comparison x86/dsputil_mmx.c @ 10633:66242b8fbd32 libavcodec

port ape dsp functions from sse2 to mmx now requires yasm

10632:54982e4c4478

#if HAVE_YASM
void ff_float_to_int16_interleave6_sse(int16_t *dst, const float **src, int len);
void ff_float_to_int16_interleave6_3dnow(int16_t *dst, const float **src, int len);
void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len);
void ff_add_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left);
int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left);
void ff_x264_deblock_v_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0);
void ff_x264_deblock_h_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0);

void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
                                   int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
                                  int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);


static void add_int16_sse2(int16_t * v1, int16_t * v2, int order)
{
    x86_reg o = -(order << 1);
    v1 += order;
    v2 += order;
    __asm__ volatile(
        "1:                          \n\t"
        "movdqu   (%1,%2),   %%xmm0  \n\t"
        "movdqu 16(%1,%2),   %%xmm1  \n\t"
        "paddw    (%0,%2),   %%xmm0  \n\t"
        "paddw  16(%0,%2),   %%xmm1  \n\t"
        "movdqa   %%xmm0,    (%0,%2) \n\t"
        "movdqa   %%xmm1,  16(%0,%2) \n\t"
        "add      $32,       %2      \n\t"
        "js       1b                 \n\t"
        : "+r"(v1), "+r"(v2), "+r"(o)
    );
}

static void sub_int16_sse2(int16_t * v1, int16_t * v2, int order)
{
    x86_reg o = -(order << 1);
    v1 += order;
    v2 += order;
    __asm__ volatile(
        "1:                           \n\t"
        "movdqa    (%0,%2),   %%xmm0  \n\t"
        "movdqa  16(%0,%2),   %%xmm2  \n\t"
        "movdqu    (%1,%2),   %%xmm1  \n\t"
        "movdqu  16(%1,%2),   %%xmm3  \n\t"
        "psubw     %%xmm1,    %%xmm0  \n\t"
        "psubw     %%xmm3,    %%xmm2  \n\t"
        "movdqa    %%xmm0,    (%0,%2) \n\t"
        "movdqa    %%xmm2,  16(%0,%2) \n\t"
        "add       $32,       %2      \n\t"
        "js        1b                 \n\t"
        : "+r"(v1), "+r"(v2), "+r"(o)
    );
}

static int32_t scalarproduct_int16_sse2(int16_t * v1, int16_t * v2, int order, int shift)
{
    int res = 0;
    DECLARE_ALIGNED_16(xmm_reg, sh);
    x86_reg o = -(order << 1);

    v1 += order;
    v2 += order;
    sh.a = shift;
    __asm__ volatile(
        "pxor      %%xmm7,  %%xmm7        \n\t"
        "1:                               \n\t"
        "movdqu    (%0,%3), %%xmm0        \n\t"
        "movdqu  16(%0,%3), %%xmm1        \n\t"
        "pmaddwd   (%1,%3), %%xmm0        \n\t"
        "pmaddwd 16(%1,%3), %%xmm1        \n\t"
        "paddd     %%xmm0,  %%xmm7        \n\t"
        "paddd     %%xmm1,  %%xmm7        \n\t"
        "add       $32,     %3            \n\t"
        "js        1b                     \n\t"
        "movhlps   %%xmm7,  %%xmm2        \n\t"
        "paddd     %%xmm2,  %%xmm7        \n\t"
        "psrad     %4,      %%xmm7        \n\t"
        "pshuflw   $0x4E,   %%xmm7,%%xmm2 \n\t"
        "paddd     %%xmm2,  %%xmm7        \n\t"
        "movd      %%xmm7,  %2            \n\t"
        : "+r"(v1), "+r"(v2), "=r"(res), "+r"(o)
        : "m"(sh)
    );
    return res;
}

void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
{
    mm_flags = mm_support();

            c->vector_fmul_window = vector_fmul_window_3dnow2;
            if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
                c->float_to_int16_interleave = float_to_int16_interleave_3dn2;
            }
        }
        if(mm_flags & FF_MM_SSE){
            c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;
            c->ac3_downmix = ac3_downmix_sse;
            c->vector_fmul = vector_fmul_sse;
            c->vector_fmul_reverse = vector_fmul_reverse_sse;

            c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse
        if(mm_flags & FF_MM_SSE2){
            c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2;
            c->float_to_int16 = float_to_int16_sse2;
            c->float_to_int16_interleave = float_to_int16_interleave_sse2;
            c->add_int16 = add_int16_sse2;
            c->sub_int16 = sub_int16_sse2;
            c->scalarproduct_int16 = scalarproduct_int16_sse2;
        }
    }

    if (CONFIG_ENCODERS)
        dsputilenc_init_mmx(c, avctx);

10633:66242b8fbd32

#if HAVE_YASM
void ff_float_to_int16_interleave6_sse(int16_t *dst, const float **src, int len);
void ff_float_to_int16_interleave6_3dnow(int16_t *dst, const float **src, int len);
void ff_float_to_int16_interleave6_3dn2(int16_t *dst, const float **src, int len);
void ff_add_int16_mmx2(int16_t * v1, int16_t * v2, int order);
void ff_add_int16_sse2(int16_t * v1, int16_t * v2, int order);
void ff_sub_int16_mmx2(int16_t * v1, int16_t * v2, int order);
void ff_sub_int16_sse2(int16_t * v1, int16_t * v2, int order);
int32_t ff_scalarproduct_int16_mmx2(int16_t * v1, int16_t * v2, int order, int shift);
int32_t ff_scalarproduct_int16_sse2(int16_t * v1, int16_t * v2, int order, int shift);
void ff_add_hfyu_median_prediction_mmx2(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top);
int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src, int w, int left);
int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src, int w, int left);
void ff_x264_deblock_v_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0);
void ff_x264_deblock_h_luma_sse2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0);

void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
                                   int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
                                  int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);


void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
{
    mm_flags = mm_support();

            c->vector_fmul_window = vector_fmul_window_3dnow2;
            if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
                c->float_to_int16_interleave = float_to_int16_interleave_3dn2;
            }
        }
        if(mm_flags & FF_MM_MMX2){
#if HAVE_YASM
            c->add_int16 = ff_add_int16_mmx2;
            c->sub_int16 = ff_sub_int16_mmx2;
            c->scalarproduct_int16 = ff_scalarproduct_int16_mmx2;
#endif
        }
        if(mm_flags & FF_MM_SSE){
            c->vorbis_inverse_coupling = vorbis_inverse_coupling_sse;
            c->ac3_downmix = ac3_downmix_sse;
            c->vector_fmul = vector_fmul_sse;
            c->vector_fmul_reverse = vector_fmul_reverse_sse;

            c->vector_fmul_add = vector_fmul_add_3dnow; // faster than sse
        if(mm_flags & FF_MM_SSE2){
            c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_sse2;
            c->float_to_int16 = float_to_int16_sse2;
            c->float_to_int16_interleave = float_to_int16_interleave_sse2;
#if HAVE_YASM
            c->add_int16 = ff_add_int16_sse2;
            c->sub_int16 = ff_sub_int16_sse2;
            c->scalarproduct_int16 = ff_scalarproduct_int16_sse2;
#endif
        }
    }

    if (CONFIG_ENCODERS)
        dsputilenc_init_mmx(c, avctx);