Specialized int8 conv kernels
- Add int8 conv kernels specialized for the following shapes:
- filter_input_depth % 32
- filter_input_depth % 32 && input_channels % 8 && output_width % 8
&& pad_width == 1
- Move the selection of kernels into the toplevel conv_per_channel_b8,
instead of inside the TF kernel.
Change-Id: Ie10bcff04cc3394aea630f7e073e63cbf239eb68
diff --git a/tflm/opt/BUILD b/tflm/opt/BUILD
index e4d533b..19eb017 100644
--- a/tflm/opt/BUILD
+++ b/tflm/opt/BUILD
@@ -18,6 +18,7 @@
name = "opt",
srcs = [
"conv.cc",
+ "conv_s8.cc",
"depthwise_conv_s16.cc",
"depthwise_conv_s8.cc",
"elementwise_add_s16.cc",
diff --git a/tflm/opt/conv.cc b/tflm/opt/conv.cc
index 8d33848..49d32d5 100644
--- a/tflm/opt/conv.cc
+++ b/tflm/opt/conv.cc
@@ -196,7 +196,6 @@
const auto input_height = input_shape.Dims(1);
const auto input_width = input_shape.Dims(2);
const auto input_depth = input_shape.Dims(3);
- const auto input_offset = params.input_offset;
const auto filter_input_depth = filter_shape.Dims(3);
const auto output_depth = output_shape.Dims(3);
const auto output_offset = params.output_offset;
@@ -271,7 +270,6 @@
const auto pad_width = params.padding_values.width;
const auto input_width = input_shape.Dims(2);
const auto input_depth = input_shape.Dims(3);
- const auto input_offset = params.input_offset;
const auto filter_width = filter_shape.Dims(2);
const auto filter_depth = filter_shape.Dims(3);
const auto output_width = output_shape.Dims(2);
@@ -370,7 +368,6 @@
const auto pad_width = params.padding_values.width;
const auto input_width = input_shape.Dims(2);
const auto input_depth = input_shape.Dims(3);
- const auto input_offset = params.input_offset;
const auto filter_width = filter_shape.Dims(2);
const auto filter_depth = filter_shape.Dims(3);
const auto output_width = output_shape.Dims(2);
@@ -584,148 +581,4 @@
output_data);
}
-#define INA0 v0
-#define FLTA0 v8
-#define FLTA1 v9
-#define FLTA2 v10
-#define FLTA3 v11
-#define FLTA4 v12
-#define FLTA5 v13
-#define FLTA6 v14
-#define FLTA7 v15
-#define ACC v48
-#define ACC0 v48
-#define OUT0 v56
-void conv_per_channel_b8(
- 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) {
- const auto batches = MatchingDim(input_shape, 0, output_shape, 0);
- const auto stride_width = params.stride_width;
- const auto stride_height = params.stride_height;
- const auto dilation_width_factor = params.dilation_width_factor;
- const auto dilation_height_factor = params.dilation_height_factor;
- const auto pad_width = params.padding_values.width;
- const auto pad_height = params.padding_values.height;
- const auto input_height = input_shape.Dims(1);
- const auto input_width = input_shape.Dims(2);
- const auto input_depth = input_shape.Dims(3);
- const auto input_offset = params.input_offset;
- const auto filter_height = filter_shape.Dims(1);
- const auto filter_width = filter_shape.Dims(2);
- const auto filter_depth = filter_shape.Dims(3);
- const auto output_height = output_shape.Dims(1);
- const auto output_width = output_shape.Dims(2);
- const auto output_depth = output_shape.Dims(3);
- const auto output_offset = params.output_offset;
- const auto output_activation_min = params.quantized_activation_min;
- const auto output_activation_max = params.quantized_activation_max;
- const auto groups = input_depth / filter_depth;
- const auto filters_per_group = output_depth / groups;
- 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;
-
- // Zero out accumulators.
