mozilla
/
aom
зеркало из
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

mips msa vpx_dsp sadx3 sadx8 optimization 

average improvement ~3x-5x

Change-Id: Ifdb4670d31ae83c4e22a4238293d1377b16c90db
This commit is contained in:
Parag Salasakar 2015-07-02 08:02:19 +05:30
Родитель e757808429
Коммит 6abf1aea63
2 изменённых файлов: 754 добавлений и 10 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

738
vpx_dsp/mips/sad_msa.c
Просмотреть файл

@ -11,6 +11,14 @@
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/mips/macros_msa.h"
#define SAD_INSVE_W4(RTYPE, in0, in1, in2, in3, out) { \
out = (RTYPE)__msa_insve_w((v4i32)out, 0, (v4i32)in0); \
out = (RTYPE)__msa_insve_w((v4i32)out, 1, (v4i32)in1); \
out = (RTYPE)__msa_insve_w((v4i32)out, 2, (v4i32)in2); \
out = (RTYPE)__msa_insve_w((v4i32)out, 3, (v4i32)in3); \
}
#define SAD_INSVE_W4_UB(...) SAD_INSVE_W4(v16u8, __VA_ARGS__)
static uint32_t sad_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
const uint8_t *ref_ptr, int32_t ref_stride,
int32_t height) {
@ -150,6 +158,640 @@ static uint32_t sad_64width_msa(const uint8_t *src, int32_t src_stride,
return sad;
}
static void sad_4width_x3_msa(const uint8_t *src_ptr, int32_t src_stride,
const uint8_t *ref_ptr, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
uint32_t src0, src1, src2, src3;
v16u8 src = { 0 };
v16u8 ref = { 0 };
v16u8 ref0, ref1, ref2, ref3, diff;
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
for (ht_cnt = (height >> 2); ht_cnt--;) {
LW4(src_ptr, src_stride, src0, src1, src2, src3);
src_ptr += (4 * src_stride);
INSERT_W4_UB(src0, src1, src2, src3, src);
LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
ref_ptr += (4 * ref_stride);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad0 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad1 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad2 += __msa_hadd_u_h(diff, diff);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
}
static void sad_8width_x3_msa(const uint8_t *src, int32_t src_stride,
const uint8_t *ref, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
v16u8 src0, src1, src2, src3;
v16u8 ref0, ref1, ref00, ref11, ref22, ref33;
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
for (ht_cnt = (height >> 2); ht_cnt--;) {
LD_UB4(src, src_stride, src0, src1, src2, src3);
src += (4 * src_stride);
LD_UB4(ref, ref_stride, ref00, ref11, ref22, ref33);
ref += (4 * ref_stride);
PCKEV_D4_UB(src1, src0, src3, src2, ref11, ref00, ref33, ref22,
src0, src1, ref0, ref1);
sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
}
static void sad_16width_x3_msa(const uint8_t *src_ptr, int32_t src_stride,
const uint8_t *ref_ptr, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
v16u8 src, ref, ref0, ref1, diff;
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
for (ht_cnt = (height >> 1); ht_cnt--;) {
src = LD_UB(src_ptr);
src_ptr += src_stride;
LD_UB2(ref_ptr, 16, ref0, ref1);
ref_ptr += ref_stride;
diff = __msa_asub_u_b(src, ref0);
sad0 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
diff = __msa_asub_u_b(src, ref);
sad1 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
diff = __msa_asub_u_b(src, ref);
sad2 += __msa_hadd_u_h(diff, diff);
src = LD_UB(src_ptr);
src_ptr += src_stride;
LD_UB2(ref_ptr, 16, ref0, ref1);
ref_ptr += ref_stride;
diff = __msa_asub_u_b(src, ref0);
sad0 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
diff = __msa_asub_u_b(src, ref);
sad1 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
