#define VARIANCE_LARGE_WxH(sse, diff, shift) \
sse - (((int64_t)diff * diff) >> shift)
+static uint32_t avg_sse_diff_4width_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ const uint8_t *sec_pred,
+ int32_t height,
+ int32_t *diff) {
+ int32_t ht_cnt;
+ uint32_t src0, src1, src2, src3;
+ uint32_t ref0, ref1, ref2, ref3;
+ v16u8 pred, src = { 0 };
+ v16u8 ref = { 0 };
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ for (ht_cnt = (height >> 2); ht_cnt--;) {
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ LW4(src_ptr, src_stride, src0, src1, src2, src3);
+ src_ptr += (4 * src_stride);
+ LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+ ref_ptr += (4 * ref_stride);
+
+ INSERT_W4_UB(src0, src1, src2, src3, src);
+ INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+ src = __msa_aver_u_b(src, pred);
+ CALC_MSE_AVG_B(src, ref, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_8width_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ const uint8_t *sec_pred,
+ int32_t height,
+ int32_t *diff) {
+ int32_t ht_cnt;
+ v16u8 src0, src1, src2, src3;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16u8 pred0, pred1;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ for (ht_cnt = (height >> 2); ht_cnt--;) {
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+ src_ptr += (4 * src_stride);
+ LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+ ref_ptr += (4 * ref_stride);
+
+ PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+ src0, src1, ref0, ref1);
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg);
+ CALC_MSE_AVG_B(src1, ref1, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_16width_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ const uint8_t *sec_pred,
+ int32_t height,
+ int32_t *diff) {
+ int32_t ht_cnt;
+ v16u8 src, ref, pred;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ for (ht_cnt = (height >> 2); ht_cnt--;) {
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ src = LD_UB(src_ptr);
+ src_ptr += src_stride;
+ ref = LD_UB(ref_ptr);
+ ref_ptr += ref_stride;
+ src = __msa_aver_u_b(src, pred);
+ CALC_MSE_AVG_B(src, ref, var, avg);
+
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ src = LD_UB(src_ptr);
+ src_ptr += src_stride;
+ ref = LD_UB(ref_ptr);
+ ref_ptr += ref_stride;
+ src = __msa_aver_u_b(src, pred);
+ CALC_MSE_AVG_B(src, ref, var, avg);
+
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ src = LD_UB(src_ptr);
+ src_ptr += src_stride;
+ ref = LD_UB(ref_ptr);
+ ref_ptr += ref_stride;
+ src = __msa_aver_u_b(src, pred);
+ CALC_MSE_AVG_B(src, ref, var, avg);
+
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ src = LD_UB(src_ptr);
+ src_ptr += src_stride;
+ ref = LD_UB(ref_ptr);
+ ref_ptr += ref_stride;
+ src = __msa_aver_u_b(src, pred);
+ CALC_MSE_AVG_B(src, ref, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_32width_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ const uint8_t *sec_pred,
+ int32_t height,
+ int32_t *diff) {
+ int32_t ht_cnt;
+ v16u8 src0, src1, ref0, ref1, pred0, pred1;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ for (ht_cnt = (height >> 2); ht_cnt--;) {
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg);
+ CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg);
+ CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg);
+ CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg);
+ CALC_MSE_AVG_B(src1, ref1, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_32x64_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ const uint8_t *sec_pred,
+ int32_t *diff) {
+ int32_t ht_cnt;
+ v16u8 src0, src1, ref0, ref1, pred0, pred1;
+ v8i16 avg0 = { 0 };
+ v8i16 avg1 = { 0 };
+ v4i32 vec, var = { 0 };
+
+ for (ht_cnt = 16; ht_cnt--;) {
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB2(src_ptr, 16, src0, src1);
+ src_ptr += src_stride;
+ LD_UB2(ref_ptr, 16, ref0, ref1);
+ ref_ptr += ref_stride;
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+ }
+
+ vec = __msa_hadd_s_w(avg0, avg0);
+ vec += __msa_hadd_s_w(avg1, avg1);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_64x32_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ const uint8_t *sec_pred,
+ int32_t *diff) {
+ int32_t ht_cnt;
+ v16u8 src0, src1, src2, src3;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16u8 pred0, pred1, pred2, pred3;
+ v8i16 avg0 = { 0 };
