May 2024 - ffmpeg-cvslog - lists.ffmpeg.org

checkasm/lpc: use fixed length to bench apply_welch_window
by James Almer 31 May '24

31 May '24

ffmpeg | branch: master | James Almer <jamrial(a)gmail.com> | Fri May 31 17:06:08 2024 -0300| [0a949aacaedb7c3a503028c4c1a1067162483073] | committer: James Almer checkasm/lpc: use fixed length to bench apply_welch_window Signed-off-by: James Almer <jamrial(a)gmail.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=0a949aacaedb7c3a5… --- tests/checkasm/lpc.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/checkasm/lpc.c b/tests/checkasm/lpc.c index cdd1b02698..514a2cdb3c 100644 --- a/tests/checkasm/lpc.c +++ b/tests/checkasm/lpc.c @@ -55,7 +55,7 @@ static void test_window(int len) } } - bench_new(src, len, dst1); + bench_new(src, 4608 + (len & 1), dst1); } static void test_compute_autocorr(ptrdiff_t len, int lag)

1 0

checkasm/lpc: test compute_autocorr
by Rémi Denis-Courmont 31 May '24

31 May '24

ffmpeg | branch: master | Rémi Denis-Courmont <remi(a)remlab.net> | Fri May 31 20:51:14 2024 +0300| [16132a810d04c680d4fa90d592a216bf1fabb348] | committer: James Almer checkasm/lpc: test compute_autocorr Signed-off-by: James Almer <jamrial(a)gmail.com> > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=16132a810d04c680d… --- tests/checkasm/lpc.c | 53 ++++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 51 insertions(+), 2 deletions(-) diff --git a/tests/checkasm/lpc.c b/tests/checkasm/lpc.c index 592e34c03d..cdd1b02698 100644 --- a/tests/checkasm/lpc.c +++ b/tests/checkasm/lpc.c @@ -16,6 +16,7 @@ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. */ +#include "libavutil/avassert.h" #include "libavutil/mem_internal.h" #include "libavcodec/lpc.h" @@ -57,10 +58,47 @@ static void test_window(int len) bench_new(src, len, dst1); } +static void test_compute_autocorr(ptrdiff_t len, int lag) +{ + const double eps = EPS * (double)len; + LOCAL_ALIGNED(32, double, src, [5000 + 2 + MAX_LPC_ORDER]); + LOCAL_ALIGNED(16, double, dst0, [MAX_LPC_ORDER + 1]); + LOCAL_ALIGNED(16, double, dst1, [MAX_LPC_ORDER + 1]); + + declare_func(void, const double *in, ptrdiff_t len, int lag, double *out); + + av_assert0(lag >= 0 && lag <= MAX_LPC_ORDER); + + for (int i = 0; i < MAX_LPC_ORDER; i++) + src[i] = 0.; + + src += MAX_LPC_ORDER; + + for (int i = 0; i < 5000 + 2; i++) { + src[i] = (double)rnd() / (double)UINT_MAX; + } + + call_ref(src, len, lag, dst0); + call_new(src, len, lag, dst1); + + for (size_t i = 0; i <= lag; i++) { + if (!double_near_abs_eps(dst0[i], dst1[i], eps)) { + fprintf(stderr, "%zu: %- .12f - %- .12f = % .12g\n", + i, dst0[i], dst1[i], dst0[i] - dst1[i]); + fail(); + break; + } + } + + bench_new(src, 4608 + (len & 1), lag, dst1); +} + void checkasm_check_lpc(void) { LPCContext ctx; - int len = rnd() % 5000; + int len = 2000 + rnd() % 3000; + static const int lags[] = { 8, 12, }; + ff_lpc_init(&ctx, 32, 16, FF_LPC_TYPE_DEFAULT); if (check_func(ctx.lpc_apply_welch_window, "apply_welch_window_even")) { @@ -72,6 +110,17 @@ void checkasm_check_lpc(void) test_window(len | 1); } report("apply_welch_window_odd"); - ff_lpc_end(&ctx); + + for (size_t i = 0; i < FF_ARRAY_ELEMS(lags); i++) { + ff_lpc_init(&ctx, len, lags[i], FF_LPC_TYPE_DEFAULT); + if (check_func(ctx.lpc_compute_autocorr, "autocorr_%d_even", lags[i])) + test_compute_autocorr(len & ~1, lags[i]); +#if !ARCH_X86 + if (check_func(ctx.lpc_compute_autocorr, "autocorr_%d_odd", lags[i])) + test_compute_autocorr(len | 1, lags[i]); +#endif + ff_lpc_end(&ctx); + } + report("compute_autocorr"); }