- vdup_b_x(v0, 0);
- acset_v(ACC, v0);
- vdup_b_x_m(ACC0, 0);
- for (int batch = 0; batch < batches; ++batch) {
- for (int out_y = 0; out_y < output_height; ++out_y) {
- const int in_y_origin = (out_y * stride_height) - pad_height;
- for (int out_x = 0; out_x < output_width; /*out_x += 32*/ ++out_x) {
- const int in_x_origin = (out_x * stride_width) - pad_width;
- for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
- auto group = out_channel / filters_per_group;
-
- for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
- const int in_y = in_y_origin + dilation_height_factor * filter_y;
- const int in_x = in_x_origin + dilation_width_factor * 0;
-
- // Zero padding by omitting the areas outside the image.
- const bool is_point_inside_image =
- (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
- (in_y < input_height);
- if (!is_point_inside_image) {
- continue;
- }
-
- int q = filter_width * filter_depth;
- for (int i = 0; i < q; i += 32) {
- int count = std::min(q - i, 32);
- count = std::min(
- count, static_cast<int>((input_width - in_x) * filter_depth));
- int input_offset = tflite::Offset(input_shape, batch, in_y, in_x,
- group * filter_depth) +
- i;
- vdup_w_x_m(vm0, 0);
- vdup_w_x_m(vm1, 0);
- vld_b_l_xx(INA0, &input_data[input_offset], count);
- int filter_offset =
- tflite::Offset(filter_shape, out_channel, filter_y, 0, 0) + i;
- vdup_w_x_m(FLTA0, 0);
- vdup_w_x_m(FLTA4, 0);
- if (count > 0) {
- vld_b_l_xx(FLTA0, &filter_data[filter_offset],
- std::min(count, 4));
- }
- if (count > 4) {
- vld_b_l_xx(FLTA1, &filter_data[filter_offset + 4],
- std::min(count - 4, 4));
- }
- if (count > 8) {
- vld_b_l_xx(FLTA2, &filter_data[filter_offset + 8],
- std::min(count - 8, 4));
- }
- if (count > 12) {
- vld_b_l_xx(FLTA3, &filter_data[filter_offset + 12],
- std::min(count - 12, 4));
- }
- if (count > 16) {
- vld_b_l_xx(FLTA4, &filter_data[filter_offset + 16],
- std::min(count - 16, 4));
- }
- if (count > 20) {
- vld_b_l_xx(FLTA5, &filter_data[filter_offset + 20],
- std::min(count - 20, 4));
- }
- if (count > 24) {
- vld_b_l_xx(FLTA6, &filter_data[filter_offset + 24],
- std::min(count - 24, 4));
- }
- if (count > 28) {
- vld_b_l_xx(FLTA7, &filter_data[filter_offset + 28],
- std::min(count - 28, 4));
- }
- aconv_vxv(ACC, INA0, cmds, FLTA0);
- }
- }
- vcget(ACC);
- vadd_w_vx_m(ACC0, ACC0, bias_data[out_channel]);
- vsll_w_vx_m(ACC0, ACC0, LEFT_SHIFT(output_shift[out_channel]));
- vdmulh_w_r_vx_m(ACC0, ACC0, output_multiplier[out_channel]);
- vsha_w_r_vx_m(ACC0, ACC0, RIGHT_SHIFT(output_shift[out_channel]));
- vadd_w_vx_m(ACC0, ACC0, output_offset);
- vmin_w_vx_m(ACC0, ACC0, output_activation_max);
- vmax_w_vx_m(ACC0, ACC0, output_activation_min);
- vsraqs_b_vx(OUT0, ACC0, 0);
- size_t output_offset =
- tflite::Offset(output_shape, batch, out_y, out_x, out_channel);
- vst_b_l_xx(OUT0, &output_data[output_offset], 1);
- }
- }
- }
- }
-}
} // namespace kelvin::opt
diff --git a/tflm/opt/conv_s8.cc b/tflm/opt/conv_s8.cc
new file mode 100644
index 0000000..2da0028
--- /dev/null
+++ b/tflm/opt/conv_s8.cc
@@ -0,0 +1,599 @@
+/*
+ * 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.