+ v8i16 avg1 = { 0 };
+ v4i32 vec, var = { 0 };
+
+ for (ht_cnt = 16; ht_cnt--;) {
+ LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+ sec_pred += 64;
+ LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+ src_ptr += src_stride;
+ LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+ ref_ptr += ref_stride;
+ AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+ src0, src1, src2, src3);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src2, ref2, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+ CALC_MSE_AVG_B(src3, ref3, var, avg1);
+
+ LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+ sec_pred += 64;
+ LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+ src_ptr += src_stride;
+ LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+ ref_ptr += ref_stride;
+ AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+ src0, src1, src2, src3);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src2, ref2, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+ CALC_MSE_AVG_B(src3, ref3, var, avg1);
+ }
+
+ vec = __msa_hadd_s_w(avg0, avg0);
+ vec += __msa_hadd_s_w(avg1, avg1);
+
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_64x64_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ const uint8_t *sec_pred,
+ int32_t *diff) {
+ int32_t ht_cnt;
+ v16u8 src0, src1, src2, src3;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16u8 pred0, pred1, pred2, pred3;
+ v8i16 avg0 = { 0 };
+ v8i16 avg1 = { 0 };
+ v8i16 avg2 = { 0 };
+ v8i16 avg3 = { 0 };
+ v4i32 vec, var = { 0 };
+
+ for (ht_cnt = 32; ht_cnt--;) {
+ LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+ sec_pred += 64;
+ LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+ src_ptr += src_stride;
+ LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+ ref_ptr += ref_stride;
+ AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+ src0, src1, src2, src3);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+ CALC_MSE_AVG_B(src2, ref2, var, avg2);
+ CALC_MSE_AVG_B(src3, ref3, var, avg3);
+
+ LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+ sec_pred += 64;
+ LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+ src_ptr += src_stride;
+ LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+ ref_ptr += ref_stride;
+ AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+ src0, src1, src2, src3);
+ CALC_MSE_AVG_B(src0, ref0, var, avg0);
+ CALC_MSE_AVG_B(src1, ref1, var, avg1);
+ CALC_MSE_AVG_B(src2, ref2, var, avg2);
+ CALC_MSE_AVG_B(src3, ref3, var, avg3);
+ }
+
+ vec = __msa_hadd_s_w(avg0, avg0);
+ vec += __msa_hadd_s_w(avg1, avg1);
+ vec += __msa_hadd_s_w(avg2, avg2);
+ vec += __msa_hadd_s_w(avg3, avg3);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
static uint32_t sub_pixel_sse_diff_4width_h_msa(const uint8_t *src,
int32_t src_stride,
const uint8_t *dst,
src0 = src4;
- /* loop runs height/4 */
CALC_MSE_AVG_B(out0, ref0, var, avg);
CALC_MSE_AVG_B(out1, ref1, var, avg);
CALC_MSE_AVG_B(out2, ref2, var, avg);
return sse;
}
-#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4);
-#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5);
-#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5);
-#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6);
-#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7);
-#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7);
-#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8);
+static uint32_t sub_pixel_avg_sse_diff_4width_h_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ uint32_t ref0, ref1, ref2, ref3;
+ v16u8 out, pred, filt0, ref = { 0 };
+ v16i8 src0, src1, src2, src3;
+ v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+ v8u16 vec0, vec1, vec2, vec3;
+ v8u16 const255;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
-#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
-#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
-#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10);
-#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
-#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
-#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12);
+ filtval = LH(filter);
+ filt0 = (v16u8)__msa_fill_h(filtval);
-#define VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(wd, ht) \
-uint32_t vp9_sub_pixel_variance##wd##x##ht##_msa(const uint8_t *src, \
