diff --git a/src/layer/mips/requantize_leakyrelu_pack4.h b/src/layer/mips/requantize_leakyrelu_pack4.h
deleted file mode 100644
index 89bc14bd08b..00000000000
--- a/src/layer/mips/requantize_leakyrelu_pack4.h
+++ /dev/null
@@ -1,267 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- int size = w * h;
- int outc = top_blob.c;
- int out_elempack = top_blob.elempack;
-
- int scale_in_data_size = scale_in_data.w;
- int scale_out_data_size = scale_out_data.w;
- int bias_data_size = bias_data.w;
-
- // int8(leakyrelu(v * scale_in, slope) * scale_out)
- // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
- // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
- // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
- if (out_elempack == 8)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < outc; q++)
- {
- const int* intptr0 = bottom_blob.channel(q * 2);
- const int* intptr1 = bottom_blob.channel(q * 2 + 1);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
- v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr0 + 64);
- __builtin_prefetch(intptr1 + 64);
- v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
- v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
- v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
- v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
- v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
- v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
- _v00 = __msa_fmul_w(_v00, _scale0);
- _v01 = __msa_fmul_w(_v01, _scale0);
- _v02 = __msa_fmul_w(_v02, _scale0);
- _v03 = __msa_fmul_w(_v03, _scale0);
- _v10 = __msa_fmul_w(_v10, _scale1);
- _v11 = __msa_fmul_w(_v11, _scale1);
- _v12 = __msa_fmul_w(_v12, _scale1);
- _v13 = __msa_fmul_w(_v13, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
- *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
- *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v02, _v12, _slope);
- *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v03, _v13, _slope);
-
- intptr0 += 16;
- intptr1 += 16;
- ptr += 32;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < outc; q++)
- {
- const int* intptr0 = bottom_blob.channel(q * 2);
- const int* intptr1 = bottom_blob.channel(q * 2 + 1);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
- _bias0 = __msa_fmul_w(_bias0, _scale_out0);
- _bias1 = __msa_fmul_w(_bias1, _scale_out1);
- v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr0 + 64);
- __builtin_prefetch(intptr1 + 64);
- v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
- v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
- v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
- v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
- v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
- v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
- _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
- _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
- _v02 = __msa_fmadd_w(_bias0, _v02, _scale0);
- _v03 = __msa_fmadd_w(_bias0, _v03, _scale0);
- _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
- _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
- _v12 = __msa_fmadd_w(_bias1, _v12, _scale1);
- _v13 = __msa_fmadd_w(_bias1, _v13, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
- *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
- *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v02, _v12, _slope);
- *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v03, _v13, _slope);
-
- intptr0 += 16;
- intptr1 += 16;
- ptr += 32;
- }
- for (; i + 1 < size; i += 2)
- {
- __builtin_prefetch(intptr0 + 32);
- __builtin_prefetch(intptr1 + 32);
- v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
- v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
- _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
- _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
- _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
- _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v00, _v10, _slope);
- *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v01, _v11, _slope);
-
- intptr0 += 8;
- intptr1 += 8;
- ptr += 16;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- }
- if (out_elempack == 1)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr0 = top_blob.channel(q * 4);
- signed char* ptr1 = top_blob.channel(q * 4 + 1);
- signed char* ptr2 = top_blob.channel(q * 4 + 2);
- signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
-
- v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
- v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
- int i = 0;
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale);
- v16i8 v = float2int8leakyrelu(_v, _slope);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
-
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr0 = top_blob.channel(q * 4);
- signed char* ptr1 = top_blob.channel(q * 4 + 1);
- signed char* ptr2 = top_blob.channel(q * 4 + 2);
- signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
- v4f32 _bias = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 4, 0);
-
- v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
- _bias = __msa_fmul_w(_bias, _scale_out);
- v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
- int i = 0;
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale);
- v16i8 v = float2int8leakyrelu(_v, _slope);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
-
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- }
-}
diff --git a/src/layer/mips/requantize_leakyrelu_pack8.h b/src/layer/mips/requantize_leakyrelu_pack8.h
deleted file mode 100644
index f7968c9df70..00000000000
--- a/src/layer/mips/requantize_leakyrelu_pack8.h
+++ /dev/null
@@ -1,188 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_leakyrelu_pack8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, float slope, const Option& opt)
-{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- int size = w * h;
-
- int scale_in_data_size = scale_in_data.w;
- int scale_out_data_size = scale_out_data.w;
- int bias_data_size = bias_data.w;
-
- // int8(leakyrelu(v * scale_in, slope) * scale_out)
- // int8_leakyrelu(v * (scale_in * scale_out), slope)
-
- // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
- // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
- v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr + 128);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
- v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
- v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
- v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- _v2 = __msa_fmul_w(_v2, _scale0);
- _v3 = __msa_fmul_w(_v3, _scale1);
- _v4 = __msa_fmul_w(_v4, _scale0);
- _v5 = __msa_fmul_w(_v5, _scale1);
- _v6 = __msa_fmul_w(_v6, _scale0);
- _v7 = __msa_fmul_w(_v7, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