+ */
+
+#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/opt.h"
+#include "tflm/opt/util.h"
+
+namespace kelvin::opt {
+
+namespace {
+constexpr int kFilterInputChannelIndex = 3;
+constexpr int kOutputWidthIndex = 2;
+constexpr int kOutputChannelIndex = 3;
+
+// Convert: input [zo][ky][kx][zi] (N,4,4,M)
+// output [ky][kx][zi_hi=M/4][zo=8][zi_lo=4]
+// output [3][3][16][8][4]
+void Filter_8_H_W_M(const int8_t* input, int8_t* output, 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 < 8; ++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];
+ }
+ }
+ }
+ }
+}
+
+void Swizzle(const int32_t* input, int32_t* output, int N) {
+ const int32_t(&in)[N] = *(int32_t(*)[N])input;
+ int32_t(&out)[N * 4] = *(int32_t(*)[N * 4]) output;
+ // Convert to accumulator swizzle pattern.
+ for (int i = 0; i < N / 8; ++i) {
+ int32_t* out0 = out + i * 32 + 0;
+ int32_t* out1 = out + i * 32 + 16;
+ int32_t* out2 = out + i * 32 + 8;
+ int32_t* out3 = out + i * 32 + 24;
+ for (int j = 0; j < 4; ++j) {
+ const int32_t* p_in = in + i * 8;
+ for (int k = 0; k < 2; ++k) {
+ *out0++ = *p_in++;
+ *out1++ = *p_in++;
+ *out2++ = *p_in++;
+ *out3++ = *p_in++;
+ }
+ }
+ }
+}
+
+} // namespace
+
+void conv_per_channel_pw1_ow8_id8_filterd32(
+ 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 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];
+
+ for (int out_channel = 0; out_channel + 8 <= output_depth; out_channel += 8) {
+ Filter_8_H_W_M(filter_data + (out_channel * filter_height * filter_width *
+ filter_input_depth),
+ p_swizzled_filter_data, filter_height, filter_width,
+ filter_input_depth);
+ Swizzle(bias_data + out_channel, swizzled_bias_data, 8);
+ Swizzle(output_multiplier + out_channel, swizzled_mult_data, 8);
+ Swizzle(output_shift + out_channel, swizzled_shift_data, 8);
+ 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) {
+ for (int out_y = 0; out_y < output_height; ++out_y) {
+ const int in_y_origin = (out_y * stride_height) - pad_height;
+ for (int out_x = 0; out_x + 8 <= output_width; out_x += 8) {
+ // 8x accumulators
+ vdup_w_x_m(v48, 0);
+ vdup_w_x_m(v52, 0);
+ acset_v(v48, v48);
+ for (int in_channel = 0; in_channel + 32 <= filter_input_depth;
+ in_channel += 32) {
+ for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+ const int in_y = in_y_origin + dilation_height_factor * filter_y;
+ const bool is_row_inside_input =
+ (in_y >= 0) && (in_y < input_height);
+ if (!is_row_inside_input) {
+ continue;
+ }
+
+ for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+ int in_x[8];
+ bool left_pad = false;
+ bool right_pad = false;
+ for (int i = 0; i < 8; ++i) {
+ const int in_x_origin =
+ ((out_x + i) * stride_width) - pad_width;
+ in_x[i] = in_x_origin + dilation_width_factor * filter_x;
+ if (in_x[i] < 0) {
+ left_pad = true;
+ }
+ if (in_x[i] >= input_width) {
+ right_pad = true;
+ }
+ }
+
+ if (left_pad) {
+ vdup_b_x(v0, -input_offset);
+ vld_b_s_xx(
+ v1,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[1], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx(
+ v2,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[2], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx(
