Improve ElementwiseAddS8
- Run up to three groups of 32 per iteration
- Break up dependencies between consecutive instructions
- Keep load units more busy than before
Change-Id: I9c41d3f083f52929cd5178911f93f1c0232bcdb3
diff --git a/tflm/opt/elementwise_add_s8.cc b/tflm/opt/elementwise_add_s8.cc
index b751ceb..e4119d4 100644
--- a/tflm/opt/elementwise_add_s8.cc
+++ b/tflm/opt/elementwise_add_s8.cc
@@ -31,82 +31,145 @@
const int32_t output_activation_max,
const int32_t block_size) {
int blocks = block_size;
- int vl;
- getmaxvl_b(vl);
const int32_t input1_shift_mul = 1 << LEFT_SHIFT(input1_shift);
const int32_t input2_shift_mul = 1 << LEFT_SHIFT(input2_shift);
+ while (blocks >= 96) {
+ vld_b_lp_xx(v0, input1, 32);
+ vld_b_lp_xx(v8, input2, 32);
+
+ vaddw_h_vx(v2, v0, 0);
+ vaddw_h_vx(v10, v8, 0);
+ vaddw_w_vx(v4, v2, input1_offset);
+ vaddw_w_vx(v6, v3, input1_offset);
+ vaddw_w_vx(v12, v10, input2_offset);
+ vaddw_w_vx(v14, v11, input2_offset);
+
+ vld_b_lp_xx(v16, input1, 32);
+ vld_b_lp_xx(v24, input2, 32);
+
+ vaddw_h_vx(v18, v16, 0);
+ vaddw_h_vx(v26, v24, 0);
+ vaddw_w_vx(v20, v18, input1_offset);
+ vaddw_w_vx(v22, v19, input1_offset);
+ vaddw_w_vx(v28, v26, input2_offset);
+ vaddw_w_vx(v30, v27, input2_offset);
+
+ vld_b_lp_xx(v32, input1, 32);
+ vld_b_lp_xx(v40, input2, 32);
+
+ vaddw_h_vx(v34, v32, 0);
+ vaddw_h_vx(v42, v40, 0);
+ vaddw_w_vx(v36, v34, input1_offset);
+ vaddw_w_vx(v38, v35, input1_offset);
+ vaddw_w_vx(v44, v42, input2_offset);
+ vaddw_w_vx(v46, v43, input2_offset);
+
+ vsll_w_vx_m(v4, v4, left_shift);
+ vsll_w_vx_m(v12, v12, left_shift);
+
+ vsll_w_vx_m(v20, v20, left_shift);
+ vsll_w_vx_m(v28, v28, left_shift);
+
+ vsll_w_vx_m(v36, v36, left_shift);
+ vsll_w_vx_m(v44, v44, left_shift);
+
+ vmul_w_vx_m(v4, v4, input1_shift_mul);
+ vmul_w_vx_m(v12, v12, input2_shift_mul);
+
+ vmul_w_vx_m(v20, v20, input1_shift_mul);
+ vmul_w_vx_m(v28, v28, input2_shift_mul);
+
+ vmul_w_vx_m(v36, v36, input1_shift_mul);
+ vmul_w_vx_m(v44, v44, input2_shift_mul);
+
+ vdmulh_w_r_vx_m(v4, v4, input1_mult);
+ vdmulh_w_r_vx_m(v12, v12, input2_mult);
+ vsha_w_r_vx_m(v4, v4, -input1_shift);
+ vsha_w_r_vx_m(v12, v12, -input2_shift);
+ vadd_w_vx_m(v4, v4, input1_offset);
+ vadd_w_vx_m(v12, v12, input2_offset);
+
+ vdmulh_w_r_vx_m(v20, v20, input1_mult);
