Optimize 1x1 group conv.
Change-Id: I92f60e1c1b0dfa255063692265176ec2fce65469
diff --git a/tflm/opt/conv.cc b/tflm/opt/conv.cc
index 9ea92bc..df4dc89 100644
--- a/tflm/opt/conv.cc
+++ b/tflm/opt/conv.cc
@@ -138,6 +138,298 @@
}
}
+// Accumulates in v0-v7. [v0-v3], [v4-v7] are sub accumulators for two outputs.
+// Load/swizzle filters use [v52-v63].
+// Input activations use [v32-v33].
+// No clobbers.
+void ukernel_s8_s16(const int16_t* input_data0,
+ const int8_t* filter_data0,
+ const int8_t* filter_data1,
+ size_t n) {
+ n = n >> 5;
+ while (n > 0) {
+ // Load filters 0 to v58, v59
+ vld_b_p_x(v52, filter_data0);
+ vaddw_h_vx(v56, v52, 0);
+ vzip_h_vv(v58, v56, v57);
+
+ // Load activations
+ vld_h_p_x(v32, input_data0);
+ vld_h_p_x(v33, input_data0);
+
+ // Multiply filters0 * activations
+ vmulw_w_vv(v16, v58, v32);
+ vmulw_w_vv(v18, v59, v33);
+
+ // Accumulate v0
+ vadd_w_vv_m(v0, v0, v16);
+
+ // Load filters 1 to v62, v63
+ vld_b_p_x(v53, filter_data1);
+ vaddw_h_vx(v60, v53, 0);
+ vzip_h_vv(v62, v60, v61);
+
+ // Multiply filters1 * activations
+ vmulw_w_vv(v20, v62, v32);
+ vmulw_w_vv(v22, v63, v33);
+
+ // Accumulate v4
+ vadd_w_vv_m(v4, v4, v20);
+ n--;
+ }
+}
+
+void conv_per_channel_b64_1x1(
+ const tflite::ConvParams& params, const int32_t* output_multiplier,
+ const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
+ const int16_t* input_data, const tflite::RuntimeShape& filter_shape,
+ const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
+ const int64_t* bias_data, const tflite::RuntimeShape& output_shape,
+ int16_t* output_data) {
+ const auto batches = MatchingDim(input_shape, 0, output_shape, 0);
+ 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;
+ const auto output_activation_min = params.quantized_activation_min;
+ const auto output_activation_max = params.quantized_activation_max;
+ const auto groups = input_depth / filter_input_depth;
+ const auto output_filters_per_group = output_depth / groups;
+
+ int32_t accumulators[8];
+ for (int bhw = 0; bhw < batches * input_height * input_width; bhw++) {
+ const int16_t* local_input = input_data + (bhw * input_depth);
+ int16_t* local_output = output_data + (bhw * output_depth);
+ for (int g = 0; g < groups; g++) {
+ const int16_t* group_input = local_input + (g * filter_input_depth);
+ for (int gc = 0; gc + 2 <= output_filters_per_group; gc += 2) {
+ int oc = (g * output_filters_per_group) + gc;
+ const int8_t* local_filters0 = filter_data + (oc * filter_input_depth);
+ const int8_t* local_filters1 = local_filters0 + filter_input_depth;
+
+ vdup_w_x_m(v0, 0);
+ vdup_w_x_m(v4, 0);
+ ukernel_s8_s16(group_input, local_filters0, local_filters1,
+ filter_input_depth);
+ // sum accumulators
+ vadd_w_vv(v0, v0, v1);
+ vadd_w_vv(v2, v2, v3);
+ vadd_w_vv(v0, v0, v2);
+ vadd_w_vv(v4, v4, v5);
+ vadd_w_vv(v6, v6, v7);
+ vadd_w_vv(v4, v4, v6);
+
+ {
+ vst_w_x(v0, accumulators);
+ int64_t acc64 = bias_data[oc];
+ for (int i = 0; i < 8; i++) {
+ acc64 += accumulators[i];
+ }
+ int32_t acc = tflite::MultiplyByQuantizedMultiplier(
+ acc64, output_multiplier[oc], output_shift[oc]);
+ acc += output_offset;
+ acc = std::clamp(acc, output_activation_min, output_activation_max);
+ local_output[oc] = static_cast<int16_t>(acc);
+ }
+
+ {
+ vst_w_x(v4, accumulators);
+ int64_t acc64 = bias_data[oc + 1];