1 0

lavu/float_dsp: R-V V scalarproduct_double
by Rémi Denis-Courmont 31 May '24

31 May '24

ffmpeg | branch: master | Rémi Denis-Courmont <remi(a)remlab.net> | Wed May 29 17:57:09 2024 +0300| [9238f6cb415d5a688e76f804e74ebddfd556bdda] | committer: Rémi Denis-Courmont lavu/float_dsp: R-V V scalarproduct_double C908: scalarproduct_double_c: 39.2 scalarproduct_double_rvv_f64: 10.5 X60: scalarproduct_double_c: 35.0 scalarproduct_double_rvv_f64: 5.2 > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=9238f6cb415d5a688… --- libavutil/riscv/float_dsp_init.c | 3 +++ libavutil/riscv/float_dsp_rvv.S | 21 +++++++++++++++++++++ 2 files changed, 24 insertions(+) diff --git a/libavutil/riscv/float_dsp_init.c b/libavutil/riscv/float_dsp_init.c index 585f237225..155496fa6b 100644 --- a/libavutil/riscv/float_dsp_init.c +++ b/libavutil/riscv/float_dsp_init.c @@ -46,6 +46,8 @@ void ff_vector_dmac_scalar_rvv(double *dst, const double *src, double mul, int len); void ff_vector_dmul_scalar_rvv(double *dst, const double *src, double mul, int len); +double ff_scalarproduct_double_rvv(const double *v1, const double *v2, + size_t len); av_cold void ff_float_dsp_init_riscv(AVFloatDSPContext *fdsp) { @@ -68,6 +70,7 @@ av_cold void ff_float_dsp_init_riscv(AVFloatDSPContext *fdsp) fdsp->vector_dmul = ff_vector_dmul_rvv; fdsp->vector_dmac_scalar = ff_vector_dmac_scalar_rvv; fdsp->vector_dmul_scalar = ff_vector_dmul_scalar_rvv; + fdsp->scalarproduct_double = ff_scalarproduct_double_rvv; } } #endif diff --git a/libavutil/riscv/float_dsp_rvv.S b/libavutil/riscv/float_dsp_rvv.S index e6ec182a7a..2f0ade6db6 100644 --- a/libavutil/riscv/float_dsp_rvv.S +++ b/libavutil/riscv/float_dsp_rvv.S @@ -249,3 +249,24 @@ NOHWD mv a2, a3 ret endfunc + +func ff_scalarproduct_double_rvv, zve64f + vsetvli t0, zero, e64, m8, ta, ma + vmv.v.x v8, zero + vmv.s.x v0, zero +1: + vsetvli t0, a2, e64, m8, tu, ma + vle64.v v16, (a0) + sub a2, a2, t0 + vle64.v v24, (a1) + sh3add a0, t0, a0 + vfmacc.vv v8, v16, v24 + sh3add a1, t0, a1 + bnez a2, 1b + + vsetvli t0, zero, e64, m8, ta, ma + vfredusum.vs v0, v8, v0 + vfmv.f.s fa0, v0 +NOHWD fmv.x.w a0, fa0 + ret +endfunc

1 0

lavfi: get rid of bespoke double scalar products
by Rémi Denis-Courmont 31 May '24