- *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
- *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v4, _v5, _slope);
- *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v6, _v7, _slope);
-
- intptr += 32;
- ptr += 32;
- }
- for (; i + 1 < size; i += 2)
- {
- __builtin_prefetch(intptr + 64);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- _v2 = __msa_fmul_w(_v2, _scale0);
- _v3 = __msa_fmul_w(_v3, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
- *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-
- intptr += 16;
- ptr += 16;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
- intptr += 8;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
- _bias0 = __msa_fmul_w(_bias0, _scale_out0);
- _bias1 = __msa_fmul_w(_bias1, _scale_out1);
- v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr + 128);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
- v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
- v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
- v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
- _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
- _v4 = __msa_fmadd_w(_bias0, _v4, _scale0);
- _v5 = __msa_fmadd_w(_bias1, _v5, _scale1);
- _v6 = __msa_fmadd_w(_bias0, _v6, _scale0);
- _v7 = __msa_fmadd_w(_bias1, _v7, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
- *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
- *((int64_t*)(ptr + 16)) = float2int8leakyrelu(_v4, _v5, _slope);
- *((int64_t*)(ptr + 24)) = float2int8leakyrelu(_v6, _v7, _slope);
-
- intptr += 32;
- ptr += 32;
- }
- for (; i + 1 < size; i += 2)
- {
- __builtin_prefetch(intptr + 64);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
- _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
- *((int64_t*)(ptr + 8)) = float2int8leakyrelu(_v2, _v3, _slope);
-
- intptr += 16;
- ptr += 16;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
-
- intptr += 8;
- ptr += 8;
- }
- }
- }
-}
diff --git a/src/layer/mips/requantize_mips.cpp b/src/layer/mips/requantize_mips.cpp
index 44e55f89477..6d3bbad70e0 100644
--- a/src/layer/mips/requantize_mips.cpp
+++ b/src/layer/mips/requantize_mips.cpp
@@ -23,13 +23,6 @@
namespace ncnn {
-#if __mips_msa
-#include "requantize_leakyrelu_pack4.h"
-#include "requantize_leakyrelu_pack8.h"
-#include "requantize_relu_pack4.h"
-#include "requantize_relu_pack8.h"
-#endif // __mips_msa
-
Requantize_mips::Requantize_mips()
{
#if __mips_msa
@@ -37,1344 +30,567 @@ Requantize_mips::Requantize_mips()
#endif
}
-int Requantize_mips::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
+static void requantize_relu(const int* intptr, signed char* ptr, const Mat& scale_in_data, const Mat& bias_data, const Mat& scale_out_data, int elemcount, int elempack)
{
- int dims = bottom_blob.dims;
- int elempack = bottom_blob.elempack;
-
-#if __mips_msa
- if (elempack == 8)
- {
- if (dims == 1)
- {
- int w = bottom_blob.w;
-
- top_blob.create(w, (size_t)8u, 8, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- if (scale_in_data_size == 1 && scale_out_data_size == 1)
- {
- v4f32 _scale_in = (v4f32)__msa_fill_w_f32(scale_in_data[0]);
- v4f32 _scale_out = (v4f32)__msa_fill_w_f32(scale_out_data[0]);
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in);
- _v1 = __msa_fmul_w(_v1, _scale_in);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out);
- _v1 = __msa_fmul_w(_v1, _scale_out);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias, _v0, _scale_in);
- _v1 = __msa_fmadd_w(_bias, _v1, _scale_in);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out);
- _v1 = __msa_fmul_w(_v1, _scale_out);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out);
- _v1 = __msa_fmul_w(_v1, _scale_out);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- }
- else if (scale_in_data_size == 1 && scale_out_data_size > 1)
- {
- v4f32 _scale_in = (v4f32)__msa_fill_w_f32(scale_in_data[0]);
+ const int scale_in_data_size = scale_in_data.w;
+ const int bias_data_size = bias_data.w;
+ const int scale_out_data_size = scale_out_data.w;
+ const int size = elemcount * elempack;
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
+ // NCNN_LOGE("requantize_relu %d %d %d %d %d", scale_in_data_size, bias_data_size, scale_out_data_size, elemcount, elempack);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in);
- _v1 = __msa_fmul_w(_v1, _scale_in);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
+ // int8(relu(v * scale_in) * scale_out)
+ // int8_relu(v * (scale_in * scale_out))
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
+ // int8(relu(v * scale_in + bias) * scale_out)
+ // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias, _v0, _scale_in);
- _v1 = __msa_fmadd_w(_bias, _v1, _scale_in);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- }
- else if (scale_in_data_size > 1 && scale_out_data_size == 1)
- {
- v4f32 _scale_out = (v4f32)__msa_fill_w_f32(scale_out_data[0]);
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in0);
- _v1 = __msa_fmul_w(_v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out);
- _v1 = __msa_fmul_w(_v1, _scale_out);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out);
- _v1 = __msa_fmul_w(_v1, _scale_out);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out);
- _v1 = __msa_fmul_w(_v1, _scale_out);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- }
- else // if (scale_in_data_size > 1 && scale_out_data_size > 1)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in0);
- _v1 = __msa_fmul_w(_v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 8;
- signed char* ptr = (signed char*)top_blob + i * 8;
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8 + 4, 0);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
- }
- }
- }
+ float scale_in = scale_in_data[0];
+#if __mips_msa
+ v4f32 _scale_in0 = (v4f32)__msa_fill_w_f32(scale_in);
+ v4f32 _scale_in1 = _scale_in0;
+ if (scale_in_data_size > 1)
+ {
+ if (elempack == 8)
+ {
+ _scale_in0 = (v4f32)__msa_ld_w((const float*)scale_in_data);
+ _scale_in1 = (v4f32)__msa_ld_w((const float*)scale_in_data + 4);
}
-
- if (dims == 2)
+ if (elempack == 4)
{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
-
- top_blob.create(w, h, (size_t)8u, 8, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < h; i++)
- {
- const int* intptr = bottom_blob.row<const int>(i);
- signed char* ptr = top_blob.row<signed char>(i);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
-
- for (int j = 0; j < w; j++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in0);
- _v1 = __msa_fmul_w(_v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
-
- intptr += 8;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < h; i++)
- {
- const int* intptr = bottom_blob.row<const int>(i);
- signed char* ptr = top_blob.row<signed char>(i);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8 + 4, 0);
