tflm/opt/conv_s8_d4.cc - sw/kelvin - Git at Google

 /*
  * Copyright 2024 Google LLC
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
  * You may obtain a copy of the License at
  *
  *     http://www.apache.org/licenses/LICENSE-2.0
  *
  * 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.
  */

 // Convolution based on Kelvin ops
 // Data types: input: s8, filter: s8, bias: s32
 // Special case for filter depth = 4n

 #include <cstdlib>
 #include <memory>

 #include "crt/kelvin.h"
 #include "tensorflow/lite/kernels/internal/common.h"
 #include "tensorflow/lite/kernels/internal/reference/integer_ops/conv.h"
 #include "tensorflow/lite/kernels/internal/runtime_shape.h"
 #include "tensorflow/lite/kernels/internal/types.h"
 #include "tflm/opt/conv_s8.h"

 #define unlikely(x) (__builtin_expect(false || (x), false))
 #define likely(x) (__builtin_expect(false || (x), true))

 namespace kelvin::opt {
 namespace {

 void Filter_N_H_W_M(const int8_t* input, int8_t* output, int N, int H, int W, int M) {
   const int8_t(&in)[8][H][W][M] = *(int8_t(*)[8][H][W][M])input;
   int8_t(&out)[H][W][M / 4][8][4] = *(int8_t(*)[H][W][M / 4][8][4]) output;
   // assert(M >= 4);
   for (int zo = 0; zo < N; ++zo) {
     for (int ky = 0; ky < H; ++ky) {
       for (int kx = 0; kx < W; ++kx) {
         for (int zi = 0; zi < M; ++zi) {
           const int zi_hi = zi >> 2;  // div4
           const int zi_lo = zi & 3;   // rem4
           out[ky][kx][zi_hi][zo][zi_lo] = in[zo][ky][kx][zi];
         }
       }
     }
   }
   // Zero out the rest of the output.
   for (int zo = N; zo < 8; ++zo) {
     for (int ky = 0; ky < H; ++ky) {
       for (int kx = 0; kx < W; ++kx) {
         for (int zi = M; zi < 4; ++zi) {
           const int zi_hi = zi >> 2;  // div4
           const int zi_lo = zi & 3;   // rem4
           out[ky][kx][zi_hi][zo][zi_lo] = 0;
         }
       }
     }
   }
 }

 void Swizzle(const int32_t* input, int32_t* output, int N) {
   assert(N <= 8);
   const int32_t(&in)[8] = *(int32_t(*)[8])input;
   int32_t(&out)[32] = *(int32_t(*)[32]) output;
   // Convert to accumulator swizzle pattern.
   memset(out, 0, 32 * sizeof(int32_t));
   int offsets[] = {0, 16, 8, 24, 1, 17, 9, 25};
   for (int i = 0; i < N; ++i) {
     int offset = offsets[i];
     out[0 + offset] = out[2 + offset] = out[4 + offset] = out[6 + offset] = in[i];
   }
 }

 }  // namespace

 void ConvS8D4(
     const tflite::ConvParams& params, const int32_t* output_multiplier,
     const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
     const int8_t* input_data, const tflite::RuntimeShape& filter_shape,
     const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
     const int32_t* bias_data, const tflite::RuntimeShape& output_shape,
     int8_t* output_data) {
   // Get parameters.
   const int32_t input_offset = params.input_offset;  // r = s(q - Z)
   const int32_t neg_input_offset = -params.input_offset;  // r = s(q - Z)
   const int stride_width = params.stride_width;
   const int stride_height = params.stride_height;
   const int dilation_width_factor = params.dilation_width_factor;
   const int dilation_height_factor = params.dilation_height_factor;
   const int pad_width = params.padding_values.width;
   const int pad_height = params.padding_values.height;
   const int32_t output_offset = params.output_offset;

   // Set min and max value of the output.
   const int32_t output_activation_min = params.quantized_activation_min;
   const int32_t output_activation_max = params.quantized_activation_max;

   // Consistency check.
   TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
   TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
   TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
   TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
   const int batches = MatchingDim(input_shape, 0, output_shape, 0);
   const int input_depth = input_shape.Dims(3);
   const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
   if (bias_data) {
     TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
   }