diff = __msa_asub_u_b(src, ref);
sad2 += __msa_hadd_u_h(diff, diff);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
}
static void sad_32width_x3_msa(const uint8_t *src, int32_t src_stride,
const uint8_t *ref, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
v16u8 src0, src1, ref0_0, ref0_1, ref0_2, ref0, ref1;
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
for (ht_cnt = height >> 1; ht_cnt--;) {
LD_UB2(src, 16, src0, src1);
src += src_stride;
LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
ref += ref_stride;
sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
LD_UB2(src, 16, src0, src1);
src += src_stride;
LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
ref += ref_stride;
sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
}
static void sad_64width_x3_msa(const uint8_t *src, int32_t src_stride,
const uint8_t *ref, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
v16u8 src0, src1, src2, src3;
v16u8 ref0_0, ref0_1, ref0_2, ref0_3, ref0_4, ref0, ref1, ref2, ref3;
v8u16 sad0_0 = { 0 };
v8u16 sad0_1 = { 0 };
v8u16 sad1_0 = { 0 };
v8u16 sad1_1 = { 0 };
v8u16 sad2_0 = { 0 };
v8u16 sad2_1 = { 0 };
v4u32 sad;
for (ht_cnt = height; ht_cnt--;) {
LD_UB4(src, 16, src0, src1, src2, src3);
src += src_stride;
LD_UB4(ref, 16, ref0_0, ref0_1, ref0_2, ref0_3);
ref0_4 = LD_UB(ref + 64);
ref += ref_stride;
sad0_0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
sad0_1 += SAD_UB2_UH(src2, src3, ref0_2, ref0_3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 1);
sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 2);
sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
}
sad = __msa_hadd_u_w(sad0_0, sad0_0);
sad += __msa_hadd_u_w(sad0_1, sad0_1);
sad_array[0] = HADD_SW_S32((v4i32)sad);
sad = __msa_hadd_u_w(sad1_0, sad1_0);
sad += __msa_hadd_u_w(sad1_1, sad1_1);
sad_array[1] = HADD_SW_S32((v4i32)sad);
sad = __msa_hadd_u_w(sad2_0, sad2_0);
sad += __msa_hadd_u_w(sad2_1, sad2_1);
sad_array[2] = HADD_SW_S32((v4i32)sad);
}
static void sad_4width_x8_msa(const uint8_t *src_ptr, int32_t src_stride,
const uint8_t *ref_ptr, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
uint32_t src0, src1, src2, src3;
v16u8 ref0, ref1, ref2, ref3, diff;
v16u8 src = { 0 };
v16u8 ref = { 0 };
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
v8u16 sad3 = { 0 };
v8u16 sad4 = { 0 };
v8u16 sad5 = { 0 };
v8u16 sad6 = { 0 };
v8u16 sad7 = { 0 };
for (ht_cnt = (height >> 2); ht_cnt--;) {
LW4(src_ptr, src_stride, src0, src1, src2, src3);
INSERT_W4_UB(src0, src1, src2, src3, src);
src_ptr += (4 * src_stride);
LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
ref_ptr += (4 * ref_stride);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad0 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad1 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad2 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad3 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad4 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad5 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad6 += __msa_hadd_u_h(diff, diff);
SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
diff = __msa_asub_u_b(src, ref);
sad7 += __msa_hadd_u_h(diff, diff);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
sad_array[3] = HADD_UH_U32(sad3);
sad_array[4] = HADD_UH_U32(sad4);
sad_array[5] = HADD_UH_U32(sad5);
sad_array[6] = HADD_UH_U32(sad6);
sad_array[7] = HADD_UH_U32(sad7);
}