+ for (int i = 0; i < 8; i++) {
+ acc64 += accumulators[i];
+ }
+ int32_t acc = tflite::MultiplyByQuantizedMultiplier(
+ acc64, output_multiplier[oc + 1], output_shift[oc + 1]);
+ acc += output_offset;
+ acc = std::clamp(acc, output_activation_min, output_activation_max);
+ local_output[oc + 1] = static_cast<int16_t>(acc);
+ }
+ }
+ }
+ }
+}
+
+// Optimized for grouped convolutions, no dilation, 1xn filter
+void conv_per_channel_b64_filter1xn_group(
+ const tflite::ConvParams& params, const int32_t* output_multiplier,
+ const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
+ const int16_t* input_data, const tflite::RuntimeShape& filter_shape,
+ const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
+ const int64_t* bias_data, const tflite::RuntimeShape& output_shape,
+ int16_t* output_data) {
+ const auto batches = MatchingDim(input_shape, 0, output_shape, 0);
+ const auto stride_width = params.stride_width;
+ 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);
+ 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 output_filters_per_group = output_depth / groups;
+
+ int32_t accumulators[8];
+ for (int g = 0; g < groups; g++) {
+ for (int gc = 0; gc + 2 <= output_filters_per_group; gc += 2) {
+ int oc = (g * output_filters_per_group) + gc;
+ for (int b = 0; b < batches; ++b) {
+ for (int out_x = 0; out_x < output_width; ++out_x) {
+ const int in_x_origin = out_x * stride_width - pad_width;
+ const int8_t* local_filters0 =
+ filter_data + (oc * filter_width * filter_depth);
+ const int8_t* local_filters1 =
+ local_filters0 + (filter_width * filter_depth);
+ const int16_t* local_input = input_data +
+ (b * input_width * input_depth) +
+ (in_x_origin * input_depth) +
+ (g * filter_depth);
+ int16_t* local_output = output_data +
+ (b * output_width * output_depth) +
+ (out_x * output_depth);
+
+ int64_t acc64_0 = 0;
+ int64_t acc64_1 = 0;
+ vdup_w_x_m(v0, 0);
+ vdup_w_x_m(v4, 0);
+ for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+ const int8_t* local_filters0x =
+ local_filters0 + (filter_x * filter_depth);
+ const int8_t* local_filters1x =
+ local_filters1 + (filter_x * filter_depth);
+ const int16_t* local_inputx =
+ local_input + (filter_x * input_depth);
+
+ ukernel_s8_s16(local_inputx, local_filters0x, local_filters1x,
+ filter_depth);
+ }
+
+ // sum accumulators
+ vadd_w_vv(v0, v0, v1);
+ vadd_w_vv(v2, v2, v3);
+ vadd_w_vv(v0, v0, v2);
+ vadd_w_vv(v4, v4, v5);
+ vadd_w_vv(v6, v6, v7);
+ vadd_w_vv(v4, v4, v6);
+
+ {
+ vst_w_x(v0, accumulators);
+ for (int i = 0; i < 8; i++) {
+ acc64_0 += accumulators[i];
+ }
+ acc64_0 += bias_data[oc];
+ int32_t acc = tflite::MultiplyByQuantizedMultiplier(
+ acc64_0, output_multiplier[oc], output_shift[oc]);
+ acc += output_offset;
+ acc = std::clamp(acc, output_activation_min, output_activation_max);
+ local_output[oc] = static_cast<int16_t>(acc);
+ }
+
+ {
+ vst_w_x(v4, accumulators);
+ for (int i = 0; i < 8; i++) {
+ acc64_1 += accumulators[i];
+ }
+ acc64_1 += bias_data[oc + 1];
+ int32_t acc = tflite::MultiplyByQuantizedMultiplier(
+ acc64_1, output_multiplier[oc + 1], output_shift[oc + 1]);
+ acc += output_offset;
+ acc = std::clamp(acc, output_activation_min, output_activation_max);
+ local_output[oc + 1] = static_cast<int16_t>(acc);
+ }
+ }
+ }
+ }
+ }
+}
+
+// Optimized for no group, no dilation, 1xn filter.