   // Check dimensions of the tensors.
   const int input_height = input_shape.Dims(1);
   const int input_width = input_shape.Dims(2);
   const int filter_height = filter_shape.Dims(1);
   const int filter_width = filter_shape.Dims(2);
   const int filter_input_depth = filter_shape.Dims(3);
   const int groups = input_depth / filter_input_depth;
   TFLITE_DCHECK_NE(groups, 0);
   TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
   const int filters_per_group = output_depth / groups;
   TFLITE_DCHECK_NE(filters_per_group, 0);
   const int output_height = output_shape.Dims(1);
   const int output_width = output_shape.Dims(2);

   union {
     vconv_u8_t conv;
     uint32_t raw;
   } cmds;
   cmds.conv.mode = 0;
   cmds.conv.start = 0;
   cmds.conv.stop = 7;
   cmds.conv.sbias1 = input_offset;
   cmds.conv.sdata1 = true;
   cmds.conv.sbias2 = 0;
   cmds.conv.sdata2 = true;

   const size_t swizzled_filter_data_size =
       8 * filter_height * filter_width * filter_input_depth;
   std::unique_ptr<int8_t> swizzled_filter_data(reinterpret_cast<int8_t*>(
       ::aligned_alloc(32, swizzled_filter_data_size)));
   int8_t* p_swizzled_filter_data = swizzled_filter_data.get();
   int32_t swizzled_bias_data[32];
   int32_t swizzled_mult_data[32];
   int32_t swizzled_shift_data[32];

   int out_channel = 0;
   do {
     int out_channels_this_iter = std::min(8, output_depth - out_channel);
     Filter_N_H_W_M(filter_data + (out_channel * filter_height * filter_width *
                                   filter_input_depth),
                    p_swizzled_filter_data, out_channels_this_iter, filter_height, filter_width,
                    filter_input_depth);
     Swizzle(bias_data + out_channel, swizzled_bias_data, out_channels_this_iter);
     Swizzle(output_multiplier + out_channel, swizzled_mult_data, out_channels_this_iter);
     Swizzle(output_shift + out_channel, swizzled_shift_data, out_channels_this_iter);
     vld_w_x_m(v16, swizzled_bias_data);
     vld_w_x_m(v20, swizzled_mult_data);
     vld_w_x_m(v24, swizzled_shift_data);
     vrsub_w_vx_m(v24, v24, 0);

     for (int batch = 0; batch < batches; ++batch) {
       const int8_t* p_output =
           output_data + (batch * output_height * output_width * output_depth) +
           out_channel;
       for (int out_y = 0; out_y < output_height; ++out_y) {
         const int in_y_origin = (out_y * stride_height) - pad_height;
         const int out_y_offset = (out_y * output_width * output_depth);
         int out_x = 0;
         do {
           int out_xs_this_iter = std::min(8, output_width - out_x);
           // 8x accumulators
           vdup_w_x_m(v48, 0);
           vdup_w_x_m(v52, 0);
           acset_v(v48, v48);
           int in_channel = 0;
           do {
             int in_channels_this_iter = std::min(filter_input_depth, 32);
             // Calculate first valid filter_y
             int filter_y = 0;
             {
               int in_y = in_y_origin;
               while (in_y < 0) {
                 ++filter_y;
                 in_y += (dilation_height_factor);
               }
             }
             for (; filter_y < filter_height; ++filter_y) {
               const int y_filter_offset =
                   (filter_y * filter_width * 8 * input_depth);
               const int in_y = in_y_origin + dilation_height_factor * filter_y;
               if (in_y >= input_height) {
                 break;
               }
               const int8_t* p_in =
                   input_data + in_channel + (in_y * input_width * input_depth) +
                   (batch * input_height * input_width * input_depth);

               int in_x[8];
 #pragma GCC unroll 8
               for (int i = 0; i < 8; ++i) {
                 in_x[i] = ((out_x + i) * stride_width) - pad_width;
               }
               for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
                 const int8_t* p_in_x[8];
                 int first_right_pad = -1;

