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

[WIP] HVX x32 transposec #7395

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

[WIP] HVX x32 transposec #7395

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a4ac981
scripts to generate x32 transposec for HVX
ejparkqc Oct 28, 2024
2099a38
[WIP] tranpose for x32, vector_width=32
ejparkqc Oct 31, 2024
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
9 changes: 8 additions & 1 deletion cmake/gen/hvx_microkernels.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,6 +106,13 @@ SET(NON_PROD_HVX_MICROKERNEL_SRCS
src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u32.c
src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u64.c
src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u96.c
src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u128.c)
src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u128.c
src/x32-transposec/gen/x32-transposec-2x32-multi-multi-hvx.c
src/x32-transposec/gen/x32-transposec-4x32-multi-multi-hvx.c
src/x32-transposec/gen/x32-transposec-8x32-multi-multi-hvx.c
src/x32-transposec/gen/x32-transposec-16x32-multi-multi-hvx.c
src/x32-transposec/gen/x32-transposec-32x32-multi-mov-hvx.c
src/x32-transposec/gen/x32-transposec-32x32-multi-multi-hvx.c
src/x32-transposec/gen/x32-transposec-32x32-multi-switch-hvx.c)

SET(ALL_HVX_MICROKERNEL_SRCS ${PROD_HVX_MICROKERNEL_SRCS} + ${NON_PROD_HVX_MICROKERNEL_SRCS})
7 changes: 7 additions & 0 deletions gen/hvx_microkernels.bzl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,6 +104,13 @@ NON_PROD_HVX_MICROKERNEL_SRCS = [
"src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u64.c",
"src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u96.c",
"src/qs8-vadd/gen/qs8-vadd-minmax-hvx-u128.c",
"src/x32-transposec/gen/x32-transposec-2x32-multi-multi-hvx.c",
"src/x32-transposec/gen/x32-transposec-4x32-multi-multi-hvx.c",
"src/x32-transposec/gen/x32-transposec-8x32-multi-multi-hvx.c",
"src/x32-transposec/gen/x32-transposec-16x32-multi-multi-hvx.c",
"src/x32-transposec/gen/x32-transposec-32x32-multi-mov-hvx.c",
"src/x32-transposec/gen/x32-transposec-32x32-multi-multi-hvx.c",
"src/x32-transposec/gen/x32-transposec-32x32-multi-switch-hvx.c",
]

ALL_HVX_MICROKERNEL_SRCS = PROD_HVX_MICROKERNEL_SRCS + NON_PROD_HVX_MICROKERNEL_SRCS
9 changes: 9 additions & 0 deletions scripts/generate-xN-transpose.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,4 +171,13 @@ tools/xngen src/x32-transposec/wasmsimd.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=S
tools/xngen src/x32-transposec/wasmsimd.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MULTI -o src/x32-transposec/gen/x32-transposec-4x4-multi-multi-wasmsimd.c &
tools/xngen src/x32-transposec/wasmsimd.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MOV -o src/x32-transposec/gen/x32-transposec-4x4-multi-mov-wasmsimd.c &

#################################### HEXAGON HVX ###############################
tools/xngen src/x32-transposec/hvx.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MULTI TILE_HEIGHT=2 -o src/x32-transposec/gen/x32-transposec-2x32-multi-multi-hvx.c &
tools/xngen src/x32-transposec/hvx.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MULTI TILE_HEIGHT=4 -o src/x32-transposec/gen/x32-transposec-4x32-multi-multi-hvx.c &
tools/xngen src/x32-transposec/hvx.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MULTI TILE_HEIGHT=8 -o src/x32-transposec/gen/x32-transposec-8x32-multi-multi-hvx.c &
tools/xngen src/x32-transposec/hvx.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MULTI TILE_HEIGHT=16 -o src/x32-transposec/gen/x32-transposec-16x32-multi-multi-hvx.c &
tools/xngen src/x32-transposec/hvx.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MULTI TILE_HEIGHT=32 -o src/x32-transposec/gen/x32-transposec-32x32-multi-multi-hvx.c &
tools/xngen src/x32-transposec/hvx.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=SWITCH TILE_HEIGHT=32 -o src/x32-transposec/gen/x32-transposec-32x32-multi-switch-hvx.c &
tools/xngen src/x32-transposec/hvx.c.in -D SIZE=32 IN_PTRS=MULTI OUT_PTRS=MOV TILE_HEIGHT=32 -o src/x32-transposec/gen/x32-transposec-32x32-multi-mov-hvx.c &

