diff --git a/libavcodec/aarch64/opusdsp_init.c b/libavcodec/aarch64/opusdsp_init.c index bb6d71b66b..a727006593 100644 --- a/libavcodec/aarch64/opusdsp_init.c +++ b/libavcodec/aarch64/opusdsp_init.c @@ -23,7 +23,7 @@ #include "libavcodec/opusdsp.h" void ff_opus_postfilter_neon(float *data, int period, float *gains, int len); -float ff_opus_deemphasis_neon(float *out, float *in, float coeff, int len); +float ff_opus_deemphasis_neon(float *out, float *in, float coeff, const float *weights, int len); av_cold void ff_opus_dsp_init_aarch64(OpusDSP *ctx) { diff --git a/libavcodec/aarch64/opusdsp_neon.S b/libavcodec/aarch64/opusdsp_neon.S index e933151ab4..253825aa61 100644 --- a/libavcodec/aarch64/opusdsp_neon.S +++ b/libavcodec/aarch64/opusdsp_neon.S @@ -18,29 +18,11 @@ #include "libavutil/aarch64/asm.S" - // 0.85..^1 0.85..^2 0.85..^3 0.85..^4 -const tab_st, align=4 - .word 0x3f599a00, 0x3f38f671, 0x3f1d382a, 0x3f05a32f -endconst -const tab_x0, align=4 - .word 0x0, 0x3f599a00, 0x3f38f671, 0x3f1d382a -endconst -const tab_x1, align=4 - .word 0x0, 0x0, 0x3f599a00, 0x3f38f671 -endconst -const tab_x2, align=4 - .word 0x0, 0x0, 0x0, 0x3f599a00 -endconst - function ff_opus_deemphasis_neon, export=1 - movrel x4, tab_st - ld1 {v4.4s}, [x4] - movrel x4, tab_x0 - ld1 {v5.4s}, [x4] - movrel x4, tab_x1 - ld1 {v6.4s}, [x4] - movrel x4, tab_x2 - ld1 {v7.4s}, [x4] + ld1 {v4.4s}, [x2], #16 + ld1 {v5.4s}, [x2], #16 + ld1 {v6.4s}, [x2], #16 + ld1 {v7.4s}, [x2] fmul v0.4s, v4.4s, v0.s[0] @@ -63,7 +45,7 @@ function ff_opus_deemphasis_neon, export=1 st1 {v1.4s, v2.4s}, [x0], #32 fmul v0.4s, v4.4s, v2.s[3] - subs w2, w2, #8 + subs w3, w3, #8 b.gt 1b mov s0, v2.s[3] diff --git a/libavcodec/opusdec_celt.c b/libavcodec/opusdec_celt.c index fd8e9929e9..b19342337d 100644 --- a/libavcodec/opusdec_celt.c +++ b/libavcodec/opusdec_celt.c @@ -460,7 +460,9 @@ int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc, /* deemphasis */ block->emph_coeff = f->opusdsp.deemphasis(output[i], &block->buf[1024 - frame_size], - block->emph_coeff, frame_size); + block->emph_coeff, + ff_opus_deemph_weights, + frame_size); } if (channels == 1) @@ -516,7 +518,7 @@ void ff_celt_flush(CeltFrame *f) * a lesser discontinuity when seeking. * The deemphasis functions differ from libopus in that they require * an initial state divided by the coefficient. */ - block->emph_coeff = 0.0f / CELT_EMPH_COEFF; + block->emph_coeff = 0.0f / ff_opus_deemph_weights[0]; } f->seed = 0; diff --git a/libavcodec/opusdsp.c b/libavcodec/opusdsp.c index 0764d712e4..e61cc36098 100644 --- a/libavcodec/opusdsp.c +++ b/libavcodec/opusdsp.c @@ -18,6 +18,7 @@ #include "config.h" #include "libavutil/attributes.h" +#include "libavutil/mem_internal.h" #include "opusdsp.h" static void postfilter_c(float *data, int period, float *gains, int len) @@ -43,10 +44,11 @@ static void postfilter_c(float *data, int period, float *gains, int len) } } -static float deemphasis_c(float *y, float *x, float