From 339f4dcaeef92c503347faacb4f07d55b1c02106 Mon Sep 17 00:00:00 2001 From: Kaustubh Raste Date: Wed, 29 Nov 2017 13:27:04 +0530 Subject: [PATCH] mips msa optimize vpx_scaled_2d function Change-Id: I638507b360c71489ab0e87bd558d2719ad995333 --- test/convolve_test.cc | 2 +- vpx_dsp/mips/macros_msa.h | 21 ++ vpx_dsp/mips/vpx_convolve8_msa.c | 596 +++++++++++++++++++++++++++++++ vpx_dsp/vpx_dsp_rtcd_defs.pl | 2 +- 4 files changed, 619 insertions(+), 2 deletions(-) diff --git a/test/convolve_test.cc b/test/convolve_test.cc index c4e6f8c54..70f0b11a7 100644 --- a/test/convolve_test.cc +++ b/test/convolve_test.cc @@ -1359,7 +1359,7 @@ const ConvolveFunctions convolve8_msa( vpx_convolve8_avg_horiz_msa, vpx_convolve8_vert_msa, vpx_convolve8_avg_vert_msa, vpx_convolve8_msa, vpx_convolve8_avg_msa, vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c, vpx_scaled_vert_c, - vpx_scaled_avg_vert_c, vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0); + vpx_scaled_avg_vert_c, vpx_scaled_2d_msa, vpx_scaled_avg_2d_c, 0); const ConvolveParam kArrayConvolve8_msa[] = { ALL_SIZES(convolve8_msa) }; INSTANTIATE_TEST_CASE_P(MSA, ConvolveTest, diff --git a/vpx_dsp/mips/macros_msa.h b/vpx_dsp/mips/macros_msa.h index 5c460ce19..f9a446e7b 100644 --- a/vpx_dsp/mips/macros_msa.h +++ b/vpx_dsp/mips/macros_msa.h @@ -555,6 +555,7 @@ #define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__) #define VSHF_B2_SB(...) VSHF_B2(v16i8, __VA_ARGS__) #define VSHF_B2_UH(...) VSHF_B2(v8u16, __VA_ARGS__) +#define VSHF_B2_SH(...) VSHF_B2(v8i16, __VA_ARGS__) #define VSHF_B4(RTYPE, in0, in1, mask0, mask1, mask2, mask3, out0, out1, out2, \ out3) \ @@ -1182,6 +1183,7 @@ out1 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1); \ } #define ILVRL_W2_UB(...) ILVRL_W2(v16u8, __VA_ARGS__) +#define ILVRL_W2_SB(...) ILVRL_W2(v16i8, __VA_ARGS__) #define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__) #define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__) @@ -1595,6 +1597,25 @@ out = (v4i32)__msa_ilvr_h(sign_m, (v8i16)in); \ } +/* Description : Sign extend byte elements from input vector and return + halfword results in pair of vectors + Arguments : Input - in (byte vector) + Outputs - out0, out1 (sign extended halfword vectors) + Return Type - signed halfword + Details : Sign bit of byte elements from input vector 'in' is + extracted and interleaved right with same vector 'in0' to + generate 8 signed halfword elements in 'out0' + Then interleaved left with same vector 'in0' to + generate 8 signed halfword elements in 'out1' +*/ +#define UNPCK_SB_SH(in, out0, out1) \ + { \ + v16i8 tmp_m; \ + \ + tmp_m = __msa_clti_s_b((v16i8)in, 0); \ + ILVRL_B2_SH(tmp_m, in, out0, out1); \ + } + /* Description : Zero extend unsigned byte elements to halfword elements Arguments : Input - in (unsigned byte vector) Outputs - out0, out1 (unsigned halfword vectors) diff --git a/vpx_dsp/mips/vpx_convolve8_msa.c b/vpx_dsp/mips/vpx_convolve8_msa.c index 8c4d19f45..d35a5a7a6 100644 --- a/vpx_dsp/mips/vpx_convolve8_msa.c +++ b/vpx_dsp/mips/vpx_convolve8_msa.c @@ -629,3 +629,599 @@ void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, } } } + +static void filter_horiz_w4_msa(const uint8_t *src_x, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *x_filter) { + uint64_t srcd0, srcd1, srcd2, srcd3; + uint32_t res; + v16u8 src0 = { 0 }, src1 = { 0 }, dst0; + v16i8 out0, out1; + v16i8 shf1 = { 0, 8, 16, 24, 4, 12, 20, 28, 1, 9, 17, 25, 5, 13, 21, 29 }; + v16i8 shf2 = shf1 + 2; + v16i8 filt_shf0 = { 0, 1, 0, 1, 0, 1, 0, 1, 8, 9, 8, 9, 8, 9, 8, 9 }; + v16i8 filt_shf1 = filt_shf0 + 2; + v16i8 filt_shf2 = filt_shf0 + 4; + v16i8 filt_shf3 = filt_shf0 + 6; + v8i16 filt, src0_h, src1_h, src2_h, src3_h, filt0, filt1, filt2, filt3; + + LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_UB(srcd0, srcd1, src0); + INSERT_D2_UB(srcd2, srcd3, src1); + VSHF_B2_SB(src0, src1, src0, src1, shf1, shf2, out0, out1); + XORI_B2_128_SB(out0, out1); + UNPCK_SB_SH(out0, src0_h, src1_h); + UNPCK_SB_SH(out1, src2_h, src3_h); + + filt = LD_SH(x_filter); + VSHF_B2_SH(filt, filt, filt, filt, filt_shf0, filt_shf1, filt0, filt1); + VSHF_B2_SH(filt, filt, filt, filt, filt_shf2, filt_shf3, filt2, filt3); + + src0_h *= filt0; + src0_h += src1_h * filt1; + src0_h += src2_h * filt2; + src0_h += src3_h * filt3; + + src1_h = (v8i16)__msa_sldi_b((v16i8)src0_h, (v16i8)src0_h, 8); + + src0_h = __msa_adds_s_h(src0_h, src1_h); + src0_h = __msa_srari_h(src0_h, FILTER_BITS); + src0_h = __msa_sat_s_h(src0_h, 7); + dst0 = PCKEV_XORI128_UB(src0_h, src0_h); + res = __msa_copy_u_w((v4i32)dst0, 0); + SW(res, dst); +} + +static void filter_horiz_w8_msa(const uint8_t *src_x, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *x_filter) { + uint64_t srcd0, srcd1, srcd2, srcd3; + v16u8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 }; + v16u8 tmp0, tmp1, tmp2, tmp3, dst0; + v16i8 out0, out1, out2, out3; + v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 }; + v16i8 shf2 = shf1 + 4; + v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; + v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; + + LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_UB(srcd0, srcd1, src0); + INSERT_D2_UB(srcd2, srcd3, src1); + LD4(src_x + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_UB(srcd0, srcd1, src2); + INSERT_D2_UB(srcd2, srcd3, src3); + + filt = LD_SH(x_filter); + SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); + + // transpose + VSHF_B2_UB(src0, src1, src0, src1, shf1, shf2, tmp0, tmp1); + VSHF_B2_UB(src2, src3, src2, src3, shf1, shf2, tmp2, tmp3); + ILVRL_W2_SB(tmp2, tmp0, out0, out1); + ILVRL_W2_SB(tmp3, tmp1, out2, out3); + + XORI_B4_128_SB(out0, out1, out2, out3); + UNPCK_SB_SH(out0, src0_h, src1_h); + UNPCK_SB_SH(out1, src2_h, src3_h); + UNPCK_SB_SH(out2, src4_h, src5_h); + UNPCK_SB_SH(out3, src6_h, src7_h); + + src0_h *= filt0; + src4_h *= filt4; + src0_h += src1_h * filt1; + src4_h += src5_h * filt5; + src0_h += src2_h * filt2; + src4_h += src6_h * filt6; + src0_h += src3_h * filt3; + src4_h += src7_h * filt7; + + src0_h = __msa_adds_s_h(src0_h, src4_h); + src0_h = __msa_srari_h(src0_h, FILTER_BITS); + src0_h = __msa_sat_s_h(src0_h, 7); + dst0 = PCKEV_XORI128_UB(src0_h, src0_h); + ST8x1_UB(dst0, dst); +} + +static void filter_horiz_w16_msa(const uint8_t *src_x, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *x_filter) { + uint64_t