-
- for (int j = 0; j < w; j++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
-
- intptr += 8;
- ptr += 8;
- }
- }
- }
+ _scale_in0 = (v4f32)__msa_ld_w((const float*)scale_in_data);
+ _scale_in1 = _scale_in0;
}
+ }
+#endif // __mips_msa
- if (dims == 3)
+ float scale_out = scale_out_data[0];
+#if __mips_msa
+ v4f32 _scale_out0 = (v4f32)__msa_fill_w_f32(scale_out);
+ v4f32 _scale_out1 = _scale_out0;
+ if (scale_out_data_size > 1)
+ {
+ if (elempack == 8)
{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- int size = w * h;
-
- top_blob.create(w, h, channels, (size_t)8u, 8, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- if (activation_type == 1)
- {
- requantize_relu_pack8_msa(bottom_blob, top_blob, scale_in_data, scale_out_data, bias_data, opt);
- return 0;
- }
-
- if (activation_type == 2 && activation_params[0] > 0.f)
- {
- requantize_leakyrelu_pack8_msa(bottom_blob, top_blob, scale_in_data, scale_out_data, bias_data, activation_params[0], opt);
- return 0;
- }
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
- for (int i = 0; i < size; i++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in0);
- _v1 = __msa_fmul_w(_v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
-
- intptr += 8;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
+ _scale_out0 = (v4f32)__msa_ld_w((const float*)scale_out_data);
+ _scale_out1 = (v4f32)__msa_ld_w((const float*)scale_out_data + 4);
+ }
+ if (elempack == 4)
+ {
+ _scale_out0 = (v4f32)__msa_ld_w((const float*)scale_out_data);
+ _scale_out1 = _scale_out0;
+ }
+ }
+#endif // __mips_msa
- for (int i = 0; i < size; i++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
+ float scale = scale_in * scale_out;
+#if __mips_msa
+ v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
+ v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
+#endif // __mips_msa
- intptr += 8;
- ptr += 8;
- }
- }
- }
+ if (bias_data_size == 0)
+ {
+ int i = 0;
+#if __mips_msa
+ for (; i + 7 < size; i += 8)
+ {
+ __builtin_prefetch(intptr + 32);
+ v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w((intptr + 4)));
+ _v0 = __msa_fmul_w(_v0, _scale0);
+ _v1 = __msa_fmul_w(_v1, _scale1);
+ *((int64_t*)ptr) = float2int8relu(_v0, _v1);
+ intptr += 8;
+ ptr += 8;
+ }
+ for (; i + 3 < size; i += 4)
+ {
+ v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ _v = __msa_fmul_w(_v, _scale0);
+ v16i8 v = float2int8relu(_v, _v);
+ ptr[0] = v[0];
+ ptr[1] = v[1];
+ ptr[2] = v[2];
+ ptr[3] = v[3];
+ intptr += 4;
+ ptr += 4;
+ }
+#endif // __mips_msa
+ for (; i < size; i++)
+ {
+ float v = *intptr * scale;
+ *ptr = float2int8(v);
+ if (*ptr < 0) *ptr = 0;
+ intptr++;
+ ptr++;
}
-
- return 0;
}
-
- if (elempack == 4)
+ else
{
- if (dims == 1)
+ float bias = bias_data[0];
+#if __mips_msa
+ v4f32 _bias0 = (v4f32)__msa_fill_w_f32(bias);
+ v4f32 _bias1 = _bias0;
+ if (bias_data_size > 1)
{
- int w = bottom_blob.w;
- int out_elempack = opt.use_packing_layout && w * elempack % 8 == 0 ? 8 : 1;
- int outw = w * elempack / out_elempack;
-
- top_blob.create(outw, (size_t)out_elempack, out_elempack, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- if (scale_in_data_size == 1 && scale_out_data_size == 1)
- {
- v4f32 _scale_in = (v4f32)__msa_fill_w_f32(scale_in_data[0]);
- v4f32 _scale_out = (v4f32)__msa_fill_w_f32(scale_out_data[0]);
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _bias = (v4f32)__msa_ld_w((const float*)bias_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- }
- else if (scale_in_data_size == 1 && scale_out_data_size > 1)
+ if (elempack == 8)
{
- v4f32 _scale_in = (v4f32)__msa_fill_w_f32(scale_in_data[0]);
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _scale_out = (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _scale_out = (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _scale_out = (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
- v4f32 _bias = (v4f32)__msa_ld_w((const float*)bias_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
+ _bias0 = (v4f32)__msa_ld_w((const float*)bias_data);
+ _bias1 = (v4f32)__msa_ld_w((const float*)bias_data + 4);
}
- else if (scale_in_data_size > 1 && scale_out_data_size == 1)
+ if (elempack == 4)
{
- v4f32 _scale_out = (v4f32)__msa_fill_w_f32(scale_out_data[0]);
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _scale_in = (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _scale_in = (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _scale_in = (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _bias = (v4f32)__msa_ld_w((const float*)bias_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
+ _bias0 = (v4f32)__msa_ld_w((const float*)bias_data);
+ _bias1 = _bias0;
}
- else // if (scale_in_data_size > 1 && scale_out_data_size > 1)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
-
- v4f32 _scale_in = (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _scale_out = (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else if (bias_data_size == 1)
- {
- v4f32 _bias = (v4f32)__msa_fill_w_f32(bias_data[0]);
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
+ }
+#endif // __mips_msa
- v4f32 _scale_in = (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _scale_out = (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- const int* intptr = (const int*)bottom_blob + i * 4;
- signed char* ptr = (signed char*)top_blob + i * 4;
+ bias = bias * scale_out;
+#if __mips_msa
+ _bias0 = __msa_fmul_w(_bias0, _scale_out0);
+ _bias1 = __msa_fmul_w(_bias1, _scale_out1);
+#endif // __mips_msa
- v4f32 _scale_in = (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _scale_out = (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
- v4f32 _bias = (v4f32)__msa_ld_w((const float*)bias_data + i * 4, 0);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr[0] = v[0];
- ptr[1] = v[1];
- ptr[2] = v[2];
- ptr[3] = v[3];
- }
- }
- }
+ int i = 0;
+#if __mips_msa
+ for (; i + 7 < size; i += 8)
+ {
+ __builtin_prefetch(intptr + 32);
+ v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w((intptr + 4)));
+ _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
+ _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
+ *((int64_t*)ptr) = float2int8relu(_v0, _v1);
+ intptr += 8;
+ ptr += 8;
}
-
- if (dims == 2)
+ for (; i + 3 < size; i += 4)
{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int out_elempack = opt.use_packing_layout && h * elempack % 8 == 0 ? 8 : 1;
- int outh = h * elempack / out_elempack;
-
- top_blob.create(w, outh, (size_t)out_elempack, out_elempack, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- if (out_elempack == 8)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < outh; i++)
- {
- const int* intptr0 = bottom_blob.row<const int>(i * 2);
- const int* intptr1 = bottom_blob.row<const int>(i * 2 + 1);