coeff, int len) +static float deemphasis_c(float *y, float *x, float coeff, const float *weights, int len) { + const float c = weights[0]; for (int i = 0; i < len; i++) - coeff = y[i] = x[i] + coeff*CELT_EMPH_COEFF; + coeff = y[i] = x[i] + coeff*c; return coeff; } diff --git a/libavcodec/opusdsp.h b/libavcodec/opusdsp.h index c2a301e832..9d93336cfe 100644 --- a/libavcodec/opusdsp.h +++ b/libavcodec/opusdsp.h @@ -19,11 +19,9 @@ #ifndef AVCODEC_OPUSDSP_H #define AVCODEC_OPUSDSP_H -#define CELT_EMPH_COEFF 0.8500061035f - typedef struct OpusDSP { void (*postfilter)(float *data, int period, float *gains, int len); - float (*deemphasis)(float *out, float *in, float coeff, int len); + float (*deemphasis)(float *out, float *in, float coeff, const float *weights, int len); } OpusDSP; void ff_opus_dsp_init(OpusDSP *ctx); diff --git a/libavcodec/opusenc.c b/libavcodec/opusenc.c index e0ce07cddc..164d535ed5 100644 --- a/libavcodec/opusenc.c +++ b/libavcodec/opusenc.c @@ -164,6 +164,7 @@ static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f) { const int subframesize = s->avctx->frame_size; const int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize; + const float c = ff_opus_deemph_weights[0]; /* Filter overlap */ for (int ch = 0; ch < f->channels; ch++) { @@ -172,7 +173,7 @@ static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f) for (int i = 0; i < CELT_OVERLAP; i++) { float sample = b->overlap[i]; b->overlap[i] = sample - m; - m = sample * CELT_EMPH_COEFF; + m = sample * c; } b->emph_coeff = m; } @@ -185,7 +186,7 @@ static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f) for (int i = 0; i < subframesize; i++) { float sample = b->samples[sf*subframesize + i]; b->samples[sf*subframesize + i] = sample - m; - m = sample * CELT_EMPH_COEFF; + m = sample * c; } if (sf != (subframes - 1)) b->emph_coeff = m; diff --git a/libavcodec/opustab.c b/libavcodec/opustab.c index 2a57511177..47624fe651 100644 --- a/libavcodec/opustab.c +++ b/libavcodec/opustab.c @@ -1159,3 +1159,31 @@ const uint32_t * const ff_celt_pvq_u_row[15] = { celt_pvq_u + 1207, celt_pvq_u + 1226, celt_pvq_u + 1240, celt_pvq_u + 1248, celt_pvq_u + 1254, celt_pvq_u + 1257 }; + +/* Deemphasis constant (alpha_p), as specified in RFC6716 as 0.8500061035. + * libopus uses a slighly rounded constant, set to 0.85 exactly, + * to simplify its fixed-point version, but it's not significant to impact + * compliance. */ +#define CELT_EMPH_COEFF 0.8500061035f + +DECLARE_ALIGNED(16, const float, ff_opus_deemph_weights)[] = { + CELT_EMPH_COEFF, + CELT_EMPH_COEFF*CELT_EMPH_COEFF, + CELT_EMPH_COEFF*CELT_EMPH_COEFF*CELT_EMPH_COEFF, + CELT_EMPH_COEFF*CELT_EMPH_COEFF*CELT_EMPH_COEFF*CELT_EMPH_COEFF, + + 0, + CELT_EMPH_COEFF, + CELT_EMPH_COEFF*CELT_EMPH_COEFF, + CELT_EMPH_COEFF*CELT_EMPH_COEFF*CELT_EMPH_COEFF, + + 0, + 0, + CELT_EMPH_COEFF, + CELT_EMPH_COEFF*CELT_EMPH_COEFF, + + 0, + 0, + 0, + CELT_EMPH_COEFF, +}; diff --git a/libavcodec/opustab.h b/libavcodec/opustab.h