Skip to content

Commit

Permalink
Replace deprecated wasm_simd128.h intrinsics with new versions
Browse files Browse the repository at this point in the history 
The final version of wasm_simd128.h renamed and deprecated some of the
intrinsics used in XNNPACK WAsm SIMD microkernels. Update them to the new names
of these intrinsics.

PiperOrigin-RevId: 382364703
  • Loading branch information
@Maratyszcza @xnnpack-bot
Maratyszcza authored and xnnpack-bot committed Jun 30, 2021
1 parent f10af6c commit ee029b2
Show file tree
Hide file tree
Showing 603 changed files with 4,985 additions and 4,985 deletions.
6 changes: 3 additions & 3 deletions src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c
View file
Expand Up @@ -27,9 +27,9 @@ void xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4(
assert(kernel_elements > 9);
assert(channels != 0);

const v128_t vscale = wasm_v32x4_load_splat(&params->scalar.scale);
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);

do {
{
Expand Down
6 changes: 3 additions & 3 deletions src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c
View file
Expand Up @@ -27,9 +27,9 @@ void xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4(
assert(kernel_elements > 9);
assert(channels != 0);

const v128_t vscale = wasm_v32x4_load_splat(&params->scalar.scale);
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);

do {
{
Expand Down
6 changes: 3 additions & 3 deletions src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c
View file
Expand Up @@ -27,9 +27,9 @@ void xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4(
assert(kernel_elements <= 9);
assert(channels != 0);

const v128_t vscale = wasm_v32x4_load_splat(&params->scalar.scale);
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);

do {
const float* i0 = input[0];
Expand Down
6 changes: 3 additions & 3 deletions src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c
View file
Expand Up @@ -27,9 +27,9 @@ void xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4(
assert(kernel_elements <= 9);
assert(channels != 0);

const v128_t vscale = wasm_v32x4_load_splat(&params->scalar.scale);
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);

do {
const float* i0 = input[0];
Expand Down
14 changes: 7 additions & 7 deletions src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c
View file
Expand Up @@ -50,8 +50,8 @@ void xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2(
}

const v128_t vzero = wasm_f32x4_splat(0.0f);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);

for (size_t output_y = output_y_start; output_y < output_y_end; output_y += 2) {
const size_t input_y2 = output_y * 2 + 2 - input_padding_top;
Expand Down Expand Up @@ -508,11 +508,11 @@ void xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2(
v128_t vi4x3 = vzero;
if (iw > 2) {
// viMx3 = ( 0.0, 0.0, 0.0, iM3c2 )
vi0x3 = wasm_v32x4_load_splat(i0 + 8);
vi1x3 = wasm_v32x4_load_splat(i1 + 8);
vi2x3 = wasm_v32x4_load_splat(i2 + 8);
vi3x3 = wasm_v32x4_load_splat(i3 + 8);
vi4x3 = wasm_v32x4_load_splat(i4 + 8);
vi0x3 = wasm_v128_load32_splat(i0 + 8);
vi1x3 = wasm_v128_load32_splat(i1 + 8);
vi2x3 = wasm_v128_load32_splat(i2 + 8);
vi3x3 = wasm_v128_load32_splat(i3 + 8);
vi4x3 = wasm_v128_load32_splat(i4 + 8);
}

vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk01c2, wasm_v32x4_shuffle(vi0x1, vi0x1, 2, 2, 2, 2)));
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c
View file
Expand Up @@ -29,8 +29,8 @@ void xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86(
assert(channels != 0);
assert(output_width != 0);

const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
do {
const float* i0 = input[0];
assert(i0 != NULL);
Expand Down
4 changes: 2 additions & 2 deletions src/f32-dwconv/up-wasmsimd.c.in
View file
Expand Up @@ -34,8 +34,8 @@ void xnn_f32_dwconv${ACTIVATION_SUFFIX}_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}
assert(output_width != 0);

$if ACTIVATION == "MINMAX":
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
$elif ACTIVATION == "RELU":
const v128_t vzero = wasm_f32x4_splat(0.0f);
do {
Expand Down
6 changes: 3 additions & 3 deletions src/f32-dwconv2d-chw/3x3p1-wasmsimd-loadsplat.c.in
View file
Expand Up @@ -32,12 +32,12 @@ void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd${ARCH_SUFFIX}_loadsplat_${ROW_
assert(padding_top == 1);

const v128_t vmask = wasm_v128_load(params->scalar.mask);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);

const v128_t vw0123 = wasm_v128_load(weights);
const v128_t vw4567 = wasm_v128_load(weights + 4);
const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);
const v128_t vw89 = wasm_v128_load64_splat(weights + 8);
const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
const v128_t vk01 = wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
Expand Down
6 changes: 3 additions & 3 deletions src/f32-dwconv2d-chw/3x3p1-wasmsimd-splat.c.in
View file
Expand Up @@ -31,12 +31,12 @@ void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd${ARCH_SUFFIX}_splat_${ROW_TILE
assert(padding_top == 1);

const v128_t vmask = wasm_v128_load(params->scalar.mask);
const v128_t vmax = wasm_v32x4_load_splat(&params->scalar.max);
const v128_t vmin = wasm_v32x4_load_splat(&params->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);

const v128_t vw0123 = wasm_v128_load(weights);
const v128_t vw4567 = wasm_v128_load(weights + 4);
const v128_t vw89 = wasm_v64x2_load_splat(weights + 8);
const v128_t vw89 = wasm_v128_load64_splat(weights + 8);

const v128_t vzero = wasm_f32x4_splat(0.0f);

Expand Down

0 comments on commit ee029b2

Please sign in to comment.