Skip to content

Commit

Permalink
Refactor Arm Neon transpose_concat_*() to not need lookup table
Browse files Browse the repository at this point in the history 
Refactor the transpose_concat_*() helper function used in the Arm Neon
DotProd and I8MM vertical convolution implementations to not use TBL
instructions. Using vzip* to achieve the same outcome (with the same
number of instructions) avoids needing/loading the lookup indices and
also increases performance on little (in-order) Arm Cortex cores.

Change-Id: Ic118bdef3fce944ea6f02c85ec52ffc9c85908ea
  • Loading branch information
@jwright-arm @jzern
jwright-arm authored and jzern committed Apr 10, 2024
1 parent f1f1bf4 commit 24c1c6c
Show file tree
Hide file tree
Showing 2 changed files with 108 additions and 120 deletions.
114 changes: 54 additions & 60 deletions aom_dsp/arm/aom_convolve8_neon_dotprod.c
View file
Expand Up @@ -30,11 +30,6 @@ DECLARE_ALIGNED(16, static const uint8_t, dot_prod_permute_tbl[48]) = {
8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14
};

DECLARE_ALIGNED(16, static const uint8_t, dot_prod_tran_concat_tbl[32]) = {
0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27,
4, 12, 20, 28, 5, 13, 21, 29, 6, 14, 22, 30, 7, 15, 23, 31
};

DECLARE_ALIGNED(16, static const uint8_t, dot_prod_merge_block_tbl[48]) = {
/* Shift left and insert new last column in transposed 4x4 block. */
1, 2, 3, 16, 5, 6, 7, 20, 9, 10, 11, 24, 13, 14, 15, 28,
Expand Down Expand Up @@ -176,46 +171,54 @@ void aom_convolve8_horiz_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,
}

static INLINE void transpose_concat_4x4(int8x8_t a0, int8x8_t a1, int8x8_t a2,
int8x8_t a3, int8x16_t *b,
const uint8x16_t permute_tbl) {
/* Transpose 8-bit elements and concatenate result rows as follows:
* a0: 00, 01, 02, 03, XX, XX, XX, XX
* a1: 10, 11, 12, 13, XX, XX, XX, XX
* a2: 20, 21, 22, 23, XX, XX, XX, XX
* a3: 30, 31, 32, 33, XX, XX, XX, XX
*
* b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
*
* The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it
* as an argument is preferable to loading it directly from memory as this
* inline helper is called many times from the same parent function.
*/

int8x16x2_t samples = { { vcombine_s8(a0, a1), vcombine_s8(a2, a3) } };
*b = vqtbl2q_s8(samples, permute_tbl);
int8x8_t a3, int8x16_t *b) {
// Transpose 8-bit elements and concatenate result rows as follows:
// a0: 00, 01, 02, 03, XX, XX, XX, XX
// a1: 10, 11, 12, 13, XX, XX, XX, XX
// a2: 20, 21, 22, 23, XX, XX, XX, XX
// a3: 30, 31, 32, 33, XX, XX, XX, XX
//
// b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33

int8x16_t a0q = vcombine_s8(a0, vdup_n_s8(0));
int8x16_t a1q = vcombine_s8(a1, vdup_n_s8(0));
int8x16_t a2q = vcombine_s8(a2, vdup_n_s8(0));
int8x16_t a3q = vcombine_s8(a3, vdup_n_s8(0));

int8x16_t a01 = vzipq_s8(a0q, a1q).val[0];
int8x16_t a23 = vzipq_s8(a2q, a3q).val[0];

int16x8_t a0123 =
vzipq_s16(vreinterpretq_s16_s8(a01), vreinterpretq_s16_s8(a23)).val[0];