srcd0, srcd1, srcd2, srcd3; + v16u8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 }; + v16u8 src4 = { 0 }, src5 = { 0 }, src6 = { 0 }, src7 = { 0 }; + v16u8 tmp0, tmp1, tmp2, tmp3, dst0; + v16i8 out0, out1, out2, out3, out4, out5, out6, out7; + v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 }; + v16i8 shf2 = shf1 + 4; + v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; + v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; + v8i16 dst0_h, dst1_h, dst2_h, dst3_h; + + LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_UB(srcd0, srcd1, src0); + INSERT_D2_UB(srcd2, srcd3, src1); + LD4(src_x + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_UB(srcd0, srcd1, src2); + INSERT_D2_UB(srcd2, srcd3, src3); + LD4(src_x + 8 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_UB(srcd0, srcd1, src4); + INSERT_D2_UB(srcd2, srcd3, src5); + LD4(src_x + 12 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_UB(srcd0, srcd1, src6); + INSERT_D2_UB(srcd2, srcd3, src7); + + filt = LD_SH(x_filter); + SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); + + // transpose + VSHF_B2_UB(src0, src1, src0, src1, shf1, shf2, tmp0, tmp1); + VSHF_B2_UB(src2, src3, src2, src3, shf1, shf2, tmp2, tmp3); + ILVRL_W2_SB(tmp2, tmp0, out0, out1); + ILVRL_W2_SB(tmp3, tmp1, out2, out3); + XORI_B4_128_SB(out0, out1, out2, out3); + + UNPCK_SB_SH(out0, src0_h, src1_h); + UNPCK_SB_SH(out1, src2_h, src3_h); + UNPCK_SB_SH(out2, src4_h, src5_h); + UNPCK_SB_SH(out3, src6_h, src7_h); + + VSHF_B2_UB(src4, src5, src4, src5, shf1, shf2, tmp0, tmp1); + VSHF_B2_UB(src6, src7, src6, src7, shf1, shf2, tmp2, tmp3); + ILVRL_W2_SB(tmp2, tmp0, out4, out5); + ILVRL_W2_SB(tmp3, tmp1, out6, out7); + XORI_B4_128_SB(out4, out5, out6, out7); + + dst0_h = src0_h * filt0; + dst1_h = src4_h * filt4; + dst0_h += src1_h * filt1; + dst1_h += src5_h * filt5; + dst0_h += src2_h * filt2; + dst1_h += src6_h * filt6; + dst0_h += src3_h * filt3; + dst1_h += src7_h * filt7; + + UNPCK_SB_SH(out4, src0_h, src1_h); + UNPCK_SB_SH(out5, src2_h, src3_h); + UNPCK_SB_SH(out6, src4_h, src5_h); + UNPCK_SB_SH(out7, src6_h, src7_h); + + dst2_h = src0_h * filt0; + dst3_h = src4_h * filt4; + dst2_h += src1_h * filt1; + dst3_h += src5_h * filt5; + dst2_h += src2_h * filt2; + dst3_h += src6_h * filt6; + dst2_h += src3_h * filt3; + dst3_h += src7_h * filt7; + + ADDS_SH2_SH(dst0_h, dst1_h, dst2_h, dst3_h, dst0_h, dst2_h); + SRARI_H2_SH(dst0_h, dst2_h, FILTER_BITS); + SAT_SH2_SH(dst0_h, dst2_h, 7); + dst0 = PCKEV_XORI128_UB(dst0_h, dst2_h); + ST_UB(dst0, dst); +} + +static void transpose4x4_to_dst(const uint8_t *src, uint8_t *dst, + ptrdiff_t dst_stride) { + v16u8 in0; + v16i8 out0 = { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 }; + + in0 = LD_UB(src); + out0 = __msa_vshf_b(out0, (v16i8)in0, (v16i8)in0); + ST4x4_UB(out0, out0, 0, 1, 2, 3, dst, dst_stride); +} + +static void transpose8x8_to_dst(const uint8_t *src, uint8_t *dst, + ptrdiff_t dst_stride) { + v16u8 in0, in1, in2, in3, out0, out1, out2, out3, tmp0, tmp1, tmp2, tmp3; + v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 }; + v16i8 shf2 = shf1 + 4; + + LD_UB4(src, 