wait
240 changes: 240 additions & 0 deletions src/x32-transposec/gen/x32-transposec-16x32-multi-multi-hvx.c
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,240 @@
// Auto-generated file. Do not edit!
// Template: src/x32-transposec/hvx.c.in
// Generator: tools/xngen
//
// Copyright 2024 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <xnnpack/simd/f32-hvx.h>

#include <assert.h>

#include "xnnpack/common.h"
#include "xnnpack/math.h"
#include "xnnpack/transpose.h"

void xnn_x32_transposec_ukernel__16x32_multi_multi_hvx(
const uint32_t* input,
uint32_t* output,
size_t input_stride,
size_t output_stride,
size_t block_width,
size_t block_height) XNN_OOB_READS
{
assert(block_width == 1 || output_stride >= block_height * sizeof(uint32_t));
assert(block_height == 1 || input_stride >= block_width * sizeof(uint32_t));

const size_t tile_height = 16;
const size_t tile_width = 32;
const size_t tile_hbytes = tile_height * sizeof(uint32_t);
const size_t tile_wbytes = tile_width * sizeof(uint32_t);
const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
const size_t input_offset = tile_height * input_stride;
const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);

const uint32_t* i0 = input;
const uint32_t* i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
const uint32_t* i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
const uint32_t* i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
const uint32_t* i4 = (const uint32_t*) ((uintptr_t) i3 + input_stride);
const uint32_t* i5 = (const uint32_t*) ((uintptr_t) i4 + input_stride);
const uint32_t* i6 = (const uint32_t*) ((uintptr_t) i5 + input_stride);
const uint32_t* i7 = (const uint32_t*) ((uintptr_t) i6 + input_stride);
const uint32_t* i8 = (const uint32_t*) ((uintptr_t) i7 + input_stride);
const uint32_t* i9 = (const uint32_t*) ((uintptr_t) i8 + input_stride);
const uint32_t* i10 = (const uint32_t*) ((uintptr_t) i9 + input_stride);
const uint32_t* i11 = (const uint32_t*) ((uintptr_t) i10 + input_stride);
const uint32_t* i12 = (const uint32_t*) ((uintptr_t) i11 + input_stride);
const uint32_t* i13 = (const uint32_t*) ((uintptr_t) i12 + input_stride);
const uint32_t* i14 = (const uint32_t*) ((uintptr_t) i13 + input_stride);
const uint32_t* i15 = (const uint32_t*) ((uintptr_t) i14 + input_stride);
uint32_t* o0 = (uint32_t*) output;
uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
uint32_t* o2 = (uint32_t*) ((uintptr_t) o1 + output_stride);
uint32_t* o3 = (uint32_t*) ((uintptr_t) o2 + output_stride);
uint32_t* o4 = (uint32_t*) ((uintptr_t) o3 + output_stride);
uint32_t* o5 = (uint32_t*) ((uintptr_t) o4 + output_stride);
uint32_t* o6 = (uint32_t*) ((uintptr_t) o5 + output_stride);
uint32_t* o7 = (uint32_t*) ((uintptr_t) o6 + output_stride);
uint32_t* o8 = (uint32_t*) ((uintptr_t) o7 + output_stride);
uint32_t* o9 = (uint32_t*) ((uintptr_t) o8 + output_stride);
uint32_t* o10 = (uint32_t*) ((uintptr_t) o9 + output_stride);
uint32_t* o11 = (uint32_t*) ((uintptr_t) o10 + output_stride);
uint32_t* o12 = (uint32_t*) ((uintptr_t) o11 + output_stride);
uint32_t* o13 = (uint32_t*) ((uintptr_t) o12 + output_stride);
uint32_t* o14 = (uint32_t*) ((uintptr_t) o13 + output_stride);
uint32_t* o15 = (uint32_t*) ((uintptr_t) o14 + output_stride);
const size_t minus_output_stride = -output_stride;