- signed char* ptr = top_blob.row<signed char>(i);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
-
- for (int j = 0; j < w; j++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in0);
- _v1 = __msa_fmul_w(_v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < outh; i++)
- {
- const int* intptr0 = bottom_blob.row<const int>(i * 2);
- const int* intptr1 = bottom_blob.row<const int>(i * 2 + 1);
- signed char* ptr = top_blob.row<signed char>(i);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 8 + 4, 0);
-
- for (int j = 0; j < w; j++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- }
- if (out_elempack == 1)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < h; i++)
- {
- const int* intptr = bottom_blob.row<const int>(i);
- signed char* ptr0 = top_blob.row<signed char>(i * 4);
- signed char* ptr1 = top_blob.row<signed char>(i * 4 + 1);
- signed char* ptr2 = top_blob.row<signed char>(i * 4 + 2);
- signed char* ptr3 = top_blob.row<signed char>(i * 4 + 3);
-
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
+ v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ _v = __msa_fmadd_w(_bias0, _v, _scale0);
+ v16i8 v = float2int8relu(_v, _v);
+ ptr[0] = v[0];
+ ptr[1] = v[1];
+ ptr[2] = v[2];
+ ptr[3] = v[3];
+ intptr += 4;
+ ptr += 4;
+ }
+#endif // __mips_msa
+ for (; i < size; i++)
+ {
+ float v = *intptr * scale + bias;
+ *ptr = float2int8(v);
+ if (*ptr < 0) *ptr = 0;
+ intptr++;
+ ptr++;
+ }
+ }
+}
- for (int j = 0; j < w; j++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
+static void requantize_leakyrelu(const int* intptr, signed char* ptr, const Mat& scale_in_data, const Mat& bias_data, const Mat& scale_out_data, float slope, int elemcount, int elempack)
+{
+ const int scale_in_data_size = scale_in_data.w;
+ const int bias_data_size = bias_data.w;
+ const int scale_out_data_size = scale_out_data.w;
+ const int size = elemcount * elempack;
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < h; i++)
- {
- const int* intptr = bottom_blob.row<const int>(i);
- signed char* ptr0 = top_blob.row<signed char>(i * 4);
- signed char* ptr1 = top_blob.row<signed char>(i * 4 + 1);
- signed char* ptr2 = top_blob.row<signed char>(i * 4 + 2);
- signed char* ptr3 = top_blob.row<signed char>(i * 4 + 3);
+ // NCNN_LOGE("requantize_leakyrelu %d %d %d %d %d", scale_in_data_size, bias_data_size, scale_out_data_size, elemcount, elempack);
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + i * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + i * 4, 0);
- v4f32 _bias = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + i * 4, 0);
+ // int8(leakyrelu(v * scale_in, slope) * scale_out)
+ // int8_leakyrelu(v * (scale_in * scale_out), slope)
- for (int j = 0; j < w; j++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
+ // int8(leakyrelu(v * scale_in + bias, slope) * scale_out)
+ // int8_leakyrelu(v * (scale_in * scale_out) + (bias * scale_out), slope)
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- }
+ float scale_in = scale_in_data[0];
+#if __mips_msa
+ v4f32 _scale_in0 = (v4f32)__msa_fill_w_f32(scale_in);
+ v4f32 _scale_in1 = _scale_in0;
+ if (scale_in_data_size > 1)
+ {
+ if (elempack == 8)
+ {
+ _scale_in0 = (v4f32)__msa_ld_w((const float*)scale_in_data);
+ _scale_in1 = (v4f32)__msa_ld_w((const float*)scale_in_data + 4);
+ }
+ if (elempack == 4)
+ {
+ _scale_in0 = (v4f32)__msa_ld_w((const float*)scale_in_data);
+ _scale_in1 = _scale_in0;
}
+ }
+#endif // __mips_msa
- if (dims == 3)
+ float scale_out = scale_out_data[0];
+#if __mips_msa
+ v4f32 _scale_out0 = (v4f32)__msa_fill_w_f32(scale_out);
+ v4f32 _scale_out1 = _scale_out0;
+ if (scale_out_data_size > 1)
+ {
+ if (elempack == 8)
+ {
+ _scale_out0 = (v4f32)__msa_ld_w((const float*)scale_out_data);
+ _scale_out1 = (v4f32)__msa_ld_w((const float*)scale_out_data + 4);
+ }
+ if (elempack == 4)
{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- int size = w * h;
- int out_elempack = opt.use_packing_layout && channels * elempack % 8 == 0 ? 8 : 1;
- int outc = channels * elempack / out_elempack;
+ _scale_out0 = (v4f32)__msa_ld_w((const float*)scale_out_data);
+ _scale_out1 = _scale_out0;
+ }
+ }
+#endif // __mips_msa
- top_blob.create(w, h, outc, (size_t)out_elempack, out_elempack, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
+ float scale = scale_in * scale_out;
+#if __mips_msa
+ v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
+ v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
+ v4f32 _slope = (v4f32)__msa_fill_w_f32(slope);
+#endif // __mips_msa
- if (activation_type == 1)
+ if (bias_data_size == 0)
+ {
+ int i = 0;
+#if __mips_msa
+ for (; i + 7 < size; i += 8)
+ {
+ __builtin_prefetch(intptr + 32);
+ v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w((intptr + 4)));
+ _v0 = __msa_fmul_w(_v0, _scale0);
+ _v1 = __msa_fmul_w(_v1, _scale1);
+ *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
+ intptr += 8;
+ ptr += 8;
+ }
+ for (; i + 3 < size; i += 4)
+ {
+ v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ _v = __msa_fmul_w(_v, _scale0);
+ v16i8 v = float2int8leakyrelu(_v, _v, _slope);
+ ptr[0] = v[0];
+ ptr[1] = v[1];
+ ptr[2] = v[2];
+ ptr[3] = v[3];
+ intptr += 4;
+ ptr += 4;
+ }
+#endif // __mips_msa
+ for (; i < size; i++)
+ {
+ float v = *intptr * scale;
+ *ptr = float2int8(v);
+ if (*ptr < 0) *ptr *= slope;
+ intptr++;
+ ptr++;
+ }
+ }
+ else
+ {
+ float bias = bias_data[0];
+#if __mips_msa
+ v4f32 _bias0 = (v4f32)__msa_fill_w_f32(bias);
+ v4f32 _bias1 = _bias0;
+ if (bias_data_size > 1)
+ {
+ if (elempack == 8)
{
- requantize_relu_pack4_msa(bottom_blob, top_blob, scale_in_data, scale_out_data, bias_data, opt);
- return 0;
+ _bias0 = (v4f32)__msa_ld_w((const float*)bias_data);
+ _bias1 = (v4f32)__msa_ld_w((const float*)bias_data + 4);
}
-
- if (activation_type == 2 && activation_params[0] > 0.f)
+ if (elempack == 4)
{
- requantize_leakyrelu_pack4_msa(bottom_blob, top_blob, scale_in_data, scale_out_data, bias_data, activation_params[0], opt);
- return 0;
+ _bias0 = (v4f32)__msa_ld_w((const float*)bias_data);
+ _bias1 = _bias0;
}
+ }
+#endif // __mips_msa
- if (out_elempack == 8)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < outc; q++)
- {
- const int* intptr0 = bottom_blob.channel(q * 2);
- const int* intptr1 = bottom_blob.channel(q * 2 + 1);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
- for (int i = 0; i < size; i++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmul_w(_v0, _scale_in0);
- _v1 = __msa_fmul_w(_v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < outc; q++)
- {
- const int* intptr0 = bottom_blob.channel(q * 2);
- const int* intptr1 = bottom_blob.channel(q * 2 + 1);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
- for (int i = 0; i < size; i++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in1);