- int32_t src_stride, \
- int32_t xoffset, \
- int32_t yoffset, \
- const uint8_t *ref, \
- int32_t ref_stride, \
- uint32_t *sse) { \
- int32_t diff; \
- uint32_t var; \
- const int8_t *h_filter = vp9_bilinear_filters_msa[xoffset - 1]; \
- const int8_t *v_filter = vp9_bilinear_filters_msa[yoffset - 1]; \
- \
- if (yoffset) { \
- if (xoffset) { \
- *sse = sub_pixel_sse_diff_##wd##width_hv_msa(src, src_stride, \
- ref, ref_stride, \
- h_filter, v_filter, \
- ht, &diff); \
- } else { \
- *sse = sub_pixel_sse_diff_##wd##width_v_msa(src, src_stride, \
- ref, ref_stride, \
- v_filter, ht, &diff); \
- } \
- \
- var = VARIANCE_##wd##Wx##ht##H(*sse, diff); \
- } else { \
- if (xoffset) { \
- *sse = sub_pixel_sse_diff_##wd##width_h_msa(src, src_stride, \
- ref, ref_stride, \
- h_filter, ht, &diff); \
- \
- var = VARIANCE_##wd##Wx##ht##H(*sse, diff); \
- } else { \
- var = vpx_variance##wd##x##ht##_msa(src, src_stride, \
- ref, ref_stride, sse); \
- } \
- } \
- \
- return var; \
+ const255 = (v8u16)__msa_ldi_h(255);
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_SB4(src, src_stride, src0, src1, src2, src3);
+ src += (4 * src_stride);
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
+
+ INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+ VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+ VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+ DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+ vec0, vec1, vec2, vec3);
+ SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+ MIN_UH4_UH(vec0, vec1, vec2, vec3, const255);
+ PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3,
+ src0, src1, src2, src3);
+ ILVEV_W2_SB(src0, src1, src2, src3, src0, src2);
+ out = (v16u8)__msa_ilvev_d((v2i64)src2, (v2i64)src0);
+ out = __msa_aver_u_b(out, pred);
+ CALC_MSE_AVG_B(out, ref, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
}
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 4);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 8);
+static uint32_t sub_pixel_avg_sse_diff_8width_h_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ v16u8 out, pred, filt0;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16i8 src0, src1, src2, src3;
+ v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+ v8u16 vec0, vec1, vec2, vec3;
+ v8u16 const255;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 4);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 8);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 16);
+ filtval = LH(filter);
+ filt0 = (v16u8)__msa_fill_h(filtval);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 8);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 16);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 32);
+ const255 = (v8u16)__msa_ldi_h(255);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 16);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 32);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 64);
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_SB4(src, src_stride, src0, src1, src2, src3);
+ src += (4 * src_stride);
+ LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 32);
-VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 64);
+ PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+ VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+ VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+ DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+ vec0, vec1, vec2, vec3);
+ SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+ MIN_UH4_UH(vec0, vec1, vec2, vec3, const255);
+ PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3,
+ src0, src1, src2, src3);
+ out = (v16u8)__msa_ilvev_d((v2i64)src1, (v2i64)src0);
+
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ out = __msa_aver_u_b(out, pred);
+ CALC_MSE_AVG_B(out, ref0, var, avg);
+ out = (v16u8)__msa_ilvev_d((v2i64)src3, (v2i64)src2);
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ out = __msa_aver_u_b(out, pred);
+ CALC_MSE_AVG_B(out, ref1, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_h_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff,
+ int32_t width) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+ v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+ v16u8 dst0, dst1, dst2, dst3;
+ v16u8 tmp0, tmp1, tmp2, tmp3;
+ v16u8 pred0, pred1, pred2, pred3, filt0;
