Skip to content
This repository has been archived by the owner on Dec 1, 2021. It is now read-only.

Fix threshold skipping if flg = -1 #917

Merged
merged 5 commits into from
Mar 18, 2020
Merged

Fix threshold skipping if flg = -1 #917

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
ea6ee15
Fix threshold skipping if flg = -1
primenumber Mar 12, 2020
dec308a
Fix greater-than to greater-or-equal
primenumber Mar 12, 2020
97682fd
Merge branch 'master' into fix-ts-arm
iizukak Mar 18, 2020
ec0a2d2
Merge branch 'master' into fix-ts-arm
bo-mergebot[bot] Mar 18, 2020
90cb687
Merge branch 'master' into fix-ts-arm
bo-mergebot[bot] Mar 18, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 12 additions & 12 deletions blueoil/converter/templates/src/func/impl/arm_neon/quantized_conv2d_tiling.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -281,7 +281,7 @@ void QuantizedConv2DTiling(const tiling_input_t& input,
if (p.thresholds != nullptr) {
#define LOAD_TH(k) \
const auto ts##k = vld4q_s16(p.thresholds + NUM_OF_A2W1_THRESHOLD * (out_ch_high * OutChUnroll2 + Om + 8 * k)); \
const auto is_neg##k = vreinterpretq_s16_u16(vcltq_s16(ts##k.val[3], vdupq_n_s16(0))); \
const auto mask##k = vreinterpretq_s16_u16(0x0003 & vcltq_s16(ts##k.val[3], vdupq_n_s16(0))); \
const auto m2_##k = vsubq_s16(ts##k.val[3], vdupq_n_s16(2)); \
const auto is_const##k = vcgeq_s16(m2_##k, vdupq_n_s16(0));
LOAD_TH(0)
Expand All @@ -292,11 +292,11 @@ void QuantizedConv2DTiling(const tiling_input_t& input,
if (col_high + col >= out_width) break;
#define APPLY(k) \
const auto d##k = vld1q_s16(out_tile + buf_index + 8 * k); \
const auto f##k##0 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[0])) & ts##k.val[3]; \
const auto f##k##1 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[1])) & ts##k.val[3]; \
const auto f##k##2 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[2])) & ts##k.val[3]; \
const auto tmp##k = f##k##0 + f##k##1 + f##k##2 + is_neg##k; \
const auto res##k = vreinterpretq_u8_s16(vbslq_s16(is_const##k, m2_##k, tmp##k));
const auto f##k##0 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[0]) & 0x0001u); \
const auto f##k##1 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[1]) & 0x0001u); \
const auto f##k##2 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[2]) & 0x0001u); \
const auto tmp##k = f##k##0 + f##k##1 + f##k##2; \
const auto res##k = vreinterpretq_u8_s16(vbslq_s16(is_const##k, m2_##k, mask##k ^ tmp##k));
const auto buf_index = row * TileWidth * OutChUnroll
+ col * OutChUnroll;
APPLY(0)
Expand Down Expand Up @@ -546,7 +546,7 @@ void QuantizedConv2DTiling(const tiling_input_t& input,
if (p.thresholds != nullptr) {
#define LOAD_TH(k) \
const auto ts##k = vld4q_s16(p.thresholds + NUM_OF_A2W1_THRESHOLD * (out_ch_high * OutChUnroll2 + Om + 8 * k)); \
const auto is_neg##k = vreinterpretq_s16_u16(vcltq_s16(ts##k.val[3], vdupq_n_s16(0))); \
const auto mask##k = vreinterpretq_s16_u16(0x0003 & vcltq_s16(ts##k.val[3], vdupq_n_s16(0))); \
const auto m2_##k = vsubq_s16(ts##k.val[3], vdupq_n_s16(2)); \
const auto is_const##k = vcgeq_s16(m2_##k, vdupq_n_s16(0));
LOAD_TH(0)
Expand All @@ -557,11 +557,11 @@ void QuantizedConv2DTiling(const tiling_input_t& input,
if (col_high + col >= out_width) break;
#define APPLY(k) \
const auto d##k = vld1q_s16(out_tile + buf_index + 8 * k); \
const auto f##k##0 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[0])) & ts##k.val[3]; \
const auto f##k##1 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[1])) & ts##k.val[3]; \
const auto f##k##2 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[2])) & ts##k.val[3]; \
const auto tmp##k = f##k##0 + f##k##1 + f##k##2 + is_neg##k; \
const auto res##k = vreinterpretq_u8_s16(vbslq_s16(is_const##k, m2_##k, tmp##k));
const auto f##k##0 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[0]) & 0x0001u); \
const auto f##k##1 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[1]) & 0x0001u); \
const auto f##k##2 = vreinterpretq_s16_u16(vcgeq_s16(d##k, ts##k.val[2]) & 0x0001u); \
const auto tmp##k = f##k##0 + f##k##1 + f##k##2; \
const auto res##k = vreinterpretq_u8_s16(vbslq_s16(is_const##k, m2_##k, mask##k ^ tmp##k));
const auto buf_index = row * TileWidth * OutChUnroll
+ col * OutChUnroll;
APPLY(0)
Expand Down