From patchwork Mon Nov 6 13:20:16 2017 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Manojkumar Bhosale X-Patchwork-Id: 5890 Delivered-To: ffmpegpatchwork@gmail.com Received: by 10.2.161.90 with SMTP id m26csp2882370jah; Mon, 6 Nov 2017 05:20:30 -0800 (PST) X-Google-Smtp-Source: ABhQp+SHRQO06chSgmvpWRtde15RnuyutJeWrdRLLYRHgi6TPIr7SOhEahBu1pQb7y1SC4au1Lib X-Received: by 10.28.199.139 with SMTP id x133mr4871745wmf.145.1509974430820; Mon, 06 Nov 2017 05:20:30 -0800 (PST) ARC-Seal: i=1; a=rsa-sha256; t=1509974430; cv=none; d=google.com; s=arc-20160816; b=Px9PazFdb1fm3kro7eNXFERp6XX0/eJg8ourrvSrhbcH/Z2l1d6AJslc8nozQ/uH6p y2idymf6X2ARqrua6MVTDAa9yb8aMvAY1oKBqTRpWpryaHfjc6vxbv1Ge0jFbhNRnJ80 KoOerfNmvKWFfxuwbfU11IyniRPoraGRPvVjAneImxdu5nK3uCof/xwKuF0yrkqYC2ei XNR7vWoLuoCjsEToNQ4kSQddPvMFTIjZe5pGC0P6h9h35wmejCfu4TPWHcpgsjRDyUFs P104nB1g+UHMtWOy8dldqkg8XsOOaCDtgRH91ISCS4JZITlGscPNGzROfrLlX2ufYhhQ SyOg== ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=google.com; s=arc-20160816; h=sender:errors-to:content-transfer-encoding:cc:reply-to :list-subscribe:list-help:list-post:list-archive:list-unsubscribe :list-id:precedence:subject:mime-version:content-language :accept-language:in-reply-to:references:message-id:date:thread-index :thread-topic:to:from:delivered-to:arc-authentication-results; bh=tYvMi9zv7TNrVn2QrAirgJUC3oYBGQzWj302Vu38dgc=; b=T8A5bU2OXdi3ecAV24cLoT1kitEvjZUFI5vut+P6Iri6qppby2vmnhQTQWlHyNRlnK ne0L0n0vhvnrDr6ukuhgNL7F7Q5uyAAPry6jkSmku3/SceO51Fwo+X3HPxvQI3pwytdZ 8U5Xx58LIayHSNhqlN01HMt4V6ReAR+6xuJVTgc65/PCaFCRx/LNLsSf0pNzcdlnqkKU 1DNoFlU5HzGnZg1ZCB06SeTMWnV6xzq5iRqHIwWmubmEpUFT8YNy1Tsjm1V8lxyeOHgW k7b6JDAEgx6aF7YRNeUM1zRADN437VQXTSSWF1YqWhTBmBgY7UvuK5/L2AdV0bLJg6ob whzA== ARC-Authentication-Results: i=1; mx.google.com; spf=pass (google.com: domain of ffmpeg-devel-bounces@ffmpeg.org designates 79.124.17.100 as permitted sender) smtp.mailfrom=ffmpeg-devel-bounces@ffmpeg.org Return-Path: Received: from ffbox0-bg.mplayerhq.hu (ffbox0-bg.ffmpeg.org. [79.124.17.100]) by mx.google.com with ESMTP id w14si7014386wmf.33.2017.11.06.05.20.29; Mon, 06 Nov 2017 05:20:30 -0800 (PST) Received-SPF: pass (google.com: domain of ffmpeg-devel-bounces@ffmpeg.org designates 79.124.17.100 as permitted sender) client-ip=79.124.17.100; Authentication-Results: mx.google.com; spf=pass (google.com: domain of ffmpeg-devel-bounces@ffmpeg.org designates 79.124.17.100 as permitted sender) smtp.mailfrom=ffmpeg-devel-bounces@ffmpeg.org Received: from [127.0.1.1] (localhost [127.0.0.1]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTP id BB1A3689C57; Mon, 6 Nov 2017 15:20:16 +0200 (EET) X-Original-To: ffmpeg-devel@ffmpeg.org Delivered-To: ffmpeg-devel@ffmpeg.org Received: from mailapp01.imgtec.com (mailapp01.imgtec.com [195.59.15.196]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTP id 17899689822 for ; Mon, 6 Nov 2017 15:20:10 +0200 (EET) Received: from hhmail02.hh.imgtec.org (unknown [10.100.10.20]) by Forcepoint Email with ESMTP id E587C8F917478 for ; Mon, 6 Nov 2017 13:20:16 +0000 (GMT) Received: from PUMAIL01.pu.imgtec.org (192.168.91.250) by hhmail02.hh.imgtec.org (10.100.10.20) with Microsoft SMTP Server (TLS) id 14.3.361.1; Mon, 6 Nov 2017 13:20:19 +0000 Received: from PUMAIL01.pu.imgtec.org ([::1]) by PUMAIL01.pu.imgtec.org ([::1]) with mapi id 14.03.0266.001; Mon, 6 Nov 2017 18:50:16 +0530 From: Manojkumar Bhosale To: FFmpeg development discussions and patches Thread-Topic: [FFmpeg-devel] [PATCH] avcodec/mips: Improve avc avg mc 10, 30, 01 and 03 msa functions Thread-Index: AQHTVr7SVPCXsRyK4kSBz8l7rcMtJaMHVlMX Date: Mon, 6 Nov 2017 13:20:16 +0000 Message-ID: <70293ACCC3BA6A4E81FFCA024C7A86E1E05971B7@PUMAIL01.pu.imgtec.org> References: <1509945563-32085-1-git-send-email-kaustubh.raste@imgtec.com> In-Reply-To: <1509945563-32085-1-git-send-email-kaustubh.raste@imgtec.com> Accept-Language: en-US Content-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [10.100.10.20] MIME-Version: 1.0 Subject: Re: [FFmpeg-devel] [PATCH] avcodec/mips: Improve avc avg mc 10, 30, 01 and 03 msa functions X-BeenThere: ffmpeg-devel@ffmpeg.org X-Mailman-Version: 2.1.20 Precedence: list List-Id: FFmpeg development discussions and patches List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Reply-To: FFmpeg development discussions and patches Cc: Kaustubh Raste Errors-To: ffmpeg-devel-bounces@ffmpeg.org Sender: "ffmpeg-devel" LGTM diff --git a/libavcodec/mips/h264qpel_msa.c b/libavcodec/mips/h264qpel_msa.c index dd11f00..9c779bd 100644 --- a/libavcodec/mips/h264qpel_msa.c +++ b/libavcodec/mips/h264qpel_msa.c @@ -21,7 +21,7 @@ #include "libavutil/mips/generic_macros_msa.h" #include "h264dsp_mips.h" -static const uint8_t luma_mask_arr[16 * 8] = { +static const uint8_t luma_mask_arr[16 * 6] __attribute__((aligned(0x40))) = { /* 8 width cases */ 0, 5, 1, 6, 2, 7, 3, 8, 4, 9, 5, 10, 6, 11, 7, 12, 1, 4, 2, 5, 3, 6, 4, 7, 5, 8, 6, 9, 7, 10, 8, 11, @@ -31,9 +31,6 @@ static const uint8_t luma_mask_arr[16 * 8] = { 0, 5, 1, 6, 2, 7, 3, 8, 16, 21, 17, 22, 18, 23, 19, 24, 1, 4, 2, 5, 3, 6, 4, 7, 17, 20, 18, 21, 19, 22, 20, 23, 2, 3, 3, 4, 4, 5, 5, 6, 18, 19, 19, 20, 20, 21, 21, 22, - - 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 24, 25, - 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26 }; #define AVC_CALC_DPADD_B_6PIX_2COEFF_SH(vec0, vec1, vec2, vec3, vec4, vec5, \ @@ -356,414 +353,6 @@ static void avc_luma_hv_qrt_16x16_msa(const uint8_t *src_x, } } -static void avc_luma_hz_qrt_and_aver_dst_4x4_msa(const uint8_t *src, - int32_t src_stride, - uint8_t *dst, - int32_t dst_stride, - uint8_t hor_offset) -{ - uint8_t slide; - v16i8 src0, src1, src2, src3; - v16u8 dst0, dst1, dst2, dst3; - v16i8 mask0, mask1, mask2; - v16i8 vec0, vec1, vec2, vec3, vec4, vec5; - v8i16 out0, out1; - v16i8 minus5b = __msa_ldi_b(-5); - v16i8 plus20b = __msa_ldi_b(20); - v16u8 res0, res1; - - LD_SB3(&luma_mask_arr[48], 16, mask0, mask1, mask2); - - if (hor_offset) { - slide = 3; - } else { - slide = 2; - } - - LD_SB4(src, src_stride, src0, src1, src2, src3); - LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); - - XORI_B4_128_SB(src0, src1, src2, src3); - VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0, vec1); - HADD_SB2_SH(vec0, vec1, out0, out1); - VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2, vec3); - DPADD_SB2_SH(vec2, vec3, minus5b, minus5b, out0, out1); - VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4, vec5); - DPADD_SB2_SH(vec4, vec5, plus20b, plus20b, out0, out1); - SRARI_H2_SH(out0, out1, 5); - SAT_SH2_SH(out0, out1, 7); - - PCKEV_B2_UB(out0, out0, out1, out1, res0, res1); - - src0 = __msa_sld_b(src0, src0, slide); - src1 = __msa_sld_b(src1, src1, slide); - src2 = __msa_sld_b(src2, src2, slide); - src3 = __msa_sld_b(src3, src3, slide); - src0 = (v16i8) __msa_insve_w((v4i32) src0, 1, (v4i32) src1); - src1 = (v16i8) __msa_insve_w((v4i32) src2, 1, (v4i32) src3); - res0 = (v16u8) __msa_aver_s_b((v16i8) res0, src0); - res1 = (v16u8) __msa_aver_s_b((v16i8) res1, src1); - - XORI_B2_128_UB(res0, res1); - - dst0 = (v16u8) __msa_insve_w((v4i32) dst0, 1, (v4i32) dst1); - dst1 = (v16u8) __msa_insve_w((v4i32) dst2, 1, (v4i32) dst3); - - AVER_UB2_UB(res0, dst0, res1, dst1, dst0, dst1); - - ST4x4_UB(dst0, dst1, 0, 1, 0, 1, dst, dst_stride); -} - -static void avc_luma_hz_qrt_and_aver_dst_8x8_msa(const uint8_t *src, - int32_t src_stride, - uint8_t *dst, - int32_t