+ v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+ v8u16 out0, out1, out2, out3, out4, out5, out6, out7;
+ v8u16 const255;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ filtval = LH(filter);
+ filt0 = (v16u8)__msa_fill_h(filtval);
+
+ const255 = (v8u16)__msa_ldi_h(255);
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_SB4(src, src_stride, src0, src2, src4, src6);
+ LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+ src += (4 * src_stride);
+ LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+ dst += (4 * dst_stride);
+ LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+ sec_pred += (4 * width);
+
+ VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+ VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+ VSHF_B2_UH(src4, src4, src5, src5, mask, mask, vec4, vec5);
+ VSHF_B2_UH(src6, src6, src7, src7, mask, mask, vec6, vec7);
+ DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+ out0, out1, out2, out3);
+ DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0,
+ out4, out5, out6, out7);
+ SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+ SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+ MIN_UH4_UH(out0, out1, out2, out3, const255);
+ MIN_UH4_UH(out4, out5, out6, out7, const255);
+ PCKEV_B4_UB(out1, out0, out3, out2, out5, out4, out7, out6,
+ tmp0, tmp1, tmp2, tmp3);
+ AVER_UB4_UB(tmp0, pred0, tmp1, pred1, tmp2, pred2, tmp3, pred3,
+ tmp0, tmp1, tmp2, tmp3);
+
+ CALC_MSE_AVG_B(tmp0, dst0, var, avg);
+ CALC_MSE_AVG_B(tmp1, dst1, var, avg);
+ CALC_MSE_AVG_B(tmp2, dst2, var, avg);
+ CALC_MSE_AVG_B(tmp3, dst3, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_h_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ return subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter, height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_h_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ uint32_t loop_cnt, sse = 0;
+ int32_t diff0[2];
+
+ for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+ sse += subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter, height,
+ &diff0[loop_cnt], 32);
+ src += 16;
+ dst += 16;
+ sec_pred += 16;
+ }
+
+ *diff = diff0[0] + diff0[1];
+
+ return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_h_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ uint32_t loop_cnt, sse = 0;
+ int32_t diff0[4];
+
+ for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+ sse += subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter, height,
+ &diff0[loop_cnt], 64);
+ src += 16;
+ dst += 16;
+ sec_pred += 16;
+ }
+
+ *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+ return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_v_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ uint32_t ref0, ref1, ref2, ref3;
+ v16u8 src0, src1, src2, src3, src4;
+ v16u8 src10_r, src32_r, src21_r, src43_r;
+ v16u8 out, pred, ref = { 0 };
+ v16u8 src2110, src4332, filt0;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+ v8u16 tmp0, tmp1;
+
+ filtval = LH(filter);
+ filt0 = (v16u8)__msa_fill_h(filtval);
+
+ src0 = LD_UB(src);
+ src += src_stride;
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_UB4(src, src_stride, src1, src2, src3, src4);
+ src += (4 * src_stride);
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
+
+ INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+ ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
+ src10_r, src21_r, src32_r, src43_r);
+ ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+ DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+
+ out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+ out = __msa_aver_u_b(out, pred);
+ CALC_MSE_AVG_B(out, ref, var, avg);
+ src0 = src4;
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_v_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ v16u8 src0, src1, src2, src3, src4;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16u8 pred0, pred1, filt0;
+ v8u16 vec0, vec1, vec2, vec3;
+ v8u16 tmp0, tmp1, tmp2, tmp3;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ filtval = LH(filter);
+ filt0 = (v16u8)__msa_fill_h(filtval);
+
+ src0 = LD_UB(src);
+ src += src_stride;
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_UB4(src, src_stride, src1, src2, src3, src4);
+ src += (4 * src_stride);
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
+ PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+ ILVR_B4_UH(src1, src0, src2, src1, src3, src2, src4, src3,
+ vec0, vec1, vec2, vec3);
+ DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+ tmp0, tmp1, tmp2, tmp3);
+ SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+ SAT_UH4_UH(tmp0, tmp1, tmp2, tmp3, 7);
+ PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1);
+ AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+ CALC_MSE_AVG_B(src0, ref0, var, avg);
+ CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+ src0 = src4;
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_v_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff,
+ int32_t width) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16u8 pred0, pred1, pred2, pred3;
+ v16u8 src0, src1, src2, src3, src4;
+ v16u8 out0, out1, out2, out3, filt0;
+ v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+ v8u16 tmp0, tmp1, tmp2, tmp3;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ filtval = LH(filter);
+ filt0 = (v16u8)__msa_fill_h(filtval);
+
+ src0 = LD_UB(src);
+ src += src_stride;
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_UB4(src, src_stride, src1, src2, src3, src4);
+ src += (4 * src_stride);
+ LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+ sec_pred += (4 * width);
+
+ ILVR_B2_UH(src1, src0, src2, src1, vec0, vec2);
+ ILVL_B2_UH(src1, src0, src2, src1, vec1, vec3);
+ DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+ ILVR_B2_UH(src3, src2, src4, src3, vec4, vec6);
+ ILVL_B2_UH(src3, src2, src4, src3, vec5, vec7);
+ DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+ SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+ SAT_UH2_UH(tmp2, tmp3, 7);
+ out1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+ DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+ DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+ SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+ SAT_UH2_UH(tmp2, tmp3, 7);
+ out3 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+ src0 = src4;
+ LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
+
+ AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3,
+ out0, out1, out2, out3);
+
+ CALC_MSE_AVG_B(out0, ref0, var, avg);
+ CALC_MSE_AVG_B(out1, ref1, var, avg);
+ CALC_MSE_AVG_B(out2, ref2, var, avg);
+ CALC_MSE_AVG_B(out3, ref3, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_v_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ return subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter, height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_v_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ uint32_t loop_cnt, sse = 0;
+ int32_t diff0[2];
+
+ for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+ sse += subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter, height,
+ &diff0[loop_cnt], 32);
+ src += 16;
+ dst += 16;
+ sec_pred += 16;
+ }
+
+ *diff = diff0[0] + diff0[1];
+
+ return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_v_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter,
+ int32_t height,
+ int32_t *diff) {
+ uint32_t loop_cnt, sse = 0;
+ int32_t diff0[4];
+
+ for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+ sse += subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter, height,
+ &diff0[loop_cnt], 64);
+ src += 16;
+ dst += 16;
+ sec_pred += 16;
+ }
+
+ *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+ return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_hv_msa(
+ const uint8_t *src, int32_t src_stride,
+ const uint8_t *dst, int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter_horiz, const int8_t *filter_vert,
+ int32_t height, int32_t *diff) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ uint32_t ref0, ref1, ref2, ref3;
+ v16u8 src0, src1, src2, src3, src4;
+ v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20 };
+ v16u8 filt_hz, filt_vt, vec0, vec1;
+ v16u8 out, pred, ref = { 0 };
+ v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ filtval = LH(filter_horiz);
+ filt_hz = (v16u8)__msa_fill_h(filtval);
+ filtval = LH(filter_vert);
+ filt_vt = (v16u8)__msa_fill_h(filtval);
+
+ src0 = LD_UB(src);
+ src += src_stride;
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_UB4(src, src_stride, src1, src2, src3, src4);
+ src += (4 * src_stride);
+ pred = LD_UB(sec_pred);
+ sec_pred += 16;
+ LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
+ INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+ hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+ hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+ hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+ hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+ hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+ ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+ DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+ out = __msa_aver_u_b(out, pred);
+ CALC_MSE_AVG_B(out, ref, var, avg);
+ src0 = src4;
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_hv_msa(
+ const uint8_t *src, int32_t src_stride,
+ const uint8_t *dst, int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter_horiz, const int8_t *filter_vert,
+ int32_t height, int32_t *diff) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16u8 src0, src1, src2, src3, src4;
+ v16u8 pred0, pred1, out0, out1;
+ v16u8 filt_hz, filt_vt, vec0;
+ v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+ v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ filtval = LH(filter_horiz);
+ filt_hz = (v16u8)__msa_fill_h(filtval);
+ filtval = LH(filter_vert);
+ filt_vt = (v16u8)__msa_fill_h(filtval);
+
+ src0 = LD_UB(src);
+ src += src_stride;
+ hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_UB4(src, src_stride, src1, src2, src3, src4);
+ src += (4 * src_stride);
+ LD_UB2(sec_pred, 16, pred0, pred1);
+ sec_pred += 32;
+ LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
+
+ PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+ hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+ vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+ tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+ hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+
+ vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+ tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+
+ vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+ tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+ hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+
+ vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+ tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+ SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+ SAT_UH2_UH(tmp2, tmp3, 7);
+ PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+ AVER_UB2_UB(out0, pred0, out1, pred1, out0, out1);
+
+ CALC_MSE_AVG_B(out0, ref0, var, avg);
+ CALC_MSE_AVG_B(out1, ref1, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_hv_msa(const uint8_t *src,
+ int32_t src_stride,
+ const uint8_t *dst,
+ int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter_horiz,
+ const int8_t *filter_vert,
+ int32_t height,
+ int32_t *diff,
+ int32_t width) {
+ int16_t filtval;
+ uint32_t loop_cnt;
+ v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+ v16u8 ref0, ref1, ref2, ref3;
+ v16u8 pred0, pred1, pred2, pred3;
+ v16u8 out0, out1, out2, out3;
+ v16u8 filt_hz, filt_vt, vec0, vec1;
+ v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+ v8u16 hz_out0, hz_out1, hz_out2, hz_out3, tmp0, tmp1;
+ v8i16 avg = { 0 };
+ v4i32 vec, var = { 0 };
+
+ filtval = LH(filter_horiz);
+ filt_hz = (v16u8)__msa_fill_h(filtval);
+ filtval = LH(filter_vert);
+ filt_vt = (v16u8)__msa_fill_h(filtval);
+
+ LD_UB2(src, 8, src0, src1);
+ src += src_stride;
+
+ hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+ hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+ for (loop_cnt = (height >> 2); loop_cnt--;) {
+ LD_UB4(src, src_stride, src0, src2, src4, src6);
+ LD_UB4(src + 8, src_stride, src1, src3, src5, src7);
+ src += (4 * src_stride);
+ LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+ sec_pred += (4 * width);
+
+ hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+ hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+ ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+ DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+ hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+ hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+ ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+ DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ out1 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+ hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+ hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+ ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+ DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+ hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+ hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+ ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+ DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+ SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+ SAT_UH2_UH(tmp0, tmp1, 7);
