static WEBP_INLINE void PredictLineGradient(const uint8_t* pinput,
const uint8_t* ppred,
uint8_t* poutput, int stride,
int size) {
int w;
const v16i8 zero = { 0 };
while (size >= 16) {
v16u8 pred0, dst0;
v8i16 a0, a1, b0, b1, c0, c1;
const v16u8 tmp0 = LD_UB(ppred - 1);
const v16u8 tmp1 = LD_UB(ppred - stride);
const v16u8 tmp2 = LD_UB(ppred - stride - 1);
const v16u8 src0 = LD_UB(pinput);
ILVRL_B2_SH(zero, tmp0, a0, a1);
ILVRL_B2_SH(zero, tmp1, b0, b1);
ILVRL_B2_SH(zero, tmp2, c0, c1);
ADD2(a0, b0, a1, b1, a0, a1);
SUB2(a0, c0, a1, c1, a0, a1);
CLIP_SH2_0_255(a0, a1);
pred0 = (v16u8)__msa_pckev_b((v16i8)a1, (v16i8)a0);
dst0 = src0 - pred0;
ST_UB(dst0, poutput);
ppred += 16;
pinput += 16;
poutput += 16;
size -= 16;
}
for (w = 0; w < size; ++w) {
const int pred = ppred[w - 1] + ppred[w - stride] - ppred[w - stride - 1];
poutput[w] = pinput[w] - (pred < 0 ? 0 : pred > 255 ? 255 : pred);
}
}
static WEBP_INLINE void PredictLineInverse0(const uint8_t* src,
const uint8_t* pred,
uint8_t* dst, int length) {
v16u8 src0, pred0, dst0;
assert(length >= 0);
while (length >= 32) {
v16u8 src1, pred1, dst1;
LD_UB2(src, 16, src0, src1);
LD_UB2(pred, 16, pred0, pred1);
SUB2(src0, pred0, src1, pred1, dst0, dst1);
ST_UB2(dst0, dst1, dst, 16);
src += 32;
pred += 32;
dst += 32;
length -= 32;
}
if (length > 0) {
int i;
if (length >= 16) {
src0 = LD_UB(src);
pred0 = LD_UB(pred);
dst0 = src0 - pred0;
ST_UB(dst0, dst);
src += 16;
pred += 16;
dst += 16;
length -= 16;
}
for (i = 0; i < length; i++) {
dst[i] = src[i] - pred[i];
}
}
}
void aom_plane_add_noise_msa(uint8_t *start_ptr, char *noise,
char blackclamp[16], char whiteclamp[16],
char bothclamp[16], uint32_t width,
uint32_t height, int32_t pitch) {
uint32_t i, j;
for (i = 0; i < height / 2; ++i) {
uint8_t *pos0_ptr = start_ptr + (2 * i) * pitch;
int8_t *ref0_ptr = (int8_t *)(noise + (rand() & 0xff));
uint8_t *pos1_ptr = start_ptr + (2 * i + 1) * pitch;
int8_t *ref1_ptr = (int8_t *)(noise + (rand() & 0xff));
for (j = width / 16; j--;) {
v16i8 temp00_s, temp01_s;
v16u8 temp00, temp01, black_clamp, white_clamp;
v16u8 pos0, ref0, pos1, ref1;
v16i8 const127 = __msa_ldi_b(127);
pos0 = LD_UB(pos0_ptr);
ref0 = LD_UB(ref0_ptr);
pos1 = LD_UB(pos1_ptr);
ref1 = LD_UB(ref1_ptr);
black_clamp = (v16u8)__msa_fill_b(blackclamp[0]);
white_clamp = (v16u8)__msa_fill_b(whiteclamp[0]);
temp00 = (pos0 < black_clamp);
pos0 = __msa_bmnz_v(pos0, black_clamp, temp00);
temp01 = (pos1 < black_clamp);
pos1 = __msa_bmnz_v(pos1, black_clamp, temp01);
XORI_B2_128_UB(pos0, pos1);
temp00_s = __msa_adds_s_b((v16i8)white_clamp, const127);
temp00 = (v16u8)(temp00_s < pos0);
pos0 = (v16u8)__msa_bmnz_v((v16u8)pos0, (v16u8)temp00_s, temp00);
temp01_s = __msa_adds_s_b((v16i8)white_clamp, const127);
temp01 = (temp01_s < pos1);
pos1 = (v16u8)__msa_bmnz_v((v16u8)pos1, (v16u8)temp01_s, temp01);
XORI_B2_128_UB(pos0, pos1);
pos0 += ref0;
ST_UB(pos0, pos0_ptr);
pos1 += ref1;
ST_UB(pos1, pos1_ptr);
pos0_ptr += 16;
pos1_ptr += 16;
ref0_ptr += 16;
ref1_ptr += 16;
}
}
}
static void intra_predict_vert_16x16_msa(const uint8_t *src, uint8_t *dst,
int32_t dst_stride) {
uint32_t row;
v16u8 src0;
src0 = LD_UB(src);
for (row = 16; row--;) {
ST_UB(src0, dst);
dst += dst_stride;
}
}
void