static void sad_8width_x8_msa(const uint8_t *src, int32_t src_stride,
const uint8_t *ref, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
v16u8 src0, src1, src2, src3;
v16u8 ref0, ref1, ref00, ref11, ref22, ref33;
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
v8u16 sad3 = { 0 };
v8u16 sad4 = { 0 };
v8u16 sad5 = { 0 };
v8u16 sad6 = { 0 };
v8u16 sad7 = { 0 };
for (ht_cnt = (height >> 2); ht_cnt--;) {
LD_UB4(src, src_stride, src0, src1, src2, src3);
src += (4 * src_stride);
LD_UB4(ref, ref_stride, ref00, ref11, ref22, ref33);
ref += (4 * ref_stride);
PCKEV_D4_UB(src1, src0, src3, src2, ref11, ref00, ref33, ref22,
src0, src1, ref0, ref1);
sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad4 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad5 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad6 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
sad7 += SAD_UB2_UH(src0, src1, ref0, ref1);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
sad_array[3] = HADD_UH_U32(sad3);
sad_array[4] = HADD_UH_U32(sad4);
sad_array[5] = HADD_UH_U32(sad5);
sad_array[6] = HADD_UH_U32(sad6);
sad_array[7] = HADD_UH_U32(sad7);
}
static void sad_16width_x8_msa(const uint8_t *src_ptr, int32_t src_stride,
const uint8_t *ref_ptr, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
v16u8 src, ref0, ref1, ref;
v16u8 diff;
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
v8u16 sad3 = { 0 };
v8u16 sad4 = { 0 };
v8u16 sad5 = { 0 };
v8u16 sad6 = { 0 };
v8u16 sad7 = { 0 };
for (ht_cnt = (height >> 1); ht_cnt--;) {
src = LD_UB(src_ptr);
src_ptr += src_stride;
LD_UB2(ref_ptr, 16, ref0, ref1);
ref_ptr += ref_stride;
diff = __msa_asub_u_b(src, ref0);
sad0 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
diff = __msa_asub_u_b(src, ref);
sad1 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
diff = __msa_asub_u_b(src, ref);
sad2 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 3);
diff = __msa_asub_u_b(src, ref);
sad3 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 4);
diff = __msa_asub_u_b(src, ref);
sad4 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 5);
diff = __msa_asub_u_b(src, ref);
sad5 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 6);
diff = __msa_asub_u_b(src, ref);
sad6 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 7);
diff = __msa_asub_u_b(src, ref);
sad7 += __msa_hadd_u_h(diff, diff);
src = LD_UB(src_ptr);
src_ptr += src_stride;
LD_UB2(ref_ptr, 16, ref0, ref1);
ref_ptr += ref_stride;
diff = __msa_asub_u_b(src, ref0);
sad0 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
diff = __msa_asub_u_b(src, ref);
sad1 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
diff = __msa_asub_u_b(src, ref);
sad2 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 3);
diff = __msa_asub_u_b(src, ref);
sad3 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 4);
diff = __msa_asub_u_b(src, ref);
sad4 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 5);
diff = __msa_asub_u_b(src, ref);
sad5 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 6);
diff = __msa_asub_u_b(src, ref);
sad6 += __msa_hadd_u_h(diff, diff);
ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 7);
diff = __msa_asub_u_b(src, ref);
sad7 += __msa_hadd_u_h(diff, diff);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
sad_array[3] = HADD_UH_U32(sad3);
sad_array[4] = HADD_UH_U32(sad4);
sad_array[5] = HADD_UH_U32(sad5);
sad_array[6] = HADD_UH_U32(sad6);
sad_array[7] = HADD_UH_U32(sad7);
}
static void sad_32width_x8_msa(const uint8_t *src, int32_t src_stride,