31 May '24

ffmpeg | branch: master | Rémi Denis-Courmont <remi(a)remlab.net> | Wed May 29 17:51:06 2024 +0300| [b1149520305839eaceeb20864f501c96d358731e] | committer: Rémi Denis-Courmont lavfi: get rid of bespoke double scalar products > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=b1149520305839eac… --- libavfilter/aap_template.c | 14 +------------- libavfilter/anlms_template.c | 16 ++-------------- libavfilter/arls_template.c | 14 +------------- 3 files changed, 4 insertions(+), 40 deletions(-) diff --git a/libavfilter/aap_template.c b/libavfilter/aap_template.c index ea9c815a89..0e0580fb32 100644 --- a/libavfilter/aap_template.c +++ b/libavfilter/aap_template.c @@ -36,18 +36,6 @@ #define fn2(a,b) fn3(a,b) #define fn(a) fn2(a, SAMPLE_FORMAT) -#if DEPTH == 64 -static double scalarproduct_double(const double *v1, const double *v2, int len) -{ - double p = 0.0; - - for (int i = 0; i < len; i++) - p += v1[i] * v2[i]; - - return p; -} -#endif - static ftype fn(fir_sample)(AudioAPContext *s, ftype sample, ftype *delay, ftype *coeffs, ftype *tmp, int *offset) { @@ -60,7 +48,7 @@ static ftype fn(fir_sample)(AudioAPContext *s, ftype sample, ftype *delay, #if DEPTH == 32 output = s->fdsp->scalarproduct_float(delay, tmp, s->kernel_size); #else - output = scalarproduct_double(delay, tmp, s->kernel_size); + output = s->fdsp->scalarproduct_double(delay, tmp, s->kernel_size); #endif if (--(*offset) < 0) diff --git a/libavfilter/anlms_template.c b/libavfilter/anlms_template.c index b25df4fa18..a8d1dbfe0f 100644 --- a/libavfilter/anlms_template.c +++ b/libavfilter/anlms_template.c @@ -33,18 +33,6 @@ #define fn2(a,b) fn3(a,b) #define fn(a) fn2(a, SAMPLE_FORMAT) -#if DEPTH == 64 -static double scalarproduct_double(const double *v1, const double *v2, int len) -{ - double p = 0.0; - - for (int i = 0; i < len; i++) - p += v1[i] * v2[i]; - - return p; -} -#endif - static ftype fn(fir_sample)(AudioNLMSContext *s, ftype sample, ftype *delay, ftype *coeffs, ftype *tmp, int *offset) { @@ -58,7 +46,7 @@ static ftype fn(fir_sample)(AudioNLMSContext *s, ftype sample, ftype *delay, #if DEPTH == 32 output = s->fdsp->scalarproduct_float(delay, tmp, s->kernel_size); #else - output = scalarproduct_double(delay, tmp, s->kernel_size); + output = s->fdsp->scalarproduct_double(delay, tmp, s->kernel_size); #endif if (--(*offset) < 0) @@ -85,7 +73,7 @@ static ftype fn(process_sample)(AudioNLMSContext *s, ftype input, ftype desired, #if DEPTH == 32 sum = s->fdsp->scalarproduct_float(delay, delay, s->kernel_size); #else - sum = scalarproduct_double(delay, delay, s->kernel_size); + sum = s->fdsp->scalarproduct_double(delay, delay, s->kernel_size); #endif norm = s->eps + sum; b = mu * e / norm; diff --git a/libavfilter/arls_template.c b/libavfilter/arls_template.c index d8b19d89a5..c67b48cf6f 100644 --- a/libavfilter/arls_template.c +++ b/libavfilter/arls_template.c @@ -39,18 +39,6 @@ #define fn2(a,b) fn3(a,b) #define fn(a) fn2(a, SAMPLE_FORMAT) -#if DEPTH == 64 -static double scalarproduct_double(const double *v1, const double *v2, int len) -{ - double p = 0.0; - - for (int i = 0; i < len; i++) - p += v1[i] * v2[i]; - - return p; -} -#endif - static ftype fn(fir_sample)(AudioRLSContext *s, ftype sample, ftype *delay, ftype *coeffs, ftype *tmp, int *offset) { @@ -64,7 +52,7 @@ static ftype fn(fir_sample)(AudioRLSContext *s, ftype sample, ftype *delay, #if DEPTH == 32 output = s->fdsp->scalarproduct_float(delay, tmp, s->kernel_size); #else - output = scalarproduct_double(delay, tmp, s->kernel_size); + output = s->fdsp->scalarproduct_double(delay, tmp, s->kernel_size); #endif if (--(*offset) < 0)

1 0

lavu/lls: use ff_scalarproduct_double_c()
by Rémi Denis-Courmont 31 May '24

31 May '24

ffmpeg | branch: master | Rémi Denis-Courmont <remi(a)remlab.net> | Wed May 29 19:02:55 2024 +0300| [73c278d2702cbf07d6c011c5867e08eab43896b0] | committer: Rémi Denis-Courmont lavu/lls: use ff_scalarproduct_double_c() > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=73c278d2702cbf07d… --- libavutil/lls.c | 9 ++------- 1 file changed, 2 insertions(+), 7 deletions(-) diff --git a/libavutil/lls.c b/libavutil/lls.c index c1e038daf1..1096ae69d5 100644 --- a/libavutil/lls.c +++ b/libavutil/lls.c @@ -30,6 +30,7 @@ #include "config.h" #include "attributes.h" +#include "float_dsp.h" #include "lls.h" static void update_lls(LLSModel *m, const double *var) @@ -102,13 +103,7 @@ void avpriv_solve_lls(LLSModel *m, double threshold, unsigned short min_order) static double evaluate_lls(LLSModel *m, const double *param, int order) { - int i; - double out = 0; - - for (i = 0; i <= order; i++) - out += param[i] * m->coeff[order][i]; - - return out; + return ff_scalarproduct_double_c(m->coeff[order], param, order + 1); } av_cold void avpriv_init_lls(LLSModel *m, int indep_count)