 #pragma GCC unroll 8
                 for (int i = 0; i < 8; ++i) {
                   p_in_x[i] = p_in + (in_x[i] * input_depth);
                 }

 #pragma GCC unroll 8
                 for (int i = 7; i >= 0; --i) {
                   if (in_x[i] < input_width) {
                     break;
                   }
                   first_right_pad = i;
                 }
                 bool left_pad = (in_x[0] < 0);
                 bool right_pad = (first_right_pad != -1);

                 int stride = input_depth * stride_width;

                 if (unlikely(left_pad)) {
                   vdup_b_x(v0, -input_offset);
                   vld_b_s_xx(v1, p_in_x[1], stride);
                   vld_b_s_xx(v2, p_in_x[2], stride);
                   vld_b_s_xx(v3, p_in_x[3], stride);
                   vld_b_s_xx_m(v4, p_in_x[4], stride);
                 } else if (unlikely(right_pad)) {
                   int first_pad = std::min(first_right_pad, out_xs_this_iter);
                   switch (first_pad) {
                     case 0:
                       vdup_b_x(v0, neg_input_offset);
                     case 1:
                       vdup_b_x(v1, neg_input_offset);
                     case 2:
                       vdup_b_x(v2, neg_input_offset);
                     case 3:
                       vdup_b_x(v3, neg_input_offset);
                     case 4:
                       vdup_b_x(v4, neg_input_offset);
                     case 5:
                       vdup_b_x(v5, neg_input_offset);
                     case 6:
                       vdup_b_x(v6, neg_input_offset);
                     case 7:
                       vdup_b_x(v7, neg_input_offset);
                   }
                   switch (8 - first_pad) { // rest (stripmines?)
                     case 0:
                       vld_b_s_xx(v7, p_in_x[7], stride);
                     case 1:
                       vld_b_s_xx(v6, p_in_x[6], stride);
                     case 2:
                       vld_b_s_xx(v5, p_in_x[5], stride);
                     case 3:
                       vld_b_s_xx(v4, p_in_x[4], stride);
                     case 4:
                       vld_b_s_xx(v3, p_in_x[3], stride);
                     case 5:
                       vld_b_s_xx(v2, p_in_x[2], stride);
                     case 6:
                       vld_b_s_xx(v1, p_in_x[1], stride);
                     case 7:
                       vld_b_s_xx(v0, p_in_x[0], stride);
                   }
                 } else if (likely(!left_pad && !right_pad)) {
                   // Inputs
                   vld_b_s_xx_m(v0, p_in_x[0], stride);
                   vld_b_s_xx_m(v4, p_in_x[4], stride);
                 } else {
                   vdup_b_x(v0, neg_input_offset);
                   vdup_b_x(v7, neg_input_offset);
                   vld_b_s_xx_m(v1, p_in_x[1], stride);
                   vld_b_s_xx(v5, p_in_x[5], stride);
                   vld_b_s_xx(v6, p_in_x[6], stride);
                 }
                 size_t local_filter_offset = y_filter_offset +
                                              (filter_x * 8 * input_depth) +
                                              (in_channel * 8);
                 int8_t* p_local_filter_start =
                     p_swizzled_filter_data + local_filter_offset;
                 vld_b_p_x_m(v8, p_local_filter_start);
                 vld_b_x_m(v12, p_local_filter_start);

                 cmds.conv.stop = (in_channels_this_iter / 4) - 1;
                 aconv_vxv(v48, v0, cmds, v8);

 #pragma GCC unroll 8
                 for (int i = 0; i < 8; ++i) {
                   in_x[i] += dilation_width_factor;
                 }
               }
             }
             in_channel += in_channels_this_iter;
           } while (in_channel < filter_input_depth);
           vcget(v48);
           vadd_w_vv_m(v48, v48, v16);
           vadd_w_vv_m(v52, v52, v16);
           vdmulh_w_rn_vv_m(v48, v48, v20);
           vdmulh_w_rn_vv_m(v52, v52, v20);
           vsha_w_r_vv_m(v48, v48, v24);
           vsha_w_r_vv_m(v52, v52, v24);
           vadd_w_vx_m(v48, v48, output_offset);
           vadd_w_vx_m(v52, v52, output_offset);
           vmin_w_vx_m(v48, v48, output_activation_max);
           vmin_w_vx_m(v52, v52, output_activation_max);
           vmax_w_vx_m(v48, v48, output_activation_min);
           vmax_w_vx_m(v52, v52, output_activation_min);
           vsraqs_b_vx(v56, v48, 0);
           vsraqs_b_vx(v57, v52, 0);