16, in0, in1, in2, in3); + VSHF_B2_UB(in0, in1, in0, in1, shf1, shf2, tmp0, tmp1); + VSHF_B2_UB(in2, in3, in2, in3, shf1, shf2, tmp2, tmp3); + ILVRL_W2_UB(tmp2, tmp0, out0, out1); + ILVRL_W2_UB(tmp3, tmp1, out2, out3); + ST8x4_UB(out0, out1, dst, dst_stride); + ST8x4_UB(out2, out3, dst + 4 * dst_stride, dst_stride); +} + +static void transpose16x16_to_dst(const uint8_t *src, uint8_t *dst, + ptrdiff_t dst_stride) { + v16u8 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11, in12; + v16u8 in13, in14, in15, out0, out1, out2, out3, out4, out5, out6, out7, out8; + v16u8 out9, out10, out11, out12, out13, out14, out15; + + LD_UB8(src, 16, in0, in1, in2, in3, in4, in5, in6, in7); + LD_UB8(src + 16 * 8, 16, in8, in9, in10, in11, in12, in13, in14, in15); + + TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, + in11, in12, in13, in14, in15, out0, out1, out2, out3, + out4, out5, out6, out7); + ST_UB8(out0, out1, out2, out3, out4, out5, out6, out7, dst, dst_stride); + dst += 8 * dst_stride; + + SLDI_B4_0_UB(in0, in1, in2, in3, in0, in1, in2, in3, 8); + SLDI_B4_0_UB(in4, in5, in6, in7, in4, in5, in6, in7, 8); + SLDI_B4_0_UB(in8, in9, in10, in11, in8, in9, in10, in11, 8); + SLDI_B4_0_UB(in12, in13, in14, in15, in12, in13, in14, in15, 8); + + TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, + in11, in12, in13, in14, in15, out8, out9, out10, out11, + out12, out13, out14, out15); + ST_UB8(out8, out9, out10, out11, out12, out13, out14, out15, dst, dst_stride); +} + +static void scaledconvolve_horiz_w4(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int h) { + DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]); + int y, z, i; + src -= SUBPEL_TAPS / 2 - 1; + + for (y = 0; y < h; y += 4) { + int x_q4 = x0_q4; + for (z = 0; z < 4; ++z) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + + if (x_q4 & SUBPEL_MASK) { + filter_horiz_w4_msa(src_x, src_stride, temp + (z * 4), x_filter); + } else { + for (i = 0; i < 4; ++i) { + temp[z * 4 + i] = src_x[i * src_stride + 3]; + } + } + + x_q4 += x_step_q4; + } + + transpose4x4_to_dst(temp, dst, dst_stride); + + src += src_stride * 4; + dst += dst_stride * 4; + } +} + +static void scaledconvolve_horiz_w8(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int h) { + DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]); + int y, z, i; + src -= SUBPEL_TAPS / 2 - 1; + + // This function processes 8x8 areas. The intermediate height is not always + // a multiple of 8, so force it to be a multiple of 8 here. + y = h + (8 - (h & 0x7)); + + do { + int x_q4 = x0_q4; + for (z = 0; z < 8; ++z) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + + if (x_q4 & SUBPEL_MASK) { + filter_horiz_w8_msa(src_x, src_stride, temp + (z * 8), x_filter); + } else { + for (i = 0; i < 8; ++i) { + temp[z * 8 + i] = src_x[3 + i * src_stride]; + } + } + + x_q4 += x_step_q4; + } + + transpose8x8_to_dst(temp, dst, dst_stride); + + src += src_stride * 8; + dst += dst_stride * 8; + } while (y -= 8); +} + +static void scaledconvolve_horiz_mul16(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h) { + DECLARE_ALIGNED(16, uint8_t, temp[16 * 16]); + int x, y, z, i; + + src -= SUBPEL_TAPS / 2 - 1; + + // This function processes 16x16 areas. The intermediate height is not always + // a multiple of 16, so force it