1 0

checkasm/float_dsp: add double-precision scalar product
by Rémi Denis-Courmont 31 May '24

31 May '24

ffmpeg | branch: master | Rémi Denis-Courmont <remi(a)remlab.net> | Wed May 29 17:50:37 2024 +0300| [98405d28fab14933a0ef0d76d61d9806bbb912ba] | committer: Rémi Denis-Courmont checkasm/float_dsp: add double-precision scalar product > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=98405d28fab14933a… --- tests/checkasm/float_dsp.c | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/tests/checkasm/float_dsp.c b/tests/checkasm/float_dsp.c index cadfa65e2a..296db1cff9 100644 --- a/tests/checkasm/float_dsp.c +++ b/tests/checkasm/float_dsp.c @@ -278,6 +278,22 @@ static void test_scalarproduct_float(const float *src0, const float *src1) bench_new(src0, src1, LEN); } +static void test_scalarproduct_double(const double *src0, const double *src1) +{ + double cprod, oprod; + + declare_func_float(double, const double *, const double *, size_t); + + cprod = call_ref(src0, src1, LEN); + oprod = call_new(src0, src1, LEN); + if (!double_near_abs_eps(cprod, oprod, ARBITRARY_SCALARPRODUCT_CONST)) { + fprintf(stderr, "%- .12f - %- .12f = % .12g\n", + cprod, oprod, cprod - oprod); + fail(); + } + bench_new(src0, src1, LEN); +} + void checkasm_check_float_dsp(void) { LOCAL_ALIGNED_32(float, src0, [LEN]); @@ -334,6 +350,9 @@ void checkasm_check_float_dsp(void) if (check_func(fdsp->scalarproduct_float, "scalarproduct_float")) test_scalarproduct_float(src3, src4); report("scalarproduct_float"); + if (check_func(fdsp->scalarproduct_double, "scalarproduct_double")) + test_scalarproduct_double(dbl_src0, dbl_src1); + report("scalarproduct_double"); av_freep(&fdsp); }

1 0

lavu/float_dsp: add double-precision scalar product
by Rémi Denis-Courmont 31 May '24