index 9c9f1b9d98..57e546aef5 100644 --- a/libavcodec/opustab.h +++ b/libavcodec/opustab.h @@ -161,6 +161,8 @@ extern const float ff_celt_window2[120]; extern const float ff_celt_window_padded[]; static const float *const ff_celt_window = &ff_celt_window_padded[8]; +extern const float ff_opus_deemph_weights[]; + extern const uint32_t * const ff_celt_pvq_u_row[15]; FF_VISIBILITY_POP_HIDDEN diff --git a/libavcodec/x86/opusdsp.asm b/libavcodec/x86/opusdsp.asm index 418cc16330..635f59f83c 100644 --- a/libavcodec/x86/opusdsp.asm +++ b/libavcodec/x86/opusdsp.asm @@ -22,16 +22,13 @@ SECTION_RODATA - ; 0.85..^1 0.85..^2 0.85..^3 0.85..^4 -tab_st: dd 0x3f599a00, 0x3f38f671, 0x3f1d382a, 0x3f05a32f - SECTION .text INIT_XMM fma3 %if UNIX64 -cglobal opus_deemphasis, 3, 3, 8, out, in, len +cglobal opus_deemphasis, 4, 4, 8, out, in, weights, len %else -cglobal opus_deemphasis, 4, 4, 8, out, in, coeff, len +cglobal opus_deemphasis, 5, 5, 8, out, in, coeff, weights, len %endif %if ARCH_X86_32 VBROADCASTSS m0, coeffm @@ -41,7 +38,7 @@ cglobal opus_deemphasis, 4, 4, 8, out, in, coeff, len shufps m0, m0, 0 %endif - movaps m4, [tab_st] + movaps m4, [weightsq] VBROADCASTSS m5, m4 shufps m6, m4, m4, q1111 shufps m7, m4, m4, q2222 diff --git a/libavcodec/x86/opusdsp_init.c b/libavcodec/x86/opusdsp_init.c index 582fbb4f0c..66d6839589 100644 --- a/libavcodec/x86/opusdsp_init.c +++ b/libavcodec/x86/opusdsp_init.c @@ -23,7 +23,7 @@ #include "libavcodec/opusdsp.h" void ff_opus_postfilter_fma3(float *data, int period, float *gains, int len); -float ff_opus_deemphasis_fma3(float *out, float *in, float coeff, int len); +float ff_opus_deemphasis_fma3(float *out, float *in, float coeff, const float *weights, int len); av_cold void ff_opus_dsp_init_x86(OpusDSP *ctx) { diff --git a/tests/checkasm/opusdsp.c b/tests/checkasm/opusdsp.c index 34ee5a4edc..b9ecc55240 100644 --- a/tests/checkasm/opusdsp.c +++ b/tests/checkasm/opusdsp.c @@ -19,6 +19,7 @@ #include "libavutil/mem_internal.h" #include "libavcodec/opusdsp.h" +#include "libavcodec/opustab.h" #include "checkasm.h" @@ -69,17 +70,17 @@ static void test_deemphasis(void) LOCAL_ALIGNED(16, float, dst1, [FFALIGN(MAX_SIZE, 4)]); float coeff0 = (float)rnd() / (UINT_MAX >> 5) - 16.0f, coeff1 = coeff0; - declare_func_float(float, float *out, float *in, float coeff, int len); + declare_func_float(float, float *out, float *in, float coeff, const float *weights, int len); randomize_float(src, MAX_SIZE); - coeff0 = call_ref(dst0, src, coeff0, MAX_SIZE); - coeff1 = call_new(dst1, src, coeff1, MAX_SIZE); + coeff0 = call_ref(dst0, src, coeff0, ff_opus_deemph_weights, MAX_SIZE); + coeff1 = call_new(dst1, src, coeff1, ff_opus_deemph_weights, MAX_SIZE); if (!float_near_abs_eps(coeff0, coeff1, EPS) || !float_near_abs_eps_array(dst0, dst1, EPS, MAX_SIZE)) fail(); - bench_new(dst1, src, coeff1, MAX_SIZE); + bench_new(dst1, src, coeff1, ff_opus_deemph_weights, MAX_SIZE); } void checkasm_check_opusdsp(void)