           const int8_t* p_out_x[8];
 #pragma GCC unroll 8
           for (int i = 0; i < 8; ++i) {
             p_out_x[i] = p_output + out_y_offset + ((out_x + i) * output_depth);
           }

           vslidep_h_4_vv(v58, v57, v57);  // x7
           vslidep_h_4_vv(v59, v58, v58);  // x6
           vslidep_h_4_vv(v60, v59, v59);  // x5
           vslidep_h_4_vv(v61, v60, v60);  // x4
           vslidep_h_4_vv(v62, v56, v56);  // x3
           vslidep_h_4_vv(v63, v62, v62);  // x2
           vslidep_h_4_vv(v57, v63, v63);  // x1
           vslidep_h_4_vv(v56, v57, v57);  // x0
           switch (out_xs_this_iter) {
             case 8:
               vst_b_l_xx(v58, p_out_x[7], out_channels_this_iter);
             case 7:
               vst_b_l_xx(v59, p_out_x[6], out_channels_this_iter);
             case 6:
               vst_b_l_xx(v60, p_out_x[5], out_channels_this_iter);
             case 5:
               vst_b_l_xx(v61, p_out_x[4], out_channels_this_iter);
             case 4:
               vst_b_l_xx(v62, p_out_x[3], out_channels_this_iter);
             case 3:
               vst_b_l_xx(v63, p_out_x[2], out_channels_this_iter);
             case 2:
               vst_b_l_xx(v57, p_out_x[1], out_channels_this_iter);
             case 1:
               vst_b_l_xx(v56, p_out_x[0], out_channels_this_iter);
           }
           out_x += out_xs_this_iter;
         } while (out_x < output_width);
       }
     }
     out_channel += out_channels_this_iter;
   } while (out_channel < output_depth);
 }
 }  // namespace kelvin::opt
	/*
	* Copyright 2024 Google LLC
	*
	* Licensed under the Apache License, Version 2.0 (the "License");
	* you may not use this file except in compliance with the License.
	* You may obtain a copy of the License at
	*
	* http://www.apache.org/licenses/LICENSE-2.0
	*
	* 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.
	*/

	// Convolution based on Kelvin ops
	// Data types: input: s8, filter: s8, bias: s32
	// Special case for filter depth = 4n

	#include <cstdlib>
	#include <memory>

	#include "crt/kelvin.h"
	#include "tensorflow/lite/kernels/internal/common.h"
	#include "tensorflow/lite/kernels/internal/reference/integer_ops/conv.h"
	#include "tensorflow/lite/kernels/internal/runtime_shape.h"
	#include "tensorflow/lite/kernels/internal/types.h"
	#include "tflm/opt/conv_s8.h"

	#define unlikely(x) (__builtin_expect(false \|\| (x), false))
	#define likely(x) (__builtin_expect(false \|\| (x), true))

	namespace kelvin::opt {
	namespace {

	void Filter_N_H_W_M(const int8_t* input, int8_t* output, int N, int H, int W, int M) {
	const int8_t(&in)[8][H][W][M] = (int8_t()[8][H][W][M])input;
	int8_t(&out)[H][W][M / 4][8][4] = (int8_t()[H][W][M / 4][8][4]) output;
	// assert(M >= 4);
	for (int zo = 0; zo < N; ++zo) {
	for (int ky = 0; ky < H; ++ky) {
	for (int kx = 0; kx < W; ++kx) {
	for (int zi = 0; zi < M; ++zi) {
	const int zi_hi = zi >> 2; // div4
	const int zi_lo = zi & 3; // rem4
	out[ky][kx][zi_hi][zo][zi_lo] = in[zo][ky][kx][zi];
	}
	}
	}
	}
	// Zero out the rest of the output.
	for (int zo = N; zo < 8; ++zo) {
	for (int ky = 0; ky < H; ++ky) {
	for (int kx = 0; kx < W; ++kx) {
	for (int zi = M; zi < 4; ++zi) {
	const int zi_hi = zi >> 2; // div4
	const int zi_lo = zi & 3; // rem4
	out[ky][kx][zi_hi][zo][zi_lo] = 0;
	}
	}
	}
	}
	}