do {
if XNN_UNPREDICTABLE(block_width < 2) {
o1 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 2) {
o2 = o0;
}
if XNN_UNPREDICTABLE(block_width < 4) {
o3 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 4) {
o4 = o0;
}
if XNN_UNPREDICTABLE(block_width < 6) {
o5 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 6) {
o6 = o0;
}
if XNN_UNPREDICTABLE(block_width < 8) {
o7 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 8) {
o8 = o0;
}
if XNN_UNPREDICTABLE(block_width < 10) {
o9 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 10) {
o10 = o0;
}
if XNN_UNPREDICTABLE(block_width < 12) {
o11 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 12) {
o12 = o0;
}
if XNN_UNPREDICTABLE(block_width < 14) {
o13 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 14) {
o14 = o0;
}
if XNN_UNPREDICTABLE(block_width < 16) {
o15 = o0;
}
size_t bh = block_height;
for (; bh >= 16; bh -= 16) {
const HVX_Vector v4_0 = *((HVX_UVector *) i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
const HVX_Vector v4_1 = *((HVX_UVector *) i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);
const HVX_Vector v4_2 = *((HVX_UVector *) i2); i2 = (uint32_t*) ((uintptr_t) i2 + input_offset);
const HVX_Vector v4_3 = *((HVX_UVector *) i3); i3 = (uint32_t*) ((uintptr_t) i3 + input_offset);
const HVX_Vector v4_4 = *((HVX_UVector *) i4); i4 = (uint32_t*) ((uintptr_t) i4 + input_offset);
const HVX_Vector v4_5 = *((HVX_UVector *) i5); i5 = (uint32_t*) ((uintptr_t) i5 + input_offset);
const HVX_Vector v4_6 = *((HVX_UVector *) i6); i6 = (uint32_t*) ((uintptr_t) i6 + input_offset);
const HVX_Vector v4_7 = *((HVX_UVector *) i7); i7 = (uint32_t*) ((uintptr_t) i7 + input_offset);
const HVX_Vector v4_8 = *((HVX_UVector *) i8); i8 = (uint32_t*) ((uintptr_t) i8 + input_offset);
const HVX_Vector v4_9 = *((HVX_UVector *) i9); i9 = (uint32_t*) ((uintptr_t) i9 + input_offset);
const HVX_Vector v4_10 = *((HVX_UVector *) i10); i10 = (uint32_t*) ((uintptr_t) i10 + input_offset);
const HVX_Vector v4_11 = *((HVX_UVector *) i11); i11 = (uint32_t*) ((uintptr_t) i11 + input_offset);
const HVX_Vector v4_12 = *((HVX_UVector *) i12); i12 = (uint32_t*) ((uintptr_t) i12 + input_offset);
const HVX_Vector v4_13 = *((HVX_UVector *) i13); i13 = (uint32_t*) ((uintptr_t) i13 + input_offset);
const HVX_Vector v4_14 = *((HVX_UVector *) i14); i14 = (uint32_t*) ((uintptr_t) i14 + input_offset);
const HVX_Vector v4_15 = *((HVX_UVector *) i15); i15 = (uint32_t*) ((uintptr_t) i15 + input_offset);

int rt = -4;
const HVX_VectorPair v3_0 = Q6_W_vshuff_VVR(v4_1, v4_0, rt);
const HVX_VectorPair v3_1 = Q6_W_vshuff_VVR(v4_3, v4_2, rt);
const HVX_VectorPair v3_2 = Q6_W_vshuff_VVR(v4_5, v4_4, rt);
const HVX_VectorPair v3_3 = Q6_W_vshuff_VVR(v4_7, v4_6, rt);
const HVX_VectorPair v3_4 = Q6_W_vshuff_VVR(v4_9, v4_8, rt);
const HVX_VectorPair v3_5 = Q6_W_vshuff_VVR(v4_11, v4_10, rt);
const HVX_VectorPair v3_6 = Q6_W_vshuff_VVR(v4_13, v4_12, rt);
const HVX_VectorPair v3_7 = Q6_W_vshuff_VVR(v4_15, v4_14, rt);