const uint8_t *ref, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
int32_t ht_cnt;
v16u8 src0, src1;
v16u8 ref0, ref1, ref0_0, ref0_1, ref0_2;
v8u16 sad0 = { 0 };
v8u16 sad1 = { 0 };
v8u16 sad2 = { 0 };
v8u16 sad3 = { 0 };
v8u16 sad4 = { 0 };
v8u16 sad5 = { 0 };
v8u16 sad6 = { 0 };
v8u16 sad7 = { 0 };
for (ht_cnt = height; ht_cnt--;) {
LD_UB2(src, 16, src0, src1);
src += src_stride;
LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
ref += ref_stride;
sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 3);
sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 4);
sad4 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 5);
sad5 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 6);
sad6 += SAD_UB2_UH(src0, src1, ref0, ref1);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 7);
sad7 += SAD_UB2_UH(src0, src1, ref0, ref1);
}
sad_array[0] = HADD_UH_U32(sad0);
sad_array[1] = HADD_UH_U32(sad1);
sad_array[2] = HADD_UH_U32(sad2);
sad_array[3] = HADD_UH_U32(sad3);
sad_array[4] = HADD_UH_U32(sad4);
sad_array[5] = HADD_UH_U32(sad5);
sad_array[6] = HADD_UH_U32(sad6);
sad_array[7] = HADD_UH_U32(sad7);
}
static void sad_64width_x8_msa(const uint8_t *src, int32_t src_stride,
const uint8_t *ref, int32_t ref_stride,
int32_t height, uint32_t *sad_array) {
const uint8_t *src_dup, *ref_dup;
int32_t ht_cnt;
v16u8 src0, src1, src2, src3;
v16u8 ref0_0, ref0_1, ref0_2, ref0_3, ref0_4;
v16u8 ref0, ref1, ref2, ref3;
v8u16 sad0_0 = { 0 };
v8u16 sad0_1 = { 0 };
v8u16 sad1_0 = { 0 };
v8u16 sad1_1 = { 0 };
v8u16 sad2_0 = { 0 };
v8u16 sad2_1 = { 0 };
v8u16 sad3_0 = { 0 };
v8u16 sad3_1 = { 0 };
v4u32 sad;
src_dup = src;
ref_dup = ref;
for (ht_cnt = height; ht_cnt--;) {
LD_UB4(src, 16, src0, src1, src2, src3);
src += src_stride;
LD_UB5(ref, 16, ref0_0, ref0_1, ref0_2, ref0_3, ref0_4);
ref += ref_stride;
sad0_0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
sad0_1 += SAD_UB2_UH(src2, src3, ref0_2, ref0_3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 1);
sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 2);
sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 3);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 3);
sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
}
sad = __msa_hadd_u_w(sad0_0, sad0_0);
sad += __msa_hadd_u_w(sad0_1, sad0_1);
sad_array[0] = HADD_SW_S32(sad);
sad = __msa_hadd_u_w(sad1_0, sad1_0);
sad += __msa_hadd_u_w(sad1_1, sad1_1);
sad_array[1] = HADD_SW_S32(sad);
sad = __msa_hadd_u_w(sad2_0, sad2_0);
sad += __msa_hadd_u_w(sad2_1, sad2_1);
sad_array[2] = HADD_SW_S32(sad);
sad = __msa_hadd_u_w(sad3_0, sad3_0);
sad += __msa_hadd_u_w(sad3_1, sad3_1);
sad_array[3] = HADD_SW_S32(sad);
sad0_0 = (v8u16)__msa_ldi_h(0);
sad0_1 = (v8u16)__msa_ldi_h(0);
sad1_0 = (v8u16)__msa_ldi_h(0);
sad1_1 = (v8u16)__msa_ldi_h(0);
sad2_0 = (v8u16)__msa_ldi_h(0);
sad2_1 = (v8u16)__msa_ldi_h(0);
sad3_0 = (v8u16)__msa_ldi_h(0);
sad3_1 = (v8u16)__msa_ldi_h(0);
for (ht_cnt = 64; ht_cnt--;) {
LD_UB4(src_dup, 16, src0, src1, src2, src3);
src_dup += src_stride;
LD_UB5(ref_dup, 16, ref0_0, ref0_1, ref0_2, ref0_3, ref0_4);
ref_dup += ref_stride;
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 4);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 4);
sad0_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad0_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 5);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 5);
sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 6);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 6);
sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 7);
SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 7);
sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
}
sad = __msa_hadd_u_w(sad0_0, sad0_0);
sad += __msa_hadd_u_w(sad0_1, sad0_1);
sad_array[4] = HADD_SW_S32(sad);
sad = __msa_hadd_u_w(sad1_0, sad1_0);
sad += __msa_hadd_u_w(sad1_1, sad1_1);
sad_array[5] = HADD_SW_S32(sad);
sad = __msa_hadd_u_w(sad2_0, sad2_0);
sad += __msa_hadd_u_w(sad2_1, sad2_1);
sad_array[6] = HADD_SW_S32(sad);
sad = __msa_hadd_u_w(sad3_0, sad3_0);
sad += __msa_hadd_u_w(sad3_1, sad3_1);
sad_array[7] = HADD_SW_S32(sad);
}
static void sad_4width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
const uint8_t * const aref_ptr[],
int32_t ref_stride,
@ -646,6 +1288,76 @@ uint32_t vpx_sad64x##height##_msa(const uint8_t *src, int32_t src_stride, \
return sad_64width_msa(src, src_stride, ref, ref_stride, height); \
}
#define VPX_SAD_4xHEIGHTx3_MSA(height) \
void vpx_sad4x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_4width_x3_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_8xHEIGHTx3_MSA(height) \
void vpx_sad8x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_8width_x3_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_16xHEIGHTx3_MSA(height) \
void vpx_sad16x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_16width_x3_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_32xHEIGHTx3_MSA(height) \
void vpx_sad32x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_32width_x3_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_64xHEIGHTx3_MSA(height) \
void vpx_sad64x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_64width_x3_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_4xHEIGHTx8_MSA(height) \
void vpx_sad4x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_4width_x8_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_8xHEIGHTx8_MSA(height) \
void vpx_sad8x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_8width_x8_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_16xHEIGHTx8_MSA(height) \
void vpx_sad16x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_16width_x8_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_32xHEIGHTx8_MSA(height) \
void vpx_sad32x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_32width_x8_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_64xHEIGHTx8_MSA(height) \
void vpx_sad64x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *ref, int32_t ref_stride, \
uint32_t *sads) { \
sad_64width_x8_msa(src, src_stride, ref, ref_stride, height, sads); \
}
#define VPX_SAD_4xHEIGHTx4D_MSA(height) \
void vpx_sad4x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \
const uint8_t *const refs[], \
@ -723,65 +1435,91 @@ uint32_t vpx_sad64x##height##_avg_msa(const uint8_t *src, int32_t src_stride, \
// 64x64
VPX_SAD_64xHEIGHT_MSA(64);
VPX_SAD_64xHEIGHTx3_MSA(64);
VPX_SAD_64xHEIGHTx8_MSA(64);
VPX_SAD_64xHEIGHTx4D_MSA(64);
VPX_AVGSAD_64xHEIGHT_MSA(64);
// 64x32
VPX_SAD_64xHEIGHT_MSA(32);
VPX_SAD_64xHEIGHTx3_MSA(32);
VPX_SAD_64xHEIGHTx8_MSA(32);
VPX_SAD_64xHEIGHTx4D_MSA(32);
VPX_AVGSAD_64xHEIGHT_MSA(32);
// 32x64
VPX_SAD_32xHEIGHT_MSA(64);
VPX_SAD_32xHEIGHTx3_MSA(64);
VPX_SAD_32xHEIGHTx8_MSA(64);
VPX_SAD_32xHEIGHTx4D_MSA(64);
VPX_AVGSAD_32xHEIGHT_MSA(64);
// 32x32
VPX_SAD_32xHEIGHT_MSA(32);
VPX_SAD_32xHEIGHTx3_MSA(32);
VPX_SAD_32xHEIGHTx8_MSA(32);
VPX_SAD_32xHEIGHTx4D_MSA(32);