*b = vreinterpretq_s8_s16(a0123);
}

static INLINE void transpose_concat_8x4(int8x8_t a0, int8x8_t a1, int8x8_t a2,
int8x8_t a3, int8x16_t *b0,
int8x16_t *b1,
const uint8x16x2_t permute_tbl) {
/* Transpose 8-bit elements and concatenate result rows as follows:
* a0: 00, 01, 02, 03, 04, 05, 06, 07
* a1: 10, 11, 12, 13, 14, 15, 16, 17
* a2: 20, 21, 22, 23, 24, 25, 26, 27
* a3: 30, 31, 32, 33, 34, 35, 36, 37
*
* b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
* b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37
*
* The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it
* as an argument is preferable to loading it directly from memory as this
* inline helper is called many times from the same parent function.
*/

int8x16x2_t samples = { { vcombine_s8(a0, a1), vcombine_s8(a2, a3) } };
*b0 = vqtbl2q_s8(samples, permute_tbl.val[0]);
*b1 = vqtbl2q_s8(samples, permute_tbl.val[1]);
int8x16_t *b1) {
// Transpose 8-bit elements and concatenate result rows as follows:
// a0: 00, 01, 02, 03, 04, 05, 06, 07
// a1: 10, 11, 12, 13, 14, 15, 16, 17
// a2: 20, 21, 22, 23, 24, 25, 26, 27
// a3: 30, 31, 32, 33, 34, 35, 36, 37
//
// b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
// b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37

int8x16_t a0q = vcombine_s8(a0, vdup_n_s8(0));
int8x16_t a1q = vcombine_s8(a1, vdup_n_s8(0));
int8x16_t a2q = vcombine_s8(a2, vdup_n_s8(0));
int8x16_t a3q = vcombine_s8(a3, vdup_n_s8(0));

int8x16_t a01 = vzipq_s8(a0q, a1q).val[0];
int8x16_t a23 = vzipq_s8(a2q, a3q).val[0];

int16x8x2_t a0123 =
vzipq_s16(vreinterpretq_s16_s8(a01), vreinterpretq_s16_s8(a23));

*b0 = vreinterpretq_s8_s16(a0123.val[0]);
*b1 = vreinterpretq_s8_s16(a0123.val[1]);
}

static INLINE int16x4_t convolve8_4_sdot_partial(const int8x16_t samples_lo,
Expand Down Expand Up @@ -277,8 +280,6 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,
src -= ((SUBPEL_TAPS / 2) - 1) * src_stride;

if (w == 4) {
const uint8x16_t tran_concat_tbl = vld1q_u8(dot_prod_tran_concat_tbl);

uint8x8_t t0, t1, t2, t3, t4, t5, t6;
load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
src += 7 * src_stride;
Expand All @@ -296,10 +297,10 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,
* (see horizontal case) required before computing the dot product.
*/
int8x16_t s0123, s1234, s2345, s3456;
transpose_concat_4x4(s0, s1, s2, s3, &s0123, tran_concat_tbl);
transpose_concat_4x4(s1, s2, s3, s4, &s1234, tran_concat_tbl);
transpose_concat_4x4(s2, s3, s4, s5, &s2345, tran_concat_tbl);
transpose_concat_4x4(s3, s4, s5, s6, &s3456, tran_concat_tbl);
transpose_concat_4x4(s0, s1, s2, s3, &s0123);
transpose_concat_4x4(s1, s2, s3, s4, &s1234);
transpose_concat_4x4(s2, s3, s4, s5, &s2345);
transpose_concat_4x4(s3, s4, s5, s6, &s3456);

do {
uint8x8_t t7, t8, t9, t10;
Expand All @@ -311,7 +312,7 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,
int8x8_t s10 = vreinterpret_s8_u8(vsub_u8(t10, range_limit));

int8x16_t s4567, s5678, s6789, s78910;
transpose_concat_4x4(s7, s8, s9, s10, &s78910, tran_concat_tbl);
transpose_concat_4x4(s7, s8, s9, s10, &s78910);

/* Merge new data into block from previous iteration. */
samples_LUT.val[0] = s3456;
Expand Down Expand Up @@ -343,8 +344,6 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,
h -= 4;
} while (h != 0);
} else {
const uint8x16x2_t tran_concat_tbl = vld1q_u8_x2(dot_prod_tran_concat_tbl);

do {
int height = h;
const uint8_t *s = src;
Expand All @@ -368,14 +367,10 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,
*/
int8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi,
s3456_lo, s3456_hi;
transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi,
tran_concat_tbl);
transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi,
tran_concat_tbl);
transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi,
tran_concat_tbl);
transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi,
tran_concat_tbl);
transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi);
transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi);
transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi);
transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi);

do {
uint8x8_t t7, t8, t9, t10;
Expand All @@ -388,8 +383,7 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride,

int8x16_t s4567_lo, s4567_hi, s5678_lo, s5678_hi, s6789_lo, s6789_hi,
s78910_lo, s78910_hi;
transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi,
tran_concat_tbl);
transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi);