+ vsha_w_r_vx_m(v20, v20, -input1_shift);
+ vadd_w_vx_m(v20, v20, input1_offset);
+ vdmulh_w_r_vx_m(v28, v28, input2_mult);
+ vsha_w_r_vx_m(v28, v28, -input2_shift);
+ vadd_w_vx_m(v28, v28, input2_offset);
+
+ vdmulh_w_r_vx_m(v36, v36, input1_mult);
+ vsha_w_r_vx_m(v36, v36, -input1_shift);
+ vadd_w_vx_m(v36, v36, input1_offset);
+ vdmulh_w_r_vx_m(v44, v44, input2_mult);
+ vsha_w_r_vx_m(v44, v44, -input2_shift);
+ vadd_w_vx_m(v44, v44, input2_offset);
+
+ vadd_w_vv_m(v12, v4, v12);
+ vadd_w_vv_m(v28, v20, v28);
+ vadd_w_vv_m(v44, v36, v44);
+
+ vdmulh_w_r_vx_m(v12, v12, output_mult);
+ vdmulh_w_r_vx_m(v28, v28, output_mult);
+ vdmulh_w_r_vx_m(v44, v44, output_mult);
+ vsha_w_r_vx_m(v12, v12, -output_shift);
+ vsha_w_r_vx_m(v28, v28, -output_shift);
+ vsha_w_r_vx_m(v44, v44, -output_shift);
+ vadd_w_vx_m(v12, v12, output_offset);
+ vadd_w_vx_m(v28, v28, output_offset);
+ vadd_w_vx_m(v44, v44, output_offset);
+
+ vmin_w_vx_m(v12, v12, output_activation_max);
+ vmin_w_vx_m(v28, v28, output_activation_max);
+ vmin_w_vx_m(v44, v44, output_activation_max);
+ vmax_w_vx_m(v12, v12, output_activation_min);
+ vmax_w_vx_m(v28, v28, output_activation_min);
+ vmax_w_vx_m(v44, v44, output_activation_min);
+
+ vsraqs_b_vx(v12, v12, 0);
+ vst_b_lp_xx(v12, output, 32);
+ vsraqs_b_vx(v28, v28, 0);
+ vst_b_lp_xx(v28, output, 32);
+ vsraqs_b_vx(v44, v44, 0);
+ vst_b_lp_xx(v44, output, 32);
+
+ blocks -= 96;
+ }
+
while (blocks) {
- int count = std::min(blocks, vl);
+ int count = std::min(blocks, 32);
+ vld_b_lp_xx(v0, input1, count);
+ vld_b_lp_xx(v8, input2, count);
- // Widen input1 to 32-bit wide values (in vm0, vm1, vm2, vm3).
- vld_b_lp_xx_m(vm0, input1, count);
- vaddw_h_vx_m(vm0, vm0, 0);
- vaddw_w_vx_m(vm2, vm1, input1_offset);
- vaddw_w_vx_m(vm0, vm0, input1_offset);
+ vaddw_h_vx(v2, v0, 0);
+ vaddw_h_vx(v10, v8, 0);
+ vaddw_w_vx(v4, v2, input1_offset);
+ vaddw_w_vx(v6, v3, input1_offset);
+ vaddw_w_vx(v12, v10, input2_offset);
+ vaddw_w_vx(v14, v11, input2_offset);
- // Widen input2 to 32-bit wide values (in vm4, vm5, vm6, vm7).
- vld_b_lp_xx_m(vm4, input2, count);
- vaddw_h_vx_m(vm4, vm4, 0);
- vaddw_w_vx_m(vm6, vm5, input2_offset);
- vaddw_w_vx_m(vm4, vm4, input2_offset);
+ vsll_w_vx_m(v4, v4, left_shift);
+ vsll_w_vx_m(v12, v12, left_shift);
- // Apply left_shift to all inputs.