31 May '24

ffmpeg | branch: master | Rémi Denis-Courmont <remi(a)remlab.net> | Wed May 29 17:43:21 2024 +0300| [6a7c4d60a1498929c2a366f2ef4ccc35621a4358] | committer: Rémi Denis-Courmont lavu/float_dsp: add double-precision scalar product The function pointer is appended to the structure for backward binary compatibility. Fortunately, this is allocated by libavutil, not by the user, so increasing the structure size is safe. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=6a7c4d60a1498929c… --- libavutil/float_dsp.c | 12 ++++++++++++ libavutil/float_dsp.h | 34 ++++++++++++++++++++++++++++++---- 2 files changed, 42 insertions(+), 4 deletions(-) diff --git a/libavutil/float_dsp.c b/libavutil/float_dsp.c index e9fb023466..08bbc85e3e 100644 --- a/libavutil/float_dsp.c +++ b/libavutil/float_dsp.c @@ -132,6 +132,17 @@ float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len) return p; } +double ff_scalarproduct_double_c(const double *v1, const double *v2, + size_t len) +{ + double p = 0.0; + + for (size_t i = 0; i < len; i++) + p += v1[i] * v2[i]; + + return p; +} + av_cold AVFloatDSPContext *avpriv_float_dsp_alloc(int bit_exact) { AVFloatDSPContext *fdsp = av_mallocz(sizeof(AVFloatDSPContext)); @@ -149,6 +160,7 @@ av_cold AVFloatDSPContext *avpriv_float_dsp_alloc(int bit_exact) fdsp->vector_fmul_reverse = vector_fmul_reverse_c; fdsp->butterflies_float = butterflies_float_c; fdsp->scalarproduct_float = avpriv_scalarproduct_float_c; + fdsp->scalarproduct_double = ff_scalarproduct_double_c; #if ARCH_AARCH64 ff_float_dsp_init_aarch64(fdsp); diff --git a/libavutil/float_dsp.h b/libavutil/float_dsp.h index 342a8715c5..ca05fd9382 100644 --- a/libavutil/float_dsp.h +++ b/libavutil/float_dsp.h @@ -19,6 +19,8 @@ #ifndef AVUTIL_FLOAT_DSP_H #define AVUTIL_FLOAT_DSP_H +#include <stddef.h> + typedef struct AVFloatDSPContext { /** * Calculate the entry wise product of two vectors of floats and store the result in @@ -177,20 +179,32 @@ typedef struct AVFloatDSPContext { * a vector of doubles. * * @param dst output vector - * constraints: 32-byte aligned * @param src0 first input vector - * constraints: 32-byte aligned * @param src1 second input vector - * constraints: 32-byte aligned * @param len number of elements in the input * constraints: multiple of 16 */ void (*vector_dmul)(double *dst, const double *src0, const double *src1, int len); + + /** + * Calculate the scalar product of two vectors of doubles. + * + * @param v1 first vector + * constraints: 32-byte aligned + * @param v2 second vector + * constraints: 32-byte aligned + * @param len length of vectors + * constraints: 32-byte aligned + * + * @return inner product of the vectors + */ + double (*scalarproduct_double)(const double *v1, const double *v2, + size_t len); } AVFloatDSPContext; /** - * Return the scalar product of two vectors. + * Return the scalar product of two vectors of floats. * * @param v1 first input vector * @param v2 first input vector @@ -200,6 +214,18 @@ typedef struct AVFloatDSPContext { */ float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len); +/** + * Return the scalar product of two vectors of doubles. + * + * @param v1 first input vector + * @param v2 first input vector + * @param len number of elements + * + * @return inner product of the vectors + */ +double ff_scalarproduct_double_c(const double *v1, const double *v2, + size_t len); + void ff_float_dsp_init_aarch64(AVFloatDSPContext *fdsp); void ff_float_dsp_init_arm(AVFloatDSPContext *fdsp); void ff_float_dsp_init_ppc(AVFloatDSPContext *fdsp, int strict);