rt = rt << 1;
HVX_VectorPair v2_0 = Q6_W_vshuff_VVR(Q6_V_lo_W(v3_1), Q6_V_lo_W(v3_0), rt);
HVX_VectorPair v2_1 = Q6_W_vshuff_VVR(Q6_V_hi_W(v3_1), Q6_V_hi_W(v3_0), rt);

HVX_VectorPair v2_2 = Q6_W_vshuff_VVR(Q6_V_lo_W(v3_3), Q6_V_lo_W(v3_2), rt);
HVX_VectorPair v2_3 = Q6_W_vshuff_VVR(Q6_V_hi_W(v3_3), Q6_V_hi_W(v3_2), rt);

HVX_VectorPair v2_4 = Q6_W_vshuff_VVR(Q6_V_lo_W(v3_5), Q6_V_lo_W(v3_4), rt);
HVX_VectorPair v2_5 = Q6_W_vshuff_VVR(Q6_V_hi_W(v3_5), Q6_V_hi_W(v3_4), rt);

HVX_VectorPair v2_6 = Q6_W_vshuff_VVR(Q6_V_lo_W(v3_7), Q6_V_lo_W(v3_6), rt);
HVX_VectorPair v2_7 = Q6_W_vshuff_VVR(Q6_V_hi_W(v3_7), Q6_V_hi_W(v3_6), rt);

rt = rt << 1;
HVX_VectorPair v1_0 = Q6_W_vshuff_VVR(Q6_V_lo_W(v2_2), Q6_V_lo_W(v2_0), rt);
HVX_VectorPair v1_1 = Q6_W_vshuff_VVR(Q6_V_hi_W(v2_2), Q6_V_hi_W(v2_0), rt);

HVX_VectorPair v1_2 = Q6_W_vshuff_VVR(Q6_V_lo_W(v2_3), Q6_V_lo_W(v2_1), rt);
HVX_VectorPair v1_3 = Q6_W_vshuff_VVR(Q6_V_hi_W(v2_3), Q6_V_hi_W(v2_1), rt);

HVX_VectorPair v1_4 = Q6_W_vshuff_VVR(Q6_V_lo_W(v2_6), Q6_V_lo_W(v2_4), rt);
HVX_VectorPair v1_5 = Q6_W_vshuff_VVR(Q6_V_hi_W(v2_6), Q6_V_hi_W(v2_4), rt);

HVX_VectorPair v1_6 = Q6_W_vshuff_VVR(Q6_V_lo_W(v2_7), Q6_V_lo_W(v2_5), rt);
HVX_VectorPair v1_7 = Q6_W_vshuff_VVR(Q6_V_hi_W(v2_7), Q6_V_hi_W(v2_5), rt);

rt = rt << 1;
HVX_VectorPair v0_0 = Q6_W_vshuff_VVR(Q6_V_lo_W(v1_4), Q6_V_lo_W(v1_0), rt);
HVX_VectorPair v0_1 = Q6_W_vshuff_VVR(Q6_V_hi_W(v1_4), Q6_V_hi_W(v1_0), rt);

HVX_VectorPair v0_2 = Q6_W_vshuff_VVR(Q6_V_lo_W(v1_5), Q6_V_lo_W(v1_1), rt);
HVX_VectorPair v0_3 = Q6_W_vshuff_VVR(Q6_V_hi_W(v1_5), Q6_V_hi_W(v1_1), rt);

HVX_VectorPair v0_4 = Q6_W_vshuff_VVR(Q6_V_lo_W(v1_6), Q6_V_lo_W(v1_2), rt);
HVX_VectorPair v0_5 = Q6_W_vshuff_VVR(Q6_V_hi_W(v1_6), Q6_V_hi_W(v1_2), rt);

