void test_vqdmulhQ_ns16 (void) { int16x8_t out_int16x8_t; int16x8_t arg0_int16x8_t; int16_t arg1_int16_t; out_int16x8_t = vqdmulhq_n_s16 (arg0_int16x8_t, arg1_int16_t); }
void vp8_short_idct4x4llm_neon(int16_t *input, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride) { int i; uint32x2_t d6u32 = vdup_n_u32(0); uint8x8_t d1u8; int16x4_t d2, d3, d4, d5, d10, d11, d12, d13; uint16x8_t q1u16; int16x8_t q1s16, q2s16, q3s16, q4s16; int32x2x2_t v2tmp0, v2tmp1; int16x4x2_t v2tmp2, v2tmp3; d2 = vld1_s16(input); d3 = vld1_s16(input + 4); d4 = vld1_s16(input + 8); d5 = vld1_s16(input + 12); // 1st for loop q1s16 = vcombine_s16(d2, d4); // Swap d3 d4 here q2s16 = vcombine_s16(d3, d5); q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2); q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1); d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // a1 d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // b1 q3s16 = vshrq_n_s16(q3s16, 1); q4s16 = vshrq_n_s16(q4s16, 1); q3s16 = vqaddq_s16(q3s16, q2s16); q4s16 = vqaddq_s16(q4s16, q2s16); d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16)); // c1 d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16)); // d1 d2 = vqadd_s16(d12, d11); d3 = vqadd_s16(d13, d10); d4 = vqsub_s16(d13, d10); d5 = vqsub_s16(d12, d11); v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4)); v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5)); v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]), vreinterpret_s16_s32(v2tmp1.val[0])); v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]), vreinterpret_s16_s32(v2tmp1.val[1])); // 2nd for loop q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp3.val[0]); q2s16 = vcombine_s16(v2tmp2.val[1], v2tmp3.val[1]); q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2); q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1); d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // a1 d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16)); // b1 q3s16 = vshrq_n_s16(q3s16, 1); q4s16 = vshrq_n_s16(q4s16, 1); q3s16 = vqaddq_s16(q3s16, q2s16); q4s16 = vqaddq_s16(q4s16, q2s16); d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16)); // c1 d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16)); // d1 d2 = vqadd_s16(d12, d11); d3 = vqadd_s16(d13, d10); d4 = vqsub_s16(d13, d10); d5 = vqsub_s16(d12, d11); d2 = vrshr_n_s16(d2, 3); d3 = vrshr_n_s16(d3, 3); d4 = vrshr_n_s16(d4, 3); d5 = vrshr_n_s16(d5, 3); v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4)); v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5)); v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]), vreinterpret_s16_s32(v2tmp1.val[0])); v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]), vreinterpret_s16_s32(v2tmp1.val[1])); q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp2.val[1]); q2s16 = vcombine_s16(v2tmp3.val[0], v2tmp3.val[1]); // dc_only_idct_add for (i = 0; i < 2; i++, q1s16 = q2s16) { d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 0); pred_ptr += pred_stride; d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 1); pred_ptr += pred_stride; q1u16 = vaddw_u8(vreinterpretq_u16_s16(q1s16), vreinterpret_u8_u32(d6u32)); d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16)); vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 0); dst_ptr += dst_stride; vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 1); dst_ptr += dst_stride; } return; }