+void conv_per_channel_b64_filter1xn_non_group(
+ const tflite::ConvParams& params, const int32_t* output_multiplier,
+ const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
+ const int16_t* input_data, const tflite::RuntimeShape& filter_shape,
+ const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
+ const int64_t* bias_data, const tflite::RuntimeShape& output_shape,
+ int16_t* output_data) {
+ const auto batches = MatchingDim(input_shape, 0, output_shape, 0);
+ const auto stride_width = params.stride_width;
+ 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);
+ 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;
+ int32_t accumulators[8];
+ for (int oc = 0; oc + 2 <= output_depth; oc += 2) {
+ for (int batch = 0; batch < batches; ++batch) {
+ for (int out_x = 0; out_x < output_width; ++out_x) {
+ const int in_x_origin = out_x * stride_width - pad_width;
+
+ const int8_t* local_filters0 =
+ filter_data + (oc * filter_width * filter_depth);
+ const int8_t* local_filters1 =
+ local_filters0 + (filter_width * filter_depth);
+ const int16_t* local_input = input_data +
+ (batch * input_width * input_depth) +
+ (in_x_origin * input_depth);
+ int16_t* local_output = output_data +
+ (batch * output_width * output_depth) +
+ (out_x * output_depth);
+
+ vdup_w_x_m(v0, 0);
+ vdup_w_x_m(v4, 0);
+ ukernel_s8_s16(local_input, local_filters0, local_filters1,
+ filter_width * filter_depth);
+ // sum accumulators
+ vadd_w_vv(v0, v0, v1);
+ vadd_w_vv(v2, v2, v3);
+ vadd_w_vv(v0, v0, v2);
+ vadd_w_vv(v4, v4, v5);
+ vadd_w_vv(v6, v6, v7);
+ vadd_w_vv(v4, v4, v6);
+ {
+ vst_w_x(v0, accumulators);
+ int64_t acc64 = bias_data[oc];
+ for (int i = 0; i < 8; i++) {
+ acc64 += accumulators[i];
+ }
+ int32_t acc = tflite::MultiplyByQuantizedMultiplier(
+ acc64, output_multiplier[oc], output_shift[oc]);
+ acc += output_offset;
+ acc = std::clamp(acc, output_activation_min, output_activation_max);
+ local_output[oc] = static_cast<int16_t>(acc);
+ }
+
+ {
+ vst_w_x(v4, accumulators);
+ int64_t acc64 = bias_data[oc + 1];
+ for (int i = 0; i < 8; i++) {
+ acc64 += accumulators[i];
+ }
+ int32_t acc = tflite::MultiplyByQuantizedMultiplier(
+ acc64, output_multiplier[oc + 1], output_shift[oc + 1]);
+ acc += output_offset;
+ acc = std::clamp(acc, output_activation_min, output_activation_max);
+ local_output[oc + 1] = static_cast<int16_t>(acc);
+ }
+ }
+ }
+ }
+}
+
void conv_per_channel_b64(
const tflite::ConvParams& params, const int32_t* output_multiplier,
const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
diff --git a/tflm/opt/opt.h b/tflm/opt/opt.h
index 17e004a..6797f36 100644
--- a/tflm/opt/opt.h
+++ b/tflm/opt/opt.h
@@ -70,6 +70,31 @@
const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
const int32_t* bias_data, const tflite::RuntimeShape& output_shape,
int16_t* output_data);
+
+void conv_per_channel_b64_1x1(
+ const tflite::ConvParams& params, const int32_t* output_multiplier,
+ const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
+ const int16_t* input_data, const tflite::RuntimeShape& filter_shape,
+ const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
+ const int64_t* bias_data, const tflite::RuntimeShape& output_shape,
+ int16_t* output_data);
+
+void conv_per_channel_b64_filter1xn_non_group(
+ const tflite::ConvParams& params, const int32_t* output_multiplier,
+ const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
+ const int16_t* input_data, const tflite::RuntimeShape& filter_shape,
+ const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
+ const int64_t* bias_data, const tflite::RuntimeShape& output_shape,
+ int16_t* output_data);
+
+void conv_per_channel_b64_filter1xn_group(
+ const tflite::ConvParams& params, const int32_t* output_multiplier,
+ const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
+ const int16_t* input_data, const tflite::RuntimeShape& filter_shape,
+ const int8_t* filter_data, const tflite::RuntimeShape& bias_shape,
+ const int64_t* bias_data, const tflite::RuntimeShape& output_shape,
+ int16_t* output_data);
+
void conv_per_channel_b64(
const tflite::ConvParams& params, const int32_t* output_multiplier,
const int32_t* output_shift, const tflite::RuntimeShape& input_shape,