	void Swizzle(const int32_t* input, int32_t* output, int N) {
	assert(N <= 8);
	const int32_t(&in)[8] = (int32_t()[8])input;
	int32_t(&out)[32] = (int32_t()[32]) output;
	// Convert to accumulator swizzle pattern.
	memset(out, 0, 32 * sizeof(int32_t));
	int offsets[] = {0, 16, 8, 24, 1, 17, 9, 25};
	for (int i = 0; i < N; ++i) {
	int offset = offsets[i];
	out[0 + offset] = out[2 + offset] = out[4 + offset] = out[6 + offset] = in[i];
	}
	}

	} // namespace

	void ConvS8D4(
	const tflite::ConvParams& params, const int32_t* output_multiplier,
	const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
	const int8_t* input_data, const tflite::RuntimeShape& filter_shape,
	const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
	const int32_t* bias_data, const tflite::RuntimeShape& output_shape,
	int8_t* output_data) {
	// Get parameters.
	const int32_t input_offset = params.input_offset; // r = s(q - Z)
	const int32_t neg_input_offset = -params.input_offset; // r = s(q - Z)
	const int stride_width = params.stride_width;
	const int stride_height = params.stride_height;
	const int dilation_width_factor = params.dilation_width_factor;
	const int dilation_height_factor = params.dilation_height_factor;
	const int pad_width = params.padding_values.width;
	const int pad_height = params.padding_values.height;
	const int32_t output_offset = params.output_offset;

	// Set min and max value of the output.
	const int32_t output_activation_min = params.quantized_activation_min;
	const int32_t output_activation_max = params.quantized_activation_max;

	// Consistency check.
	TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
	TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
	TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
	TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
	const int batches = MatchingDim(input_shape, 0, output_shape, 0);
	const int input_depth = input_shape.Dims(3);
	const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
	if (bias_data) {
	TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
	}

	// Check dimensions of the tensors.
	const int input_height = input_shape.Dims(1);
	const int input_width = input_shape.Dims(2);
	const int filter_height = filter_shape.Dims(1);
	const int filter_width = filter_shape.Dims(2);
	const int filter_input_depth = filter_shape.Dims(3);
	const int groups = input_depth / filter_input_depth;
	TFLITE_DCHECK_NE(groups, 0);
	TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
	const int filters_per_group = output_depth / groups;
	TFLITE_DCHECK_NE(filters_per_group, 0);
	const int output_height = output_shape.Dims(1);
	const int output_width = output_shape.Dims(2);

	union {
	vconv_u8_t conv;
	uint32_t raw;
	} cmds;
	cmds.conv.mode = 0;
	cmds.conv.start = 0;
	cmds.conv.stop = 7;
	cmds.conv.sbias1 = input_offset;
	cmds.conv.sdata1 = true;
	cmds.conv.sbias2 = 0;
	cmds.conv.sdata2 = true;

	const size_t swizzled_filter_data_size =
	8 * filter_height * filter_width * filter_input_depth;
	std::unique_ptr<int8_t> swizzled_filter_data(reinterpret_cast<int8_t*>(
	::aligned_alloc(32, swizzled_filter_data_size)));
	int8_t* p_swizzled_filter_data = swizzled_filter_data.get();
	int32_t swizzled_bias_data[32];
	int32_t swizzled_mult_data[32];
	int32_t swizzled_shift_data[32];

