INT16xF32xF32_RECORD) {
+    TaskRecordChecker<ElemwiseMultiType> checker(0);
+    checker.set_param({ElemwiseMultiType::Mode::MUL_INT16xF32xF32});
+    checker.set_dtype(0, dtype::Int16());
+    checker.set_dtype(1, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.execs({{5, 7, 16}, {1, 1, 16}, {}})
+            .execs({{2, 700, 600}, {1, 1, 600}, {}})
+            .execs({{2, 700, 600}, {2, 700, 600}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {}});
+}
+
+TEST_F(ARM_COMMON, ELEMWISE_FMA3_UINT8xF32xF32xF32) {
+    Checker<ElemwiseMultiType> checker(handle());
+    checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_UINT8xF32xF32xF32});
+    checker.set_dtype(0, dtype::Uint8());
+    checker.set_dtype(1, dtype::Float32());
+    checker.set_dtype(2, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.set_rng(2, &rng);
+    checker.execs({{5, 7, 16}, {1, 1, 16}, {1, 1, 16}, {}})
+            .execs({{2, 700, 600}, {1, 1, 600}, {1, 1, 600}, {}})
+            .execs({{2, 700, 600}, {2, 700, 600}, {2, 700, 600}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {1, 1, 1, 1}, {}});
+}
+
+TEST_F(ARM_COMMON, ELEMWISE_FMA3_UINT8xF32xF32xF32_RECORD) {
+    TaskRecordChecker<ElemwiseMultiType> checker(0);
+    checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_UINT8xF32xF32xF32});
+    checker.set_dtype(0, dtype::Uint8());
+    checker.set_dtype(1, dtype::Float32());
+    checker.set_dtype(2, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.set_rng(2, &rng);
+    checker.execs({{5, 7, 16}, {1, 1, 16}, {1, 1, 16}, {}})
+            .execs({{2, 700, 600}, {1, 1, 600}, {1, 1, 600}, {}})
+            .execs({{2, 700, 600}, {2, 700, 600}, {2, 700, 600}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {1, 1, 1, 1}, {}});
+}
+
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

dnn/test/common/elemwise_multi_type.cpp

@@ -79,6 +79,73 @@ DEF_TEST(fuse_mul_add3_int16x32x32x32) {
             .execs({{102, 67, 71}, {1, 67, 1}, {1, 67, 1}, {}});
 }
 
+DEF_TEST(fuse_mul_add3_int16xf32xf32xf32) {
+    // This is not implemented on CUDA.
+    if (handle->type() == Handle::HandleType::CUDA) {
+        return;
+    }
+    Checker<ElemwiseMultiType> checker(handle);
+    checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_INT16xF32xF32xF32});
+    checker.set_dtype(0, dtype::Int16());
+    checker.set_dtype(1, dtype::Float32());
+    checker.set_dtype(2, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.set_rng(2, &rng);
+    checker.execs({{5, 7, 6}, {1, 1, 6}, {1, 1, 6}, {}})
+            .execs({{1, 700, 600}, {1, 1, 600}, {1, 1, 600}, {}})
+            .execs({{1, 700, 600}, {1, 700, 600}, {1, 700, 600}, {}})
+            .execs({{102, 71, 67}, {1, 1, 67}, {1, 1, 67}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {1, 1, 1, 1}, {}});
+}
+
+DEF_TEST(fuse_mul_add3_uint8xf32xf32xf32) {
+    // This is not implemented on CUDA.
+    if (handle->type() == Handle::HandleType::CUDA) {
+        return;
+    }
+    Checker<ElemwiseMultiType> checker(handle);
+    checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_UINT8xF32xF32xF32});
+    checker.set_dtype(0, dtype::Uint8());
+    checker.set_dtype(1, dtype::Float32());
+    checker.set_dtype(2, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.set_rng(2, &rng);
+    checker.execs({{5, 7, 6}, {1, 1, 6}, {1, 1, 6}, {}})
+            .execs({{1, 700, 600}, {1, 1, 600}, {1, 1, 600}, {}})
+            .execs({{1, 700, 600}, {1, 700, 600}, {1, 700, 600}, {}})
+            .execs({{102, 71, 67}, {1, 1, 67}, {1, 1, 67}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {1, 1, 1, 1}, {}});
+}
+
+DEF_TEST(fuse_mul_add3_int16xf32xf32) {
+    // This is not implemented on CUDA.
+    if (handle->type() == Handle::HandleType::CUDA) {
+        return;
+    }
+    Checker<ElemwiseMultiType> checker(handle);
+    checker.set_param({ElemwiseMultiType::Mode::MUL_INT16xF32xF32});
+    checker.set_dtype(0, dtype::Int16());
+    checker.set_dtype(1, dtype::Float32());
+    UniformIntRNG rng{-100, 100};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.execs({{5, 7, 6}, {1, 1, 6}, {}})
+            .execs({{1, 700, 600}, {1, 1, 600}, {}})
+            .execs({{1, 700, 600}, {1, 700, 600}, {}})
+            .execs({{102, 71, 67}, {1, 1, 67}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {}});
+}
+
 DEF_TEST(fuse_mul_add3_iXxf32xf32xi8) {
     Checker<ElemwiseMultiType> checker(handle);
     checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_IXxF32xF32xI8});