+ v3,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[3], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx_m(
+ v4,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[4], in_channel)],
+ input_depth * stride_width);
+ } else if (right_pad) {
+ vld_b_s_xx_m(
+ v0,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[0], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx(
+ v4,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[4], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx(
+ v5,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[5], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx(
+ v6,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[6], in_channel)],
+ input_depth * stride_width);
+ vdup_b_x(v7, -input_offset);
+ } else if (!left_pad && !right_pad) {
+ // Inputs
+ vld_b_s_xx_m(
+ v0,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[0], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx_m(
+ v4,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[4], in_channel)],
+ input_depth * stride_width);
+ } else {
+ vdup_b_x(v0, -input_offset);
+ vdup_b_x(v7, -input_offset);
+ vld_b_s_xx_m(
+ v1,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[1], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx(
+ v5,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[5], in_channel)],
+ input_depth * stride_width);
+ vld_b_s_xx(
+ v6,
+ &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x[6], in_channel)],
+ input_depth * stride_width);
+ }
+ size_t local_filter_offset =
+ (filter_y * filter_width * 8 * input_depth) +
+ (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);
+
+ aconv_vxv(v48, v0, cmds, v8);
+ }
+ }
+ }
+ vcget(v48);
+ vadd_w_vv_m(v48, v48, v16);
+ vadd_w_vv_m(v52, v52, v16);
+ vdmulh_w_r_vv_m(v48, v48, v20);
+ vdmulh_w_r_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);
+ vstq_b_s_xx(v56,
+ &output_data[tflite::Offset(output_shape, batch, out_y,
+ out_x, out_channel)],
+ output_depth);
+ vstq_b_s_xx(v57,
+ &output_data[tflite::Offset(output_shape, batch, out_y,
+ out_x + 4, out_channel)],
+ output_depth);
+ }
+ }
+ }
+ }
+}
+
+// Fixed-point per-channel-quantization convolution reference kernel.
+void conv_per_channel_filterd32(
+ 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 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);
+ for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+ for (int batch = 0; batch < batches; ++batch) {
+ for (int out_y = 0; out_y < output_height; ++out_y) {
+ const int in_y_origin = (out_y * stride_height) - pad_height;
+ for (int out_x = 0; out_x < output_width; ++out_x) {
+ const int in_x_origin = (out_x * stride_width) - pad_width;
+ vdup_w_x_m(v60, 0);
+ int32_t acc = 0;
+ for (int in_channel = 0; in_channel + 32 <= filter_input_depth;
+ in_channel += 32) {
+ for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+ const int in_y = in_y_origin + dilation_height_factor * filter_y;
+ for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+ const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+ // Zero padding by omitting the areas outside the image.