	int out_channel = 0;
	do {
	int out_channels_this_iter = std::min(8, output_depth - out_channel);
	Filter_N_H_W_M(filter_data + (out_channel * filter_height * filter_width *
	filter_input_depth),
	p_swizzled_filter_data, out_channels_this_iter, filter_height, filter_width,
	filter_input_depth);
	Swizzle(bias_data + out_channel, swizzled_bias_data, out_channels_this_iter);
	Swizzle(output_multiplier + out_channel, swizzled_mult_data, out_channels_this_iter);
	Swizzle(output_shift + out_channel, swizzled_shift_data, out_channels_this_iter);
	vld_w_x_m(v16, swizzled_bias_data);
	vld_w_x_m(v20, swizzled_mult_data);
	vld_w_x_m(v24, swizzled_shift_data);
	vrsub_w_vx_m(v24, v24, 0);

	for (int batch = 0; batch < batches; ++batch) {
	const int8_t* p_output =
	output_data + (batch * output_height * output_width * output_depth) +
	out_channel;
	for (int out_y = 0; out_y < output_height; ++out_y) {
	const int in_y_origin = (out_y * stride_height) - pad_height;
	const int out_y_offset = (out_y * output_width * output_depth);
	int out_x = 0;
	do {
	int out_xs_this_iter = std::min(8, output_width - out_x);
	// 8x accumulators
	vdup_w_x_m(v48, 0);
	vdup_w_x_m(v52, 0);
	acset_v(v48, v48);
	int in_channel = 0;
	do {
	int in_channels_this_iter = std::min(filter_input_depth, 32);
	// Calculate first valid filter_y
	int filter_y = 0;
	{
	int in_y = in_y_origin;
	while (in_y < 0) {
	++filter_y;
	in_y += (dilation_height_factor);
	}
	}
	for (; filter_y < filter_height; ++filter_y) {
	const int y_filter_offset =
	(filter_y * filter_width * 8 * input_depth);
	const int in_y = in_y_origin + dilation_height_factor * filter_y;
	if (in_y >= input_height) {
	break;
	}
	const int8_t* p_in =
	input_data + in_channel + (in_y * input_width * input_depth) +
	(batch * input_height * input_width * input_depth);

	int in_x[8];
	#pragma GCC unroll 8
	for (int i = 0; i < 8; ++i) {
	in_x[i] = ((out_x + i) * stride_width) - pad_width;
	}
	for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
	const int8_t* p_in_x[8];
	int first_right_pad = -1;

	#pragma GCC unroll 8
	for (int i = 0; i < 8; ++i) {
	p_in_x[i] = p_in + (in_x[i] * input_depth);
	}

	#pragma GCC unroll 8
	for (int i = 7; i >= 0; --i) {
	if (in_x[i] < input_width) {
	break;
	}
	first_right_pad = i;
	}
	bool left_pad = (in_x[0] < 0);
	bool right_pad = (first_right_pad != -1);

	int stride = input_depth * stride_width;

	if (unlikely(left_pad)) {
	vdup_b_x(v0, -input_offset);
	vld_b_s_xx(v1, p_in_x[1], stride);
	vld_b_s_xx(v2, p_in_x[2], stride);
	vld_b_s_xx(v3, p_in_x[3], stride);
	vld_b_s_xx_m(v4, p_in_x[4], stride);
	} else if (unlikely(right_pad)) {
	int first_pad = std::min(first_right_pad, out_xs_this_iter);
	switch (first_pad) {
	case 0:
	vdup_b_x(v0, neg_input_offset);
	case 1:
	vdup_b_x(v1, neg_input_offset);
	case 2:
	vdup_b_x(v2, neg_input_offset);
	case 3:
	vdup_b_x(v3, neg_input_offset);
	case 4:
	vdup_b_x(v4, neg_input_offset);
	case 5:
	vdup_b_x(v5, neg_input_offset);
	case 6:
	vdup_b_x(v6, neg_input_offset);
	case 7:
	vdup_b_x(v7, neg_input_offset);
	}
	switch (8 - first_pad) { // rest (stripmines?)
	case 0:
	vld_b_s_xx(v7, p_in_x[7], stride);
	case 1:
	vld_b_s_xx(v6, p_in_x[6], stride);
	case 2:
	vld_b_s_xx(v5, p_in_x[5], stride);
	case 3:
	vld_b_s_xx(v4, p_in_x[4], stride);
	case 4:
	vld_b_s_xx(v3, p_in_x[3], stride);
	case 5:
	vld_b_s_xx(v2, p_in_x[2], stride);
	case 6:
	vld_b_s_xx(v1, p_in_x[1], stride);
	case 7:
	vld_b_s_xx(v0, p_in_x[0], stride);
	}
	} else if (likely(!left_pad && !right_pad)) {
	// Inputs
	vld_b_s_xx_m(v0, p_in_x[0], stride);
	vld_b_s_xx_m(v4, p_in_x[4], stride);
	} else {
	vdup_b_x(v0, neg_input_offset);
	vdup_b_x(v7, neg_input_offset);
	vld_b_s_xx_m(v1, p_in_x[1], stride);
	vld_b_s_xx(v5, p_in_x[5], stride);
	vld_b_s_xx(v6, p_in_x[6], stride);
	}
	size_t local_filter_offset = y_filter_offset +
	(filter_x * 8 * input_depth) +
	(in_channel * 8);
	int8_t* p_local_filter_start =
	p_swizzled_filter_data + local_filter_offset;
	vld_b_p_x_m(v8, p_local_filter_start);
	vld_b_x_m(v12, p_local_filter_start);

