Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[arm] add v7 fp16 scale implement #9048

Merged
merged 3 commits into from
May 26, 2022
Merged

[arm] add v7 fp16 scale implement #9048

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f6472a7
add v7 fp16 scale implement
chenjiaoAngel May 20, 2022
ecd6248
Update .pre-commit-config.yaml
chenjiaoAngel May 20, 2022
f96362d
Update cpplint_pre_commit.hook
chenjiaoAngel May 20, 2022
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
157 changes: 157 additions & 0 deletions lite/backends/arm/math/scale.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -863,6 +863,40 @@ inline void scale_compute_fp16(const flaot16_t* din,
[remain_cnt] "+r"(remain_cnt)
: [vscale] "w"(vscale), [vbias] "w"(vbias)
: "cc", "memory", "v4", "v5", "v8", "v9");
#else
asm volatile(
"cmp %[cnt], #1 \n"
"blt 0f \n"
"1: \n"
"vld1.16 {d8-d9}, [%[din]]! \n"
"vmov q8, %q[vbias] \n"
"vld1.16 {d10-d11}, [%[din]]! \n"
"vmov q9, %q[vbias] \n"

"vmla.f16 q8, q4, %q[vscale] \n"
"vmla.f16 q9, q5, %q[vscale] \n"

"subs %[cnt], %[cnt], #1 \n"
"vst1.16 {d16-d19}, [%[dout]]! \n"
"bne 1b \n"
"0: \n"
"cmp %[remain_cnt], #1 \n"
"blt 2f \n"
"3: \n"
"vld1.16 {d8}, [%[din]]! \n"
"vmov d16, %e[vbias] \n"
"vmla.f16 d16, d8, %e[vscale] \n"
"subs %[remain_cnt], %[remain_cnt], #1 \n"
"vst1.16 {d16}, [%[dout]]! \n"
"bne 3b \n"
"2: \n"
: [dout] "+r"(dout),
[din] "+r"(din),
[cnt] "+r"(cnt),
[remain_cnt] "+r"(remain_cnt)
: [vscale] "w"(vscale), [vbias] "w"(vbias)
: "cc", "memory", "q4", "q5", "q8", "q9");

#endif
for (int j = 0; j < remain_rem; j++) {
*dout = *din * vscale[0] + vbias[0];
Expand Down Expand Up @@ -1013,6 +1047,42 @@ void scale_relu<float16_t>(const float16_t* din,
: [vscale] "w"(vscale), [vbias] "w"(vbias), [vzero] "w"(vzero)
: "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11");
#else
asm volatile(
"cmp %[cnt], #1 \n"
"blt 0f \n"
"1: \n"
"vld1.16 {d8-d9}, [%[din]]! \n"
"vmov q8, %q[vbias] \n"
"vld1.16 {d10-d11}, [%[din]]! \n"
"vmov q9, %q[vbias] \n"

"vmla.f16 q8, q4, %q[vscale] \n"
"vmla.f16 q9, q5, %q[vscale] \n"
"vmax.f16 q8, q8, %q[vzero] \n"
"vmax.f16 q9, q9, %q[vzero] \n"

"subs %[cnt], %[cnt], #1 \n"
"vst1.16 {d16-d19}, [%[dout]]! \n"
"bne 1b \n"
"0: \n"
"cmp %[remain_cnt], #1 \n"
"blt 2f \n"
"3: \n"
"vld1.16 {d8}, [%[din]]! \n"
"vmov d16, %e[vbias] \n"
"vmla.f16 d16, d8, %e[vscale] \n"
"vmax.f16 d16, d16, %e[vzero] \n"
"subs %[remain_cnt], %[remain_cnt], #1 \n"
"vst1.16 {d16}, [%[dout]]! \n"
"bne 3b \n"
"2: \n"
: [dout] "+r"(dout),
[din] "+r"(din),
[cnt] "+r"(cnt),
[remain_cnt] "+r"(remain_num)
: [vscale] "w"(vscale), [vbias] "w"(vbias), [vzero] "w"(vzero)
: "cc", "memory", "q4", "q5", "q8", "q9");

