From: Andrew Russell Date: Mon, 3 Mar 2014 15:38:02 +0000 (-0800) Subject: improved speed of 4x4 sse2 fdct. X-Git-Tag: v1.4.0~2154^2 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=a46f5459c3f5da8700e1ed34d4ed0fd4480e6ff3;p=platform%2Fupstream%2Flibvpx.git improved speed of 4x4 sse2 fdct. * speed improvment of 30 percent achieved * multiplies and adds remain the same * non-arithmetic instructions minimized by hand, by: -expanding 2 pass loop -removing irrelivant "shuffles" -combining last two rounding steps * further improvments may be possible Change-Id: Idec2c3f52910c48e6a0e0f9aefed5cae31b0b8c0 --- diff --git a/vp9/encoder/x86/vp9_dct_sse2.c b/vp9/encoder/x86/vp9_dct_sse2.c index f3735eb..6865822 100644 --- a/vp9/encoder/x86/vp9_dct_sse2.c +++ b/vp9/encoder/x86/vp9_dct_sse2.c @@ -13,39 +13,80 @@ #include "vpx_ports/mem.h" void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) { - // The 2D transform is done with two passes which are actually pretty - // similar. In the first one, we transform the columns and transpose - // the results. In the second one, we transform the rows. To achieve that, - // as the first pass results are transposed, we transpose the columns (that - // is the transposed rows) and transpose the results (so that it goes back - // in normal/row positions). - int pass; + // This 2D transform implements 4 vertical 1D transforms followed + // by 4 horizontal 1D transforms. The multiplies and adds are as given + // by Chen, Smith and Fralick ('77). The commands for moving the data + // around have been minimized by hand. + // For the purposes of the comments, the 16 inputs are referred to at i0 + // through iF (in raster order), intermediate variables are a0, b0, c0 + // through f, and correspond to the in-place computations mapped to input + // locations. The outputs, o0 through oF are labeled according to the + // output locations. + // Constants - // When we use them, in one case, they are all the same. In all others - // it's a pair of them that we need to repeat four times. This is done - // by constructing the 32 bit constant corresponding to that pair. - const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64); - const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); - const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); - const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); + // These are the coefficients used for the multiplies. + // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64), + // where cospi_N_64 = cos(N pi /64) + const __m128i k__cospi_A = _mm_setr_epi16(cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64); + const __m128i k__cospi_B = _mm_setr_epi16(cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64); + const __m128i k__cospi_C = _mm_setr_epi16(cospi_8_64, cospi_24_64, + cospi_8_64, cospi_24_64, + cospi_24_64, -cospi_8_64, + cospi_24_64, -cospi_8_64); + const __m128i k__cospi_D = _mm_setr_epi16(cospi_24_64, -cospi_8_64, + cospi_24_64, -cospi_8_64, + cospi_8_64, cospi_24_64, + cospi_8_64, cospi_24_64); + const __m128i k__cospi_E = _mm_setr_epi16(cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64); + const __m128i k__cospi_F = _mm_setr_epi16(cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64); + const __m128i k__cospi_G = _mm_setr_epi16(cospi_8_64, cospi_24_64, + cospi_8_64, cospi_24_64, + -cospi_8_64, -cospi_24_64, + -cospi_8_64, -cospi_24_64); + const __m128i k__cospi_H = _mm_setr_epi16(cospi_24_64, -cospi_8_64, + cospi_24_64, -cospi_8_64, + -cospi_24_64, cospi_8_64, + -cospi_24_64, cospi_8_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + // This second rounding constant saves doing some extra adds at the end + const __m128i k__DCT_CONST_ROUNDING2 = _mm_set1_epi32(DCT_CONST_ROUNDING + +(DCT_CONST_ROUNDING << 1)); + const int DCT_CONST_BITS2 = DCT_CONST_BITS+2; const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0); - const __m128i kOne = _mm_set1_epi16(1); __m128i in0, in1; + // Load inputs. { in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); + in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); + in1 = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *) + (input + 2 * stride))); in0 = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *) - (input + 1 * stride))); - in1 = _mm_loadl_epi64((const __m128i *)(input + 2 * stride)); - in1 = _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *) - (input + 3 * stride)), in1); + (input + 3 * stride))); + // in0 = [i0 i1 i2 i3 iC iD iE iF] + // in1 = [i4 i5 i6 i7 i8 i9 iA iB] + - // x = x << 4 + // multiply by 16 to give some extra precision in0 = _mm_slli_epi16(in0, 4); in1 = _mm_slli_epi16(in1, 4); // if (i == 0 && input[0]) input[0] += 1; + // add 1 to the upper left pixel if it is non-zero, which helps reduce + // the round-trip error { // The mask will only contain whether the first value is zero, all // other comparison will fail as something shifted by 4 (above << 4) @@ -58,57 +99,119 @@ void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride) { in0 = _mm_add_epi16(in0, k__nonzero_bias_b); } } - // Do the two transform/transpose passes - for (pass = 0; pass < 2; ++pass) { - // Transform 1/2: Add/subtract - const __m128i r0 = _mm_add_epi16(in0, in1); - const __m128i r1 = _mm_sub_epi16(in0, in1); - const __m128i r2 = _mm_unpacklo_epi64(r0, r1); - const __m128i r3 = _mm_unpackhi_epi64(r0, r1); - // Transform 1/2: Interleave to do the multiply by constants which gets us - // into 32 bits. - const __m128i t0 = _mm_unpacklo_epi16(r2, r3); - const __m128i t2 = _mm_unpackhi_epi16(r2, r3); - const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); - const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); - const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p08_p24); - const __m128i u6 = _mm_madd_epi16(t2, k__cospi_p24_m08); + // There are 4 total stages, alternating between an add/subtract stage + // followed by an multiply-and-add stage. + { + // Stage 1: Add/subtract + + // in0 = [i0 i1 i2 i3 iC iD iE iF] + // in1 = [i4 i5 i6 i7 i8 i9 iA iB] + const __m128i r0 = _mm_unpacklo_epi16(in0, in1); + const __m128i r1 = _mm_unpackhi_epi16(in0, in1); + // r0 = [i0 i4 i1 i5 i2 i6 i3 i7] + // r1 = [iC i8 iD i9 iE iA iF iB] + const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4); + const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4); + // r2 = [i0 i4 i1 i5 i3 i7 i2 i6] + // r3 = [iC i8 iD i9 iF iB iE iA] + + const __m128i t0 = _mm_add_epi16(r2, r3); + const __m128i t1 = _mm_sub_epi16(r2, r3); + // t0 = [a0 a4 a1 a5 a3 a7 a2 a6] + // t1 = [aC a8 aD a9 aF aB aE aA] + + // Stage 2: multiply by constants (which gets us into 32 bits). + // The constants needed here are: + // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16] + // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16] + // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08] + // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24] + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D); + // Then add and right-shift to get back to 16-bit range const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); - const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); - const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); - const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); - const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); - // Combine and transpose - const __m128i res0 = _mm_packs_epi32(w0, w2); - const __m128i res1 = _mm_packs_epi32(w4, w6); - // 00 01 02 03 20 21 22 23 - // 10 11 12 13 30 31 32 33 - const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); - const __m128i tr0_1 = _mm_unpackhi_epi16(res0, res1); - // 00 10 01 11 02 12 03 13 - // 20 30 21 31 22 32 23 33 - in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); - in1 = _mm_unpackhi_epi32(tr0_0, tr0_1); - in1 = _mm_shuffle_epi32(in1, 0x4E); - // 00 10 20 30 01 11 21 31 in0 contains 0 followed by 1 - // 02 12 22 32 03 13 23 33 in1 contains 2 followed by 3 + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + // w0 = [b0 b1 b7 b6] + // w1 = [b8 b9 bF bE] + // w2 = [b4 b5 b3 b2] + // w3 = [bC bD bB bA] + const __m128i x0 = _mm_packs_epi32(w0, w1); + const __m128i x1 = _mm_packs_epi32(w2, w3); + // x0 = [b0 b1 b7 b6 b8 b9 bF bE] + // x1 = [b4 b5 b3 b2 bC bD bB bA] + in0 = _mm_shuffle_epi32(x0, 0xD8); + in1 = _mm_shuffle_epi32(x1, 0x8D); + // in0 = [b0 b1 b8 b9 b7 b6 bF bE] + // in1 = [b3 b2 bB bA b4 b5 bC bD] } - in1 = _mm_shuffle_epi32(in1, 0x4E); - // Post-condition output and store it (v + 1) >> 2, taking advantage - // of the fact 1/3 are stored just after 0/2. { - __m128i out01 = _mm_add_epi16(in0, kOne); - __m128i out23 = _mm_add_epi16(in1, kOne); - out01 = _mm_srai_epi16(out01, 2); - out23 = _mm_srai_epi16(out23, 2); - _mm_storeu_si128((__m128i *)(output + 0 * 4), out01); - _mm_storeu_si128((__m128i *)(output + 2 * 4), out23); + // vertical DCTs finished. Now we do the horizontal DCTs. + // Stage 3: Add/subtract + + const __m128i t0 = _mm_add_epi16(in0, in1); + const __m128i t1 = _mm_sub_epi16(in0, in1); + // t0 = [c0 c1 c8 c9 c4 c5 cC cD] + // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE] + + // Stage 4: multiply by constants (which gets us into 32 bits). + // The constants needed here are: + // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16] + // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16] + // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24] + // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08] + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E); + const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F); + const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H); + // Then add and right-shift to get back to 16-bit range + // but this combines the final right-shift as well to save operations + // This unusual rounding operations is to maintain bit-accurate + // compatibility with the c version of this function which has two + // rounding steps in a row. + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2); + // w0 = [o0 o4 o8 oC] + // w1 = [o2 o6 oA oE] + // w2 = [o1 o5 o9 oD] + // w3 = [o3 o7 oB oF] + // remember the o's are numbered according to the correct output location + const __m128i x0 = _mm_packs_epi32(w0, w1); + const __m128i x1 = _mm_packs_epi32(w2, w3); + // x0 = [o0 o4 o8 oC o2 o6 oA oE] + // x1 = [o1 o5 o9 oD o3 o7 oB oF] + const __m128i y0 = _mm_unpacklo_epi16(x0, x1); + const __m128i y1 = _mm_unpackhi_epi16(x0, x1); + // y0 = [o0 o1 o4 o5 o8 o9 oC oD] + // y1 = [o2 o3 o6 o7 oA oB oE oF] + in0 = _mm_unpacklo_epi32(y0, y1); + // in0 = [o0 o1 o2 o3 o4 o5 o6 o7] + in1 = _mm_unpackhi_epi32(y0, y1); + // in1 = [o8 o9 oA oB oC oD oE oF] + } + // Post-condition (v + 1) >> 2 is now incorporated into previous + // add and right-shift commands. Only 2 store instructions needed + // because we are using the fact that 1/3 are stored just after 0/2. + { + _mm_storeu_si128((__m128i *)(output + 0 * 4), in0); + _mm_storeu_si128((__m128i *)(output + 2 * 4), in1); } } + static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in, int stride) { const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);