void vp9_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
int stride, int tx_type) {
__m128i in0[16], in1[16];
load_buffer_8x16(input, in0);
input += 8;
load_buffer_8x16(input, in1);
switch (tx_type) {
case DCT_DCT:
idct16_sse2(in0, in1);
idct16_sse2(in0, in1);
break;
case ADST_DCT:
idct16_sse2(in0, in1);
iadst16_sse2(in0, in1);
break;
case DCT_ADST:
iadst16_sse2(in0, in1);
idct16_sse2(in0, in1);
break;
default:
assert(tx_type == ADST_ADST);
iadst16_sse2(in0, in1);
iadst16_sse2(in0, in1);
break;
}
write_buffer_8x16(dest, in0, stride);
dest += 8;
write_buffer_8x16(dest, in1, stride);
}
void vp9_iht16x16_256_add_sse2(const int16_t *input, uint8_t *dest, int stride,
int tx_type) {
__m128i in0[16], in1[16];
load_buffer_8x16(input, in0);
input += 8;
load_buffer_8x16(input, in1);
switch (tx_type) {
case 0: // DCT_DCT
idct16_sse2(in0, in1);
idct16_sse2(in0, in1);
break;
case 1: // ADST_DCT
idct16_sse2(in0, in1);
iadst16_sse2(in0, in1);
break;
case 2: // DCT_ADST
iadst16_sse2(in0, in1);
idct16_sse2(in0, in1);
break;
case 3: // ADST_ADST
iadst16_sse2(in0, in1);
iadst16_sse2(in0, in1);
break;
default:
assert(0);
break;
}
write_buffer_8x16(dest, in0, stride);
dest += 8;
write_buffer_8x16(dest, in1, stride);
}