diff options
Diffstat (limited to 'third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c')
-rw-r--r-- | third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c | 514 |
1 files changed, 297 insertions, 217 deletions
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c index 49c20b467..9d6b4c2f7 100644 --- a/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c +++ b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c @@ -20,6 +20,7 @@ #include "aom_dsp/x86/synonyms.h" #include "aom_dsp/x86/blend_sse4.h" +#include "aom_dsp/x86/blend_mask_sse4.h" #include "config/aom_dsp_rtcd.h" @@ -32,19 +33,13 @@ static void blend_a64_mask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - (void)w; - + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { const __m128i v_m0_b = xx_loadl_32(mask); - const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); - - const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); - - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); - + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_32(dst, v_res_b); dst += dst_stride; @@ -59,19 +54,13 @@ static void blend_a64_mask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - (void)w; - + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { const __m128i v_m0_b = xx_loadl_64(mask); - const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); - - const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); - - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); - + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_64(dst, v_res_b); dst += dst_stride; @@ -85,23 +74,17 @@ static void blend_a64_mask_w16n_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { int c; for (c = 0; c < w; c += 16) { - const __m128i v_m0l_b = xx_loadl_64(mask + c); - const __m128i v_m0h_b = xx_loadl_64(mask + c + 8); - const __m128i v_m0l_w = _mm_cvtepu8_epi16(v_m0l_b); - const __m128i v_m0h_w = _mm_cvtepu8_epi16(v_m0h_b); - const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); - const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); - - const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); - const __m128i v_resh_w = - blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); + const __m128i v_m0_b = xx_loadu_128(mask + c); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); - const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); xx_storeu_128(dst + c, v_res_b); } @@ -120,23 +103,20 @@ static void blend_a64_mask_sx_w4_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, - 0xff, 0, 0xff, 0, 0xff, 0, 0xff); - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - (void)w; + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { const __m128i v_r_b = xx_loadl_64(mask); - const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); - - const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); - - const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); - - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_32(dst, v_res_b); dst += dst_stride; @@ -150,22 +130,20 @@ static void blend_a64_mask_sx_w8_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, - 0xff, 0, 0xff, 0, 0xff, 0, 0xff); - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - (void)w; + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { const __m128i v_r_b = xx_loadu_128(mask); - const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); - - const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); - - const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_64(dst, v_res_b); @@ -180,28 +158,24 @@ static void blend_a64_mask_sx_w16n_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, - 0xff, 0, 0xff, 0, 0xff, 0, 0xff); - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { int c; for (c = 0; c < w; c += 16) { - const __m128i v_rl_b = xx_loadu_128(mask + 2 * c); - const __m128i v_rh_b = xx_loadu_128(mask + 2 * c + 16); - const __m128i v_al_b = _mm_avg_epu8(v_rl_b, _mm_srli_si128(v_rl_b, 1)); - const __m128i v_ah_b = _mm_avg_epu8(v_rh_b, _mm_srli_si128(v_rh_b, 1)); - - const __m128i v_m0l_w = _mm_and_si128(v_al_b, v_zmask_b); - const __m128i v_m0h_w = _mm_and_si128(v_ah_b, v_zmask_b); - const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); - const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); - - const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); - const __m128i v_resh_w = - blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); - - const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + const __m128i v_r0_b = xx_loadu_128(mask + 2 * c); + const __m128i v_r1_b = xx_loadu_128(mask + 2 * c + 16); + const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r0_b, v_shuffle_b); + const __m128i v_r1_s_b = _mm_shuffle_epi8(v_r1_b, v_shuffle_b); + const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r1_s_b); + const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r1_s_b); + const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); xx_storeu_128(dst + c, v_res_b); } @@ -220,21 +194,18 @@ static void blend_a64_mask_sy_w4_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - (void)w; + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); + do { const __m128i v_ra_b = xx_loadl_32(mask); const __m128i v_rb_b = xx_loadl_32(mask + mask_stride); - const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); - const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); - - const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); - - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_32(dst, v_res_b); @@ -249,21 +220,16 @@ static void blend_a64_mask_sy_w8_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - (void)w; + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { const __m128i v_ra_b = xx_loadl_64(mask); const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); - const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); - - const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); - - const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); - - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_64(dst, v_res_b); @@ -278,26 +244,18 @@ static void blend_a64_mask_sy_w16n_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_zero = _mm_setzero_si128(); - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { int c; for (c = 0; c < w; c += 16) { const __m128i v_ra_b = xx_loadu_128(mask + c); const __m128i v_rb_b = xx_loadu_128(mask + c + mask_stride); - const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); - const __m128i v_m0l_w = _mm_cvtepu8_epi16(v_a_b); - const __m128i v_m0h_w = _mm_unpackhi_epi8(v_a_b, v_zero); - const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); - const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); - - const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); - const __m128i v_resh_w = - blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); - - const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); xx_storeu_128(dst + c, v_res_b); } @@ -316,27 +274,24 @@ static void blend_a64_mask_sx_sy_w4_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, - 0xff, 0, 0xff, 0, 0xff, 0, 0xff); - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); (void)w; do { const __m128i v_ra_b = xx_loadl_64(mask); const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); - const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); - const __m128i v_rvsb_w = - _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); - const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); - + const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b); + const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b); + const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8)); + const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w); const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); - - const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); + const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_32(dst, v_res_b); @@ -351,27 +306,25 @@ static void blend_a64_mask_sx_sy_w8_sse4_1( uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h) { - const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, - 0xff, 0, 0xff, 0, 0xff, 0, 0xff); - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - + const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle); + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); (void)w; do { const __m128i v_ra_b = xx_loadu_128(mask); const __m128i v_rb_b = xx_loadu_128(mask + mask_stride); - const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); - const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); - const __m128i v_rvsb_w = - _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); - const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b); + const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b); + const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8)); + const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w); const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); - const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); - const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); - - const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r); xx_storel_64(dst, v_res_b); @@ -388,8 +341,8 @@ static void blend_a64_mask_sx_sy_w16n_sse4_1( const uint8_t *mask, uint32_t mask_stride, int w, int h) { const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff); - const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - + const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA); + const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS)); do { int c; for (c = 0; c < w; c += 16) { @@ -410,14 +363,11 @@ static void