1 0

lavc/sbrdsp: add support for 256-bit vectors
by Rémi Denis-Courmont 31 May '24

31 May '24

ffmpeg | branch: master | Rémi Denis-Courmont <remi(a)remlab.net> | Sat May 25 21:09:46 2024 +0300| [06fc919aadd18f5cfb4e5c60fe31bc4dbccede77] | committer: Rémi Denis-Courmont lavc/sbrdsp: add support for 256-bit vectors hf_apply_noise_0_c: 35.7 hf_apply_noise_0_rvv_f32: 9.5 hf_apply_noise_1_c: 38.5 hf_apply_noise_1_rvv_f32: 10.0 hf_apply_noise_2_c: 35.5 hf_apply_noise_2_rvv_f32: 9.7 hf_apply_noise_3_c: 38.5 hf_apply_noise_3_rvv_f32: 10.0 Maybe extending the noise table manually is not such great idea, but I not quite sure how to deal with that otherwise? Allocating the table dynamically is possible but would require an ELF destructor to clean up. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=06fc919aadd18f5cf… --- libavcodec/riscv/sbrdsp_init.c | 2 +- libavcodec/sbrdsp_template.c | 8 ++++++++ 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/libavcodec/riscv/sbrdsp_init.c b/libavcodec/riscv/sbrdsp_init.c index 6c17b12ae0..562c75e6b4 100644 --- a/libavcodec/riscv/sbrdsp_init.c +++ b/libavcodec/riscv/sbrdsp_init.c @@ -52,7 +52,7 @@ av_cold void ff_sbrdsp_init_riscv(SBRDSPContext *c) c->sum_square = ff_sbr_sum_square_rvv; c->hf_gen = ff_sbr_hf_gen_rvv; c->hf_g_filt = ff_sbr_hf_g_filt_rvv; - if (ff_get_rv_vlenb() <= 16) { + if (ff_get_rv_vlenb() <= 32) { c->hf_apply_noise[0] = ff_sbr_hf_apply_noise_0_rvv; c->hf_apply_noise[2] = ff_sbr_hf_apply_noise_2_rvv; if (flags & AV_CPU_FLAG_RVB_BASIC) { diff --git a/libavcodec/sbrdsp_template.c b/libavcodec/sbrdsp_template.c index 0f731ba50d..9a94af8670 100644 --- a/libavcodec/sbrdsp_template.c +++ b/libavcodec/sbrdsp_template.c @@ -376,5 +376,13 @@ const attribute_visibility_hidden DECLARE_ALIGNED(16, INTFLOAT, AAC_RENAME(ff_sb {Q31(-0.99867974711855f), Q31(-0.88147068645358f)}, {Q31(-0.95531076805040f), Q31( 0.90908757154593f)}, {Q31(-0.45725933317144f), Q31(-0.56716323646760f)}, {Q31(-0.72929675029275f), Q31(-0.98008272727324f)}, {Q31( 0.75622801399036f), Q31( 0.20950329995549f)}, {Q31( 0.07069442601050f), Q31(-0.78247898470706f)}, +{Q31( 0.74496252926055f), Q31(-0.91169004445807f)}, {Q31(-0.96440182703856f), Q31(-0.94739918296622f)}, +{Q31( 0.30424629369539f), Q31(-0.49438267012479f)}, {Q31( 0.66565033746925f), Q31( 0.64652935542491f)}, +{Q31( 0.91697008020594f), Q31( 0.17514097332009f)}, {Q31(-0.70774918760427f), Q31( 0.52548653416543f)}, +{Q31(-0.70051415345560f), Q31(-0.45340028808763f)}, {Q31(-0.99496513054797f), Q31(-0.90071908066973f)}, +{Q31( 0.98164490790123f), Q31(-0.77463155528697f)}, {Q31(-0.54671580548181f), Q31(-0.02570928536004f)}, +{Q31(-0.01689629065389f), Q31( 0.00287506445732f)}, {Q31(-0.86110349531986f), Q31( 0.42548583726477f)}, +{Q31(-0.98892980586032f), Q31(-0.87881132267556f)}, {Q31( 0.51756627678691f), Q31( 0.66926784710139f)}, +{Q31(-0.99635026409640f), Q31(-0.58107730574765f)}, {Q31(-0.99969370862163f), Q31( 0.98369989360250f)}, #endif };