- _v0 = activation_ps(_v0, activation_type, activation_params);
- _v1 = activation_ps(_v1, activation_type, activation_params);
- _v0 = __msa_fmul_w(_v0, _scale_out0);
- _v1 = __msa_fmul_w(_v1, _scale_out1);
- *((int64_t*)ptr) = float2int8(_v0, _v1);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- }
- if (out_elempack == 1)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr0 = top_blob.channel(q * 4);
- signed char* ptr1 = top_blob.channel(q * 4 + 1);
- signed char* ptr2 = top_blob.channel(q * 4 + 2);
- signed char* ptr3 = top_blob.channel(q * 4 + 3);
+ bias = bias * scale_out;
+#if __mips_msa
+ _bias0 = __msa_fmul_w(_bias0, _scale_out0);
+ _bias1 = __msa_fmul_w(_bias1, _scale_out1);
+#endif // __mips_msa
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
+ int i = 0;
+#if __mips_msa
+ for (; i + 7 < size; i += 8)
+ {
+ __builtin_prefetch(intptr + 32);
+ v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w((intptr + 4)));
+ _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
+ _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
+ *((int64_t*)ptr) = float2int8leakyrelu(_v0, _v1, _slope);
+ intptr += 8;
+ ptr += 8;
+ }
+ for (; i + 3 < size; i += 4)
+ {
+ v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ _v = __msa_fmadd_w(_bias0, _v, _scale0);
+ v16i8 v = float2int8leakyrelu(_v, _v, _slope);
+ ptr[0] = v[0];
+ ptr[1] = v[1];
+ ptr[2] = v[2];
+ ptr[3] = v[3];
+ intptr += 4;
+ ptr += 4;
+ }
+#endif // __mips_msa
+ for (; i < size; i++)
+ {
+ float v = *intptr * scale + bias;
+ *ptr = float2int8(v);
+ if (*ptr < 0) *ptr *= slope;
+ intptr++;
+ ptr++;
+ }
+ }
+}
- for (int i = 0; i < size; i++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
+static void requantize(const int* intptr, signed char* ptr, const Mat& scale_in_data, const Mat& bias_data, const Mat& scale_out_data, int activation_type, const Mat& activation_params, int elemcount, int elempack)
+{
+ if (activation_type == 1)
+ {
+ requantize_relu(intptr, ptr, scale_in_data, bias_data, scale_out_data, elemcount, elempack);
+ return;
+ }
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr0 = top_blob.channel(q * 4);
- signed char* ptr1 = top_blob.channel(q * 4 + 1);
- signed char* ptr2 = top_blob.channel(q * 4 + 2);
- signed char* ptr3 = top_blob.channel(q * 4 + 3);
+ if (activation_type == 2 && activation_params[0] > 0.f)
+ {
+ const float slope = activation_params[0];
+ requantize_leakyrelu(intptr, ptr, scale_in_data, bias_data, scale_out_data, slope, elemcount, elempack);
+ return;
+ }
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
- v4f32 _bias = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 4, 0);
+ const int scale_in_data_size = scale_in_data.w;
+ const int bias_data_size = bias_data.w;
+ const int scale_out_data_size = scale_out_data.w;
+ const int size = elemcount * elempack;
- for (int i = 0; i < size; i++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale_in);
- _v = activation_ps(_v, activation_type, activation_params);
- _v = __msa_fmul_w(_v, _scale_out);
- v16i8 v = float2int8(_v);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
+ // NCNN_LOGE("requantize %d %d %d %d %d", scale_in_data_size, bias_data_size, scale_out_data_size, elemcount, elempack);
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- }
+ float scale_in = scale_in_data[0];
+#if __mips_msa
+ v4f32 _scale_in0 = (v4f32)__msa_fill_w_f32(scale_in);
+ v4f32 _scale_in1 = _scale_in0;
+ if (scale_in_data_size > 1)
+ {
+ if (elempack == 8)
+ {
+ _scale_in0 = (v4f32)__msa_ld_w((const float*)scale_in_data);
+ _scale_in1 = (v4f32)__msa_ld_w((const float*)scale_in_data + 4);
+ }
+ if (elempack == 4)
+ {
+ _scale_in0 = (v4f32)__msa_ld_w((const float*)scale_in_data);
+ _scale_in1 = _scale_in0;
}
-
- return 0;
}
#endif // __mips_msa
- if (dims == 1)
+ float scale_out = scale_out_data[0];
+#if __mips_msa
+ v4f32 _scale_out0 = (v4f32)__msa_fill_w_f32(scale_out);
+ v4f32 _scale_out1 = _scale_out0;
+ if (scale_out_data_size > 1)
{
- int w = bottom_blob.w;
-
- top_blob.create(w, (size_t)1u, opt.blob_allocator);
- if (top_blob.empty())
- return -100;
-
- const int* intptr = bottom_blob;
- signed char* ptr = top_blob;
-
- if (scale_in_data_size == 1 && scale_out_data_size == 1)
+ if (elempack == 8)
{
- const float scale_in = scale_in_data[0];
- const float scale_out = scale_out_data[0];
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
- else if (bias_data_size == 1)
- {
- const float bias = bias_data[0];
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in + bias;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in + bias_data[i];
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
+ _scale_out0 = (v4f32)__msa_ld_w((const float*)scale_out_data);
+ _scale_out1 = (v4f32)__msa_ld_w((const float*)scale_out_data + 4);
}
- else if (scale_in_data_size == 1 && scale_out_data_size > 1)
+ if (elempack == 4)
{
- const float scale_in = scale_in_data[0];
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out_data[i]);
- }
- }
- else if (bias_data_size == 1)
- {
- const float bias = bias_data[0];
+ _scale_out0 = (v4f32)__msa_ld_w((const float*)scale_out_data);
+ _scale_out1 = _scale_out0;
+ }
+ }
+#endif // __mips_msa
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in + bias;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out_data[i]);
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in + bias_data[i];
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out_data[i]);
- }
- }
+ if (bias_data_size == 0)
+ {
+ int i = 0;
+#if __mips_msa
+ for (; i + 7 < size; i += 8)
+ {
+ __builtin_prefetch(intptr + 32);
+ v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w((intptr + 4)));
+ _v0 = __msa_fmul_w(_v0, _scale_in0);
+ _v1 = __msa_fmul_w(_v1, _scale_in1);
+ _v0 = activation_ps(_v0, activation_type, activation_params);
+ _v1 = activation_ps(_v1, activation_type, activation_params);
+ _v0 = __msa_fmul_w(_v0, _scale_out0);
+ _v1 = __msa_fmul_w(_v1, _scale_out1);
+ *((int64_t*)ptr) = float2int8(_v0, _v1);
+ intptr += 8;
+ ptr += 8;
}
- else if (scale_in_data_size > 1 && scale_out_data_size == 1)
+ for (; i + 3 < size; i += 4)
{
- const float scale_out = scale_out_data[0];
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in_data[i];
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
- else if (bias_data_size == 1)
- {
- const float bias = bias_data[0];
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in_data[i] + bias;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in_data[i] + bias_data[i];
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
+ v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ _v = __msa_fmul_w(_v, _scale_in0);
+ _v = activation_ps(_v, activation_type, activation_params);
+ _v = __msa_fmul_w(_v, _scale_out0);
+ v16i8 v = float2int8(_v, _v);
+ ptr[0] = v[0];