dnn/test/common/elemwise_multi_type.h

@@ -20,10 +20,13 @@ namespace test {
 namespace elemwise_multi_type {
 #define FIRST_ELEMWISE_MULTI_TYPE_CASE fuse_mul_add3_int16x32x32x32
 
-#define FOREACH_ELEMWISE_MULTI_TYPE_NONFIRST_CASE(cb)               \
-    cb(fuse_mul_add3_iXxf32xf32xi8) cb(round_shr_saturate_iXxi8xi8) \
-            cb(fuse_add_rmulh_round_shr_saturate_int16)             \
-                    cb(fuse_add_rmulh_round_shr_saturate_int32)
+#define FOREACH_ELEMWISE_MULTI_TYPE_NONFIRST_CASE(cb)                   \
+    cb(fuse_mul_add3_iXxf32xf32xi8) cb(round_shr_saturate_iXxi8xi8)     \
+            cb(fuse_add_rmulh_round_shr_saturate_int16)                 \
+                    cb(fuse_add_rmulh_round_shr_saturate_int32)         \
+                            cb(fuse_mul_add3_int16xf32xf32xf32)         \
+                                    cb(fuse_mul_add3_uint8xf32xf32xf32) \
+                                            cb(fuse_mul_add3_int16xf32xf32)
 
 #define FOREACH_ELEMWISE_MULTI_TYPE_CASE(cb) \
     cb(FIRST_ELEMWISE_MULTI_TYPE_CASE) FOREACH_ELEMWISE_MULTI_TYPE_NONFIRST_CASE(cb)

dnn/test/fallback/elemwise_multi_type.cpp

@@ -40,6 +40,24 @@ TEST_F(FALLBACK, ELEMWISE_MULTI_TYPE_RECORD_FMA3_INT16x32x32x32) {
     checker.execs({{A, B, C}, {1, B, 1}, {1, B, 1}, {}});
 }
 
+TEST_F(FALLBACK, ELEMWISE_MULTI_TYPE_RECORD_FMA3_INT16xF32xF32xF32) {
+    TaskRecordChecker<ElemwiseMultiType> checker{1};
+    checker.set_param({ElemwiseMultiType::Mode::FUSE_MUL_ADD3_INT16xF32xF32xF32});
+    checker.set_dtype(0, dtype::Int16());
+    checker.set_dtype(1, dtype::Float32());
+    checker.set_dtype(2, dtype::Float32());
+    UniformIntRNG rng{-10, 10};
+    checker.set_rng(0, &rng);
+    checker.set_rng(1, &rng);
+    checker.set_rng(2, &rng);
+    checker.execs({{5, 7, 16}, {1, 1, 16}, {1, 1, 16}, {}})
+            .execs({{2, 700, 600}, {1, 1, 600}, {1, 1, 600}, {}})
+            .execs({{2, 700, 600}, {2, 700, 600}, {2, 700, 600}, {}})
+            .execs({{16, 16, 128}, {16, 16, 128}, {16, 16, 128}, {}})
+            .execs({{16, 128, 16, 16}, {1, 128, 1, 1}, {1, 128, 1, 1}, {}})
+            .execs({{16, 128, 16, 16}, {1, 1, 1, 1}, {1, 1, 1, 1}, {}});
+}
+
 #if MEGDNN_WITH_BENCHMARK
 TEST_F(FALLBACK, ELEMWISE_MULTI_TYPE_BENCHMARK_FMA3_INT16x32x32x32) {
     Benchmarker<ElemwiseMultiType> bench{handle()};