/* u32x4 mm mul */
void mw_neon_mm_mul_u32x4(unsigned int * A, int Row, int T, unsigned int * B, int Col, unsigned int * C)
{
int i, k, j;
uint32x4_t neon_b, neon_c;
uint32x4_t neon_a0, neon_a1, neon_a2, neon_a3;
uint32x4_t neon_b0, neon_b1, neon_b2, neon_b3;
for (i = 0; i < Row; i+=4)
{
for (k = 0; k < Col; k+=1)
{
neon_c = vmovq_n_u32(0);
for (j = 0; j < T; j+=4)
{
int j_T = j * T + i;
int k_Row = k * Row;
neon_a0 = vld1q_u32(A + j_T);
j_T+=Row;
neon_a1 = vld1q_u32(A + j_T);
j_T+=Row;
neon_a2 = vld1q_u32(A + j_T);
j_T+=Row;
neon_a3 = vld1q_u32(A + j_T);
neon_b = vld1q_u32(B + k_Row + j);
neon_b0 = vdupq_n_u32(vgetq_lane_u32(neon_b, 0));
neon_b1 = vdupq_n_u32(vgetq_lane_u32(neon_b, 1));
neon_b2 = vdupq_n_u32(vgetq_lane_u32(neon_b, 2));
neon_b3 = vdupq_n_u32(vgetq_lane_u32(neon_b, 3));
neon_c = vaddq_u32(vmulq_u32(neon_a0, neon_b0), neon_c);
neon_c = vaddq_u32(vmulq_u32(neon_a1, neon_b1), neon_c);
neon_c = vaddq_u32(vmulq_u32(neon_a2, neon_b2), neon_c);
neon_c = vaddq_u32(vmulq_u32(neon_a3, neon_b3), neon_c);
vst1q_lane_u32(C + k_Row + i, neon_c, 0);
vst1q_lane_u32(C + k_Row + i + 1, neon_c, 1);
vst1q_lane_u32(C + k_Row + i + 2, neon_c, 2);
vst1q_lane_u32(C + k_Row + i + 3, neon_c, 3);
}
}
}
}
void test_vmulQu32 (void)
{
uint32x4_t out_uint32x4_t;
uint32x4_t arg0_uint32x4_t;
uint32x4_t arg1_uint32x4_t;
out_uint32x4_t = vmulq_u32 (arg0_uint32x4_t, arg1_uint32x4_t);
}
/* u32x4 mv mul */
void mw_neon_mv_mul_u32x4(unsigned int * A, int Row, int T, unsigned int * B, unsigned int * C)
{
int i = 0;
int k = 0;
uint32x4_t neon_b, neon_c;
uint32x4_t neon_a0, neon_a1, neon_a2, neon_a3;
uint32x4_t neon_b0, neon_b1, neon_b2, neon_b3;
for (i = 0; i < Row; i+=4)
{
neon_c = vmovq_n_u32(0);
for (k = 0; k < T; k+=4)
{
int j = k * T + i;
neon_a0 = vld1q_u32(A + j);
j+=Row;
neon_a1 = vld1q_u32(A + j);
j+=Row;
neon_a2 = vld1q_u32(A + j);
j+=Row;
neon_a3 = vld1q_u32(A + j);
neon_b = vld1q_u32(B + k);
neon_b0 = vdupq_n_u32(vgetq_lane_u32(neon_b, 0));
neon_b1 = vdupq_n_u32(vgetq_lane_u32(neon_b, 1));
neon_b2 = vdupq_n_u32(vgetq_lane_u32(neon_b, 2));
neon_b3 = vdupq_n_u32(vgetq_lane_u32(neon_b, 3));
neon_c = vaddq_u32(vmulq_u32(neon_a0, neon_b0), neon_c);
neon_c = vaddq_u32(vmulq_u32(neon_a1, neon_b1), neon_c);
neon_c = vaddq_u32(vmulq_u32(neon_a2, neon_b2), neon_c);
neon_c = vaddq_u32(vmulq_u32(neon_a3, neon_b3), neon_c);
}
vst1q_u32(C + i, neon_c);
}
}
void SkRGB16BlitterBlitV_neon(uint16_t* device,
int height,
size_t deviceRB,
unsigned scale,
uint32_t src32) {
if (height >= 8)
{
uint16_t* dst = device;
// prepare constants
uint16x8_t vdev = vdupq_n_u16(0);
uint16x8_t vmaskq_g16 = vdupq_n_u16(SK_G16_MASK_IN_PLACE);
uint16x8_t vmaskq_ng16 = vdupq_n_u16(~SK_G16_MASK_IN_PLACE);
uint32x4_t vsrc32 = vdupq_n_u32(src32);
uint32x4_t vscale5 = vdupq_n_u32((uint32_t)scale);
while (height >= 8){
LOAD_LANE_16(vdev, 0)
LOAD_LANE_16(vdev, 1)
LOAD_LANE_16(vdev, 2)
LOAD_LANE_16(vdev, 3)
LOAD_LANE_16(vdev, 4)
LOAD_LANE_16(vdev, 5)
LOAD_LANE_16(vdev, 6)
LOAD_LANE_16(vdev, 7)
// Expand_rgb_16
uint16x8x2_t vdst = vzipq_u16((vdev & vmaskq_ng16), (vdev & vmaskq_g16));
uint32x4_t vdst32_lo = vmulq_u32(vreinterpretq_u32_u16(vdst.val[0]), vscale5);
uint32x4_t vdst32_hi = vmulq_u32(vreinterpretq_u32_u16(vdst.val[1]), vscale5);
// Compact_rgb_16
vdst32_lo = vaddq_u32(vdst32_lo, vsrc32);
vdst32_hi = vaddq_u32(vdst32_hi, vsrc32);
vdst32_lo = vshrq_n_u32(vdst32_lo, 5);
vdst32_hi = vshrq_n_u32(vdst32_hi, 5);
uint16x4_t vtmp_lo = vmovn_u32(vdst32_lo) & vget_low_u16(vmaskq_ng16);
uint16x4_t vtmp_hi = vshrn_n_u32(vdst32_lo, 16) & vget_low_u16(vmaskq_g16);
uint16x4_t vdst16_lo = vorr_u16(vtmp_lo, vtmp_hi);
vtmp_lo = vmovn_u32(vdst32_hi) & vget_low_u16(vmaskq_ng16);
vtmp_hi = vshrn_n_u32(vdst32_hi, 16) & vget_low_u16(vmaskq_g16);
uint16x4_t vdst16_hi = vorr_u16(vtmp_lo, vtmp_hi);
STORE_LANE_16(vdst16_lo, 0)
STORE_LANE_16(vdst16_lo, 1)
STORE_LANE_16(vdst16_lo, 2)
STORE_LANE_16(vdst16_lo, 3)
STORE_LANE_16(vdst16_hi, 0)
STORE_LANE_16(vdst16_hi, 1)
STORE_LANE_16(vdst16_hi, 2)
STORE_LANE_16(vdst16_hi, 3)
height -= 8;
}
}
while (height != 0){
uint32_t dst32 = SkExpand_rgb_16(*device) * scale;
*device = SkCompact_rgb_16((src32 + dst32) >> 5);
device = (uint16_t*)((char*)device + deviceRB);
height--;
}
}