void idct_dequant_full_2x_neon( int16_t *q, int16_t *dq, unsigned char *dst, int stride) { unsigned char *dst0, *dst1; int32x2_t d28, d29, d30, d31; int16x8_t q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; int16x8_t qEmpty = vdupq_n_s16(0); int32x4x2_t q2tmp0, q2tmp1; int16x8x2_t q2tmp2, q2tmp3; int16x4_t dLow0, dLow1, dHigh0, dHigh1; d28 = d29 = d30 = d31 = vdup_n_s32(0); // load dq q0 = vld1q_s16(dq); dq += 8; q1 = vld1q_s16(dq); // load q q2 = vld1q_s16(q); vst1q_s16(q, qEmpty); q += 8; q3 = vld1q_s16(q); vst1q_s16(q, qEmpty); q += 8; q4 = vld1q_s16(q); vst1q_s16(q, qEmpty); q += 8; q5 = vld1q_s16(q); vst1q_s16(q, qEmpty); // load src from dst dst0 = dst; dst1 = dst + 4; d28 = vld1_lane_s32((const int32_t *)dst0, d28, 0); dst0 += stride; d28 = vld1_lane_s32((const int32_t *)dst1, d28, 1); dst1 += stride; d29 = vld1_lane_s32((const int32_t *)dst0, d29, 0); dst0 += stride; d29 = vld1_lane_s32((const int32_t *)dst1, d29, 1); dst1 += stride; d30 = vld1_lane_s32((const int32_t *)dst0, d30, 0); dst0 += stride; d30 = vld1_lane_s32((const int32_t *)dst1, d30, 1); dst1 += stride; d31 = vld1_lane_s32((const int32_t *)dst0, d31, 0); d31 = vld1_lane_s32((const int32_t *)dst1, d31, 1); q2 = vmulq_s16(q2, q0); q3 = vmulq_s16(q3, q1); q4 = vmulq_s16(q4, q0); q5 = vmulq_s16(q5, q1); // vswp dLow0 = vget_low_s16(q2); dHigh0 = vget_high_s16(q2); dLow1 = vget_low_s16(q4); dHigh1 = vget_high_s16(q4); q2 = vcombine_s16(dLow0, dLow1); q4 = vcombine_s16(dHigh0, dHigh1); dLow0 = vget_low_s16(q3); dHigh0 = vget_high_s16(q3); dLow1 = vget_low_s16(q5); dHigh1 = vget_high_s16(q5); q3 = vcombine_s16(dLow0, dLow1); q5 = vcombine_s16(dHigh0, dHigh1); q6 = vqdmulhq_n_s16(q4, sinpi8sqrt2); q7 = vqdmulhq_n_s16(q5, sinpi8sqrt2); q8 = vqdmulhq_n_s16(q4, cospi8sqrt2minus1); q9 = vqdmulhq_n_s16(q5, cospi8sqrt2minus1); q10 = vqaddq_s16(q2, q3); q11 = vqsubq_s16(q2, q3); q8 = vshrq_n_s16(q8, 1); q9 = vshrq_n_s16(q9, 1); q4 = vqaddq_s16(q4, q8); q5 = vqaddq_s16(q5, q9); q2 = vqsubq_s16(q6, q5); q3 = vqaddq_s16(q7, q4); q4 = vqaddq_s16(q10, q3); q5 = vqaddq_s16(q11, q2); q6 = vqsubq_s16(q11, q2); q7 = vqsubq_s16(q10, q3); q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6)); q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7)); q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]), vreinterpretq_s16_s32(q2tmp1.val[0])); q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]), vreinterpretq_s16_s32(q2tmp1.val[1])); // loop 2 q8 = vqdmulhq_n_s16(q2tmp2.val[1], sinpi8sqrt2); q9 = vqdmulhq_n_s16(q2tmp3.val[1], sinpi8sqrt2); q10 = vqdmulhq_n_s16(q2tmp2.val[1], cospi8sqrt2minus1); q11 = vqdmulhq_n_s16(q2tmp3.val[1], cospi8sqrt2minus1); q2 = vqaddq_s16(q2tmp2.val[0], q2tmp3.val[0]); q3 = vqsubq_s16(q2tmp2.val[0], q2tmp3.val[0]); q10 = vshrq_n_s16(q10, 1); q11 = vshrq_n_s16(q11, 1); q10 = vqaddq_s16(q2tmp2.val[1], q10); q11 = vqaddq_s16(q2tmp3.val[1], q11); q8 = vqsubq_s16(q8, q11); q9 = vqaddq_s16(q9, q10); q4 = vqaddq_s16(q2, q9); q5 = vqaddq_s16(q3, q8); q6 = vqsubq_s16(q3, q8); q7 = vqsubq_s16(q2, q9); q4 = vrshrq_n_s16(q4, 3); q5 = vrshrq_n_s16(q5, 3); q6 = vrshrq_n_s16(q6, 3); q7 = vrshrq_n_s16(q7, 3); q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6)); q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7)); q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]), vreinterpretq_s16_s32(q2tmp1.val[0])); q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]), vreinterpretq_s16_s32(q2tmp1.val[1])); q4 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[0]), vreinterpret_u8_s32(d28))); q5 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[1]), vreinterpret_u8_s32(d29))); q6 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[0]), vreinterpret_u8_s32(d30))); q7 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[1]), vreinterpret_u8_s32(d31))); d28 = vreinterpret_s32_u8(vqmovun_s16(q4)); d29 = vreinterpret_s32_u8(vqmovun_s16(q5)); d30 = vreinterpret_s32_u8(vqmovun_s16(q6)); d31 = vreinterpret_s32_u8(vqmovun_s16(q7)); dst0 = dst; dst1 = dst + 4; vst1_lane_s32((int32_t *)dst0, d28, 0); dst0 += stride; vst1_lane_s32((int32_t *)dst1, d28, 1); dst1 += stride; vst1_lane_s32((int32_t *)dst0, d29, 0); dst0 += stride; vst1_lane_s32((int32_t *)dst1, d29, 1); dst1 += stride; vst1_lane_s32((int32_t *)dst0, d30, 0); dst0 += stride; vst1_lane_s32((int32_t *)dst1, d30, 1); dst1 += stride; vst1_lane_s32((int32_t *)dst0, d31, 0); vst1_lane_s32((int32_t *)dst1, d31, 1); return; }