	cmds.conv.stop = (in_channels_this_iter / 4) - 1;
	aconv_vxv(v48, v0, cmds, v8);

	#pragma GCC unroll 8
	for (int i = 0; i < 8; ++i) {
	in_x[i] += dilation_width_factor;
	}
	}
	}
	in_channel += in_channels_this_iter;
	} while (in_channel < filter_input_depth);
	vcget(v48);
	vadd_w_vv_m(v48, v48, v16);
	vadd_w_vv_m(v52, v52, v16);
	vdmulh_w_rn_vv_m(v48, v48, v20);
	vdmulh_w_rn_vv_m(v52, v52, v20);
	vsha_w_r_vv_m(v48, v48, v24);
	vsha_w_r_vv_m(v52, v52, v24);
	vadd_w_vx_m(v48, v48, output_offset);
	vadd_w_vx_m(v52, v52, output_offset);
	vmin_w_vx_m(v48, v48, output_activation_max);
	vmin_w_vx_m(v52, v52, output_activation_max);
	vmax_w_vx_m(v48, v48, output_activation_min);
	vmax_w_vx_m(v52, v52, output_activation_min);
	vsraqs_b_vx(v56, v48, 0);
	vsraqs_b_vx(v57, v52, 0);

	const int8_t* p_out_x[8];
	#pragma GCC unroll 8
	for (int i = 0; i < 8; ++i) {
	p_out_x[i] = p_output + out_y_offset + ((out_x + i) * output_depth);
	}

	vslidep_h_4_vv(v58, v57, v57); // x7
	vslidep_h_4_vv(v59, v58, v58); // x6
	vslidep_h_4_vv(v60, v59, v59); // x5
	vslidep_h_4_vv(v61, v60, v60); // x4
	vslidep_h_4_vv(v62, v56, v56); // x3
	vslidep_h_4_vv(v63, v62, v62); // x2
	vslidep_h_4_vv(v57, v63, v63); // x1
	vslidep_h_4_vv(v56, v57, v57); // x0
	switch (out_xs_this_iter) {
	case 8:
	vst_b_l_xx(v58, p_out_x[7], out_channels_this_iter);
	case 7:
	vst_b_l_xx(v59, p_out_x[6], out_channels_this_iter);
	case 6:
	vst_b_l_xx(v60, p_out_x[5], out_channels_this_iter);
	case 5:
	vst_b_l_xx(v61, p_out_x[4], out_channels_this_iter);
	case 4:
	vst_b_l_xx(v62, p_out_x[3], out_channels_this_iter);
	case 3:
	vst_b_l_xx(v63, p_out_x[2], out_channels_this_iter);
	case 2:
	vst_b_l_xx(v57, p_out_x[1], out_channels_this_iter);
	case 1:
	vst_b_l_xx(v56, p_out_x[0], out_channels_this_iter);
	}
	out_x += out_xs_this_iter;
	} while (out_x < output_width);
	}
	}
	out_channel += out_channels_this_iter;
	} while (out_channel < output_depth);
	}
	} // namespace kelvin::opt