#endif
for (int i = 0; i < remain_rem; i++) {
*dout = *din * scale + bias;
Expand Down Expand Up @@ -1082,6 +1152,48 @@ void scale_relu6<float16_t>(const float16_t* din,
[valpha] "w"(valpha)
: "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11");
#else
asm volatile(
"cmp %[cnt], #1 \n"
"blt 0f \n"
"1: \n"
"vld1.16 {d8-d9}, [%[din]]! \n"
"vmov q8, %q[vbias] \n"
"vld1.16 {d10-d11}, [%[din]]! \n"
"vmov q9, %q[vbias] \n"

"vmla.f16 q8, q4, %q[vscale] \n"
"vmla.f16 q9, q5, %q[vscale] \n"
"vmax.f16 q8, q8, %q[vzero] \n"
"vmax.f16 q9, q9, %q[vzero] \n"
"vmin.f16 q8, q8, %q[valpha] \n"
"vmin.f16 q9, q9, %q[valpha] \n"

"subs %[cnt], %[cnt], #1 \n"
"vst1.16 {d16-d19}, [%[dout]]! \n"
"bne 1b \n"
"0: \n"
"cmp %[remain_cnt], #1 \n"
"blt 2f \n"
"3: \n"
"vld1.16 {d8}, [%[din]]! \n"
"vmov d16, %e[vbias] \n"
"vmla.f16 d16, d8, %e[vscale] \n"
"vmax.f16 d16, d16, %e[vzero] \n"
"vmin.f16 d16, d16, %e[valpha] \n"
"subs %[remain_cnt], %[remain_cnt], #1 \n"
"vst1.16 {d16}, [%[dout]]! \n"
"bne 3b \n"
"2: \n"
: [dout] "+r"(dout),
[din] "+r"(din),
[cnt] "+r"(cnt),
[remain_cnt] "+r"(remain_num)
: [vscale] "w"(vscale),
[vbias] "w"(vbias),
[vzero] "w"(vzero),
[valpha] "w"(valpha)
: "cc", "memory", "q4", "q5", "q8", "q9");

#endif
for (int i = 0; i < remain_rem; i++) {
*dout = *din * scale + bias;
Expand Down Expand Up @@ -1153,6 +1265,51 @@ void scale_leaky_relu<float16_t>(const float16_t* din,
[valpha] "w"(valpha)
: "cc", "memory", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11");
#else
asm volatile(
"cmp %[cnt], #1 \n"
"blt 0f \n"
"1: \n"
"vld1.16 {d8-d9}, [%[din]]! \n"
"vmov q8, %q[vbias] \n"
"vld1.16 {d10-d11}, [%[din]]! \n"
"vmov q9, %q[vbias] \n"

"vmla.f16 q8, q4, %q[vscale] \n"
"vmla.f16 q9, q5, %q[vscale] \n"
"vcge.f16 q10, q8, %q[vzero] \n"
"vmul.f16 q11, q8, %q[valpha]\n"
"vcge.f16 q12, q9, %q[vzero] \n"
"vmul.f16 q13, q9, %q[valpha]\n"

"subs %[cnt], %[cnt], #1 \n"
"vbif q8, q11, q10 \n"
"vbif q9, q13, q12 \n"
"vst1.16 {d16-d19}, [%[dout]]! \n"
"bne 1b \n"
"0: \n"
"cmp %[remain_cnt], #1 \n"
"blt 2f \n"
"3: \n"
"vld1.16 {d8}, [%[din]]! \n"
"vmov d16, %e[vbias] \n"
"vmla.f16 d16, d8, %e[vscale] \n"
"vcge.f16 d20, d16, %e[vzero] \n"
"vmul.f16 d22, d16, %e[valpha]\n"
"subs %[remain_cnt], %[remain_cnt], #1 \n"
"vbif d16, d22, d20 \n"
"vst1.16 {d16}, [%[dout]]! \n"
"bne 3b \n"
"2: \n"
: [dout] "+r"(dout),
[din] "+r"(din),
[cnt] "+r"(cnt),
[remain_cnt] "+r"(remain_num)
: [vscale] "w"(vscale),
[vbias] "w"(vbias),
[vzero] "w"(vzero),
[valpha] "w"(valpha)
: "cc", "memory", "q4", "q5", "q8", "q9", "q10", "q11");

#endif
for (int i = 0; i < remain_rem; i++) {
*dout = *din * scale + bias;
Expand Down