dst_stride, - uint8_t hor_offset) -{ - uint8_t slide; - uint32_t loop_cnt; - v16i8 src0, src1, src2, src3; - v16i8 mask0, mask1, mask2; - v16u8 dst0, dst1, dst2, dst3; - v16i8 vec0, vec1, vec2, vec3, vec4, vec5; - v16i8 vec6, vec7, vec8, vec9, vec10, vec11; - v8i16 out0, out1, out2, out3; - v16i8 minus5b = __msa_ldi_b(-5); - v16i8 plus20b = __msa_ldi_b(20); - v16i8 res0, res1, res2, res3; - - LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); - - if (hor_offset) { - slide = 3; - } else { - slide = 2; - } - - for (loop_cnt = 2; loop_cnt--;) { - LD_SB4(src, src_stride, src0, src1, src2, src3); - src += (4 * src_stride); - - LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); - - XORI_B4_128_SB(src0, src1, src2, src3); - VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec1); - VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2, vec3); - HADD_SB4_SH(vec0, vec1, vec2, vec3, out0, out1, out2, out3); - VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4, vec5); - VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6, vec7); - DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b, - out0, out1, out2, out3); - VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec8, vec9); - VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec10, vec11); - DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, - plus20b, out0, out1, out2, out3); - - src0 = __msa_sld_b(src0, src0, slide); - src1 = __msa_sld_b(src1, src1, slide); - src2 = __msa_sld_b(src2, src2, slide); - src3 = __msa_sld_b(src3, src3, slide); - - SRARI_H4_SH(out0, out1, out2, out3, 5); - SAT_SH4_SH(out0, out1, out2, out3, 7); - - PCKEV_B4_SB(out0, out0, out1, out1, out2, out2, out3, out3, - res0, res1, res2, res3); - - res0 = __msa_aver_s_b(res0, src0); - res1 = __msa_aver_s_b(res1, src1); - res2 = __msa_aver_s_b(res2, src2); - res3 = __msa_aver_s_b(res3, src3); - - XORI_B4_128_SB(res0, res1, res2, res3); - AVER_ST8x4_UB(res0, dst0, res1, dst1, res2, dst2, res3, dst3, - dst, dst_stride); - - dst += (4 * dst_stride); - } -} - -static void avc_luma_hz_qrt_and_aver_dst_16x16_msa(const uint8_t *src, - int32_t src_stride, - uint8_t *dst, - int32_t dst_stride, - uint8_t hor_offset) -{ - uint32_t loop_cnt; - v16i8 out0, out1; - v16i8 src0, src1, src2, src3; - v16i8 mask0, mask1, mask2, vshf; - v16u8 dst0, dst1; - v8i16 res0, res1, res2, res3; - v16i8 vec0, vec1, vec2, vec3, vec4, vec5; - v16i8 vec6, vec7, vec8, vec9, vec10, vec11; - v16i8 minus5b = __msa_ldi_b(-5); - v16i8 plus20b = __msa_ldi_b(20); - - LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); - - if (hor_offset) { - vshf = LD_SB(&luma_mask_arr[16 + 96]); - } else { - vshf = LD_SB(&luma_mask_arr[96]); - } - - for (loop_cnt = 8; loop_cnt--;) { - LD_SB2(src, 8, src0, src1); - src += src_stride; - LD_SB2(src, 8, src2, src3); - src += src_stride; - - LD_UB2(dst, dst_stride, dst0, dst1); - - XORI_B4_128_SB(src0, src1, src2, src3); - VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec3); - VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec6, vec9); - VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec1, vec4); - VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec7, vec10); - VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec2, vec5); - VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec8, vec11); - HADD_SB4_SH(vec0, vec3, vec6, vec9, res0, res1, res2, res3); - DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b, - minus5b, res0, res1, res2, res3); - DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b, - plus20b, res0, res1, res2, res3); - VSHF_B2_SB(src0, src1, src2, src3, vshf, vshf, src0, src2); - SRARI_H4_SH(res0, res1, res2, res3, 5); - SAT_SH4_SH(res0, res1, res2, res3, 7); - PCKEV_B2_SB(res1, res0, res3, res2, out0, out1); - - out0 = __msa_aver_s_b(out0, src0); - out1 = __msa_aver_s_b(out1, src2); - - XORI_B2_128_SB(out0, out1); - AVER_UB2_UB(out0, dst0, out1, dst1, dst0, dst1); - ST_UB2(dst0, dst1, dst, dst_stride); - dst += (2 * dst_stride); - } -} - -static void avc_luma_vt_qrt_and_aver_dst_4x4_msa(const uint8_t *src, - int32_t src_stride, - uint8_t *dst, - int32_t dst_stride, - uint8_t ver_offset) -{ - int16_t filt_const0 = 0xfb01; - int16_t filt_const1 = 0x1414; - int16_t filt_const2 = 0x1fb; - v16u8 dst0, dst1, dst2, dst3; - v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8; - v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r; - v16i8 src87_r, src2110, src4332, src6554, src8776; - v8i16 out10, out32; - v16i8 filt0, filt1, filt2; - v16u8 res; - - filt0 = (v16i8) __msa_fill_h(filt_const0); - filt1 = (v16i8) __msa_fill_h(filt_const1); - filt2 = (v16i8) __msa_fill_h(filt_const2); - - LD_SB5(src, src_stride, src0, src1, src2, src3, src4); - src += (5 * src_stride); - - ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, - src10_r, src21_r, src32_r, src43_r); - ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332); - XORI_B2_128_SB(src2110, src4332); - LD_SB4(src, src_stride, src5, src6, src7, src8); - ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, - src54_r, src65_r, src76_r, src87_r); - ILVR_D2_SB(src65_r, src54_r, src87_r, src76_r, src6554, src8776); - XORI_B2_128_SB(src6554, src8776); - out10 = DPADD_SH3_SH(src2110, src4332, src6554, filt0, filt1, filt2); - out32 = DPADD_SH3_SH(src4332, src6554, src8776, filt0, filt1, filt2); - SRARI_H2_SH(out10, out32, 5); - SAT_SH2_SH(out10, out32, 7); - LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); - res = PCKEV_XORI128_UB(out10, out32); - - if (ver_offset) { - src32_r = (v16i8) __msa_insve_w((v4i32) src3, 1, (v4i32) src4); - src54_r = (v16i8) __msa_insve_w((v4i32) src5, 1, (v4i32) src6); - } else { - src32_r = (v16i8) __msa_insve_w((v4i32) src2, 1, (v4i32) src3); - src54_r = (v16i8) __msa_insve_w((v4i32) src4, 1, (v4i32) src5); - } - - src32_r = (v16i8) __msa_insve_d((v2i64) src32_r, 1, (v2i64) src54_r); - res = __msa_aver_u_b(res, (v16u8) src32_r); - - ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst1); - - dst0 = (v16u8) __msa_pckev_d((v2i64) dst1, (v2i64) dst0); - dst0 = __msa_aver_u_b(res, dst0); - - ST4x4_UB(dst0, dst0, 0, 1, 2, 3, dst, dst_stride); -} - -static void avc_luma_vt_qrt_and_aver_dst_8x8_msa(const uint8_t *src, - int32_t src_stride, - uint8_t *dst, - int32_t dst_stride, - uint8_t ver_offset) -{ - int32_t loop_cnt; - int16_t filt_const0 = 0xfb01; - int16_t filt_const1 = 0x1414; - int16_t filt_const2 = 0x1fb; - v16u8 dst0, dst1, dst2, dst3; - v16i8 src0, src1, src2, src3, src4, src7, src8, src9, src10; - v16i8 src10_r, src32_r, src76_r, src98_r; - v16i8 src21_r, src43_r, src87_r, src109_r; - v8i16 out0_r, out1_r, out2_r, out3_r; - v16i8 res0, res1; - v16u8 vec0, vec1; - v16i8 filt0, filt1, filt2; - - filt0 = (v16i8) __msa_fill_h(filt_const0); - filt1 = (v16i8) __msa_fill_h(filt_const1); - filt2 = (v16i8) __msa_fill_h(filt_const2); - - LD_SB5(src, src_stride, src0, src1, src2, src3, src4); - src += (5 * src_stride); - - XORI_B5_128_SB(src0, src1, src2, src3, src4); - ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, - src10_r, src21_r, src32_r, src43_r); - - for (loop_cnt = 2; loop_cnt--;) { - LD_SB4(src, src_stride, src7, src8, src9, src10); - src += (4 * src_stride); - - XORI_B4_128_SB(src7, src8, src9, src10); - ILVR_B4_SB(src7, src4, src8, src7, src9, src8, src10, src9, - src76_r, src87_r, src98_r, src109_r); - out0_r = DPADD_SH3_SH(src10_r, src32_r, src76_r, filt0, filt1, filt2); - out1_r = DPADD_SH3_SH(src21_r, src43_r, src87_r, filt0, filt1, filt2); - out2_r = DPADD_SH3_SH(src32_r, src76_r, src98_r, filt0, filt1, filt2); - out3_r = DPADD_SH3_SH(src43_r, src87_r, src109_r, filt0, filt1, filt2); - SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, 5); - SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); - PCKEV_B2_SB(out1_r, out0_r, out3_r, out2_r, res0, res1); - - if (ver_offset) { - PCKEV_D2_SB(src4, src3, src8, src7, src10_r, src32_r); - } else { - PCKEV_D2_SB(src3, src2, src7, src4, src10_r, src32_r); - } - - LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); - ILVR_D2_UB(dst1, dst0, dst3, dst2, dst0, dst1); - - vec0 = (v16u8) __msa_aver_s_b(res0, src10_r); - vec1 = (v16u8) __msa_aver_s_b(res1, src32_r); - - XORI_B2_128_UB(vec0, vec1); - AVER_UB2_UB(vec0, dst0, vec1, dst1, vec0, vec1); - ST8x4_UB(vec0, vec1, dst, dst_stride); - dst += (4 * dst_stride); - - src10_r = src76_r; - src32_r = src98_r; - src21_r = src87_r; - src43_r = src109_r; - src2 = src8; - src3 = src9; - src4 = src10; - } -} - -static void avc_luma_vt_qrt_and_aver_dst_16x16_msa(const uint8_t *src, - int32_t src_stride, - uint8_t *dst, - int32_t dst_stride, - uint8_t ver_offset) -{ - int32_t loop_cnt; - int16_t filt_const0 = 0xfb01; - int16_t filt_const1 = 0x1414; - int16_t filt_const2 = 0x1fb; - v16u8 dst0, dst1, dst2, dst3; - v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8; - v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r; - v16i8 src87_r, src10_l, src32_l, src54_l, src76_l, src21_l, src43_l; - v16i8 src65_l, src87_l; - v8i16 out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l; - v16i8 out0, out1, out2, out3; - v16i8 filt0, filt1, filt2; - v16u8 res0, res1, res2, res3; - - filt0 = (v16i8) __msa_fill_h(filt_const0); - filt1 = (v16i8) __msa_fill_h(filt_const1); - filt2 = (v16i8) __msa_fill_h(filt_const2); - - LD_SB5(src, src_stride, src0, src1, src2, src3, src4); - src += (5 * src_stride); - - XORI_B5_128_SB(src0, src1, src2, src3, src4); - ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, - src10_r, src21_r, src32_r, src43_r); - ILVL_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, - src10_l, src21_l, src32_l, src43_l); - - for (loop_cnt = 4; loop_cnt--;) { - LD_SB4(src, src_stride, src5, src6, src7, src8); - src += (4 * src_stride); - - XORI_B4_128_SB(src5, src6, src7, src8); - ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, - src54_r, src65_r, src76_r, src87_r); - ILVL_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, - src54_l, src65_l, src76_l, src87_l); - out0_r = DPADD_SH3_SH(src10_r, src32_r, src54_r, filt0, filt1, filt2); - out1_r = DPADD_SH3_SH(src21_r, src43_r, src65_r, filt0, filt1, filt2); - out2_r = DPADD_SH3_SH(src32_r, src54_r, src76_r, filt0, filt1, filt2); - out3_r = DPADD_SH3_SH(src43_r, src65_r, src87_r, filt0, filt1, filt2); - out0_l = DPADD_SH3_SH(src10_l, src32_l, src54_l, filt0, filt1, filt2); - out1_l = DPADD_SH3_SH(src21_l, src43_l, src65_l, filt0, filt1, filt2); - out2_l = DPADD_SH3_SH(src32_l, src54_l, src76_l, filt0, filt1, filt2); - out3_l = DPADD_SH3_SH(src43_l, src65_l, src87_l, filt0, filt1, filt2); - SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, 5); - SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, 5); - SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); - SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7); - PCKEV_B4_SB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l, - out3_r, out0, out1, out2, out3); - LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); - - if (ver_offset) { - res0 = (v16u8) __msa_aver_s_b(out0, src3); - res1 = (v16u8) __msa_aver_s_b(out1, src4); - res2 = (v16u8) __msa_aver_s_b(out2, src5); - res3 = (v16u8) __msa_aver_s_b(out3, src6); - } else { - res0 = (v16u8) __msa_aver_s_b(out0, src2); - res1 = (v16u8) __msa_aver_s_b(out1, src3); - res2 = (v16u8) __msa_aver_s_b(out2, src4); - res3 = (v16u8) __msa_aver_s_b(out3, src5); - } - - XORI_B4_128_UB(res0, res1, res2, res3); - AVER_UB4_UB(res0, dst0, res1, dst1, res2, dst2, res3, dst3, - dst0, dst1, dst2, dst3); - ST_UB4(dst0, dst1, dst2, dst3, dst, dst_stride); - dst += (4 * dst_stride); - - src10_r = src54_r; - src32_r = src76_r; - src21_r = src65_r; - src43_r = src87_r; - src10_l = src54_l; - src32_l = src76_l; - src21_l = src65_l; - src43_l = src87_l; - src2 = src6; - src3 = src7; - src4 = src8; - } -} - static void avc_luma_hv_qrt_and_aver_dst_4x4_msa(const uint8_t *src_x, const uint8_t *src_y, uint8_t *dst, @@ -3558,117 +3147,225 @@ void ff_put_h264_qpel4_mc22_msa(uint8_t *dst, const uint8_t *src, void ff_avg_h264_qpel16_mc10_msa(uint8_t *dst, const uint8_t *src, ptrdiff_t stride) { - avc_luma_hz_qrt_and_aver_dst_16x16_msa(src - 2, stride, dst, stride, 0); -} - -void ff_avg_h264_qpel16_mc30_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) -{ - avc_luma_hz_qrt_and_aver_dst_16x16_msa(src - 2, stride, dst, stride, 1); -} - -void ff_avg_h264_qpel8_mc10_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) -{ - avc_luma_hz_qrt_and_aver_dst_8x8_msa(src - 2, stride, dst, stride, 0); -} - -void ff_avg_h264_qpel8_mc30_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) -{ - avc_luma_hz_qrt_and_aver_dst_8x8_msa(src - 2, stride, dst, stride, 1); -} - -void ff_avg_h264_qpel4_mc10_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) -{ - avc_luma_hz_qrt_and_aver_dst_4x4_msa(src - 2, stride, dst, stride, 0); -} - -void ff_avg_h264_qpel4_mc30_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) -{ - avc_luma_hz_qrt_and_aver_dst_4x4_msa(src - 2, stride, dst, stride, 1); -} - -void ff_avg_h264_qpel16_mc20_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) -{ uint32_t loop_cnt; v16u8 dst0, dst1, dst2, dst3; - v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask0, mask1, mask2; + v16i8 out0, out1, out2, out3, src0, src1, src2, src3, src4, src5, src6; + v16i8 mask0, mask1, mask2, mask3, mask4, mask5, src7, vec11; v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10; - v16i8 vec11; v8i16 res0, res1, res2, res3, res4, res5, res6, res7; v16i8 minus5b = __msa_ldi_b(-5); v16i8 plus20b = __msa_ldi_b(20); LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); + mask3 = mask0 + 8; + mask4 = mask1 + 8; + mask5 = mask2 + 8; src -= 2; for (loop_cnt = 4; loop_cnt--;) { - LD_SB2(src, 8, src0, src1); + LD_SB2(src, 16, src0, src1); src += stride; - LD_SB2(src, 8, src2, src3); + LD_SB2(src, 16, src2, src3); src += stride; - LD_SB2(src, 8, src4, src5); + LD_SB2(src, 16, src4, src5); src += stride; - LD_SB2(src, 8, src6, src7); + LD_SB2(src, 16, src6, src7); src += stride; LD_UB4(dst, stride, dst0, dst1, dst2, dst3); XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7); - VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec3); - VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec6, vec9); - VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec1, vec4); - VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec7, vec10); - VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec2, vec5); - VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec8, vec11); + VSHF_B2_SB(src0, src0, src0, src1, mask0, mask3, vec0, vec3); + VSHF_B2_SB(src2, src2, src2, src3, mask0, mask3, vec6, vec9); + VSHF_B2_SB(src0, src0, src0, src1, mask1, mask4, vec1, vec4); + VSHF_B2_SB(src2, src2, src2, src3, mask1, mask4, vec7, vec10); + VSHF_B2_SB(src0, src0, src0, src1, mask2, mask5, vec2, vec5); + VSHF_B2_SB(src2, src2, src2, src3, mask2, mask5, vec8, vec11); HADD_SB4_SH(vec0, vec3, vec6, vec9, res0, res1, res2, res3); DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b, minus5b, res0, res1, res2, res3); DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b, plus20b, res0, res1, res2, res3); - VSHF_B2_SB(src4, src4, src5, src5, mask0, mask0, vec0, vec3); - VSHF_B2_SB(src6, src6, src7, src7, mask0, mask0, vec6, vec9); - VSHF_B2_SB(src4, src4, src5, src5, mask1, mask1, vec1, vec4); - VSHF_B2_SB(src6, src6, src7, src7, mask1, mask1, vec7, vec10); - VSHF_B2_SB(src4, src4, src5, src5, mask2, mask2, vec2, vec5); - VSHF_B2_SB(src6, src6, src7, src7, mask2, mask2, vec8, vec11); + VSHF_B2_SB(src4, src4, src4, src5, mask0, mask3, vec0, vec3); + VSHF_B2_SB(src6, src6, src6, src7, mask0, mask3, vec6, vec9); + VSHF_B2_SB(src4, src4, src4, src5, mask1, mask4, vec1, vec4); + VSHF_B2_SB(src6, src6, src6, src7, mask1, mask4, vec7, vec10); + VSHF_B2_SB(src4, src4, src4, src5, mask2, mask5, vec2, vec5); + VSHF_B2_SB(src6, src6, src6, src7, mask2, mask5, vec8, vec11); HADD_SB4_SH(vec0, vec3, vec6, vec9, res4, res5, res6, res7); DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b, minus5b, res4, res5, res6, res7); DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b, plus20b, res4, res5, res6, res7); + SLDI_B2_SB(src1, src3, src0, src2, src0, src2, 2); + SLDI_B2_SB(src5, src7, src4, src6, src4, src6, 2); SRARI_H4_SH(res0, res1, res2, res3, 5); SRARI_H4_SH(res4, res5, res6, res7, 5); SAT_SH4_SH(res0, res1, res2, res3, 7); SAT_SH4_SH(res4, res5, res6, res7, 7); - PCKEV_B4_SB(res1, res0, res3, res2, res5, res4, res7, res6, vec0, vec1, - vec2, vec3); - XORI_B4_128_SB(vec0, vec1, vec2, vec3); - AVER_UB2_UB(vec0, dst0, vec1, dst1, dst0, dst1); - AVER_UB2_UB(vec2, dst2, vec3, dst3, dst2, dst3); + PCKEV_B2_SB(res1, res0, res3, res2, out0, out1); + PCKEV_B2_SB(res5, res4, res7, res6, out2, out3); + out0 = __msa_aver_s_b(out0, src0); + out1 = __msa_aver_s_b(out1, src2); + out2 = __msa_aver_s_b(out2, src4); + out3 = __msa_aver_s_b(out3, src6); + XORI_B4_128_SB(out0, out1, out2, out3); + AVER_UB2_UB(out0, dst0, out1, dst1, dst0, dst1); + AVER_UB2_UB(out2, dst2, out3, dst3, dst2, dst3); ST_UB4(dst0, dst1, dst2, dst3, dst, stride); dst += (4 * stride); } } -void ff_avg_h264_qpel8_mc20_msa(uint8_t *dst, const uint8_t *src, +void ff_avg_h264_qpel16_mc30_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) +{ + uint32_t loop_cnt; + v16u8 dst0, dst1, dst2, dst3; + v16i8 out0, out1, out2, out3, src0, src1, src2, src3, src4, src5, src6; + v16i8 mask0, mask1, mask2, mask3, mask4, mask5, src7, vec11; + v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10; + v8i16 res0, res1, res2, res3, res4, res5, res6, res7; + v16i8 minus5b = __msa_ldi_b(-5); + v16i8 plus20b = __msa_ldi_b(20); + + LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); + mask3 = mask0 + 8; + mask4 = mask1 + 8; + mask5 = mask2 + 8; + src -= 2; + + for (loop_cnt = 4; loop_cnt--;) { + LD_SB2(src, 16, src0, src1); + src += stride; + LD_SB2(src, 16, src2, src3); + src += stride; + LD_SB2(src, 16, src4, src5); + src += stride; + LD_SB2(src, 16, src6, src7); + src += stride; + + LD_UB4(dst, stride, dst0, dst1, dst2, dst3); + XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7); + VSHF_B2_SB(src0, src0, src0, src1, mask0, mask3, vec0, vec3); + VSHF_B2_SB(src2, src2, src2, src3, mask0, mask3, vec6, vec9); + VSHF_B2_SB(src0, src0, src0, src1, mask1, mask4, vec1, vec4); + VSHF_B2_SB(src2, src2, src2, src3, mask1, mask4, vec7, vec10); + VSHF_B2_SB(src0, src0, src0, src1, mask2, mask5, vec2, vec5); + VSHF_B2_SB(src2, src2, src2, src3, mask2, mask5, vec8, vec11); + HADD_SB4_SH(vec0, vec3, vec6, vec9, res0, res1, res2, res3); + DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b, + minus5b, res0, res1, res2, res3); + DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b, + plus20b, res0, res1, res2, res3); + VSHF_B2_SB(src4, src4, src4, src5, mask0, mask3, vec0, vec3); + VSHF_B2_SB(src6, src6, src6, src7, mask0, mask3, vec6, vec9); + VSHF_B2_SB(src4, src4, src4, src5, mask1, mask4, vec1, vec4); + VSHF_B2_SB(src6, src6, src6, src7, mask1, mask4, vec7, vec10); + VSHF_B2_SB(src4, src4, src4, src5, mask2, mask5, vec2, vec5); + VSHF_B2_SB(src6, src6, src6, src7, mask2, mask5, vec8, vec11); + HADD_SB4_SH(vec0, vec3, vec6, vec9, res4, res5, res6, res7); + DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b, + minus5b, res4, res5, res6, res7); + DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b, + plus20b, res4, res5, res6, res7); + SLDI_B2_SB(src1, src3, src0, src2, src0, src2, 3); + SLDI_B2_SB(src5, src7, src4, src6, src4, src6, 3); + SRARI_H4_SH(res0, res1, res2, res3, 5); + SRARI_H4_SH(res4, res5, res6, res7, 5); + SAT_SH4_SH(res0, res1, res2, res3, 7); + SAT_SH4_SH(res4, res5, res6, res7, 7); + PCKEV_B2_SB(res1, res0, res3, res2, out0, out1); + PCKEV_B2_SB(res5, res4, res7, res6, out2, out3); + out0 = __msa_aver_s_b(out0, src0); + out1 = __msa_aver_s_b(out1, src2); + out2 = __msa_aver_s_b(out2, src4); + out3 = __msa_aver_s_b(out3, src6); + XORI_B4_128_SB(out0, out1, out2, out3); + AVER_UB2_UB(out0, dst0, out1, dst1, dst0, dst1); + AVER_UB2_UB(out2, dst2, out3, dst3, dst2, dst3); + ST_UB4(dst0, dst1, dst2, dst3, dst, stride); + dst += (4 * stride); + } +} + +void ff_avg_h264_qpel8_mc10_msa(uint8_t *dst, const uint8_t *src, ptrdiff_t stride) { uint64_t tp0, tp1, tp2, tp3; - v16u8 out0, out1, out2 = { 0 }, out3 = { 0 }; - v16u8 out4, out5, out6 = { 0 }, out7 = { 0 }; + v16u8 dst0 = { 0 }, dst1 = { 0 }, dst2 = { 0 }, dst3 = { 0 }; v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask0, mask1, mask2; + v16i8 tmp0, tmp1, tmp2, tmp3, vec11; v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10; - v16i8 vec11; v8i16 res0, res1, res2, res3, res4, res5, res6, res7; v16i8 minus5b = __msa_ldi_b(-5); v16i8 plus20b = __msa_ldi_b(20); LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); + LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7); + XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7); + VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec1); + VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2, vec3); + HADD_SB4_SH(vec0, vec1, vec2, vec3, res0, res1, res2, res3); + VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4, vec5); + VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6, vec7); + DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b, + res0, res1, res2, res3); + VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec8, vec9); + VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec10, vec11); + DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b, + res0, res1, res2, res3); + VSHF_B2_SB(src4, src4, src5, src5, mask0, mask0, vec0, vec1); + VSHF_B2_SB(src6, src6, src7, src7, mask0, mask0, vec2, vec3); + HADD_SB4_SH(vec0, vec1, vec2, vec3, res4, res5, res6, res7); + VSHF_B2_SB(src4, src4, src5, src5, mask1, mask1, vec4, vec5); + VSHF_B2_SB(src6, src6, src7, src7, mask1, mask1, vec6, vec7); + DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b, + res4, res5, res6, res7); + VSHF_B2_SB(src4, src4, src5, src5, mask2, mask2, vec8, vec9); + VSHF_B2_SB(src6, src6, src7, src7, mask2, mask2, vec10, vec11); + DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b, + res4, res5, res6, res7); + SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 2); + SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 2); + SLDI_B2_SB(src4, src5, src4, src5, src4, src5, 2); + SLDI_B2_SB(src6, src7, src6, src7, src6, src7, 2); + PCKEV_D2_SB(src1, src0, src3, src2, src0, src1); + PCKEV_D2_SB(src5, src4, src7, src6, src4, src5); + SRARI_H4_SH(res0, res1, res2, res3, 5); + SRARI_H4_SH(res4, res5, res6, res7, 5); + SAT_SH4_SH(res0, res1, res2, res3, 7); + SAT_SH4_SH(res4, res5, res6, res7, 7); + PCKEV_B2_SB(res1, res0, res3, res2, tmp0, tmp1); + PCKEV_B2_SB(res5, res4, res7, res6, tmp2, tmp3); + tmp0 = __msa_aver_s_b(tmp0, src0); + tmp1 = __msa_aver_s_b(tmp1, src1); + tmp2 = __msa_aver_s_b(tmp2, src4); + tmp3 = __msa_aver_s_b(tmp3, src5); + XORI_B4_128_SB(tmp0, tmp1, tmp2, tmp3); + LD4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, dst0); + INSERT_D2_UB(tp2, tp3, dst1); + LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, dst2); + INSERT_D2_UB(tp2, tp3, dst3); + AVER_UB2_UB(tmp0, dst0, tmp1, dst1, dst0, dst1); + AVER_UB2_UB(tmp2, dst2, tmp3, dst3, dst2, dst3); + ST8x8_UB(dst0, dst1, dst2, dst3, dst, stride); +} + +void ff_avg_h264_qpel8_mc30_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) +{ + uint64_t tp0, tp1, tp2, tp3; + v16u8 dst0 = { 0 }, dst1 = { 0 }, dst2 = { 0 }, dst3 = { 0 }; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask0, mask1, mask2; + v16i8 tmp0, tmp1, tmp2, tmp3, vec11; + v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10; + v8i16 res0, res1, res2, res3, res4, res5, res6, res7; + v16i8 minus5b = __msa_ldi_b(-5); + v16i8 plus20b = __msa_ldi_b(20); + LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7); XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7); VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec1); @@ -3693,33 +3390,42 @@ void ff_avg_h264_qpel8_mc20_msa(uint8_t *dst, const uint8_t *src, VSHF_B2_SB(src6, src6, src7, src7, mask2, mask2, vec10, vec11); DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b, res4, res5, res6, res7); + SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 3); + SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 3); + SLDI_B2_SB(src4, src5, src4, src5, src4, src5, 3); + SLDI_B2_SB(src6, src7, src6, src7, src6, src7, 3); + PCKEV_D2_SB(src1, src0, src3, src2, src0, src1); + PCKEV_D2_SB(src5, src4, src7, src6, src4, src5); SRARI_H4_SH(res0, res1, res2, res3, 5); SRARI_H4_SH(res4, res5, res6, res7, 5); SAT_SH4_SH(res0, res1, res2, res3, 7); SAT_SH4_SH(res4, res5, res6, res7, 7); - out0 = PCKEV_XORI128_UB(res0, res1); - out1 = PCKEV_XORI128_UB(res2, res3); - out4 = PCKEV_XORI128_UB(res4, res5); - out5 = PCKEV_XORI128_UB(res6, res7); + PCKEV_B2_SB(res1, res0, res3, res2, tmp0, tmp1); + PCKEV_B2_SB(res5, res4, res7, res6, tmp2, tmp3); + tmp0 = __msa_aver_s_b(tmp0, src0); + tmp1 = __msa_aver_s_b(tmp1, src1); + tmp2 = __msa_aver_s_b(tmp2, src4); + tmp3 = __msa_aver_s_b(tmp3, src5); + XORI_B4_128_SB(tmp0, tmp1, tmp2, tmp3); LD4(dst, stride, tp0, tp1, tp2, tp3); - INSERT_D2_UB(tp0, tp1, out2); - INSERT_D2_UB(tp2, tp3, out3); + INSERT_D2_UB(tp0, tp1, dst0); + INSERT_D2_UB(tp2, tp3, dst1); LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3); - INSERT_D2_UB(tp0, tp1, out6); - INSERT_D2_UB(tp2, tp3, out7); - AVER_UB2_UB(out0, out2, out1, out3, out0, out1); - AVER_UB2_UB(out4, out6, out5, out7, out4, out5); - ST8x8_UB(out0, out1, out4, out5, dst, stride); + INSERT_D2_UB(tp0, tp1, dst2); + INSERT_D2_UB(tp2, tp3, dst3); + AVER_UB2_UB(tmp0, dst0, tmp1, dst1, dst0, dst1); + AVER_UB2_UB(tmp2, dst2, tmp3, dst3, dst2, dst3); + ST8x8_UB(dst0, dst1, dst2, dst3, dst, stride); } -void ff_avg_h264_qpel4_mc20_msa(uint8_t *dst, const uint8_t *src, +void ff_avg_h264_qpel4_mc10_msa(uint8_t *dst, const uint8_t *src, ptrdiff_t stride) { uint32_t tp0, tp1, tp2, tp3; - v16u8 res, dst0 = { 0 }; - v16i8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, vec4, vec5; + v16u8 dst0 = { 0 }; + v16i8 src0, src1, src2, src3, res, vec0, vec1, vec2, vec3, vec4, vec5; v16i8 mask0, mask1, mask2; - v8i16 res0, res1; + v8i16 out0, out1; v16i8 minus5b = __msa_ldi_b(-5); v16i8 plus20b = __msa_ldi_b(20); @@ -3727,60 +3433,587 @@ void ff_avg_h264_qpel4_mc20_msa(uint8_t *dst, const uint8_t *src, LD_SB4(src - 2, stride, src0, src1, src2, src3); XORI_B4_128_SB(src0, src1, src2, src3); VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0, vec1); - HADD_SB2_SH(vec0, vec1, res0, res1); + HADD_SB2_SH(vec0, vec1, out0, out1); VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2, vec3); - DPADD_SB2_SH(vec2, vec3, minus5b, minus5b, res0, res1); + DPADD_SB2_SH(vec2, vec3, minus5b, minus5b, out0, out1); VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4, vec5); - DPADD_SB2_SH(vec4, vec5, plus20b, plus20b, res0, res1); - SRARI_H2_SH(res0, res1, 5); - SAT_SH2_SH(res0, res1, 7); - res = PCKEV_XORI128_UB(res0, res1); + DPADD_SB2_SH(vec4, vec5, plus20b, plus20b, out0, out1); + SRARI_H2_SH(out0, out1, 5); + SAT_SH2_SH(out0, out1, 7); + res = __msa_pckev_b((v16i8) out1, (v16i8) out0); + SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 2); + SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 2); + src0 = (v16i8) __msa_insve_w((v4i32) src0, 1, (v4i32) src1); + src1 = (v16i8) __msa_insve_w((v4i32) src2, 1, (v4i32) src3); + src0 = (v16i8) __msa_insve_d((v2i64) src0, 1, (v2i64) src1); + res = __msa_aver_s_b(res, src0); + res = (v16i8) __msa_xori_b((v16u8) res, 128); LW4(dst, stride, tp0, tp1, tp2, tp3); INSERT_W4_UB(tp0, tp1, tp2, tp3, dst0); - res = __msa_aver_u_b(res, dst0); - ST4x4_UB(res, res, 0, 1, 2, 3, dst, stride); + dst0 = __msa_aver_u_b((v16u8) res, dst0); + ST4x4_UB(dst0, dst0, 0, 1, 2, 3, dst, stride); } -void ff_avg_h264_qpel16_mc01_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) +void ff_avg_h264_qpel4_mc30_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) { - avc_luma_vt_qrt_and_aver_dst_16x16_msa(src - (stride * 2), - stride, dst, stride, 0); + uint32_t tp0, tp1, tp2, tp3; + v16u8 dst0 = { 0 }; + v16i8 src0, src1, src2, src3, res, vec0, vec1, vec2, vec3, vec4, vec5; + v16i8 mask0, mask1, mask2; + v8i16 out0, out1; + v16i8 minus5b = __msa_ldi_b(-5); + v16i8 plus20b = __msa_ldi_b(20); + + LD_SB3(&luma_mask_arr[48], 16, mask0, mask1, mask2); + LD_SB4(src - 2, stride, src0, src1, src2, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0, vec1); + HADD_SB2_SH(vec0, vec1, out0, out1); + VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2, vec3); + DPADD_SB2_SH(vec2, vec3, minus5b, minus5b, out0, out1); + VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4, vec5); + DPADD_SB2_SH(vec4, vec5, plus20b, plus20b, out0, out1); + SRARI_H2_SH(out0, out1, 5); + SAT_SH2_SH(out0, out1, 7); + res = __msa_pckev_b((v16i8) out1, (v16i8) out0); + SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 3); + SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 3); + src0 = (v16i8) __msa_insve_w((v4i32) src0, 1, (v4i32) src1); + src1 = (v16i8) __msa_insve_w((v4i32) src2, 1, (v4i32) src3); + src0 = (v16i8) __msa_insve_d((v2i64) src0, 1, (v2i64) src1); + res = __msa_aver_s_b(res, src0); + res = (v16i8) __msa_xori_b((v16u8) res, 128); + LW4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_W4_UB(tp0, tp1, tp2, tp3, dst0); + dst0 = __msa_aver_u_b((v16u8) res, dst0); + ST4x4_UB(dst0, dst0, 0, 1, 2, 3, dst, stride); } -void ff_avg_h264_qpel16_mc03_msa(uint8_t *dst, const uint8_t *src, +void ff_avg_h264_qpel16_mc20_msa(uint8_t *dst, const uint8_t *src, ptrdiff_t stride) { - avc_luma_vt_qrt_and_aver_dst_16x16_msa(src - (stride * 2), - stride, dst, stride, 1); -} + uint32_t loop_cnt; + v16u8 dst0, dst1, dst2, dst3; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask0, mask1, mask2; + v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10; + v16i8 vec11; + v8i16 res0, res1, res2, res3, res4, res5, res6, res7; + v16i8 minus5b = __msa_ldi_b(-5); + v16i8 plus20b = __msa_ldi_b(20); -void ff_avg_h264_qpel8_mc01_msa(uint8_t *dst, const uint8_t *src, - ptrdiff_t stride) -{ - avc_luma_vt_qrt_and_aver_dst_8x8_msa(src - (stride * 2), - stride, dst, stride, 0); -} + LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); + src -= 2; + + for (loop_cnt = 4; loop_cnt--;) { + LD_SB2(src, 8, src0, src1); + src += stride; + LD_SB2(src, 8, src2, src3); + src += stride; + LD_SB2(src, 8, src4, src5); + src += stride; + LD_SB2(src, 8, src6, src7); + src += stride; + + LD_UB4(dst, stride, dst0, dst1, dst2, dst3); + XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7); + VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec3); + VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec6, vec9); + VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec1, vec4); + VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec7, vec10); + VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec2, vec5); + VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec8, vec11); + HADD_SB4_SH(vec0, vec3, vec6, vec9, res0, res1, res2, res3); + DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b, + minus5b, res0, res1, res2, res3); + DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b, + plus20b, res0, res1, res2, res3); + VSHF_B2_SB(src4, src4, src5, src5, mask0, mask0, vec0, vec3); + VSHF_B2_SB(src6, src6, src7, src7, mask0, mask0, vec6, vec9); + VSHF_B2_SB(src4, src4, src5, src5, mask1, mask1, vec1, vec4); + VSHF_B2_SB(src6, src6, src7, src7, mask1, mask1, vec7, vec10); + VSHF_B2_SB(src4, src4, src5, src5, mask2, mask2, vec2, vec5); + VSHF_B2_SB(src6, src6, src7, src7, mask2, mask2, vec8, vec11); + HADD_SB4_SH(vec0, vec3, vec6, vec9, res4, res5, res6, res7); + DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b, + minus5b, res4, res5, res6, res7); + DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b, + plus20b, res4, res5, res6, res7); + SRARI_H4_SH(res0, res1, res2, res3, 5); + SRARI_H4_SH(res4, res5, res6, res7, 5); + SAT_SH4_SH(res0, res1, res2, res3, 7); + SAT_SH4_SH(res4, res5, res6, res7, 7); + PCKEV_B4_SB(res1, res0, res3, res2, res5, res4, res7, res6, vec0, vec1, + vec2, vec3); + XORI_B4_128_SB(vec0, vec1, vec2, vec3); + AVER_UB2_UB(vec0, dst0, vec1, dst1, dst0, dst1); + AVER_UB2_UB(vec2, dst2, vec3, dst3, dst2, dst3); + ST_UB4(dst0, dst1, dst2, dst3, dst, stride); + dst += (4 * stride); + } +} + +void ff_avg_h264_qpel8_mc20_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) +{ + uint64_t tp0, tp1, tp2, tp3; + v16u8 out0, out1, out2 = { 0 }, out3 = { 0 }; + v16u8 out4, out5, out6 = { 0 }, out7 = { 0 }; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask0, mask1, mask2; + v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10; + v16i8 vec11; + v8i16 res0, res1, res2, res3, res4, res5, res6, res7; + v16i8 minus5b = __msa_ldi_b(-5); + v16i8 plus20b = __msa_ldi_b(20); + + LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2); + + LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7); + XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7); + VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec1); + VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2, vec3); + HADD_SB4_SH(vec0, vec1, vec2, vec3, res0, res1, res2, res3); + VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4, vec5); + VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6, vec7); + DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b, + res0, res1, res2, res3); + VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec8, vec9); + VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec10, vec11); + DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b, + res0, res1, res2, res3); + VSHF_B2_SB(src4, src4, src5, src5, mask0, mask0, vec0, vec1); + VSHF_B2_SB(src6, src6, src7, src7, mask0, mask0, vec2, vec3); + HADD_SB4_SH(vec0, vec1, vec2, vec3, res4, res5, res6, res7); + VSHF_B2_SB(src4, src4, src5, src5, mask1, mask1, vec4, vec5); + VSHF_B2_SB(src6, src6, src7, src7, mask1, mask1, vec6, vec7); + DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b, + res4, res5, res6, res7); + VSHF_B2_SB(src4, src4, src5, src5, mask2, mask2, vec8, vec9); + VSHF_B2_SB(src6, src6, src7, src7, mask2, mask2, vec10, vec11); + DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b, + res4, res5, res6, res7); + SRARI_H4_SH(res0, res1, res2, res3, 5); + SRARI_H4_SH(res4, res5, res6, res7, 5); + SAT_SH4_SH(res0, res1, res2, res3, 7); + SAT_SH4_SH(res4, res5, res6, res7, 7); + out0 = PCKEV_XORI128_UB(res0, res1); + out1 = PCKEV_XORI128_UB(res2, res3); + out4 = PCKEV_XORI128_UB(res4, res5); + out5 = PCKEV_XORI128_UB(res6, res7); + LD4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, out2); + INSERT_D2_UB(tp2, tp3, out3); + LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, out6); + INSERT_D2_UB(tp2, tp3, out7); + AVER_UB2_UB(out0, out2, out1, out3, out0, out1); + AVER_UB2_UB(out4, out6, out5, out7, out4, out5); + ST8x8_UB(out0, out1, out4, out5, dst, stride); +} + +void ff_avg_h264_qpel4_mc20_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) +{ + uint32_t tp0, tp1, tp2, tp3; + v16u8 res, dst0 = { 0 }; + v16i8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, vec4, vec5; + v16i8 mask0, mask1, mask2; + v8i16 res0, res1; + v16i8 minus5b = __msa_ldi_b(-5); + v16i8 plus20b = __msa_ldi_b(20); + + LD_SB3(&luma_mask_arr[48], 16, mask0, mask1, mask2); + LD_SB4(src - 2, stride, src0, src1, src2, src3); + XORI_B4_128_SB(src0, src1, src2, src3); + VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0, vec1); + HADD_SB2_SH(vec0, vec1, res0, res1); + VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2, vec3); + DPADD_SB2_SH(vec2, vec3, minus5b, minus5b, res0, res1); + VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4, vec5); + DPADD_SB2_SH(vec4, vec5, plus20b, plus20b, res0, res1); + SRARI_H2_SH(res0, res1, 5); + SAT_SH2_SH(res0, res1, 7); + res = PCKEV_XORI128_UB(res0, res1); + LW4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_W4_UB(tp0, tp1, tp2, tp3, dst0); + res = __msa_aver_u_b(res, dst0); + ST4x4_UB(res, res, 0, 1, 2, 3, dst, stride); +} + +void ff_avg_h264_qpel16_mc01_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) +{ + int32_t loop_cnt; + int16_t filt_const0 = 0xfb01; + int16_t filt_const1 = 0x1414; + int16_t filt_const2 = 0x1fb; + v16u8 res0, res1, res2, res3, dst0, dst1, dst2, dst3; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8; + v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r; + v16i8 src87_r, src10_l, src32_l, src54_l, src76_l, src21_l, src43_l; + v16i8 src65_l, src87_l, filt0, filt1, filt2; + v8i16 out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l; + + filt0 = (v16i8) __msa_fill_h(filt_const0); + filt1 = (v16i8) __msa_fill_h(filt_const1); + filt2 = (v16i8) __msa_fill_h(filt_const2); + + src -= (stride * 2); + + LD_SB5(src, stride, src0, src1, src2, src3, src4); + src += (5 * stride); + + XORI_B5_128_SB(src0, src1, src2, src3, src4); + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVL_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_l, src21_l, + src32_l, src43_l); + + for (loop_cnt = 4; loop_cnt--;) { + LD_SB4(src, stride, src5, src6, src7, src8); + src += (4 * stride); + + XORI_B4_128_SB(src5, src6, src7, src8); + ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, + src65_r, src76_r, src87_r); + ILVL_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_l, + src65_l, src76_l, src87_l); + out0_r = AVC_DOT_SH3_SH(src10_r, src32_r, src54_r, filt0, filt1, filt2); + out1_r = AVC_DOT_SH3_SH(src21_r, src43_r, src65_r, filt0, filt1, filt2); + out2_r = AVC_DOT_SH3_SH(src32_r, src54_r, src76_r, filt0, filt1, filt2); + out3_r = AVC_DOT_SH3_SH(src43_r, src65_r, src87_r, filt0, filt1, filt2); + out0_l = AVC_DOT_SH3_SH(src10_l, src32_l, src54_l, filt0, filt1, filt2); + out1_l = AVC_DOT_SH3_SH(src21_l, src43_l, src65_l, filt0, filt1, filt2); + out2_l = AVC_DOT_SH3_SH(src32_l, src54_l, src76_l, filt0, filt1, filt2); + out3_l = AVC_DOT_SH3_SH(src43_l, src65_l, src87_l, filt0, filt1, filt2); + SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, 5); + SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); + SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, 5); + SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7); + PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l, + out3_r, res0, res1, res2, res3); + res0 = (v16u8) __msa_aver_s_b((v16i8) res0, src2); + res1 = (v16u8) __msa_aver_s_b((v16i8) res1, src3); + res2 = (v16u8) __msa_aver_s_b((v16i8) res2, src4); + res3 = (v16u8) __msa_aver_s_b((v16i8) res3, src5); + LD_UB4(dst, stride, dst0, dst1, dst2, dst3); + XORI_B4_128_UB(res0, res1, res2, res3); + AVER_UB2_UB(res0, dst0, res1, dst1, dst0, dst1); + AVER_UB2_UB(res2, dst2, res3, dst3, dst2, dst3); + ST_UB4(dst0, dst1, dst2, dst3, dst, stride); + dst += (4 * stride); + + src10_r = src54_r; + src32_r = src76_r; + src21_r = src65_r; + src43_r = src87_r; + src10_l = src54_l; + src32_l = src76_l; + src21_l = src65_l; + src43_l = src87_l; + src2 = src6; + src3 = src7; + src4 = src8; + } +} + +void ff_avg_h264_qpel16_mc03_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) +{ + int32_t loop_cnt; + int16_t filt_const0 = 0xfb01; + int16_t filt_const1 = 0x1414; + int16_t filt_const2 = 0x1fb; + v16u8 res0, res1, res2, res3, dst0, dst1, dst2, dst3; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8; + v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r; + v16i8 src87_r, src10_l, src32_l, src54_l, src76_l, src21_l, src43_l; + v16i8 src65_l, src87_l, filt0, filt1, filt2; + v8i16 out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l; + + filt0 = (v16i8) __msa_fill_h(filt_const0); + filt1 = (v16i8) __msa_fill_h(filt_const1); + filt2 = (v16i8) __msa_fill_h(filt_const2); + + src -= (stride * 2); + + LD_SB5(src, stride, src0, src1, src2, src3, src4); + src += (5 * stride); + + XORI_B5_128_SB(src0, src1, src2, src3, src4); + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVL_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_l, src21_l, + src32_l, src43_l); + + for (loop_cnt = 4; loop_cnt--;) { + LD_SB4(src, stride, src5, src6, src7, src8); + src += (4 * stride); + + XORI_B4_128_SB(src5, src6, src7, src8); + ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, + src65_r, src76_r, src87_r); + ILVL_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_l, + src65_l, src76_l, src87_l); + out0_r = AVC_DOT_SH3_SH(src10_r, src32_r, src54_r, filt0, filt1, filt2); + out1_r = AVC_DOT_SH3_SH(src21_r, src43_r, src65_r, filt0, filt1, filt2); + out2_r = AVC_DOT_SH3_SH(src32_r, src54_r, src76_r, filt0, filt1, filt2); + out3_r = AVC_DOT_SH3_SH(src43_r, src65_r, src87_r, filt0, filt1, filt2); + out0_l = AVC_DOT_SH3_SH(src10_l, src32_l, src54_l, filt0, filt1, filt2); + out1_l = AVC_DOT_SH3_SH(src21_l, src43_l, src65_l, filt0, filt1, filt2); + out2_l = AVC_DOT_SH3_SH(src32_l, src54_l, src76_l, filt0, filt1, filt2); + out3_l = AVC_DOT_SH3_SH(src43_l, src65_l, src87_l, filt0, filt1, filt2); + SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, 5); + SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); + SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, 5); + SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7); + PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l, + out3_r, res0, res1, res2, res3); + res0 = (v16u8) __msa_aver_s_b((v16i8) res0, src3); + res1 = (v16u8) __msa_aver_s_b((v16i8) res1, src4); + res2 = (v16u8) __msa_aver_s_b((v16i8) res2, src5); + res3 = (v16u8) __msa_aver_s_b((v16i8) res3, src6); + LD_UB4(dst, stride, dst0, dst1, dst2, dst3); + XORI_B4_128_UB(res0, res1, res2, res3); + AVER_UB2_UB(res0, dst0, res1, dst1, dst0, dst1); + AVER_UB2_UB(res2, dst2, res3, dst3, dst2, dst3); + ST_UB4(dst0, dst1, dst2, dst3, dst, stride); + dst += (4 * stride); + + src10_r = src54_r; + src32_r = src76_r; + src21_r = src65_r; + src43_r = src87_r; + src10_l = src54_l; + src32_l = src76_l; + src21_l = src65_l; + src43_l = src87_l; + src3 = src7; + src4 = src8; + } +} + +void ff_avg_h264_qpel8_mc01_msa(uint8_t *dst, const uint8_t *src, + ptrdiff_t stride) +{ + uint64_t tp0, tp1, tp2, tp3; + const int16_t filt_const0 = 0xfb01; + const int16_t filt_const1 = 0x1414; + const int16_t filt_const2 = 0x1fb; + v16u8 dst0 = { 0 }, dst1 = { 0 }, dst2 = { 0 }, dst3 = { 0 }; + v16i8 src0, src1, src2, src3, src4, src7, src8, src9, src10, src11, src12; + v16i8 src13, src14, tmp0, tmp1, tmp2, tmp3, src109_r; + v16i8 src10_r, src32_r, src76_r, src98_r, src21_r, src43_r, src87_r; + v16i8 filt0, filt1, filt2, out0, out1, out2, out3; + v8i16 out0_r, out1_r, out2_r, out3_r, out4_r, out5_r, out6_r, out7_r; + + filt0 = (v16i8) __msa_fill_h(filt_const0); + filt1 = (v16i8) __msa_fill_h(filt_const1); + filt2 = (v16i8) __msa_fill_h(filt_const2); + + src -= (stride * 2); + + LD_SB5(src, stride, src0, src1, src2, src3, src4); + src += (5 * stride); + + XORI_B5_128_SB(src0, src1, src2, src3, src4); + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + LD_SB8(src, stride, src7, src8, src9, src10, src11, src12, src13, src14); + XORI_B8_128_SB(src7, src8, src9, src10, src11, src12, src13, src14); + ILVR_B4_SB(src7, src4, src8, src7, src9, src8, src10, src9, src76_r, + src87_r, src98_r, src109_r); + out0_r = AVC_DOT_SH3_SH(src10_r, src32_r, src76_r, filt0, filt1, filt2); + out1_r = AVC_DOT_SH3_SH(src21_r, src43_r, src87_r, filt0, filt1, filt2); + out2_r = AVC_DOT_SH3_SH(src32_r, src76_r, src98_r, filt0, filt1, filt2); + out3_r = AVC_DOT_SH3_SH(src43_r, src87_r, src109_r, filt0, filt1, filt2); + PCKEV_D2_SB(src3, src2, src7, src4, tmp0, tmp1); + ILVR_B4_SB(src11, src10, src12, src11, src13, src12, src14, src13, src10_r, + src21_r, src32_r, src43_r); + out4_r = AVC_DOT_SH3_SH(src76_r, src98_r, src10_r, filt0, filt1, filt2); + out5_r = AVC_DOT_SH3_SH(src87_r, src109_r, src21_r, filt0, filt1, filt2); + out6_r = AVC_DOT_SH3_SH(src98_r, src10_r, src32_r, filt0, filt1, filt2); + out7_r = AVC_DOT_SH3_SH(src109_r, src21_r, src43_r, filt0, filt1, filt2); + PCKEV_D2_SB(src9, src8, src11, src10, tmp2, tmp3); + SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, 5); + SRARI_H4_SH(out4_r, out5_r, out6_r, out7_r, 5); + SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); + SAT_SH4_SH(out4_r, out5_r, out6_r, out7_r, 