diff --git a/src/dsp/lossless_msa.c b/src/dsp/lossless_msa.c index b9516fa4..26c5ab62 100644 --- a/src/dsp/lossless_msa.c +++ b/src/dsp/lossless_msa.c @@ -78,6 +78,37 @@ (pdst)[2] = b; \ } while (0) +#define TRANSFORM_COLOR_INVERSE_8(src0, src1, dst0, dst1, \ + c0, c1, mask0, mask1) do { \ + v8i16 g0, g1, t0, t1, t2, t3; \ + v4i32 t4, t5; \ + VSHF_B2_SH(src0, src0, src1, src1, mask0, mask0, g0, g1); \ + DOTP_SB2_SH(g0, g1, c0, c0, t0, t1); \ + SRAI_H2_SH(t0, t1, 5); \ + t0 = __msa_addv_h(t0, (v8i16)src0); \ + t1 = __msa_addv_h(t1, (v8i16)src1); \ + t4 = __msa_srli_w((v4i32)t0, 16); \ + t5 = __msa_srli_w((v4i32)t1, 16); \ + DOTP_SB2_SH(t4, t5, c1, c1, t2, t3); \ + SRAI_H2_SH(t2, t3, 5); \ + ADD2(t0, t2, t1, t3, t0, t1); \ + VSHF_B2_UB(src0, t0, src1, t1, mask1, mask1, dst0, dst1); \ +} while (0) + +#define TRANSFORM_COLOR_INVERSE_4(src, dst, c0, c1, mask0, mask1) do { \ + const v16i8 g0 = VSHF_SB(src, src, mask0); \ + v8i16 t0 = __msa_dotp_s_h(c0, g0); \ + v8i16 t1; \ + v4i32 t2; \ + t0 = SRAI_H(t0, 5); \ + t0 = __msa_addv_h(t0, (v8i16)src); \ + t2 = __msa_srli_w((v4i32)t0, 16); \ + t1 = __msa_dotp_s_h(c1, (v16i8)t2); \ + t1 = SRAI_H(t1, 5); \ + t0 = t0 + t1; \ + dst = VSHF_UB(src, t0, mask1); \ +} while (0) + static void ConvertBGRAToRGBA(const uint32_t* src, int num_pixels, uint8_t* dst) { int i; @@ -213,6 +244,88 @@ static void ConvertBGRAToRGB(const uint32_t* src, } } +static void AddGreenToBlueAndRed(uint32_t* data, int num_pixels) { + int i; + uint8_t* ptemp_data = (uint8_t*)data; + v16u8 src0, dst0, tmp0; + const v16u8 mask = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, + 13, 255, 13, 255 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1, tmp1; + LD_UB2(ptemp_data, 16, src0, src1); + VSHF_B2_UB(src0, src1, src1, src0, mask, mask, tmp0, tmp1); + ADD2(src0, tmp0, src1, tmp1, dst0, dst1); + ST_UB2(dst0, dst1, ptemp_data, 16); + ptemp_data += 32; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(ptemp_data); + tmp0 = VSHF_UB(src0, src0, mask); + dst0 = src0 + tmp0; + ST_UB(dst0, ptemp_data); + ptemp_data += 16; + num_pixels -= 4; + } + for (i = 0; i < num_pixels; i++) { + const uint8_t b = ptemp_data[0]; + const uint8_t g = ptemp_data[1]; + const uint8_t r = ptemp_data[2]; + ptemp_data[0] = (b + g) & 0xff; + ptemp_data[2] = (r + g) & 0xff; + ptemp_data += 4; + } + } +} + +static void TransformColorInverse(const VP8LMultipliers* const m, + uint32_t* data, int num_pixels) { + v16u8 src0, dst0; + const v16i8 g2br = (v16i8)__msa_fill_w(m->green_to_blue_ | + (m->green_to_red_ << 16)); + const v16i8 r2b = (v16i8)__msa_fill_w(m->red_to_blue_); + const v16u8 mask0 = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, + 13, 255, 13, 255 }; + const v16u8 mask1 = { 16, 1, 18, 3, 20, 5, 22, 7, 24, 9, 26, 11, + 28, 13, 30, 15 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1; + LD_UB2(data, 4, src0, src1); + TRANSFORM_COLOR_INVERSE_8(src0, src1, dst0, dst1, g2br, r2b, mask0, mask1); + ST_UB2(dst0, dst1, data, 4); + data += 8; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(data); + TRANSFORM_COLOR_INVERSE_4(src0, dst0, g2br, r2b, mask0, mask1); + ST_UB(dst0, data); + data += 4; + num_pixels -= 4; + } + if (num_pixels > 0) { + src0 = LD_UB(data); + TRANSFORM_COLOR_INVERSE_4(src0, dst0, g2br, r2b, mask0, mask1); + if (num_pixels == 3) { + const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); + const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2); + SD(pix_d, data + 0); + SW(pix_w, data + 2); + } else if (num_pixels == 2) { + const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); + SD(pix_d, data); + } else { + const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 0); + SW(pix_w, data); + } + } + } +} + //------------------------------------------------------------------------------ // Entry point @@ -222,6 +335,8 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMSA(void) { VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; VP8LConvertBGRAToBGR = ConvertBGRAToBGR; VP8LConvertBGRAToRGB = ConvertBGRAToRGB; + VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; + VP8LTransformColorInverse = TransformColorInverse; } #else // !WEBP_USE_MSA diff --git a/src/dsp/msa_macro.h b/src/dsp/msa_macro.h index 6a1cfb26..328aea21 100644 --- a/src/dsp/msa_macro.h +++ b/src/dsp/msa_macro.h @@ -361,6 +361,22 @@ #define VSHF_H2_UH(...) VSHF_H2(v8u16, __VA_ARGS__) #define VSHF_H2_SH(...) VSHF_H2(v8i16, __VA_ARGS__) +/* Description : Dot product of byte vector elements + * Arguments : Inputs - mult0, mult1, cnst0, cnst1 + * Outputs - out0, out1 + * Return Type - as per RTYPE + * Details : Signed byte elements from 'mult0' are multiplied with + * signed byte elements from 'cnst0' producing a result + * twice the size of input i.e. signed halfword. + * The multiplication result of adjacent odd-even elements + * are added together and written to the 'out0' vector +*/ +#define DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) do { \ + out0 = (RTYPE)__msa_dotp_s_h((v16i8)mult0, (v16i8)cnst0); \ + out1 = (RTYPE)__msa_dotp_s_h((v16i8)mult1, (v16i8)cnst1); \ +} while (0) +#define DOTP_SB2_SH(...) DOTP_SB2(v8i16, __VA_ARGS__) + /* Description : Clips all signed halfword elements of input vector * between 0 & 255 * Arguments : Input/output - val @@ -678,7 +694,7 @@ static WEBP_INLINE uint32_t func_hadd_uh_u32(v8u16 in) { out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1); \ out1 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1); \ } while (0) -#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__) +#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__) #define ILVRL_B2_SB(...) ILVRL_B2(v16i8, __VA_ARGS__) #define ILVRL_B2_UH(...) ILVRL_B2(v8u16, __VA_ARGS__) #define ILVRL_B2_SH(...) ILVRL_B2(v8i16, __VA_ARGS__)