diff options
author | Anupam Pandey <anupam.pandey@ittiam.com> | 2023-04-04 17:24:27 +0530 |
---|---|---|
committer | Anupam Pandey <anupam.pandey@ittiam.com> | 2023-04-17 15:23:51 +0530 |
commit | e15c2e34451b4177e4119cce47bf73dac9864de8 (patch) | |
tree | 3ea54e0fbb82c6985f4877bc0fdc62d75936df3f /vpx_dsp/x86/fwd_txfm_avx2.c | |
parent | 0f42bd3fb81cd79dfaef5c6aef26c643c48be909 (diff) | |
download | libvpx-e15c2e34451b4177e4119cce47bf73dac9864de8.tar libvpx-e15c2e34451b4177e4119cce47bf73dac9864de8.tar.gz libvpx-e15c2e34451b4177e4119cce47bf73dac9864de8.tar.bz2 libvpx-e15c2e34451b4177e4119cce47bf73dac9864de8.zip |
Add AVX2 intrinsic for vpx_fdct16x16() function
Introduced AVX2 intrinsic to compute FDCT for block size
16x16 case. This is a bit-exact change.
Please check the module level scaling w.r.t C function (timer based)
for existing (SSE2) and new AVX2 intrinsics:
Scaling
SSE2 AVX2
3.88x 5.95x
Change-Id: I02299c3746fcb52d808e2a75d30aa62652c816dc
Diffstat (limited to 'vpx_dsp/x86/fwd_txfm_avx2.c')
-rw-r--r-- | vpx_dsp/x86/fwd_txfm_avx2.c | 373 |
1 files changed, 373 insertions, 0 deletions
diff --git a/vpx_dsp/x86/fwd_txfm_avx2.c b/vpx_dsp/x86/fwd_txfm_avx2.c index a2ed420e3..c8f54a49c 100644 --- a/vpx_dsp/x86/fwd_txfm_avx2.c +++ b/vpx_dsp/x86/fwd_txfm_avx2.c @@ -8,9 +8,382 @@ * be found in the AUTHORS file in the root of the source tree. */ +#include <immintrin.h> // AVX2 #include "./vpx_config.h" #include "./vpx_dsp_rtcd.h" +#include "vpx_dsp/txfm_common.h" +#define ADD256_EPI16 _mm256_add_epi16 +#define SUB256_EPI16 _mm256_sub_epi16 + +static INLINE void load_buffer_16bit_to_16bit_avx2(const int16_t *in, + int stride, __m256i *out, + int out_size, int pass) { + int i; + const __m256i kOne = _mm256_set1_epi16(1); + if (pass == 0) { + for (i = 0; i < out_size; i++) { + out[i] = _mm256_loadu_si256((const __m256i *)(in + i * stride)); + // x = x << 2 + out[i] = _mm256_slli_epi16(out[i], 2); + } + } else { + for (i = 0; i < out_size; i++) { + out[i] = _mm256_loadu_si256((const __m256i *)(in + i * 16)); + // x = (x + 1) >> 2 + out[i] = _mm256_add_epi16(out[i], kOne); + out[i] = _mm256_srai_epi16(out[i], 2); + } + } +} + +static INLINE void transpose2_8x8_avx2(const __m256i *const in, + __m256i *const out) { + int i; + __m256i t[16], u[16]; + // (1st, 2nd) ==> (lo, hi) + // (0, 1) ==> (0, 1) + // (2, 3) ==> (2, 3) + // (4, 5) ==> (4, 5) + // (6, 7) ==> (6, 7) + for (i = 0; i < 4; i++) { + t[2 * i] = _mm256_unpacklo_epi16(in[2 * i], in[2 * i + 1]); + t[2 * i + 1] = _mm256_unpackhi_epi16(in[2 * i], in[2 * i + 1]); + } + + // (1st, 2nd) ==> (lo, hi) + // (0, 2) ==> (0, 2) + // (1, 3) ==> (1, 3) + // (4, 6) ==> (4, 6) + // (5, 7) ==> (5, 7) + for (i = 0; i < 2; i++) { + u[i] = _mm256_unpacklo_epi32(t[i], t[i + 2]); + u[i + 2] = _mm256_unpackhi_epi32(t[i], t[i + 2]); + + u[i + 4] = _mm256_unpacklo_epi32(t[i + 4], t[i + 6]); + u[i + 6] = _mm256_unpackhi_epi32(t[i + 4], t[i + 6]); + } + + // (1st, 2nd) ==> (lo, hi) + // (0, 4) ==> (0, 1) + // (1, 5) ==> (4, 5) + // (2, 6) ==> (2, 3) + // (3, 7) ==> (6, 7) + for (i = 0; i < 2; i++) { + out[2 * i] = _mm256_unpacklo_epi64(u[2 * i], u[2 * i + 4]); + out[2 * i + 1] = _mm256_unpackhi_epi64(u[2 * i], u[2 * i + 4]); + + out[2 * i + 4] = _mm256_unpacklo_epi64(u[2 * i + 1], u[2 * i + 5]); + out[2 * i + 5] = _mm256_unpackhi_epi64(u[2 * i + 1], u[2 * i + 5]); + } +} + +static INLINE void transpose_16bit_16x16_avx2(const __m256i *const in, + __m256i *const out) { + __m256i t[16]; + +#define LOADL(idx) \ + t[idx] = _mm256_castsi128_si256(_mm_load_si128((__m128i const *)&in[idx])); \ + t[idx] = _mm256_inserti128_si256( \ + t[idx], _mm_load_si128((__m128i const *)&in[idx + 8]), 1); + +#define LOADR(idx) \ + t[8 + idx] = \ + _mm256_castsi128_si256(_mm_load_si128((__m128i const *)&in[idx] + 1)); \ + t[8 + idx] = _mm256_inserti128_si256( \ + t[8 + idx], _mm_load_si128((__m128i const *)&in[idx + 8] + 1), 1); + + // load left 8x16 + LOADL(0) + LOADL(1) + LOADL(2) + LOADL(3) + LOADL(4) + LOADL(5) + LOADL(6) + LOADL(7) + + // load right 8x16 + LOADR(0) + LOADR(1) + LOADR(2) + LOADR(3) + LOADR(4) + LOADR(5) + LOADR(6) + LOADR(7) + + // get the top 16x8 result + transpose2_8x8_avx2(t, out); + // get the bottom 16x8 result + transpose2_8x8_avx2(&t[8], &out[8]); +} + +// Store 8 16-bit values. Sign extend the values. +static INLINE void store_buffer_16bit_to_32bit_w16_avx2(const __m256i *const in, + tran_low_t *out, + const int stride, + const int out_size) { + int i; + for (i = 0; i < out_size; ++i) { + _mm256_storeu_si256((__m256i *)(out), in[i]); + out += stride; + } +} + +#define PAIR256_SET_EPI16(a, b) \ + _mm256_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \ + (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \ + (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \ + (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a)) + +static INLINE __m256i mult256_round_shift(const __m256i *pin0, + const __m256i *pin1, + const __m256i *pmultiplier, + const __m256i *prounding, + const int shift) { + const __m256i u0 = _mm256_madd_epi16(*pin0, *pmultiplier); + const __m256i u1 = _mm256_madd_epi16(*pin1, *pmultiplier); + const __m256i v0 = _mm256_add_epi32(u0, *prounding); + const __m256i v1 = _mm256_add_epi32(u1, *prounding); + const __m256i w0 = _mm256_srai_epi32(v0, shift); + const __m256i w1 = _mm256_srai_epi32(v1, shift); + return _mm256_packs_epi32(w0, w1); +} + +static INLINE void fdct16x16_1D_avx2(__m256i *input, __m256i *output) { + int i; + __m256i step2[4]; + __m256i in[8]; + __m256i step1[8]; + __m256i step3[8]; + + const __m256i k__cospi_p16_p16 = _mm256_set1_epi16(cospi_16_64); + const __m256i k__cospi_p16_m16 = PAIR256_SET_EPI16(cospi_16_64, -cospi_16_64); + const __m256i k__cospi_p24_p08 = PAIR256_SET_EPI16(cospi_24_64, cospi_8_64); + const __m256i k__cospi_p08_m24 = PAIR256_SET_EPI16(cospi_8_64, -cospi_24_64); + const __m256i k__cospi_m08_p24 = PAIR256_SET_EPI16(-cospi_8_64, cospi_24_64); + const __m256i k__cospi_p28_p04 = PAIR256_SET_EPI16(cospi_28_64, cospi_4_64); + const __m256i k__cospi_m04_p28 = PAIR256_SET_EPI16(-cospi_4_64, cospi_28_64); + const __m256i k__cospi_p12_p20 = PAIR256_SET_EPI16(cospi_12_64, cospi_20_64); + const __m256i k__cospi_m20_p12 = PAIR256_SET_EPI16(-cospi_20_64, cospi_12_64); + const __m256i k__cospi_p30_p02 = PAIR256_SET_EPI16(cospi_30_64, cospi_2_64); + const __m256i k__cospi_p14_p18 = PAIR256_SET_EPI16(cospi_14_64, cospi_18_64); + const __m256i k__cospi_m02_p30 = PAIR256_SET_EPI16(-cospi_2_64, cospi_30_64); + const __m256i k__cospi_m18_p14 = PAIR256_SET_EPI16(-cospi_18_64, cospi_14_64); + const __m256i k__cospi_p22_p10 = PAIR256_SET_EPI16(cospi_22_64, cospi_10_64); + const __m256i k__cospi_p06_p26 = PAIR256_SET_EPI16(cospi_6_64, cospi_26_64); + const __m256i k__cospi_m10_p22 = PAIR256_SET_EPI16(-cospi_10_64, cospi_22_64); + const __m256i k__cospi_m26_p06 = PAIR256_SET_EPI16(-cospi_26_64, cospi_6_64); + const __m256i k__DCT_CONST_ROUNDING = _mm256_set1_epi32(DCT_CONST_ROUNDING); + + // Calculate input for the first 8 results. + for (i = 0; i < 8; i++) { + in[i] = ADD256_EPI16(input[i], input[15 - i]); + } + + // Calculate input for the next 8 results. + for (i = 0; i < 8; i++) { + step1[i] = SUB256_EPI16(input[7 - i], input[8 + i]); + } + + // Work on the first eight values; fdct8(input, even_results); + { + // Add/subtract + const __m256i q0 = ADD256_EPI16(in[0], in[7]); + const __m256i q1 = ADD256_EPI16(in[1], in[6]); + const __m256i q2 = ADD256_EPI16(in[2], in[5]); + const __m256i q3 = ADD256_EPI16(in[3], in[4]); + const __m256i q4 = SUB256_EPI16(in[3], in[4]); + const __m256i q5 = SUB256_EPI16(in[2], in[5]); + const __m256i q6 = SUB256_EPI16(in[1], in[6]); + const __m256i q7 = SUB256_EPI16(in[0], in[7]); + + // Work on first four results + { + // Add/subtract + const __m256i r0 = ADD256_EPI16(q0, q3); + const __m256i r1 = ADD256_EPI16(q1, q2); + const __m256i r2 = SUB256_EPI16(q1, q2); + const __m256i r3 = SUB256_EPI16(q0, q3); + + // Interleave to do the multiply by constants which gets us + // into 32 bits. + { + const __m256i t0 = _mm256_unpacklo_epi16(r0, r1); + const __m256i t1 = _mm256_unpackhi_epi16(r0, r1); + const __m256i t2 = _mm256_unpacklo_epi16(r2, r3); + const __m256i t3 = _mm256_unpackhi_epi16(r2, r3); + + output[0] = mult256_round_shift(&t0, &t1, &k__cospi_p16_p16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[8] = mult256_round_shift(&t0, &t1, &k__cospi_p16_m16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[4] = mult256_round_shift(&t2, &t3, &k__cospi_p24_p08, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[12] = + mult256_round_shift(&t2, &t3, &k__cospi_m08_p24, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + } + } + + // Work on next four results + { + // Interleave to do the multiply by constants which gets us + // into 32 bits. + const __m256i d0 = _mm256_unpacklo_epi16(q6, q5); + const __m256i d1 = _mm256_unpackhi_epi16(q6, q5); + const __m256i r0 = mult256_round_shift( + &d0, &d1, &k__cospi_p16_m16, &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + const __m256i r1 = mult256_round_shift( + &d0, &d1, &k__cospi_p16_p16, &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + + { + // Add/subtract + const __m256i x0 = ADD256_EPI16(q4, r0); + const __m256i x1 = SUB256_EPI16(q4, r0); + const __m256i x2 = SUB256_EPI16(q7, r1); + const __m256i x3 = ADD256_EPI16(q7, r1); + + // Interleave to do the multiply by constants which gets us + // into 32 bits. + { + const __m256i t0 = _mm256_unpacklo_epi16(x0, x3); + const __m256i t1 = _mm256_unpackhi_epi16(x0, x3); + const __m256i t2 = _mm256_unpacklo_epi16(x1, x2); + const __m256i t3 = _mm256_unpackhi_epi16(x1, x2); + output[2] = + mult256_round_shift(&t0, &t1, &k__cospi_p28_p04, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[14] = + mult256_round_shift(&t0, &t1, &k__cospi_m04_p28, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[10] = + mult256_round_shift(&t2, &t3, &k__cospi_p12_p20, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[6] = + mult256_round_shift(&t2, &t3, &k__cospi_m20_p12, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + } + } + } + } + // Work on the next eight values; step1 -> odd_results + { // step 2 + { + const __m256i t0 = _mm256_unpacklo_epi16(step1[5], step1[2]); + const __m256i t1 = _mm256_unpackhi_epi16(step1[5], step1[2]); + const __m256i t2 = _mm256_unpacklo_epi16(step1[4], step1[3]); + const __m256i t3 = _mm256_unpackhi_epi16(step1[4], step1[3]); + step2[0] = mult256_round_shift(&t0, &t1, &k__cospi_p16_m16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2[1] = mult256_round_shift(&t2, &t3, &k__cospi_p16_m16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2[2] = mult256_round_shift(&t0, &t1, &k__cospi_p16_p16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2[3] = mult256_round_shift(&t2, &t3, &k__cospi_p16_p16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + } + // step 3 + { + step3[0] = ADD256_EPI16(step1[0], step2[1]); + step3[1] = ADD256_EPI16(step1[1], step2[0]); + step3[2] = SUB256_EPI16(step1[1], step2[0]); + step3[3] = SUB256_EPI16(step1[0], step2[1]); + step3[4] = SUB256_EPI16(step1[7], step2[3]); + step3[5] = SUB256_EPI16(step1[6], step2[2]); + step3[6] = ADD256_EPI16(step1[6], step2[2]); + step3[7] = ADD256_EPI16(step1[7], step2[3]); + } + // step 