to be a multiple of 8 here. + y = h + (16 - (h & 0xF)); + + do { + int x_q4 = x0_q4; + for (x = 0; x < w; x += 16) { + for (z = 0; z < 16; ++z) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + + if (x_q4 & SUBPEL_MASK) { + filter_horiz_w16_msa(src_x, src_stride, temp + (z * 16), x_filter); + } else { + for (i = 0; i < 16; ++i) { + temp[z * 16 + i] = src_x[3 + i * src_stride]; + } + } + + x_q4 += x_step_q4; + } + + transpose16x16_to_dst(temp, dst + x, dst_stride); + } + + src += src_stride * 16; + dst += dst_stride * 16; + } while (y -= 16); +} + +static void filter_vert_w4_msa(const uint8_t *src_y, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *y_filter) { + uint32_t srcw0, srcw1, srcw2, srcw3, srcw4, srcw5, srcw6, srcw7; + uint32_t res; + v16u8 src0 = { 0 }, src1 = { 0 }, dst0; + v16i8 out0, out1; + v16i8 shf1 = { 0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23 }; + v16i8 shf2 = shf1 + 8; + v16i8 filt_shf0 = { 0, 1, 0, 1, 0, 1, 0, 1, 8, 9, 8, 9, 8, 9, 8, 9 }; + v16i8 filt_shf1 = filt_shf0 + 2; + v16i8 filt_shf2 = filt_shf0 + 4; + v16i8 filt_shf3 = filt_shf0 + 6; + v8i16 filt, src0_h, src1_h, src2_h, src3_h; + v8i16 filt0, filt1, filt2, filt3; + + LW4(src_y, src_pitch, srcw0, srcw1, srcw2, srcw3); + LW4(src_y + 4 * src_pitch, src_pitch, srcw4, srcw5, srcw6, srcw7); + INSERT_W4_UB(srcw0, srcw1, srcw2, srcw3, src0); + INSERT_W4_UB(srcw4, srcw5, srcw6, srcw7, src1); + VSHF_B2_SB(src0, src1, src0, src1, shf1, shf2, out0, out1); + XORI_B2_128_SB(out0, out1); + UNPCK_SB_SH(out0, src0_h, src1_h); + UNPCK_SB_SH(out1, src2_h, src3_h); + + filt = LD_SH(y_filter); + VSHF_B2_SH(filt, filt, filt, filt, filt_shf0, filt_shf1, filt0, filt1); + VSHF_B2_SH(filt, filt, filt, filt, filt_shf2, filt_shf3, filt2, filt3); + + src0_h *= filt0; + src0_h += src1_h * filt1; + src0_h += src2_h * filt2; + src0_h += src3_h * filt3; + + src1_h = (v8i16)__msa_sldi_b((v16i8)src0_h, (v16i8)src0_h, 8); + + src0_h = __msa_adds_s_h(src0_h, src1_h); + src0_h = __msa_srari_h(src0_h, FILTER_BITS); + src0_h = __msa_sat_s_h(src0_h, 7); + dst0 = PCKEV_XORI128_UB(src0_h, src0_h); + res = __msa_copy_u_w((v4i32)dst0, 0); + SW(res, dst); +} + +static void filter_vert_w8_msa(const uint8_t *src_y, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *y_filter) { + uint64_t srcd0, srcd1, srcd2, srcd3; + v16u8 dst0; + v16i8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 }; + v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; + v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; + + LD4(src_y, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_SB(srcd0, srcd1, src0); + INSERT_D2_SB(srcd2, srcd3, src1); + LD4(src_y + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3); + INSERT_D2_SB(srcd0, srcd1, src2); + INSERT_D2_SB(srcd2, srcd3, src3); + + filt = LD_SH(y_filter); + SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); + + XORI_B4_128_SB(src0, src1, src2, src3); + UNPCK_SB_SH(src0, src0_h, src1_h); + UNPCK_SB_SH(src1, src2_h, src3_h); + UNPCK_SB_SH(src2, src4_h, src5_h); + UNPCK_SB_SH(src3, src6_h, src7_h); + + src0_h *= filt0; + src4_h *= filt4; + src0_h += src1_h * filt1; + src4_h += src5_h * filt5; + src0_h += src2_h * filt2; + src4_h += src6_h * filt6; + src0_h += src3_h * filt3; + src4_h += src7_h * filt7; + + src0_h = __msa_adds_s_h(src0_h, src4_h); + src0_h = __msa_srari_h(src0_h, FILTER_BITS); + src0_h = __msa_sat_s_h(src0_h, 7); + dst0 = PCKEV_XORI128_UB(src0_h, src0_h); + ST8x1_UB(dst0, dst); +} + +static void filter_vert_mul_w16_msa(const uint8_t *src_y, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *y_filter, + int w) { + int x; + v16u8 dst0; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7; + v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h; + v8i16 src8_h, src9_h, src10_h, src11_h, src12_h, src13_h, src14_h, src15_h; + v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7; + + filt = LD_SH(y_filter); + SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3); + SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7); + + for (x = 0; x < w; x += 16) { + LD_SB8(src_y, src_pitch, src0, src1, src2, src3, src4, src5, src6, src7); + src_y += 16; + + XORI_B4_128_SB(src0, src1, src2, src3); + XORI_B4_128_SB(src4, src5, src6, src7); + UNPCK_SB_SH(src0, src0_h, src1_h); + UNPCK_SB_SH(src1, src2_h, src3_h); + UNPCK_SB_SH(src2, src4_h, src5_h); + UNPCK_SB_SH(src3, src6_h, src7_h); + UNPCK_SB_SH(src4, src8_h, src9_h); + UNPCK_SB_SH(src5, src10_h, src11_h); + UNPCK_SB_SH(src6, src12_h, src13_h); + UNPCK_SB_SH(src7, src14_h, src15_h); + + src0_h *= filt0; + src1_h *= filt0; + src8_h *= filt4; + src9_h *= filt4; + src0_h += src2_h * filt1; + src1_h += src3_h * filt1; + src8_h += src10_h * filt5; + src9_h += src11_h * filt5; + src0_h += src4_h * filt2; + src1_h += src5_h * filt2; + src8_h += src12_h * filt6; + src9_h += src13_h * filt6; + src0_h += src6_h * filt3; + src1_h += src7_h * filt3; + src8_h += src14_h * filt7; + src9_h += src15_h * filt7; + + ADDS_SH2_SH(src0_h, src8_h, src1_h, src9_h, src0_h, src1_h); + SRARI_H2_SH(src0_h, src1_h, FILTER_BITS); + SAT_SH2_SH(src0_h, src1_h, 7); + dst0 = PCKEV_XORI128_UB(src0_h, src1_h); + ST_UB(dst0, dst); + dst += 16; + } +} + +static void scaledconvolve_vert_w4(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int h) { + int y; + int y_q4 = y0_q4; + + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + + for (y = 0; y < h; ++y) { + const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + + if (y_q4 & SUBPEL_MASK) { + filter_vert_w4_msa(src_y, src_stride, &dst[y * dst_stride], y_filter); + } else { + uint32_t srcd = LW(src_y + 3 * src_stride); + SW(srcd, dst + y * dst_stride); + } + + y_q4 += y_step_q4; + } +} + +static void scaledconvolve_vert_w8(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int h) { + int y; + int y_q4 = y0_q4; + + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + + for (y = 0; y < h; ++y) { + const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + + if (y_q4 & SUBPEL_MASK) { + filter_vert_w8_msa(src_y, src_stride, &dst[y * dst_stride], y_filter); + } else { + uint64_t srcd = LD(src_y + 3 * src_stride); + SD(srcd, dst + y * dst_stride); + } + + y_q4 += y_step_q4; + } +} + +static void scaledconvolve_vert_mul16(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h) { + int x, y; + int y_q4 = y0_q4; + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + + for (y = 0; y < h; ++y) { + const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + + if (y_q4 & SUBPEL_MASK) { + filter_vert_mul_w16_msa(src_y, src_stride, &dst[y * dst_stride], y_filter, + w); + } else { + for (x = 0; x < w; ++x) { + dst[x + y * dst_stride] = src_y[x + 3 * src_stride]; + } + } + + y_q4 += y_step_q4; + } +} + +void vpx_scaled_2d_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const InterpKernel *filter, + int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, + int w, int h) { + // Note: Fixed size intermediate buffer, temp, places limits on parameters. + // 2d filtering proceeds in 2 steps: + // (1) Interpolate horizontally into an intermediate buffer, temp. + // (2) Interpolate temp vertically to derive the sub-pixel result. + // Deriving the maximum number of rows in the temp buffer (135): + // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative). + // --Largest block size is 64x64 pixels. + // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the + // original frame (in 1/16th pixel units). + // --Must round-up because block may be located at sub-pixel position. + // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails. + // --((64 - 1) * 32 + 15) >> 4 + 8 = 135. + // --Require an additional 8 rows for the horiz_w8 transpose tail. + DECLARE_ALIGNED(16, uint8_t, temp[(135 + 8) * 64]); + const int intermediate_height = + (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS; + + assert(w <= 64); + assert(h <= 64); + assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32)); + assert(x_step_q4 <= 64); + + if ((0 == x0_q4) && (16 == x_step_q4) && (0 == y0_q4) && (16 == y_step_q4)) { + vpx_convolve_copy_msa(src, src_stride, dst, dst_stride, filter, x0_q4, + x_step_q4, y0_q4, y_step_q4, w, h); + } else { + if (w >= 16) { + scaledconvolve_horiz_mul16(src - src_stride * (SUBPEL_TAPS / 2 - 1), + src_stride, temp, 64, filter, x0_q4, x_step_q4, + w, intermediate_height); + } else if (w == 8) { + scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1), + src_stride, temp, 64, filter, x0_q4, x_step_q4, + intermediate_height); + } else { + scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1), + src_stride, temp, 64, filter, x0_q4, x_step_q4, + intermediate_height); + } + + if (w >= 16) { + scaledconvolve_vert_mul16(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, + dst_stride, filter, y0_q4, y_step_q4, w, h); + } else if (w == 8) { + scaledconvolve_vert_w8(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, + dst_stride, filter, y0_q4, y_step_q4, h); + } else { + scaledconvolve_vert_w4(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, + dst_stride, filter, y0_q4, y_step_q4, h); + } + } +} diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index e117b9d15..04b9f22ba 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -374,7 +374,7 @@ add_proto qw/void vpx_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_st specialize qw/vpx_convolve8_avg_vert sse2 ssse3 avx2 neon dspr2 msa vsx/; add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h"; -specialize qw/vpx_scaled_2d ssse3 neon/; +specialize qw/vpx_scaled_2d ssse3 neon msa/; add_proto qw/void vpx_scaled_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h";