+ out3 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+ LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+ dst += (4 * dst_stride);
+
+ AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3,
+ out0, out1, out2, out3);
+
+ CALC_MSE_AVG_B(out0, ref0, var, avg);
+ CALC_MSE_AVG_B(out1, ref1, var, avg);
+ CALC_MSE_AVG_B(out2, ref2, var, avg);
+ CALC_MSE_AVG_B(out3, ref3, var, avg);
+ }
+
+ vec = __msa_hadd_s_w(avg, avg);
+ *diff = HADD_SW_S32(vec);
+
+ return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_hv_msa(
+ const uint8_t *src, int32_t src_stride,
+ const uint8_t *dst, int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter_horiz, const int8_t *filter_vert,
+ int32_t height, int32_t *diff) {
+ return subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter_horiz, filter_vert,
+ height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_hv_msa(
+ const uint8_t *src, int32_t src_stride,
+ const uint8_t *dst, int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter_horiz, const int8_t *filter_vert,
+ int32_t height, int32_t *diff) {
+ uint32_t loop_cnt, sse = 0;
+ int32_t diff0[2];
+
+ for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+ sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter_horiz, filter_vert,
+ height, &diff0[loop_cnt], 32);
+ src += 16;
+ dst += 16;
+ sec_pred += 16;
+ }
+
+ *diff = diff0[0] + diff0[1];
+
+ return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_hv_msa(
+ const uint8_t *src, int32_t src_stride,
+ const uint8_t *dst, int32_t dst_stride,
+ const uint8_t *sec_pred,
+ const int8_t *filter_horiz, const int8_t *filter_vert,
+ int32_t height, int32_t *diff) {
+ uint32_t loop_cnt, sse = 0;
+ int32_t diff0[4];
+
+ for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+ sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+ sec_pred, filter_horiz, filter_vert,
+ height, &diff0[loop_cnt], 64);
+ src += 16;
+ dst += 16;
+ sec_pred += 16;
+ }
+
+ *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+ return sse;
+}
+
+#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4);
+#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6);
+#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8);
+
+#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10);
+#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12);
+
+#define VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(wd, ht) \
+uint32_t vp9_sub_pixel_variance##wd##x##ht##_msa(const uint8_t *src, \
+ int32_t src_stride, \
+ int32_t xoffset, \
+ int32_t yoffset, \
+ const uint8_t *ref, \
+ int32_t ref_stride, \
+ uint32_t *sse) { \
+ int32_t diff; \
+ uint32_t var; \
+ const int8_t *h_filter = vp9_bilinear_filters_msa[xoffset - 1]; \
+ const int8_t *v_filter = vp9_bilinear_filters_msa[yoffset - 1]; \
+ \
+ if (yoffset) { \
+ if (xoffset) { \
+ *sse = sub_pixel_sse_diff_##wd##width_hv_msa(src, src_stride, \
+ ref, ref_stride, \
+ h_filter, v_filter, \
+ ht, &diff); \
+ } else { \
+ *sse = sub_pixel_sse_diff_##wd##width_v_msa(src, src_stride, \
+ ref, ref_stride, \
+ v_filter, ht, &diff); \
+ } \
+ \
+ var = VARIANCE_##wd##Wx##ht##H(*sse, diff); \
+ } else { \
+ if (xoffset) { \
+ *sse = sub_pixel_sse_diff_##wd##width_h_msa(src, src_stride, \
+ ref, ref_stride, \
+ h_filter, ht, &diff); \
+ \
+ var = VARIANCE_##wd##Wx##ht##H(*sse, diff); \
+ } else { \
+ var = vpx_variance##wd##x##ht##_msa(src, src_stride, \
+ ref, ref_stride, sse); \
+ } \
+ } \
+ \
+ return var; \
+}
+
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 4);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 8);
+
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 4);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 8);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 16);
+
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 8);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 16);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 32);
+