/* Merge new data into block from previous iteration. */
samples_LUT.val[0] = s3456_lo;
Expand Down
114 changes: 54 additions & 60 deletions aom_dsp/arm/aom_convolve8_neon_i8mm.c
View file
Expand Up @@ -29,11 +29,6 @@ DECLARE_ALIGNED(16, static const uint8_t, dot_prod_permute_tbl[48]) = {
8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14
};

DECLARE_ALIGNED(16, static const uint8_t, dot_prod_tran_concat_tbl[32]) = {
0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27,
4, 12, 20, 28, 5, 13, 21, 29, 6, 14, 22, 30, 7, 15, 23, 31
};

DECLARE_ALIGNED(16, static const uint8_t, dot_prod_merge_block_tbl[48]) = {
/* Shift left and insert new last column in transposed 4x4 block. */
1, 2, 3, 16, 5, 6, 7, 20, 9, 10, 11, 24, 13, 14, 15, 28,
Expand Down Expand Up @@ -162,46 +157,54 @@ void aom_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,

static INLINE void transpose_concat_4x4(uint8x8_t a0, uint8x8_t a1,
uint8x8_t a2, uint8x8_t a3,
uint8x16_t *b,
const uint8x16_t permute_tbl) {
/* Transpose 8-bit elements and concatenate result rows as follows:
* a0: 00, 01, 02, 03, XX, XX, XX, XX
* a1: 10, 11, 12, 13, XX, XX, XX, XX
* a2: 20, 21, 22, 23, XX, XX, XX, XX
* a3: 30, 31, 32, 33, XX, XX, XX, XX
*
* b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
*
* The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it
* as an argument is preferable to loading it directly from memory as this
* inline helper is called many times from the same parent function.
*/

uint8x16x2_t samples = { { vcombine_u8(a0, a1), vcombine_u8(a2, a3) } };
*b = vqtbl2q_u8(samples, permute_tbl);
uint8x16_t *b) {
// Transpose 8-bit elements and concatenate result rows as follows:
// a0: 00, 01, 02, 03, XX, XX, XX, XX
// a1: 10, 11, 12, 13, XX, XX, XX, XX
// a2: 20, 21, 22, 23, XX, XX, XX, XX
// a3: 30, 31, 32, 33, XX, XX, XX, XX
//
// b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33

uint8x16_t a0q = vcombine_u8(a0, vdup_n_u8(0));
uint8x16_t a1q = vcombine_u8(a1, vdup_n_u8(0));
uint8x16_t a2q = vcombine_u8(a2, vdup_n_u8(0));
uint8x16_t a3q = vcombine_u8(a3, vdup_n_u8(0));

uint8x16_t a01 = vzipq_u8(a0q, a1q).val[0];
uint8x16_t a23 = vzipq_u8(a2q, a3q).val[0];

uint16x8_t a0123 =
vzipq_u16(vreinterpretq_u16_u8(a01), vreinterpretq_u16_u8(a23)).val[0];

*b = vreinterpretq_u8_u16(a0123);
}

static INLINE void transpose_concat_8x4(uint8x8_t a0, uint8x8_t a1,
uint8x8_t a2, uint8x8_t a3,
uint8x16_t *b0, uint8x16_t *b1,
const uint8x16x2_t permute_tbl) {
/* Transpose 8-bit elements and concatenate result rows as follows:
* a0: 00, 01, 02, 03, 04, 05, 06, 07
* a1: 10, 11, 12, 13, 14, 15, 16, 17
* a2: 20, 21, 22, 23, 24, 25, 26, 27
* a3: 30, 31, 32, 33, 34, 35, 36, 37
*
* b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
* b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37
*
* The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it
* as an argument is preferable to loading it directly from memory as this
* inline helper is called many times from the same parent function.
*/