+ ptr[1] = v[1];
+ ptr[2] = v[2];
+ ptr[3] = v[3];
+ intptr += 4;
+ ptr += 4;
}
- else // if (scale_in_data_size > 1 && scale_out_data_size > 1)
+#endif // __mips_msa
+ for (; i < size; i++)
{
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in_data[i];
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out_data[i]);
- }
- }
- else if (bias_data_size == 1)
+ float v = *intptr * scale_in;
+ v = activation_ss(v, activation_type, activation_params);
+ *ptr = float2int8(v * scale_out);
+ intptr++;
+ ptr++;
+ }
+ }
+ else
+ {
+ float bias = bias_data[0];
+#if __mips_msa
+ v4f32 _bias0 = (v4f32)__msa_fill_w_f32(bias);
+ v4f32 _bias1 = _bias0;
+ if (bias_data_size > 1)
+ {
+ if (elempack == 8)
{
- const float bias = bias_data[0];
-
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in_data[i] + bias;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out_data[i]);
- }
+ _bias0 = (v4f32)__msa_ld_w((const float*)bias_data);
+ _bias1 = (v4f32)__msa_ld_w((const float*)bias_data + 4);
}
- else
+ if (elempack == 4)
{
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < w; i++)
- {
- float v = intptr[i] * scale_in_data[i] + bias_data[i];
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out_data[i]);
- }
+ _bias0 = (v4f32)__msa_ld_w((const float*)bias_data);
+ _bias1 = _bias0;
}
}
+#endif // __mips_msa
+
+ int i = 0;
+#if __mips_msa
+ for (; i + 7 < size; i += 8)
+ {
+ __builtin_prefetch(intptr + 32);
+ v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w((intptr + 4)));
+ _v0 = __msa_fmadd_w(_bias0, _v0, _scale_in0);
+ _v1 = __msa_fmadd_w(_bias1, _v1, _scale_in1);
+ _v0 = activation_ps(_v0, activation_type, activation_params);
+ _v1 = activation_ps(_v1, activation_type, activation_params);
+ _v0 = __msa_fmul_w(_v0, _scale_out0);
+ _v1 = __msa_fmul_w(_v1, _scale_out1);
+ *((int64_t*)ptr) = float2int8(_v0, _v1);
+ intptr += 8;
+ ptr += 8;
+ }
+ for (; i + 3 < size; i += 4)
+ {
+ v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr));
+ _v = __msa_fmadd_w(_bias0, _v, _scale_in0);
+ _v = activation_ps(_v, activation_type, activation_params);
+ _v = __msa_fmul_w(_v, _scale_out0);
+ v16i8 v = float2int8(_v, _v);
+ ptr[0] = v[0];
+ ptr[1] = v[1];
+ ptr[2] = v[2];
+ ptr[3] = v[3];
+ intptr += 4;
+ ptr += 4;
+ }
+#endif // __mips_msa
+ for (; i < size; i++)
+ {
+ float v = *intptr * scale_in + bias;
+ v = activation_ss(v, activation_type, activation_params);
+ *ptr = float2int8(v * scale_out);
+ intptr++;
+ ptr++;
+ }
}
+}
- if (dims == 2)
- {
- int w = bottom_blob.w;
- int h = bottom_blob.h;
+int Requantize_mips::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
+{
+ const int dims = bottom_blob.dims;
+ const int w = bottom_blob.w;
+ const int h = bottom_blob.h;
+ const int channels = bottom_blob.c;
+ const int elempack = bottom_blob.elempack;
+ const size_t out_elemsize = elempack * 1u;
- top_blob.create(w, h, (size_t)1u, opt.blob_allocator);
+ if (dims == 1)
+ {
+ top_blob.create(w, out_elemsize, elempack, opt.blob_allocator);
if (top_blob.empty())
return -100;
- if (bias_data_size == 0)
+ const int wp = std::max(1, w / opt.num_threads);
+ const int nn_w = (w + wp - 1) / wp;
+
+ #pragma omp parallel for num_threads(opt.num_threads)
+ for (int ii = 0; ii < nn_w; ii++)
{
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < h; i++)
- {
- const int* intptr = bottom_blob.row<const int>(i);
- signed char* ptr = top_blob.row<signed char>(i);
+ const int i = ii * wp;
- const float scale_in = scale_in_data_size == 1 ? scale_in_data[0] : scale_in_data[i];
- const float scale_out = scale_out_data_size == 1 ? scale_out_data[0] : scale_out_data[i];
+ const int* intptr = (const int*)bottom_blob + i * elempack;
+ signed char* ptr = (signed char*)top_blob + i * elempack;
- for (int j = 0; j < w; j++)
- {
- float v = intptr[j] * scale_in;
- ptr[j] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int i = 0; i < h; i++)
- {
- const int* intptr = bottom_blob.row<const int>(i);
- signed char* ptr = top_blob.row<signed char>(i);
+ // assert scale_in_data_size == 1
+ // assert bias_data_size == 0 || bias_data_size == 1
+ // assert scale_out_data_size == 1
- const float scale_in = scale_in_data_size == 1 ? scale_in_data[0] : scale_in_data[i];
- const float scale_out = scale_out_data_size == 1 ? scale_out_data[0] : scale_out_data[i];
- const float bias = bias_data_size == 1 ? bias_data[0] : bias_data[i];
+ const int size = std::min(w - i, wp) * elempack;
- for (int j = 0; j < w; j++)
- {
- float v = intptr[j] * scale_in + bias;
- ptr[j] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
+ requantize(intptr, ptr, scale_in_data, bias_data, scale_out_data, activation_type, activation_params, size, 1);
}
}
- if (dims == 3)
+ if (dims == 2)
{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- int size = w * h;
-
- top_blob.create(w, h, channels, (size_t)1u, opt.blob_allocator);
+ top_blob.create(w, h, out_elemsize, elempack, opt.blob_allocator);
if (top_blob.empty())
return -100;
- if (bias_data_size == 0)
+ #pragma omp parallel for num_threads(opt.num_threads)
+ for (int i = 0; i < h; i++)
{
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
+ const int* intptr = bottom_blob.row<const int>(i);
+ signed char* ptr = top_blob.row<signed char>(i);
- const float scale_in = scale_in_data_size == 1 ? scale_in_data[0] : scale_in_data[q];
- const float scale_out = scale_out_data_size == 1 ? scale_out_data[0] : scale_out_data[q];
+ const Mat scale_in_data_i = scale_in_data_size > 1 ? scale_in_data.range(i * elempack, elempack) : scale_in_data;
+ const Mat bias_data_i = bias_data_size > 1 ? bias_data.range(i * elempack, elempack) : bias_data;
+ const Mat scale_out_data_i = scale_out_data_size > 1 ? scale_out_data.range(i * elempack, elempack) : scale_out_data;
- for (int i = 0; i < size; i++)
- {
- float v = intptr[i] * scale_in;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
+ requantize(intptr, ptr, scale_in_data_i, bias_data_i, scale_out_data_i, activation_type, activation_params, w, elempack);
}
- else
+ }
+
+ if (dims == 3)
+ {
+ top_blob.create(w, h, channels, out_elemsize, elempack, opt.blob_allocator);
+ if (top_blob.empty())
+ return -100;
+
+ #pragma omp parallel for num_threads(opt.num_threads)
+ for (int q = 0; q < channels; q++)
{
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
+ const int* intptr = bottom_blob.channel(q);
+ signed char* ptr = top_blob.channel(q);
- const float scale_in = scale_in_data_size == 1 ? scale_in_data[0] : scale_in_data[q];
- const float scale_out = scale_out_data_size == 1 ? scale_out_data[0] : scale_out_data[q];
- const float bias = bias_data_size == 1 ? bias_data[0] : bias_data[q];
+ const Mat scale_in_data_q = scale_in_data_size > 1 ? scale_in_data.range(q * elempack, elempack) : scale_in_data;