4 + { + const __m256i t0 = _mm256_unpacklo_epi16(step3[1], step3[6]); + const __m256i t1 = _mm256_unpackhi_epi16(step3[1], step3[6]); + const __m256i t2 = _mm256_unpacklo_epi16(step3[2], step3[5]); + const __m256i t3 = _mm256_unpackhi_epi16(step3[2], step3[5]); + step2[0] = mult256_round_shift(&t0, &t1, &k__cospi_m08_p24, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2[1] = mult256_round_shift(&t2, &t3, &k__cospi_p24_p08, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2[2] = mult256_round_shift(&t0, &t1, &k__cospi_p24_p08, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2[3] = mult256_round_shift(&t2, &t3, &k__cospi_p08_m24, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + } + // step 5 + { + step1[0] = ADD256_EPI16(step3[0], step2[0]); + step1[1] = SUB256_EPI16(step3[0], step2[0]); + step1[2] = ADD256_EPI16(step3[3], step2[1]); + step1[3] = SUB256_EPI16(step3[3], step2[1]); + step1[4] = SUB256_EPI16(step3[4], step2[3]); + step1[5] = ADD256_EPI16(step3[4], step2[3]); + step1[6] = SUB256_EPI16(step3[7], step2[2]); + step1[7] = ADD256_EPI16(step3[7], step2[2]); + } + // step 6 + { + const __m256i t0 = _mm256_unpacklo_epi16(step1[0], step1[7]); + const __m256i t1 = _mm256_unpackhi_epi16(step1[0], step1[7]); + const __m256i t2 = _mm256_unpacklo_epi16(step1[1], step1[6]); + const __m256i t3 = _mm256_unpackhi_epi16(step1[1], step1[6]); + output[1] = mult256_round_shift(&t0, &t1, &k__cospi_p30_p02, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[9] = mult256_round_shift(&t2, &t3, &k__cospi_p14_p18, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[15] = mult256_round_shift(&t0, &t1, &k__cospi_m02_p30, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[7] = mult256_round_shift(&t2, &t3, &k__cospi_m18_p14, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + } + { + const __m256i t0 = _mm256_unpacklo_epi16(step1[2], step1[5]); + const __m256i t1 = _mm256_unpackhi_epi16(step1[2], step1[5]); + const __m256i t2 = _mm256_unpacklo_epi16(step1[3], step1[4]); + const __m256i t3 = _mm256_unpackhi_epi16(step1[3], step1[4]); + output[5] = mult256_round_shift(&t0, &t1, &k__cospi_p22_p10, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[13] = mult256_round_shift(&t2, &t3, &k__cospi_p06_p26, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[11] = mult256_round_shift(&t0, &t1, &k__cospi_m10_p22, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + output[3] = mult256_round_shift(&t2, &t3, &k__cospi_m26_p06, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + } + } +} + +void vpx_fdct16x16_avx2(const int16_t *input, tran_low_t *output, int stride) { + int pass; + DECLARE_ALIGNED(32, int16_t, intermediate[256]); + int16_t *out0 = intermediate; + tran_low_t *out1 = output; + const int width = 16; + const int height = 16; + __m256i buf0[16], buf1[16]; + + // Two transform and transpose passes + // Process 16 columns (transposed rows in second pass) at a time. + for (pass = 0; pass < 2; ++pass) { + // Load and pre-condition input. + load_buffer_16bit_to_16bit_avx2(input, stride, buf1, height, pass); + + // Calculate dct for 16x16 values + fdct16x16_1D_avx2(buf1, buf0); + + // Transpose the results. + transpose_16bit_16x16_avx2(buf0, buf1); + + if (pass == 0) { + store_buffer_16bit_to_32bit_w16_avx2(buf1, out0, width, height); + } else { + store_buffer_16bit_to_32bit_w16_avx2(buf1, out1, width, height); + } + // Setup in/out for next pass. + input = intermediate; + } +} + #if !CONFIG_VP9_HIGHBITDEPTH #define FDCT32x32_2D_AVX2 vpx_fdct32x32_rd_avx2 #define FDCT32x32_HIGH_PRECISION 0 |