uint8x16x2_t samples = { { vcombine_u8(a0, a1), vcombine_u8(a2, a3) } };
*b0 = vqtbl2q_u8(samples, permute_tbl.val[0]);
*b1 = vqtbl2q_u8(samples, permute_tbl.val[1]);
uint8x16_t *b0, uint8x16_t *b1) {
// Transpose 8-bit elements and concatenate result rows as follows:
// a0: 00, 01, 02, 03, 04, 05, 06, 07
// a1: 10, 11, 12, 13, 14, 15, 16, 17
// a2: 20, 21, 22, 23, 24, 25, 26, 27
// a3: 30, 31, 32, 33, 34, 35, 36, 37
//
// b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
// b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37

uint8x16_t a0q = vcombine_u8(a0, vdup_n_u8(0));
uint8x16_t a1q = vcombine_u8(a1, vdup_n_u8(0));
uint8x16_t a2q = vcombine_u8(a2, vdup_n_u8(0));
uint8x16_t a3q = vcombine_u8(a3, vdup_n_u8(0));

uint8x16_t a01 = vzipq_u8(a0q, a1q).val[0];
uint8x16_t a23 = vzipq_u8(a2q, a3q).val[0];

uint16x8x2_t a0123 =
vzipq_u16(vreinterpretq_u16_u8(a01), vreinterpretq_u16_u8(a23));

*b0 = vreinterpretq_u8_u16(a0123.val[0]);
*b1 = vreinterpretq_u8_u16(a0123.val[1]);
}

static INLINE int16x4_t convolve8_4_usdot_partial(const uint8x16_t samples_lo,
Expand Down Expand Up @@ -257,8 +260,6 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
src -= ((SUBPEL_TAPS / 2) - 1) * src_stride;

if (w == 4) {
const uint8x16_t tran_concat_tbl = vld1q_u8(dot_prod_tran_concat_tbl);

uint8x8_t s0, s1, s2, s3, s4, s5, s6;
load_u8_8x7(src, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
src += 7 * src_stride;
Expand All @@ -267,17 +268,17 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
* (see horizontal case) required before computing the dot product.
*/
uint8x16_t s0123, s1234, s2345, s3456;
transpose_concat_4x4(s0, s1, s2, s3, &s0123, tran_concat_tbl);
transpose_concat_4x4(s1, s2, s3, s4, &s1234, tran_concat_tbl);
transpose_concat_4x4(s2, s3, s4, s5, &s2345, tran_concat_tbl);
transpose_concat_4x4(s3, s4, s5, s6, &s3456, tran_concat_tbl);
transpose_concat_4x4(s0, s1, s2, s3, &s0123);
transpose_concat_4x4(s1, s2, s3, s4, &s1234);
transpose_concat_4x4(s2, s3, s4, s5, &s2345);
transpose_concat_4x4(s3, s4, s5, s6, &s3456);

do {
uint8x8_t s7, s8, s9, s10;
load_u8_8x4(src, src_stride, &s7, &s8, &s9, &s10);

uint8x16_t s4567, s5678, s6789, s78910;
transpose_concat_4x4(s7, s8, s9, s10, &s78910, tran_concat_tbl);
transpose_concat_4x4(s7, s8, s9, s10, &s78910);

/* Merge new data into block from previous iteration. */
samples_LUT.val[0] = s3456;
Expand Down Expand Up @@ -308,8 +309,6 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
h -= 4;
} while (h != 0);
} else {
const uint8x16x2_t tran_concat_tbl = vld1q_u8_x2(dot_prod_tran_concat_tbl);

do {
int height = h;
const uint8_t *s = src;
Expand All @@ -324,23 +323,18 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
*/
uint8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi,
s3456_lo, s3456_hi;
transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi,
tran_concat_tbl);
transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi,
tran_concat_tbl);
transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi,
tran_concat_tbl);
transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi,
tran_concat_tbl);
transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi);
transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi);
transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi);
transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi);

do {
uint8x8_t s7, s8, s9, s10;
load_u8_8x4(s, src_stride, &s7, &s8, &s9, &s10);

uint8x16_t s4567_lo, s4567_hi, s5678_lo, s5678_hi, s6789_lo, s6789_hi,
s78910_lo, s78910_hi;
transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi,
tran_concat_tbl);
transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi);

/* Merge new data into block from previous iteration. */
samples_LUT.val[0] = s3456_lo;
Expand Down

0 comments on commit 24c1c6c

Please sign in to comment.