+ const Mat bias_data_q = bias_data_size > 1 ? bias_data.range(q * elempack, elempack) : bias_data;
+ const Mat scale_out_data_q = scale_out_data_size > 1 ? scale_out_data.range(q * elempack, elempack) : scale_out_data;
- for (int i = 0; i < size; i++)
- {
- float v = intptr[i] * scale_in + bias;
- ptr[i] = float2int8(activation_ss(v, activation_type, activation_params) * scale_out);
- }
- }
+ requantize(intptr, ptr, scale_in_data_q, bias_data_q, scale_out_data_q, activation_type, activation_params, w * h, elempack);
}
}
diff --git a/src/layer/mips/requantize_relu_pack4.h b/src/layer/mips/requantize_relu_pack4.h
deleted file mode 100644
index e43797bd8b8..00000000000
--- a/src/layer/mips/requantize_relu_pack4.h
+++ /dev/null
@@ -1,263 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack4_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- int size = w * h;
- int outc = top_blob.c;
- int out_elempack = top_blob.elempack;
-
- int scale_in_data_size = scale_in_data.w;
- int scale_out_data_size = scale_out_data.w;
- int bias_data_size = bias_data.w;
-
- // int8(relu(v * scale_in) * scale_out)
- // int8_relu(v * (scale_in * scale_out))
-
- // int8(relu(v * scale_in + bias) * scale_out)
- // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
- if (out_elempack == 8)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < outc; q++)
- {
- const int* intptr0 = bottom_blob.channel(q * 2);
- const int* intptr1 = bottom_blob.channel(q * 2 + 1);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr0 + 64);
- __builtin_prefetch(intptr1 + 64);
- v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
- v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
- v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
- v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
- v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
- v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
- _v00 = __msa_fmul_w(_v00, _scale0);
- _v01 = __msa_fmul_w(_v01, _scale0);
- _v02 = __msa_fmul_w(_v02, _scale0);
- _v03 = __msa_fmul_w(_v03, _scale0);
- _v10 = __msa_fmul_w(_v10, _scale1);
- _v11 = __msa_fmul_w(_v11, _scale1);
- _v12 = __msa_fmul_w(_v12, _scale1);
- _v13 = __msa_fmul_w(_v13, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v00, _v10);
- *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
- *((int64_t*)(ptr + 16)) = float2int8relu(_v02, _v12);
- *((int64_t*)(ptr + 24)) = float2int8relu(_v03, _v13);
-
- intptr0 += 16;
- intptr1 += 16;
- ptr += 32;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < outc; q++)
- {
- const int* intptr0 = bottom_blob.channel(q * 2);
- const int* intptr1 = bottom_blob.channel(q * 2 + 1);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
- _bias0 = __msa_fmul_w(_bias0, _scale_out0);
- _bias1 = __msa_fmul_w(_bias1, _scale_out1);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr0 + 64);
- __builtin_prefetch(intptr1 + 64);
- v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
- v4f32 _v02 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 8, 0));
- v4f32 _v03 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 12, 0));
- v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
- v4f32 _v12 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 8, 0));
- v4f32 _v13 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 12, 0));
- _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
- _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
- _v02 = __msa_fmadd_w(_bias0, _v02, _scale0);
- _v03 = __msa_fmadd_w(_bias0, _v03, _scale0);
- _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
- _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
- _v12 = __msa_fmadd_w(_bias1, _v12, _scale1);
- _v13 = __msa_fmadd_w(_bias1, _v13, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v00, _v10);
- *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
- *((int64_t*)(ptr + 16)) = float2int8relu(_v02, _v12);
- *((int64_t*)(ptr + 24)) = float2int8relu(_v03, _v13);
-
- intptr0 += 16;
- intptr1 += 16;
- ptr += 32;
- }
- for (; i + 1 < size; i += 2)
- {
- __builtin_prefetch(intptr0 + 32);
- __builtin_prefetch(intptr1 + 32);
- v4f32 _v00 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v01 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0 + 4, 0));
- v4f32 _v10 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- v4f32 _v11 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1 + 4, 0));
- _v00 = __msa_fmadd_w(_bias0, _v00, _scale0);
- _v01 = __msa_fmadd_w(_bias0, _v01, _scale0);
- _v10 = __msa_fmadd_w(_bias1, _v10, _scale1);
- _v11 = __msa_fmadd_w(_bias1, _v11, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v00, _v10);
- *((int64_t*)(ptr + 8)) = float2int8relu(_v01, _v11);
-
- intptr0 += 8;
- intptr1 += 8;
- ptr += 16;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr0 + 16);
- __builtin_prefetch(intptr1 + 16);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr0, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr1, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
- intptr0 += 4;
- intptr1 += 4;
- ptr += 8;
- }
- }
- }
- }
- if (out_elempack == 1)
- {
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr0 = top_blob.channel(q * 4);
- signed char* ptr1 = top_blob.channel(q * 4 + 1);
- signed char* ptr2 = top_blob.channel(q * 4 + 2);
- signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
-
- v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
-
- int i = 0;
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmul_w(_v, _scale);
- v16i8 v = float2int8relu(_v);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
-
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr0 = top_blob.channel(q * 4);
- signed char* ptr1 = top_blob.channel(q * 4 + 1);
- signed char* ptr2 = top_blob.channel(q * 4 + 2);
- signed char* ptr3 = top_blob.channel(q * 4 + 3);
-
- v4f32 _scale_in = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 4, 0);
- v4f32 _scale_out = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 4, 0);
- v4f32 _bias = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 4, 0);
-
- v4f32 _scale = __msa_fmul_w(_scale_in, _scale_out);
- _bias = __msa_fmul_w(_bias, _scale_out);
-
- int i = 0;
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 16);
- v4f32 _v = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- _v = __msa_fmadd_w(_bias, _v, _scale);
- v16i8 v = float2int8relu(_v);
- ptr0[0] = v[0];
- ptr1[0] = v[1];
- ptr2[0] = v[2];
- ptr3[0] = v[3];
-
- intptr += 4;
- ptr0 += 1;
- ptr1 += 1;
- ptr2 += 1;
- ptr3 += 1;
- }
- }
- }
- }
-}
diff --git a/src/layer/mips/requantize_relu_pack8.h b/src/layer/mips/requantize_relu_pack8.h
deleted file mode 100644
index 824b668cb73..00000000000
--- a/src/layer/mips/requantize_relu_pack8.h
+++ /dev/null
@@ -1,186 +0,0 @@
-// Tencent is pleased to support the open source community by making ncnn available.