blend_a64_mask_sx_sy_w16n_sse4_1( const __m128i v_m0l_w = xx_roundn_epu16(v_rsl_w, 2); const __m128i v_m0h_w = xx_roundn_epu16(v_rsh_w, 2); - const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); - const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); - - const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); - const __m128i v_resh_w = - blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); + const __m128i v_m0_b = _mm_packus_epi16(v_m0l_w, v_m0h_w); + const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b); - const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + const __m128i v_res_b = + blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r); xx_storeu_128(dst + c, v_res_b); } @@ -921,24 +871,140 @@ void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst_8, uint32_t dst_stride, } } -static INLINE void blend_a64_d16_mask(uint8_t *dst, const CONV_BUF_TYPE *src0, - const CONV_BUF_TYPE *src1, - const __m128i *m, - const __m128i *v_round_offset, - const __m128i *v_maxval, int round_bits) { - const __m128i max_minus_m = _mm_sub_epi16(*v_maxval, *m); - const __m128i s0 = xx_loadl_64(src0); - const __m128i s1 = xx_loadl_64(src1); - const __m128i s0_s1 = _mm_unpacklo_epi16(s0, s1); - const __m128i m_max_minus_m = _mm_unpacklo_epi16(*m, max_minus_m); - const __m128i res_a = _mm_madd_epi16(s0_s1, m_max_minus_m); - const __m128i res_b = _mm_srli_epi32(res_a, AOM_BLEND_A64_ROUND_BITS); - const __m128i res_c = _mm_sub_epi32(res_b, *v_round_offset); - const __m128i res_d = xx_roundn_epi32(res_c, round_bits); - const __m128i res_e = _mm_packs_epi32(res_d, res_d); - const __m128i res = _mm_packus_epi16(res_e, res_e); - - xx_storel_32(dst, res); +static INLINE void blend_a64_d16_mask_w16_sse41( + uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1, + const __m128i *m0, const __m128i *m1, const __m128i *v_round_offset, + const __m128i *v_maxval, int shift) { + const __m128i max_minus_m0 = _mm_sub_epi16(*v_maxval, *m0); + const __m128i max_minus_m1 = _mm_sub_epi16(*v_maxval, *m1); + const __m128i s0_0 = xx_loadu_128(src0); + const __m128i s0_1 = xx_loadu_128(src0 + 8); + const __m128i s1_0 = xx_loadu_128(src1); + const __m128i s1_1 = xx_loadu_128(src1 + 8); + __m128i res0_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0_0, s1_0), + _mm_unpacklo_epi16(*m0, max_minus_m0)); + __m128i res0_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0_0, s1_0), + _mm_unpackhi_epi16(*m0, max_minus_m0)); + __m128i res1_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0_1, s1_1), + _mm_unpacklo_epi16(*m1, max_minus_m1)); + __m128i res1_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0_1, s1_1), + _mm_unpackhi_epi16(*m1, max_minus_m1)); + res0_lo = _mm_srai_epi32(_mm_sub_epi32(res0_lo, *v_round_offset), shift); + res0_hi = _mm_srai_epi32(_mm_sub_epi32(res0_hi, *v_round_offset), shift); + res1_lo = _mm_srai_epi32(_mm_sub_epi32(res1_lo, *v_round_offset), shift); + res1_hi = _mm_srai_epi32(_mm_sub_epi32(res1_hi, *v_round_offset), shift); + const __m128i res0 = _mm_packs_epi32(res0_lo, res0_hi); + const __m128i res1 = _mm_packs_epi32(res1_lo, res1_hi); + const __m128i res = _mm_packus_epi16(res0, res1); + + _mm_storeu_si128((__m128i *)(dst), res); +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m = xx_loadu_128(mask + j); + const __m128i m0 = _mm_cvtepu8_epi16(m); + const __m128i m1 = _mm_cvtepu8_epi16(_mm_srli_si128(m, 8)); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i two_w = _mm_set1_epi16(2); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m_i00 = xx_loadu_128(mask + 2 * j); + const __m128i m_i01 = xx_loadu_128(mask + 2 * j + 16); + const __m128i m_i10 = xx_loadu_128(mask + mask_stride + 2 * j); + const __m128i m_i11 = xx_loadu_128(mask + mask_stride + 2 * j + 16); + + const __m128i m0_ac = _mm_adds_epu8(m_i00, m_i10); + const __m128i m1_ac = _mm_adds_epu8(m_i01, m_i11); + const __m128i m0_acbd = _mm_maddubs_epi16(m0_ac, one_b); + const __m128i m1_acbd = _mm_maddubs_epi16(m1_ac, one_b); + const __m128i m0 = _mm_srli_epi16(_mm_add_epi16(m0_acbd, two_w), 2); + const __m128i m1 = _mm_srli_epi16(_mm_add_epi16(m1_acbd, two_w), 2); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw1_subh0_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i one_b = _mm_set1_epi8(1); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m_i00 = xx_loadu_128(mask + 2 * j); + const __m128i m_i01 = xx_loadu_128(mask + 2 * j + 16); + const __m128i m0_ac = _mm_maddubs_epi16(m_i00, one_b); + const __m128i m1_ac = _mm_maddubs_epi16(m_i01, one_b); + const __m128i m0 = _mm_avg_epu16(m0_ac, zeros); + const __m128i m1 = _mm_avg_epu16(m1_ac, zeros); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } +} + +static INLINE void lowbd_blend_a64_d16_mask_subw0_subh1_w16_sse4_1( + uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, + uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + const __m128i *round_offset, int shift) { + const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + const __m128i zeros = _mm_setzero_si128(); + for (int i = 0; i < h; ++i) { + for (int j = 0; j < w; j += 16) { + const __m128i m_i00 = xx_loadu_128(mask + j); + const __m128i m_i10 = xx_loadu_128(mask + mask_stride + j); + + const __m128i m_ac = _mm_avg_epu8(_mm_adds_epu8(m_i00, m_i10), zeros); + const __m128i m0 = _mm_cvtepu8_epi16(m_ac); + const __m128i m1 = _mm_cvtepu8_epi16(_mm_srli_si128(m_ac, 8)); + + blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1, + round_offset, &v_maxval, shift); + } + mask += mask_stride << 1; + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + } } void aom_lowbd_blend_a64_d16_mask_sse4_1( @@ -947,12 +1013,15 @@ void aom_lowbd_blend_a64_d16_mask_sse4_1( const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh, ConvolveParams *conv_params) { const int bd = 8; - const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0; - const int round_offset = (1 << (offset_bits - conv_params->round_1)) + - (1 << (offset_bits - conv_params->round_1 - 1)); const int round_bits = 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; + const int round_offset = + ((1 << (round_bits + bd)) + (1 << (round_bits + bd - 1)) - + (1 << (round_bits - 1))) + << AOM_BLEND_A64_ROUND_BITS; + + const int shift = round_bits + AOM_BLEND_A64_ROUND_BITS; assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride)); assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride)); @@ -961,69 +1030,80 @@ void aom_lowbd_blend_a64_d16_mask_sse4_1( assert(IS_POWER_OF_TWO(h)); assert(IS_POWER_OF_TWO(w)); - const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); - const __m128i v_ro_a = xx_loadl_32(&round_offset); - const __m128i v_round_offset = _mm_shuffle_epi32(v_ro_a, 0); - const __m128i one_w = _mm_set1_epi16(1); - const __m128i one_b = _mm_set1_epi8(1); - const __m128i two_w = _mm_set1_epi16(2); + const __m128i v_round_offset = _mm_set1_epi32(round_offset); if (subw == 0 && subh == 0) { - for (int i = 0; i < h; ++i) { - for (int j = 0; j < w; j += 4) { - const __m128i m0 = xx_loadl_32(&mask[i * mask_stride + j]); - const __m128i m = _mm_cvtepu8_epi16(m0); - - blend_a64_d16_mask(&dst[i * dst_stride + j], &src0[i * src0_stride + j], - &src1[i * src1_stride + j], &m, &v_round_offset, - &v_maxval, round_bits); - } + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; } + } else if (subw == 1 && subh == 1) { - for (int i = 0; i < h; ++i) { - for (int j = 0; j < w; j += 4) { - const __m128i m_i0 = - xx_loadl_64(&mask[(2 * i) * mask_stride + (2 * j)]); - const __m128i m_i1 = - xx_loadl_64(&mask[(2 * i + 1) * mask_stride + (2 * j)]); - const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b); - const __m128i m_bd = _mm_maddubs_epi16(m_i1, one_b); - const __m128i m_acbd = _mm_add_epi16(m_ac, m_bd); - const __m128i m_acbd_2 = _mm_add_epi16(m_acbd, two_w); - const __m128i m = _mm_srli_epi16(m_acbd_2, 2); - - blend_a64_d16_mask(&dst[i * dst_stride + j], &src0[i * src0_stride + j], - &src1[i * src1_stride + j], &m, &v_round_offset, - &v_maxval, round_bits); - } + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; } } else if (subw == 1 && subh == 0) { - for (int i = 0; i < h; ++i) { - for (int j = 0; j < w; j += 4) { - const __m128i m_i0 = xx_loadl_64(&mask[i * mask_stride + (2 * j)]); - const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b); - const __m128i m_ac_1 = _mm_add_epi16(m_ac, one_w); - const __m128i m = _mm_srli_epi16(m_ac_1, 1); - - blend_a64_d16_mask(&dst[i * dst_stride + j], &src0[i * src0_stride + j], - &src1[i * src1_stride + j], &m, &v_round_offset, - &v_maxval, round_bits); - } + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw1_subh0_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; } } else { - for (int i = 0; i < h; ++i) { - for (int j = 0; j < w; j += 4) { - const __m128i m_i0 = xx_loadl_64(&mask[(2 * i) * mask_stride + j]); - const __m128i m_i1 = xx_loadl_64(&mask[(2 * i + 1) * mask_stride + j]); - const __m128i m_i01 = _mm_unpacklo_epi8(m_i0, m_i1); - const __m128i m_ac = _mm_maddubs_epi16(m_i01, one_b); - const __m128i m_ac_1 = _mm_add_epi16(m_ac, one_w); - const __m128i m = _mm_srli_epi16(m_ac_1, 1); - - blend_a64_d16_mask(&dst[i * dst_stride + j], &src0[i * src0_stride + j], - &src1[i * src1_stride + j], &m, &v_round_offset, - &v_maxval, round_bits); - } + switch (w) { + case 4: + aom_lowbd_blend_a64_d16_mask_subw0_subh1_w4_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + case 8: + aom_lowbd_blend_a64_d16_mask_subw0_subh1_w8_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, &v_round_offset, shift); + break; + default: + lowbd_blend_a64_d16_mask_subw0_subh1_w16_sse4_1( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w, &v_round_offset, shift); + break; } } } |