7); + + LD4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, dst0); + INSERT_D2_UB(tp2, tp3, dst1); + LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, dst2); + INSERT_D2_UB(tp2, tp3, dst3); + + PCKEV_B2_SB(out1_r, out0_r, out3_r, out2_r, out0, out1); + PCKEV_B2_SB(out5_r, out4_r, out7_r, out6_r, out2, out3); + out0 = __msa_aver_s_b(out0, tmp0); + out1 = __msa_aver_s_b(out1, tmp1); + out2 = __msa_aver_s_b(out2, tmp2); + out3 = __msa_aver_s_b(out3, tmp3); + XORI_B4_128_SB(out0, out1, out2, out3); + AVER_UB4_UB(out0, dst0, out1, dst1, out2, dst2, out3, dst3, dst0, dst1, + dst2, dst3); + ST8x8_UB(dst0, dst1, dst2, dst3, dst, stride); +} void ff_avg_h264_qpel8_mc03_msa(uint8_t *dst, const uint8_t *src, ptrdiff_t stride) { - avc_luma_vt_qrt_and_aver_dst_8x8_msa(src - (stride * 2), - stride, dst, stride, 1); + uint64_t tp0, tp1, tp2, tp3; + const int16_t filt_const0 = 0xfb01; + const int16_t filt_const1 = 0x1414; + const int16_t filt_const2 = 0x1fb; + v16u8 dst0 = { 0 }, dst1 = { 0 }, dst2 = { 0 }, dst3 = { 0 }; + v16i8 src0, src1, src2, src3, src4, src7, src8, src9, src10, src11, src12; + v16i8 src13, src14, tmp0, tmp1, tmp2, tmp3, src109_r; + v16i8 src10_r, src32_r, src76_r, src98_r, src21_r, src43_r, src87_r; + v16i8 filt0, filt1, filt2, out0, out1, out2, out3; + v8i16 out0_r, out1_r, out2_r, out3_r, out4_r, out5_r, out6_r, out7_r; + + filt0 = (v16i8) __msa_fill_h(filt_const0); + filt1 = (v16i8) __msa_fill_h(filt_const1); + filt2 = (v16i8) __msa_fill_h(filt_const2); + + src -= (stride * 2); + + LD_SB5(src, stride, src0, src1, src2, src3, src4); + src += (5 * stride); + + XORI_B5_128_SB(src0, src1, src2, src3, src4); + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + LD_SB8(src, stride, src7, src8, src9, src10, src11, src12, src13, src14); + XORI_B8_128_SB(src7, src8, src9, src10, src11, src12, src13, src14); + ILVR_B4_SB(src7, src4, src8, src7, src9, src8, src10, src9, src76_r, + src87_r, src98_r, src109_r); + out0_r = AVC_DOT_SH3_SH(src10_r, src32_r, src76_r, filt0, filt1, filt2); + out1_r = AVC_DOT_SH3_SH(src21_r, src43_r, src87_r, filt0, filt1, filt2); + out2_r = AVC_DOT_SH3_SH(src32_r, src76_r, src98_r, filt0, filt1, filt2); + out3_r = AVC_DOT_SH3_SH(src43_r, src87_r, src109_r, filt0, filt1, filt2); + PCKEV_D2_SB(src4, src3, src8, src7, tmp0, tmp1); + ILVR_B4_SB(src11, src10, src12, src11, src13, src12, src14, src13, src10_r, + src21_r, src32_r, src43_r); + out4_r = AVC_DOT_SH3_SH(src76_r, src98_r, src10_r, filt0, filt1, filt2); + out5_r = AVC_DOT_SH3_SH(src87_r, src109_r, src21_r, filt0, filt1, filt2); + out6_r = AVC_DOT_SH3_SH(src98_r, src10_r, src32_r, filt0, filt1, filt2); + out7_r = AVC_DOT_SH3_SH(src109_r, src21_r, src43_r, filt0, filt1, filt2); + PCKEV_D2_SB(src10, src9, src12, src11, tmp2, tmp3); + SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, 5); + SRARI_H4_SH(out4_r, out5_r, out6_r, out7_r, 5); + SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7); + SAT_SH4_SH(out4_r, out5_r, out6_r, out7_r, 7); + + LD4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, dst0); + INSERT_D2_UB(tp2, tp3, dst1); + LD4(dst + 4 * stride, stride, tp0, tp1, tp2, tp3); + INSERT_D2_UB(tp0, tp1, dst2); + INSERT_D2_UB(tp2, tp3, dst3); + + PCKEV_B2_SB(out1_r, out0_r, out3_r, out2_r, out0, out1); + PCKEV_B2_SB(out5_r, out4_r, out7_r, out6_r, out2, out3); + out0 = __msa_aver_s_b(out0, tmp0); + out1 = __msa_aver_s_b(out1, tmp1); + out2 = __msa_aver_s_b(out2, tmp2); + out3 = __msa_aver_s_b(out3, tmp3); + XORI_B4_128_SB(out0, out1, out2, out3); + AVER_UB4_UB(out0, dst0, out1, dst1, out2, dst2, out3, dst3, dst0, dst1, + dst2, dst3); + ST8x8_UB(dst0, dst1, dst2, dst3, dst, stride); } void ff_avg_h264_qpel4_mc01_msa(uint8_t *dst, const uint8_t *src, ptrdiff_t stride) { - avc_luma_vt_qrt_and_aver_dst_4x4_msa(src - (stride * 2), - stride, dst, stride, 0); + uint32_t tp0, tp1, tp2, tp3; + int16_t filt_const0 = 0xfb01; + int16_t filt_const1 = 0x1414; + int16_t filt_const2 = 0x1fb; + v16u8 res, dst0 = { 0 }; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8; + v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r; + v16i8 src87_r, src2110, src4332, src6554, src8776, filt0, filt1, filt2; + v8i16 out10, out32; + + filt0 = (v16i8) __msa_fill_h(filt_const0); + filt1 = (v16i8) __msa_fill_h(filt_const1); + filt2 = (v16i8) __msa_fill_h(filt_const2); + + src -= (stride * 2); + LD_SB5(src, stride, src0, src1, src2, src3, src4); + src += (5 * stride); + + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332); + XORI_B2_128_SB(src2110, src4332); + LD_SB4(src, stride, src5, src6, src7, src8); + ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r, + src76_r, src87_r); + ILVR_D2_SB(src65_r, src54_r, src87_r, src76_r, src6554, src8776); + XORI_B2_128_SB(src6554, src8776); + src32_r = (v16i8) __msa_insve_w((v4i32) src2, 1, (v4i32) src3); + src54_r = (v16i8) __msa_insve_w((v4i32) src4, 1, (v4i32) src5); + src32_r = (v16i8) __msa_insve_d((v2i64) src32_r, 1, (v2i64) src54_r); + out10 = AVC_DOT_SH3_SH(src2110, src4332, src6554, filt0, filt1, filt2); + out32 = AVC_DOT_SH3_SH(src4332, src6554, src8776, filt0, filt1, filt2); + SRARI_H2_SH(out10, out32, 5); + SAT_SH2_SH(out10, out32, 7); + LW4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_W4_UB(tp0, tp1, tp2, tp3, dst0); + res = PCKEV_XORI128_UB(out10, out32); + res = __msa_aver_u_b(res, (v16u8) src32_r); + dst0 = __msa_aver_u_b(res, dst0); + ST4x4_UB(dst0, dst0, 0, 1, 2, 3, dst, stride); } void ff_avg_h264_qpel4_mc03_msa(uint8_t *dst, const uint8_t *src, ptrdiff_t stride) { - avc_luma_vt_qrt_and_aver_dst_4x4_msa(src - (stride * 2), - stride, dst, stride, 1); + uint32_t tp0, tp1, tp2, tp3; + int16_t filt_const0 = 0xfb01; + int16_t filt_const1 = 0x1414; + int16_t filt_const2 = 0x1fb; + v16u8 res, dst0 = { 0 }; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8; + v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r; + v16i8 src87_r, src2110, src4332, src6554, src8776, filt0, filt1, filt2; + v8i16 out10, out32; + + filt0 = (v16i8) __msa_fill_h(filt_const0); + filt1 = (v16i8) __msa_fill_h(filt_const1); + filt2 = (v16i8) __msa_fill_h(filt_const2); + + src -= (stride * 2); + + LD_SB5(src, stride, src0, src1, src2, src3, src4); + src += (5 * stride); + + ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r, + src32_r, src43_r); + ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332); + XORI_B2_128_SB(src2110, src4332); + LD_SB4(src, stride, src5, src6, src7, src8); + ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r, + src76_r, src87_r); + ILVR_D2_SB(src65_r, src54_r, src87_r, src76_r, src6554, src8776); + XORI_B2_128_SB(src6554, src8776); + out10 = AVC_DOT_SH3_SH(src2110, src4332, src6554, filt0, filt1, filt2); + out32 = AVC_DOT_SH3_SH(src4332, src6554, src8776, filt0, filt1, filt2); + SRARI_H2_SH(out10, out32, 5); + SAT_SH2_SH(out10, out32, 7); + LW4(dst, stride, tp0, tp1, tp2, tp3); + INSERT_W4_UB(tp0, tp1, tp2, tp3, dst0); + res = PCKEV_XORI128_UB(out10, out32); + src32_r = (v16i8) __msa_insve_w((v4i32) src3, 1, (v4i32) src4); + src54_r = (v16i8) __msa_insve_w((v4i32) src5, 1, (v4i32) src6); + src32_r = (v16i8) __msa_insve_d((v2i64) src32_r, 1, (v2i64) src54_r); + res = __msa_aver_u_b(res, (v16u8) src32_r); + dst0 = __msa_aver_u_b(res, dst0); + ST4x4_UB(dst0, dst0, 0, 1, 2, 3, dst, stride); } void ff_avg_h264_qpel16_mc11_msa(uint8_t *dst, const uint8_t *src,