HVX_VectorPair v0_6 = Q6_W_vshuff_VVR(Q6_V_lo_W(v1_7), Q6_V_lo_W(v1_3), rt);
HVX_VectorPair v0_7 = Q6_W_vshuff_VVR(Q6_V_hi_W(v1_7), Q6_V_hi_W(v1_3), rt);

xnn_storeu_f32(o15, Q6_V_hi_W(v0_7)); o15 = (uint32_t*) ((uintptr_t) o15 + tile_hbytes);
xnn_storeu_f32(o14, Q6_V_lo_W(v0_7)); o14 = (uint32_t*) ((uintptr_t) o14 + tile_hbytes);
xnn_storeu_f32(o13, Q6_V_hi_W(v0_6)); o13 = (uint32_t*) ((uintptr_t) o13 + tile_hbytes);
xnn_storeu_f32(o12, Q6_V_lo_W(v0_6)); o12 = (uint32_t*) ((uintptr_t) o12 + tile_hbytes);
xnn_storeu_f32(o11, Q6_V_hi_W(v0_5)); o11 = (uint32_t*) ((uintptr_t) o11 + tile_hbytes);
xnn_storeu_f32(o10, Q6_V_lo_W(v0_5)); o10 = (uint32_t*) ((uintptr_t) o10 + tile_hbytes);
xnn_storeu_f32(o9, Q6_V_hi_W(v0_4)); o9 = (uint32_t*) ((uintptr_t) o9 + tile_hbytes);
xnn_storeu_f32(o8, Q6_V_lo_W(v0_4)); o8 = (uint32_t*) ((uintptr_t) o8 + tile_hbytes);
xnn_storeu_f32(o7, Q6_V_hi_W(v0_3)); o7 = (uint32_t*) ((uintptr_t) o7 + tile_hbytes);
xnn_storeu_f32(o6, Q6_V_lo_W(v0_3)); o6 = (uint32_t*) ((uintptr_t) o6 + tile_hbytes);
xnn_storeu_f32(o5, Q6_V_hi_W(v0_2)); o5 = (uint32_t*) ((uintptr_t) o5 + tile_hbytes);
xnn_storeu_f32(o4, Q6_V_lo_W(v0_2)); o4 = (uint32_t*) ((uintptr_t) o4 + tile_hbytes);
xnn_storeu_f32(o3, Q6_V_hi_W(v0_1)); o3 = (uint32_t*) ((uintptr_t) o3 + tile_hbytes);
xnn_storeu_f32(o2, Q6_V_lo_W(v0_1)); o2 = (uint32_t*) ((uintptr_t) o2 + tile_hbytes);
xnn_storeu_f32(o1, Q6_V_hi_W(v0_0)); o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
xnn_storeu_f32(o0, Q6_V_lo_W(v0_0)); o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
}