+ const bool is_point_inside_image =
+ (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+ (in_y < input_height);
+
+ if (!is_point_inside_image) {
+ continue;
+ }
+
+ vld_b_x(v0, &input_data[tflite::Offset(input_shape, batch, in_y,
+ in_x, in_channel)]);
+ vaddw_h_vx(v0, v0, 0);
+ vadd_h_vx(v0, v0, static_cast<int16_t>(input_offset));
+ vadd_h_vx(v1, v1, static_cast<int16_t>(input_offset));
+ vld_b_x(v2, &filter_data[tflite::Offset(filter_shape,
+ out_channel, filter_y,
+ filter_x, in_channel)]);
+ vaddw_h_vx(v2, v2, 0);
+ vmulw_w_vv(v48, v0, v2);
+ vmulw_w_vv(v50, v1, v3);
+ vadd_w_vv_m(v60, v60, v48);
+ }
+ }
+ }
+ int32_t accumulators[32];
+ vst_w_x_m(v60, accumulators);
+ for (int i = 0; i < 32; ++i) {
+ acc += accumulators[i];
+ }
+
+ if (bias_data) {
+ acc += bias_data[out_channel];
+ }
+ acc = tflite::MultiplyByQuantizedMultiplier(
+ acc, output_multiplier[out_channel], output_shift[out_channel]);
+ acc += output_offset;
+ acc = std::max(acc, output_activation_min);
+ acc = std::min(acc, output_activation_max);
+ output_data[tflite::Offset(output_shape, batch, out_y, out_x,
+ out_channel)] = static_cast<int8_t>(acc);
+ }
+ }
+ }
+ }
+}
+
+void conv_per_channel_generic(
+ 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) {
+ const auto batches = MatchingDim(input_shape, 0, output_shape, 0);
+ const auto stride_width = params.stride_width;
+ const auto stride_height = params.stride_height;
+ const auto dilation_width_factor = params.dilation_width_factor;
+ const auto dilation_height_factor = params.dilation_height_factor;
+ const auto pad_width = params.padding_values.width;
+ const auto pad_height = params.padding_values.height;
+ const auto input_height = input_shape.Dims(1);
+ const auto input_width = input_shape.Dims(2);
+ const auto input_depth = input_shape.Dims(3);
+ const auto input_offset = params.input_offset;
+ const auto filter_height = filter_shape.Dims(1);
+ const auto filter_width = filter_shape.Dims(2);
+ const auto filter_depth = filter_shape.Dims(3);
+ const auto output_height = output_shape.Dims(1);
+ const auto output_width = output_shape.Dims(2);
+ const auto output_depth = output_shape.Dims(3);
+ const auto output_offset = params.output_offset;
+ const auto output_activation_min = params.quantized_activation_min;
+ const auto output_activation_max = params.quantized_activation_max;
+ const auto groups = input_depth / filter_depth;
+ const auto filters_per_group = output_depth / groups;
+ 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;
+
+ // Zero out accumulators.
+ vdup_b_x(v0, 0);
+ acset_v(v48, v0);
+ vdup_b_x_m(v48, 0);
+ for (int batch = 0; batch < batches; ++batch) {
+ for (int out_y = 0; out_y < output_height; ++out_y) {
+ const int in_y_origin = (out_y * stride_height) - pad_height;
+ for (int out_x = 0; out_x < output_width; /*out_x += 32*/ ++out_x) {
+ const int in_x_origin = (out_x * stride_width) - pad_width;
+ for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+ auto group = out_channel / filters_per_group;
+
+ for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+ const int in_y = in_y_origin + dilation_height_factor * filter_y;
+ const int in_x = in_x_origin + dilation_width_factor * 0;
+
+ // Zero padding by omitting the areas outside the image.
+ const bool is_point_inside_image =
+ (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+ (in_y < input_height);
+ if (!is_point_inside_image) {
+ continue;
+ }
+
+ int q = filter_width * filter_depth;
+ for (int i = 0; i < q; i += 32) {
+ int count = std::min(q - i, 32);
+ count = std::min(