yuv_bgr_convert_msa (JSAMPROW p_in_y, JSAMPROW p_in_cb, JSAMPROW p_in_cr,
JSAMPROW p_rgb, JDIMENSION out_width)
{
int32_t y, cb, cr;
uint32_t col, num_cols_mul_16 = out_width >> 4;
uint32_t remaining_wd = out_width & 0xF;
v16u8 mask_rgb0 = {0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, 8, 9, 20, 10};
v16u8 mask_rgb1 = {11, 21, 12, 13, 22, 14, 15, 23, 0, 1, 24, 2, 3, 25, 4, 5};
v16u8 mask_rgb2 = {26, 6, 7, 27, 8, 9, 28, 10, 11, 29, 12, 13, 30, 14, 15, 31};
v16u8 tmp0, tmp1, out0, out1, out2, input_y = {0};
v16i8 input_cb, input_cr, out_rgb0, out_rgb1, const_128 = __msa_ldi_b(128);
v8i16 y_h0, y_h1, cb_h0, cb_h1, cr_h0, cr_h1;
v4i32 cb_w0, cb_w1, cb_w2, cb_w3, cr_w0, cr_w1, cr_w2, cr_w3, zero = {0};
v16i8 out_r0, out_g0, out_b0;
for (col = num_cols_mul_16; col--;) {
input_y = LD_UB(p_in_y);
input_cb = LD_SB(p_in_cb);
input_cr = LD_SB(p_in_cr);
p_in_y += 16;
p_in_cb += 16;
p_in_cr += 16;
input_cb -= const_128;
input_cr -= const_128;
UNPCK_UB_SH(input_y, y_h0, y_h1);
UNPCK_SB_SH(input_cb, cb_h0, cb_h1);
UNPCK_SB_SH(input_cr, cr_h0, cr_h1);
CALC_G4_FRM_YUV(y_h0, y_h1, cb_h0, cb_h1, cr_h0, cr_h1, out_g0);
UNPCK_SH_SW(cr_h0, cr_w0, cr_w1);
UNPCK_SH_SW(cr_h1, cr_w2, cr_w3);
CALC_R4_FRM_YUV(y_h0, y_h1, cr_w0, cr_w1, cr_w2, cr_w3, out_r0);
UNPCK_SH_SW(cb_h0, cb_w0, cb_w1);
UNPCK_SH_SW(cb_h1, cb_w2, cb_w3);
CALC_B4_FRM_YUV(y_h0, y_h1, cb_w0, cb_w1, cb_w2, cb_w3, out_b0);
ILVRL_B2_SB(out_g0, out_b0, out_rgb0, out_rgb1);
VSHF_B2_UB(out_rgb0, out_r0, out_rgb0, out_r0, mask_rgb0, mask_rgb1,
out0, tmp0);
VSHF_B2_UB(out_rgb1, out_r0, out_rgb1, out_r0, mask_rgb1, mask_rgb2,
tmp1, out2);
out1 = (v16u8) __msa_sldi_b((v16i8) zero, (v16i8) tmp1, 8);
out1 = (v16u8) __msa_pckev_d((v2i64) out1, (v2i64) tmp0);
ST_UB(out0, p_rgb);
p_rgb += 16;
ST_UB(out1, p_rgb);
p_rgb += 16;
ST_UB(out2, p_rgb);
p_rgb += 16;
}
if (remaining_wd >= 8) {
uint64_t in_y, in_cb, in_cr;
v16i8 input_cbcr = {0};
in_y = LD(p_in_y);
in_cb = LD(p_in_cb);
in_cr = LD(p_in_cr);
p_in_y += 8;
p_in_cb += 8;
p_in_cr += 8;
input_y = (v16u8) __msa_insert_d((v2i64) input_y, 0, in_y);
input_cbcr = (v16i8) __msa_insert_d((v2i64) input_cbcr, 0, in_cb);
input_cbcr = (v16i8) __msa_insert_d((v2i64) input_cbcr, 1, in_cr);
input_cbcr -= const_128;
y_h0 = (v8i16) __msa_ilvr_b((v16i8) zero, (v16i8) input_y);
UNPCK_SB_SH(input_cbcr, cb_h0, cr_h0);
UNPCK_SH_SW(cb_h0, cb_w0, cb_w1);
UNPCK_SH_SW(cr_h0, cr_w0, cr_w1);
CALC_R2_FRM_YUV(y_h0, cr_w0, cr_w1, out_r0);
CALC_G2_FRM_YUV(y_h0, cb_h0, cr_h0, out_g0);
CALC_B2_FRM_YUV(y_h0, cb_w0, cb_w1, out_b0);
out_rgb0 = (v16i8) __msa_ilvr_b((v16i8) out_g0, (v16i8) out_b0);
VSHF_B2_UB(out_rgb0, out_r0, out_rgb0, out_r0, mask_rgb0, mask_rgb1,
out0, out1);
ST_UB(out0, p_rgb);
p_rgb += 16;
ST8x1_UB(out1, p_rgb);
p_rgb += 8;
remaining_wd -= 8;
}
for (col = 0; col < remaining_wd; col++) {
y = (int) (p_in_y[col]);
cb = (int) (p_in_cb[col]) - 128;
cr = (int) (p_in_cr[col]) - 128;
/* Range-limiting is essential due to noise introduced by DCT losses. */
p_rgb[0] = clip_pixel(y + ROUND_POWER_OF_TWO(FIX_1_77200 * cb, 16));
p_rgb[1] = clip_pixel(y + ROUND_POWER_OF_TWO(((-FIX_0_34414) * cb -
FIX_0_71414 * cr), 16));
p_rgb[2] = clip_pixel(y + ROUND_POWER_OF_TWO(FIX_1_40200 * cr, 16));
p_rgb += 3;
}
}