1 0

lavc/hevc_ps: compactify ShortTermRPS
by Anton Khirnov 31 May '24

31 May '24

ffmpeg | branch: master | Anton Khirnov <anton(a)khirnov.net> | Wed Apr 10 14:42:37 2024 +0200| [63a96dbcced2a67e96ee7306874dd2574e2d7d74] | committer: Anton Khirnov lavc/hevc_ps: compactify ShortTermRPS Do not use larger fields than needed, use size-1 bitfields for flags. Reduces sizeof(HEVCSPS) by 1280 bytes. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=63a96dbcced2a67e9… --- libavcodec/hevc_ps.c | 6 +++--- libavcodec/hevc_ps.h | 19 +++++++++++-------- libavcodec/vulkan_hevc.c | 2 +- 3 files changed, 15 insertions(+), 12 deletions(-) diff --git a/libavcodec/hevc_ps.c b/libavcodec/hevc_ps.c index 76fe507e7b..7b486ce0af 100644 --- a/libavcodec/hevc_ps.c +++ b/libavcodec/hevc_ps.c @@ -143,11 +143,11 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, for (i = 0; i <= rps_ridx->num_delta_pocs; i++) { used[k] = get_bits1(gb); - rps->use_delta_flag = 0; + rps->use_delta = 0; if (!used[k]) - rps->use_delta_flag = get_bits1(gb); + rps->use_delta = get_bits1(gb); - if (used[k] || rps->use_delta_flag) { + if (used[k] || rps->use_delta) { if (i < rps_ridx->num_delta_pocs) delta_poc = delta_rps + rps_ridx->delta_poc[i]; else diff --git a/libavcodec/hevc_ps.h b/libavcodec/hevc_ps.h index ed6372c747..d06d7cf1d4 100644 --- a/libavcodec/hevc_ps.h +++ b/libavcodec/hevc_ps.h @@ -70,16 +70,19 @@ typedef struct HEVCHdrParams { } HEVCHdrParams; typedef struct ShortTermRPS { - uint8_t rps_predict; - unsigned int delta_idx; - uint8_t use_delta_flag; - uint8_t delta_rps_sign; - unsigned int abs_delta_rps; - unsigned int num_negative_pics; - int num_delta_pocs; - int rps_idx_num_delta_pocs; int32_t delta_poc[32]; uint32_t used; + + uint8_t delta_idx; + uint8_t num_negative_pics; + uint8_t num_delta_pocs; + uint8_t rps_idx_num_delta_pocs; + + uint16_t abs_delta_rps; + unsigned delta_rps_sign:1; + + unsigned rps_predict:1; + unsigned use_delta:1; } ShortTermRPS; typedef struct HEVCWindow { diff --git a/libavcodec/vulkan_hevc.c b/libavcodec/vulkan_hevc.c index a35f3d992d..5583f56285 100644 --- a/libavcodec/vulkan_hevc.c +++ b/libavcodec/vulkan_hevc.c @@ -359,7 +359,7 @@ static void set_sps(const HEVCSPS *sps, int sps_idx, .delta_rps_sign = sps->st_rps[i].delta_rps_sign, }, .delta_idx_minus1 = sps->st_rps[i].delta_idx - 1, - .use_delta_flag = sps->st_rps[i].use_delta_flag, + .use_delta_flag = sps->st_rps[i].use_delta, .abs_delta_rps_minus1 = sps->st_rps[i].abs_delta_rps - 1, .used_by_curr_pic_flag = 0x0, .used_by_curr_pic_s0_flag = 0x0,