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 16);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 32);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 64);
+
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 32);
+VP9_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 64);
+
+#define VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(wd, ht) \
+uint32_t vp9_sub_pixel_avg_variance##wd##x##ht##_msa( \
+ const uint8_t *src_ptr, int32_t src_stride, \
+ int32_t xoffset, int32_t yoffset, \
+ const uint8_t *ref_ptr, int32_t ref_stride, \
+ uint32_t *sse, const uint8_t *sec_pred) { \
+ int32_t diff; \
+ const int8_t *h_filter = vp9_bilinear_filters_msa[xoffset - 1]; \
+ const int8_t *v_filter = vp9_bilinear_filters_msa[yoffset - 1]; \
+ \
+ if (yoffset) { \
+ if (xoffset) { \
+ *sse = sub_pixel_avg_sse_diff_##wd##width_hv_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, h_filter, \
+ v_filter, ht, &diff); \
+ } else { \
+ *sse = sub_pixel_avg_sse_diff_##wd##width_v_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, v_filter, \
+ ht, &diff); \
+ } \
+ } else { \
+ if (xoffset) { \
+ *sse = sub_pixel_avg_sse_diff_##wd##width_h_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, h_filter, \
+ ht, &diff); \
+ } else { \
+ *sse = avg_sse_diff_##wd##width_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, ht, &diff); \
+ } \
+ } \
+ \
+ return VARIANCE_##wd##Wx##ht##H(*sse, diff); \
+}
+
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 4);
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 8);
+
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 4);
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 8);
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 16);
+
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 8);
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 16);
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 32);
+
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 16);
+VP9_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 32);
+
+uint32_t vp9_sub_pixel_avg_variance32x64_msa(const uint8_t *src_ptr,
+ int32_t src_stride,
+ int32_t xoffset,
+ int32_t yoffset,
+ const uint8_t *ref_ptr,
+ int32_t ref_stride,
+ uint32_t *sse,
+ const uint8_t *sec_pred) {
+ int32_t diff;
+ const int8_t *h_filter = vp9_bilinear_filters_msa[xoffset - 1];
+ const int8_t *v_filter = vp9_bilinear_filters_msa[yoffset - 1];
+
+ if (yoffset) {
+ if (xoffset) {
+ *sse = sub_pixel_avg_sse_diff_32width_hv_msa(src_ptr, src_stride,
+ ref_ptr, ref_stride,
+ sec_pred, h_filter,
+ v_filter, 64, &diff);
+ } else {
+ *sse = sub_pixel_avg_sse_diff_32width_v_msa(src_ptr, src_stride,
+ ref_ptr, ref_stride,
+ sec_pred, v_filter,
+ 64, &diff);
+ }
+ } else {
+ if (xoffset) {
+ *sse = sub_pixel_avg_sse_diff_32width_h_msa(src_ptr, src_stride,
+ ref_ptr, ref_stride,
+ sec_pred, h_filter,
+ 64, &diff);
+ } else {
+ *sse = avg_sse_diff_32x64_msa(src_ptr, src_stride, ref_ptr, ref_stride,
+ sec_pred, &diff);
+ }
+ }
+
+ return VARIANCE_32Wx64H(*sse, diff);
+}
+
+#define VP9_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(ht) \
+uint32_t vp9_sub_pixel_avg_variance64x##ht##_msa(const uint8_t *src_ptr, \
+ int32_t src_stride, \
+ int32_t xoffset, \
+ int32_t yoffset, \
+ const uint8_t *ref_ptr, \
+ int32_t ref_stride, \
+ uint32_t *sse, \
+ const uint8_t *sec_pred) { \
+ int32_t diff; \
+ const int8_t *h_filter = vp9_bilinear_filters_msa[xoffset - 1]; \
+ const int8_t *v_filter = vp9_bilinear_filters_msa[yoffset - 1]; \
+ \
+ if (yoffset) { \
+ if (xoffset) { \
+ *sse = sub_pixel_avg_sse_diff_64width_hv_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, h_filter, \
+ v_filter, ht, &diff); \
+ } else { \
+ *sse = sub_pixel_avg_sse_diff_64width_v_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, v_filter, \
+ ht, &diff); \
+ } \
+ } else { \
+ if (xoffset) { \
+ *sse = sub_pixel_avg_sse_diff_64width_h_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, h_filter, \
+ ht, &diff); \
+ } else { \
+ *sse = avg_sse_diff_64x##ht##_msa(src_ptr, src_stride, \
+ ref_ptr, ref_stride, \
+ sec_pred, &diff); \
+ } \
+ } \
+ \
+ return VARIANCE_64Wx##ht##H(*sse, diff); \
+}
+
+VP9_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(32);
+VP9_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(64);
projects / libvpx / commitdiff