+ count, static_cast<int>((input_width - in_x) * filter_depth));
+ int input_offset = tflite::Offset(input_shape, batch, in_y, in_x,
+ group * filter_depth) +
+ i;
+ vdup_w_x_m(vm0, 0);
+ vdup_w_x_m(vm1, 0);
+ vld_b_l_xx(v0, &input_data[input_offset], count);
+ int filter_offset =
+ tflite::Offset(filter_shape, out_channel, filter_y, 0, 0) + i;
+ vdup_w_x_m(v8, 0);
+ vdup_w_x_m(v12, 0);
+ if (count > 0) {
+ vld_b_l_xx(v8, &filter_data[filter_offset], std::min(count, 4));
+ }
+ if (count > 4) {
+ vld_b_l_xx(v9, &filter_data[filter_offset + 4],
+ std::min(count - 4, 4));
+ }
+ if (count > 8) {
+ vld_b_l_xx(v10, &filter_data[filter_offset + 8],
+ std::min(count - 8, 4));
+ }
+ if (count > 12) {
+ vld_b_l_xx(v11, &filter_data[filter_offset + 12],
+ std::min(count - 12, 4));
+ }
+ if (count > 16) {
+ vld_b_l_xx(v12, &filter_data[filter_offset + 16],
+ std::min(count - 16, 4));
+ }
+ if (count > 20) {
+ vld_b_l_xx(v13, &filter_data[filter_offset + 20],
+ std::min(count - 20, 4));
+ }
+ if (count > 24) {
+ vld_b_l_xx(v14, &filter_data[filter_offset + 24],
+ std::min(count - 24, 4));
+ }
+ if (count > 28) {
+ vld_b_l_xx(v15, &filter_data[filter_offset + 28],
+ std::min(count - 28, 4));
+ }
+ aconv_vxv(v48, v0, cmds, v8);
+ }
+ }
+ vcget(v48);
+ vadd_w_vx_m(v48, v48, bias_data[out_channel]);
+ vsll_w_vx_m(v48, v48, LEFT_SHIFT(output_shift[out_channel]));
+ vdmulh_w_r_vx_m(v48, v48, output_multiplier[out_channel]);
+ vsha_w_r_vx_m(v48, v48, RIGHT_SHIFT(output_shift[out_channel]));
+ vadd_w_vx_m(v48, v48, output_offset);
+ vmin_w_vx_m(v48, v48, output_activation_max);
+ vmax_w_vx_m(v48, v48, output_activation_min);
+ vsraqs_b_vx(v56, v48, 0);
+ size_t output_offset =
+ tflite::Offset(output_shape, batch, out_y, out_x, out_channel);
+ vst_b_l_xx(v56, &output_data[output_offset], 1);
+ }
+ }
+ }
+ }
+}
+
+void conv_per_channel_b8(
+ 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) {
+ const auto stride_width = params.stride_width;
+ const auto stride_height = params.stride_height;
+ const auto dilation_width_factor = params.dilation_width_factor;
+ const auto dilation_height_factor = params.dilation_height_factor;
+ const int pad_width = params.padding_values.width;
+ const int pad_height = params.padding_values.height;
+
+ if (dilation_width_factor == 1 && dilation_height_factor == 1 &&
+ stride_width <= 2 && stride_height <= 2) {
+ if (filter_shape.Dims(kFilterInputChannelIndex) % 32 == 0 &&
+ output_shape.Dims(kOutputChannelIndex) % 8 == 0 &&
+ output_shape.Dims(kOutputWidthIndex) % 8 == 0 && pad_width <= 1) {
+ conv_per_channel_pw1_ow8_id8_filterd32(
+ params, output_multiplier, output_shift, input_shape, input_data,
+ filter_shape, filter_data, bias_shape, bias_data, output_shape,
+ output_data);
+ return;
+ } else if (filter_shape.Dims(kFilterInputChannelIndex) % 32 == 0) {
+ conv_per_channel_filterd32(params, output_multiplier, output_shift,
+ input_shape, input_data, filter_shape,
+ filter_data, bias_shape, bias_data,
+ output_shape, output_data);
+ return;
+ }
+ }
+
+ if (stride_width == 1 && stride_height == 1 && dilation_width_factor == 1 &&
+ dilation_height_factor == 1) {
+ if (pad_width == 0 && pad_height == 0) {
+ conv_per_channel_generic(params, output_multiplier, output_shift,
+ input_shape, input_data, filter_shape,
+ filter_data, bias_shape, bias_data, output_shape,
+ output_data);
+ return;
+ }
+ }
+
+ tflite::reference_integer_ops::ConvPerChannel(
+ params, output_multiplier, output_shift, input_shape, input_data,
+ filter_shape, filter_data, bias_shape, bias_data, output_shape,
+ output_data);
+}
+
+} // namespace kelvin::opt