i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
i2 = (const uint32_t*) ((uintptr_t) i1 + input_stride);
i3 = (const uint32_t*) ((uintptr_t) i2 + input_stride);
i4 = (const uint32_t*) ((uintptr_t) i3 + input_stride);
i5 = (const uint32_t*) ((uintptr_t) i4 + input_stride);
i6 = (const uint32_t*) ((uintptr_t) i5 + input_stride);
i7 = (const uint32_t*) ((uintptr_t) i6 + input_stride);
i8 = (const uint32_t*) ((uintptr_t) i7 + input_stride);
i9 = (const uint32_t*) ((uintptr_t) i8 + input_stride);
i10 = (const uint32_t*) ((uintptr_t) i9 + input_stride);
i11 = (const uint32_t*) ((uintptr_t) i10 + input_stride);
i12 = (const uint32_t*) ((uintptr_t) i11 + input_stride);
i13 = (const uint32_t*) ((uintptr_t) i12 + input_stride);
i14 = (const uint32_t*) ((uintptr_t) i13 + input_stride);
i15 = (const uint32_t*) ((uintptr_t) i14 + input_stride);
o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
o2 = (uint32_t*) ((uintptr_t) o2 + output_reset);
o3 = (uint32_t*) ((uintptr_t) o3 + output_reset);
o4 = (uint32_t*) ((uintptr_t) o4 + output_reset);
o5 = (uint32_t*) ((uintptr_t) o5 + output_reset);
o6 = (uint32_t*) ((uintptr_t) o6 + output_reset);
o7 = (uint32_t*) ((uintptr_t) o7 + output_reset);
o8 = (uint32_t*) ((uintptr_t) o8 + output_reset);
o9 = (uint32_t*) ((uintptr_t) o9 + output_reset);
o10 = (uint32_t*) ((uintptr_t) o10 + output_reset);
o11 = (uint32_t*) ((uintptr_t) o11 + output_reset);
o12 = (uint32_t*) ((uintptr_t) o12 + output_reset);
o13 = (uint32_t*) ((uintptr_t) o13 + output_reset);
o14 = (uint32_t*) ((uintptr_t) o14 + output_reset);
o15 = (uint32_t*) ((uintptr_t) o15 + output_reset);
block_width = doz(block_width, tile_width);
} while (block_width != 0);
}


68 changes: 68 additions & 0 deletions src/x32-transposec/gen/x32-transposec-2x32-multi-multi-hvx.c
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
// Auto-generated file. Do not edit!
// Template: src/x32-transposec/hvx.c.in
// Generator: tools/xngen
//
// Copyright 2024 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <xnnpack/simd/f32-hvx.h>

#include <assert.h>

#include "xnnpack/common.h"
#include "xnnpack/math.h"
#include "xnnpack/transpose.h"

void xnn_x32_transposec_ukernel__2x32_multi_multi_hvx(
const uint32_t* input,
uint32_t* output,
size_t input_stride,
size_t output_stride,
size_t block_width,
size_t block_height) XNN_OOB_READS
{
assert(block_width == 1 || output_stride >= block_height * sizeof(uint32_t));
assert(block_height == 1 || input_stride >= block_width * sizeof(uint32_t));

const size_t tile_height = 2;
const size_t tile_width = 32;
const size_t tile_hbytes = tile_height * sizeof(uint32_t);
const size_t tile_wbytes = tile_width * sizeof(uint32_t);
const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
const size_t input_offset = tile_height * input_stride;
const size_t output_reset = tile_width * output_stride - round_down_po2(block_height, 2) * sizeof(uint32_t);

const uint32_t* i0 = input;
const uint32_t* i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
uint32_t* o0 = (uint32_t*) output;
uint32_t* o1 = (uint32_t*) ((uintptr_t) o0 + output_stride);
const size_t minus_output_stride = -output_stride;

do {
if XNN_UNPREDICTABLE(block_width < 2) {
o1 = o0;
}
size_t bh = block_height;
for (; bh >= 2; bh -= 2) {
const HVX_Vector v1_0 = *((HVX_UVector *) i0); i0 = (uint32_t*) ((uintptr_t) i0 + input_offset);
const HVX_Vector v1_1 = *((HVX_UVector *) i1); i1 = (uint32_t*) ((uintptr_t) i1 + input_offset);

int rt = -4;
const HVX_VectorPair v0_0 = Q6_W_vshuff_VVR(v1_1, v1_0, rt);

xnn_storeu_f32(o1, Q6_V_hi_W(v0_0)); o1 = (uint32_t*) ((uintptr_t) o1 + tile_hbytes);
xnn_storeu_f32(o0, Q6_V_lo_W(v0_0)); o0 = (uint32_t*) ((uintptr_t) o0 + tile_hbytes);
}


i0 = (const uint32_t*) ((uintptr_t) i0 + input_reset);
i1 = (const uint32_t*) ((uintptr_t) i0 + input_stride);
o0 = (uint32_t*) ((uintptr_t) o0 + output_reset);
o1 = (uint32_t*) ((uintptr_t) o1 + output_reset);
block_width = doz(block_width, tile_width);
} while (block_width != 0);
}


Loading