VPX_AVGSAD_32xHEIGHT_MSA(32);
// 32x16
VPX_SAD_32xHEIGHT_MSA(16);
VPX_SAD_32xHEIGHTx3_MSA(16);
VPX_SAD_32xHEIGHTx8_MSA(16);
VPX_SAD_32xHEIGHTx4D_MSA(16);
VPX_AVGSAD_32xHEIGHT_MSA(16);
// 16x32
VPX_SAD_16xHEIGHT_MSA(32);
VPX_SAD_16xHEIGHTx3_MSA(32);
VPX_SAD_16xHEIGHTx8_MSA(32);
VPX_SAD_16xHEIGHTx4D_MSA(32);
VPX_AVGSAD_16xHEIGHT_MSA(32);
// 16x16
VPX_SAD_16xHEIGHT_MSA(16);
VPX_SAD_16xHEIGHTx3_MSA(16);
VPX_SAD_16xHEIGHTx8_MSA(16);
VPX_SAD_16xHEIGHTx4D_MSA(16);
VPX_AVGSAD_16xHEIGHT_MSA(16);
// 16x8
VPX_SAD_16xHEIGHT_MSA(8);
VPX_SAD_16xHEIGHTx3_MSA(8);
VPX_SAD_16xHEIGHTx8_MSA(8);
VPX_SAD_16xHEIGHTx4D_MSA(8);
VPX_AVGSAD_16xHEIGHT_MSA(8);
// 8x16
VPX_SAD_8xHEIGHT_MSA(16);
VPX_SAD_8xHEIGHTx3_MSA(16);
VPX_SAD_8xHEIGHTx8_MSA(16);
VPX_SAD_8xHEIGHTx4D_MSA(16);
VPX_AVGSAD_8xHEIGHT_MSA(16);
// 8x8
VPX_SAD_8xHEIGHT_MSA(8);
VPX_SAD_8xHEIGHTx3_MSA(8);
VPX_SAD_8xHEIGHTx8_MSA(8);
VPX_SAD_8xHEIGHTx4D_MSA(8);
VPX_AVGSAD_8xHEIGHT_MSA(8);
// 8x4
VPX_SAD_8xHEIGHT_MSA(4);
VPX_SAD_8xHEIGHTx3_MSA(4);
VPX_SAD_8xHEIGHTx8_MSA(4);
VPX_SAD_8xHEIGHTx4D_MSA(4);
VPX_AVGSAD_8xHEIGHT_MSA(4);
// 4x8
VPX_SAD_4xHEIGHT_MSA(8);
VPX_SAD_4xHEIGHTx3_MSA(8);
VPX_SAD_4xHEIGHTx8_MSA(8);
VPX_SAD_4xHEIGHTx4D_MSA(8);
VPX_AVGSAD_4xHEIGHT_MSA(8);
// 4x4
VPX_SAD_4xHEIGHT_MSA(4);
VPX_SAD_4xHEIGHTx3_MSA(4);
VPX_SAD_4xHEIGHTx8_MSA(4);
VPX_SAD_4xHEIGHTx4D_MSA(4);
VPX_AVGSAD_4xHEIGHT_MSA(4);

26
vpx_dsp/vpx_dsp_rtcd_defs.pl
Просмотреть файл

@ -125,47 +125,53 @@ specialize qw/vpx_sad4x4_avg msa/, "$sse_x86inc";
#
# Blocks of 3
add_proto qw/void vpx_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad64x64x3 msa/;
add_proto qw/void vpx_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad32x32x3 msa/;
add_proto qw/void vpx_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad16x16x3 sse3 ssse3/;
specialize qw/vpx_sad16x16x3 sse3 ssse3 msa/;
add_proto qw/void vpx_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad16x8x3 sse3 ssse3/;
specialize qw/vpx_sad16x8x3 sse3 ssse3 msa/;
add_proto qw/void vpx_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad8x16x3 sse3/;
specialize qw/vpx_sad8x16x3 sse3 msa/;
add_proto qw/void vpx_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad8x8x3 sse3/;
specialize qw/vpx_sad8x8x3 sse3 msa/;
add_proto qw/void vpx_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad4x4x3 sse3/;
specialize qw/vpx_sad4x4x3 sse3 msa/;
# Blocks of 8
add_proto qw/void vpx_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad64x64x8 msa/;
add_proto qw/void vpx_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad32x32x8 msa/;
add_proto qw/void vpx_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad16x16x8 sse4_1/;
specialize qw/vpx_sad16x16x8 sse4_1 msa/;
add_proto qw/void vpx_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad16x8x8 sse4_1/;
specialize qw/vpx_sad16x8x8 sse4_1 msa/;
add_proto qw/void vpx_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad8x16x8 sse4_1/;
specialize qw/vpx_sad8x16x8 sse4_1 msa/;
add_proto qw/void vpx_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad8x8x8 sse4_1/;
specialize qw/vpx_sad8x8x8 sse4_1 msa/;
add_proto qw/void vpx_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad8x4x8 msa/;
add_proto qw/void vpx_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad4x8x8 msa/;
add_proto qw/void vpx_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
specialize qw/vpx_sad4x4x8 sse4_1/;
specialize qw/vpx_sad4x4x8 sse4_1 msa/;
#
# Multi-block SAD, comparing a reference to N independent blocks