- vsll_w_vx_m(vm0, vm0, left_shift);
- vsll_w_vx_m(vm1, vm1, left_shift);
- vsll_w_vx_m(vm2, vm2, left_shift);
- vsll_w_vx_m(vm3, vm3, left_shift);
- vsll_w_vx_m(vm4, vm4, left_shift);
- vsll_w_vx_m(vm5, vm5, left_shift);
- vsll_w_vx_m(vm6, vm6, left_shift);
- vsll_w_vx_m(vm7, vm7, left_shift);
+ vmul_w_vx_m(v4, v4, input1_shift_mul);
+ vmul_w_vx_m(v12, v12, input2_shift_mul);
- vmul_w_vx_m(vm0, vm0, input1_shift_mul);
- vmul_w_vx_m(vm1, vm1, input1_shift_mul);
- vmul_w_vx_m(vm2, vm2, input1_shift_mul);
- vmul_w_vx_m(vm3, vm3, input1_shift_mul);
- vmul_w_vx_m(vm4, vm4, input2_shift_mul);
- vmul_w_vx_m(vm5, vm5, input2_shift_mul);
- vmul_w_vx_m(vm6, vm6, input2_shift_mul);
- vmul_w_vx_m(vm7, vm7, input2_shift_mul);
+ vdmulh_w_r_vx_m(v4, v4, input1_mult);
+ vdmulh_w_r_vx_m(v12, v12, input2_mult);
+ vsha_w_r_vx_m(v4, v4, -input1_shift);
+ vsha_w_r_vx_m(v12, v12, -input2_shift);
+ vadd_w_vx_m(v4, v4, input1_offset);
+ vadd_w_vx_m(v12, v12, input2_offset);
- rescale_m(vm0, vm0, input1_mult, input1_shift, input1_offset);
- rescale_m(vm1, vm1, input1_mult, input1_shift, input1_offset);
- rescale_m(vm2, vm2, input1_mult, input1_shift, input1_offset);
- rescale_m(vm3, vm3, input1_mult, input1_shift, input1_offset);
- rescale_m(vm4, vm4, input2_mult, input2_shift, input2_offset);
- rescale_m(vm5, vm5, input2_mult, input2_shift, input2_offset);
- rescale_m(vm6, vm6, input2_mult, input2_shift, input2_offset);
- rescale_m(vm7, vm7, input2_mult, input2_shift, input2_offset);
+ vadd_w_vv_m(v16, v4, v12);
- // Sum the rescaled inputs.
- vadd_w_vv_m(vm0, vm0, vm4);
- vadd_w_vv_m(vm1, vm1, vm5);
- vadd_w_vv_m(vm2, vm2, vm6);
- vadd_w_vv_m(vm3, vm3, vm7);
+ rescale_m(v16, v16, output_mult, output_shift, output_offset);
- // Rescale the summed output.
- rescale_m(vm0, vm0, output_mult, output_shift, output_offset);
- rescale_m(vm1, vm1, output_mult, output_shift, output_offset);
- rescale_m(vm2, vm2, output_mult, output_shift, output_offset);
- rescale_m(vm3, vm3, output_mult, output_shift, output_offset);
+ vmin_w_vx_m(v16, v16, output_activation_max);
+ vmax_w_vx_m(v16, v16, output_activation_min);
- // Clamp to the provided range.
- vmin_w_vx_m(vm0, vm0, output_activation_max);
- vmin_w_vx_m(vm1, vm1, output_activation_max);
- vmin_w_vx_m(vm2, vm2, output_activation_max);
- vmin_w_vx_m(vm3, vm3, output_activation_max);
- vmax_w_vx_m(vm0, vm0, output_activation_min);
- vmax_w_vx_m(vm1, vm1, output_activation_min);
- vmax_w_vx_m(vm2, vm2, output_activation_min);
- vmax_w_vx_m(vm3, vm3, output_activation_min);
-
- // Swizzle and narrow back to bytes.
- vsraqs_b_vx_m(vm0, vm0, 0);
-
- // Store to memory.
- vst_b_lp_xx_m(vm0, output, count);
+ vsraqs_b_vx(v16, v16, 0);
+ vst_b_lp_xx(v16, output, count);
blocks -= count;
}