1 0

lavc/hevc_ps: reduce the size of ShortTermRPS.used
by Anton Khirnov 31 May '24

31 May '24

ffmpeg | branch: master | Anton Khirnov <anton(a)khirnov.net> | Wed Apr 10 11:27:51 2024 +0200| [9127819d514a3a82031f7448b40714f2880d7804] | committer: Anton Khirnov lavc/hevc_ps: reduce the size of ShortTermRPS.used It is currently an array of 32 uint8_t, each storing a single flag. A single uint32_t is sufficient. Reduces sizeof(HEVCSPS) by 1792 bytes. > http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=9127819d514a3a820… --- libavcodec/hevc_ps.c | 33 +++++++++++++++++++-------------- libavcodec/hevc_ps.h | 2 +- libavcodec/hevc_refs.c | 6 +++--- libavcodec/vulkan_hevc.c | 13 +++++-------- 4 files changed, 28 insertions(+), 26 deletions(-) diff --git a/libavcodec/hevc_ps.c b/libavcodec/hevc_ps.c index a6b0021bc3..76fe507e7b 100644 --- a/libavcodec/hevc_ps.c +++ b/libavcodec/hevc_ps.c @@ -107,6 +107,7 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, int k = 0; int i; + rps->used = 0; rps->rps_predict = 0; if (rps != sps->st_rps && sps->nb_st_rps) @@ -114,6 +115,7 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, if (rps->rps_predict) { const ShortTermRPS *rps_ridx; + uint8_t used[32] = { 0 }; int delta_rps; if (is_slice_header) { @@ -139,13 +141,13 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, } delta_rps = (1 - (rps->delta_rps_sign << 1)) * rps->abs_delta_rps; for (i = 0; i <= rps_ridx->num_delta_pocs; i++) { - int used = rps->used[k] = get_bits1(gb); + used[k] = get_bits1(gb); rps->use_delta_flag = 0; - if (!used) + if (!used[k]) rps->use_delta_flag = get_bits1(gb); - if (used || rps->use_delta_flag) { + if (used[k] || rps->use_delta_flag) { if (i < rps_ridx->num_delta_pocs) delta_poc = delta_rps + rps_ridx->delta_poc[i]; else @@ -157,7 +159,7 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, } } - if (k >= FF_ARRAY_ELEMS(rps->used)) { + if (k >= FF_ARRAY_ELEMS(used)) { av_log(avctx, AV_LOG_ERROR, "Invalid num_delta_pocs: %d\n", k); return AVERROR_INVALIDDATA; @@ -167,35 +169,38 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, rps->num_negative_pics = k0; // sort in increasing order (smallest first) if (rps->num_delta_pocs != 0) { - int used, tmp; + int u, tmp; for (i = 1; i < rps->num_delta_pocs; i++) { delta_poc = rps->delta_poc[i]; - used = rps->used[i]; + u = used[i]; for (k = i - 1; k >= 0; k--) { tmp = rps->delta_poc[k]; if (delta_poc < tmp) { rps->delta_poc[k + 1] = tmp; - rps->used[k + 1] = rps->used[k]; + used[k + 1] = used[k]; rps->delta_poc[k] = delta_poc; - rps->used[k] = used; + used[k] = u; } } } } if ((rps->num_negative_pics >> 1) != 0) { - int used; + int u; k = rps->num_negative_pics - 1; // flip the negative values to largest first for (i = 0; i < rps->num_negative_pics >> 1; i++) { delta_poc = rps->delta_poc[i]; - used = rps->used[i]; + u = used[i]; rps->delta_poc[i] = rps->delta_poc[k]; - rps->used[i] = rps->used[k]; + used[i] = used[k]; rps->delta_poc[k] = delta_poc; - rps->used[k] = used; + used[k] = u; k--; } } + + for (unsigned i = 0; i < FF_ARRAY_ELEMS(used); i++) + rps->used |= used[i] * (1 << i); } else { unsigned int nb_positive_pics; @@ -222,7 +227,7 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, } prev -= delta_poc; rps->delta_poc[i] = prev; - rps->used[i] = get_bits1(gb); + rps->used |= get_bits1(gb) * (1 << i); } prev = 0; for (i = 0; i < nb_positive_pics; i++) { @@ -235,7 +240,7 @@ int ff_hevc_decode_short_term_rps(GetBitContext *gb, AVCodecContext *avctx, } prev += delta_poc; rps->delta_poc[rps->num_negative_pics + i] = prev; - rps->used[rps->num_negative_pics + i] = get_bits1(gb); + rps->used |= get_bits1(gb) * (1 << (rps->num_negative_pics + i)); } } } diff --git a/libavcodec/hevc_ps.h b/libavcodec/hevc_ps.h index 1d3bdca4c6..ed6372c747 100644 --- a/libavcodec/hevc_ps.h +++ b/libavcodec/hevc_ps.h @@ -79,7 +79,7 @@ typedef struct ShortTermRPS { int num_delta_pocs; int rps_idx_num_delta_pocs; int32_t delta_poc[32]; - uint8_t used[32]; + uint32_t used; } ShortTermRPS; typedef struct HEVCWindow { diff --git a/libavcodec/hevc_refs.c b/libavcodec/hevc_refs.c index 8da9ec982a..d6dc2f9e0a 100644 --- a/libavcodec/hevc_refs.c +++ b/libavcodec/hevc_refs.c @@ -497,7 +497,7 @@ int ff_hevc_frame_rps(HEVCContext *s) int poc = s->poc + short_rps->delta_poc[i]; int list; - if (!short_rps->used[i]) + if (!(short_rps->used & (1 << i))) list = ST_FOLL; else if (i < short_rps->num_negative_pics) list = ST_CURR_BEF; @@ -536,9 +536,9 @@ int ff_hevc_frame_nb_refs(const HEVCContext *s) if (rps) { for (i = 0; i < rps->num_negative_pics; i++) - ret += !!rps->used[i]; + ret += !!(rps->used & (1 << i)); for (; i < rps->num_delta_pocs; i++) - ret += !!rps->used[i]; + ret += !!(rps->used & (1 << i)); } if (long_rps) { diff --git a/libavcodec/vulkan_hevc.c b/libavcodec/vulkan_hevc.c index 21cf49c0ec..a35f3d992d 100644 --- a/libavcodec/vulkan_hevc.c +++ b/libavcodec/vulkan_hevc.c @@ -373,19 +373,16 @@ static void set_sps(const HEVCSPS *sps, int sps_idx, /* NOTE: This is the predicted, and *reordered* version. * Probably incorrect, but the spec doesn't say which version to use. */ - for (int j = 0; j < sps->st_rps[i].num_delta_pocs; j++) - str[i].used_by_curr_pic_flag |= sps->st_rps[i].used[j] << j; + str[i].used_by_curr_pic_flag = st_rps->used; + str[i].used_by_curr_pic_s0_flag = av_mod_uintp2(st_rps->used, str[i].num_negative_pics); + str[i].used_by_curr_pic_s1_flag = st_rps->used >> str[i].num_negative_pics; - for (int j = 0; j < str[i].num_negative_pics; j++) { + for (int j = 0; j < str[i].num_negative_pics; j++) str[i].delta_poc_s0_minus1[j] = st_rps->delta_poc[j] - (j ? st_rps->delta_poc[j - 1] : 0) - 1; - str[i].used_by_curr_pic_s0_flag |= sps->st_rps[i].used[j] << j; - } - for (int j = 0; j < str[i].num_positive_pics; j++) { + for (int j = 0; j < str[i].num_positive_pics; j++) str[i].delta_poc_s1_minus1[j] = st_rps->delta_poc[st_rps->num_negative_pics + j] - (j ? st_rps->delta_poc[st_rps->num_negative_pics + j - 1] : 0) - 1; - str[i].used_by_curr_pic_s1_flag |= sps->st_rps[i].used[str[i].num_negative_pics + j] << j; - } } *ltr = (StdVideoH265LongTermRefPicsSps) {

1 0