-//
-// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
-//
-// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
-// in compliance with the License. You may obtain a copy of the License at
-//
-// https://opensource.org/licenses/BSD-3-Clause
-//
-// Unless required by applicable law or agreed to in writing, software distributed
-// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
-// CONDITIONS OF ANY KIND, either express or implied. See the License for the
-// specific language governing permissions and limitations under the License.
-
-static void requantize_relu_pack8_msa(const Mat& bottom_blob, Mat& top_blob, const Mat& scale_in_data, const Mat& scale_out_data, const Mat& bias_data, const Option& opt)
-{
- int w = bottom_blob.w;
- int h = bottom_blob.h;
- int channels = bottom_blob.c;
- int size = w * h;
-
- int scale_in_data_size = scale_in_data.w;
- int scale_out_data_size = scale_out_data.w;
- int bias_data_size = bias_data.w;
-
- // int8(relu(v * scale_in) * scale_out)
- // int8_relu(v * (scale_in * scale_out))
-
- // int8(relu(v * scale_in + bias) * scale_out)
- // int8_relu(v * (scale_in * scale_out) + (bias * scale_out))
-
- if (bias_data_size == 0)
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr + 128);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
- v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
- v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
- v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- _v2 = __msa_fmul_w(_v2, _scale0);
- _v3 = __msa_fmul_w(_v3, _scale1);
- _v4 = __msa_fmul_w(_v4, _scale0);
- _v5 = __msa_fmul_w(_v5, _scale1);
- _v6 = __msa_fmul_w(_v6, _scale0);
- _v7 = __msa_fmul_w(_v7, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
- *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
- *((int64_t*)(ptr + 16)) = float2int8relu(_v4, _v5);
- *((int64_t*)(ptr + 24)) = float2int8relu(_v6, _v7);
-
- intptr += 32;
- ptr += 32;
- }
- for (; i + 1 < size; i += 2)
- {
- __builtin_prefetch(intptr + 64);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- _v2 = __msa_fmul_w(_v2, _scale0);
- _v3 = __msa_fmul_w(_v3, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
- *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-
- intptr += 16;
- ptr += 16;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmul_w(_v0, _scale0);
- _v1 = __msa_fmul_w(_v1, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
- intptr += 8;
- ptr += 8;
- }
- }
- }
- else
- {
- #pragma omp parallel for num_threads(opt.num_threads)
- for (int q = 0; q < channels; q++)
- {
- const int* intptr = bottom_blob.channel(q);
- signed char* ptr = top_blob.channel(q);
-
- v4f32 _scale_in0 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8, 0);
- v4f32 _scale_in1 = scale_in_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_in_data[0]) : (v4f32)__msa_ld_w((const float*)scale_in_data + q * 8 + 4, 0);
- v4f32 _scale_out0 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8, 0);
- v4f32 _scale_out1 = scale_out_data_size == 1 ? (v4f32)__msa_fill_w_f32(scale_out_data[0]) : (v4f32)__msa_ld_w((const float*)scale_out_data + q * 8 + 4, 0);
- v4f32 _bias0 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8, 0);
- v4f32 _bias1 = bias_data_size == 1 ? (v4f32)__msa_fill_w_f32(bias_data[0]) : (v4f32)__msa_ld_w((const float*)bias_data + q * 8 + 4, 0);
-
- v4f32 _scale0 = __msa_fmul_w(_scale_in0, _scale_out0);
- v4f32 _scale1 = __msa_fmul_w(_scale_in1, _scale_out1);
- _bias0 = __msa_fmul_w(_bias0, _scale_out0);
- _bias1 = __msa_fmul_w(_bias1, _scale_out1);
-
- int i = 0;
- for (; i + 3 < size; i += 4)
- {
- __builtin_prefetch(intptr + 128);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- v4f32 _v4 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 16, 0));
- v4f32 _v5 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 20, 0));
- v4f32 _v6 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 24, 0));
- v4f32 _v7 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 28, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
- _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
- _v4 = __msa_fmadd_w(_bias0, _v4, _scale0);
- _v5 = __msa_fmadd_w(_bias1, _v5, _scale1);
- _v6 = __msa_fmadd_w(_bias0, _v6, _scale0);
- _v7 = __msa_fmadd_w(_bias1, _v7, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
- *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
- *((int64_t*)(ptr + 16)) = float2int8relu(_v4, _v5);
- *((int64_t*)(ptr + 24)) = float2int8relu(_v6, _v7);
-
- intptr += 32;
- ptr += 32;
- }
- for (; i + 1 < size; i += 2)
- {
- __builtin_prefetch(intptr + 64);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- v4f32 _v2 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 8, 0));
- v4f32 _v3 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 12, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- _v2 = __msa_fmadd_w(_bias0, _v2, _scale0);
- _v3 = __msa_fmadd_w(_bias1, _v3, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
- *((int64_t*)(ptr + 8)) = float2int8relu(_v2, _v3);
-
- intptr += 16;
- ptr += 16;
- }
- for (; i < size; i++)
- {
- __builtin_prefetch(intptr + 32);
- v4f32 _v0 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr, 0));
- v4f32 _v1 = (v4f32)__msa_ffint_s_w(__msa_ld_w(intptr + 4, 0));
- _v0 = __msa_fmadd_w(_bias0, _v0, _scale0);
- _v1 = __msa_fmadd_w(_bias1, _v1, _scale1);
- *((int64_t*)ptr) = float2int8relu(_v0, _v1);
-
- intptr += 8;
- ptr += 8;
- }
- }
- }
-}