From patchwork Tue Jan 8 09:11:56 2019 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Lauri Kasanen X-Patchwork-Id: 11679 Return-Path: X-Original-To: patchwork@ffaux-bg.ffmpeg.org Delivered-To: patchwork@ffaux-bg.ffmpeg.org Received: from ffbox0-bg.mplayerhq.hu (ffbox0-bg.ffmpeg.org [79.124.17.100]) by ffaux.localdomain (Postfix) with ESMTP id 267E644DC2C for ; Tue, 8 Jan 2019 11:14:31 +0200 (EET) Received: from [127.0.1.1] (localhost [127.0.0.1]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTP id D201B68A551; Tue, 8 Jan 2019 11:14:27 +0200 (EET) X-Original-To: ffmpeg-devel@ffmpeg.org Delivered-To: ffmpeg-devel@ffmpeg.org Received: from mout.gmx.net (mout.gmx.net [212.227.17.20]) by ffbox0-bg.mplayerhq.hu (Postfix) with ESMTPS id 87200689E65 for ; Tue, 8 Jan 2019 11:14:21 +0200 (EET) Received: from Valinor ([84.250.81.169]) by mail.gmx.com (mrgmx103 [212.227.17.174]) with ESMTPSA (Nemesis) id 0Ma2Lr-1h0FW01cmV-00LnHT for ; Tue, 08 Jan 2019 10:09:23 +0100 Date: Tue, 8 Jan 2019 11:11:56 +0200 From: Lauri Kasanen To: ffmpeg-devel@ffmpeg.org Message-Id: <20190108111156.0503b5c1485589c8407f7b73@gmx.com> X-Mailer: Sylpheed 3.5.0 (GTK+ 2.18.6; x86_64-unknown-linux-gnu) Mime-Version: 1.0 X-Provags-ID: V03:K1:38TByXT68kitQLaS7WpohYUMeXywbVwSkBPl/aBxWhVxHmJV2t+ ECy2wu8Tc1o0dttQSB9chtsFzdSpaSThDktL04mG4V6MLgjmgXpzdTqq20jyrDh9XcYVnh3 +A1wDo0zmwIP+R13VdMejWgd2uny/y1DqtPSb3XZysZa9tM72WXLBXGYzZ7H+PSDxrdBkYr o/OHSpIbQE98CyEWsqHkQ== X-Spam-Flag: NO X-UI-Out-Filterresults: notjunk:1; V03:K0:IXUuh3NuR7M=:skbFtlwy/H5gVbH5cHrf4l pViG2nBKs6sSr43q0bf96nPuFsSX17kldsnYH5IGekc9dVGNthq3yNAol5n+pEoadeOPYSS3C kA3csJQGxZCX1ztb8ltbgAdOwzs+OFyv0ILqnK4d4nRlVpZ3cVekrJsWkICLm/CSgg4xYlJ/Q joQZodi8lRF3Y5Q1zw7D7T26zCsewVowSIZvJhEXirjE8HrPBM9Ueg4npCFHfUPYkjcDnyo9X ZEix4D8r2EWMCEE9wvb6UIHbmT0jc/EBRKpxLi4k87JA58hXwfSgUrVFTvfClR6uM0jwlbbW2 CchRxp28mH1GfwFp98ORhoqALSy8vGR8U1jPVtQIQ8HfYse3vIja74N4L5XFIj3nYa62qQzEM 4BnxTyUIYmJg5gAsysPSNIAuDN3ATLdp7vARguSvRnw9NTklwSh9UKlc1dKZfiLKqNG2EO4XR dAy4tfkvNg5us3fG5DA27xX/xfDKxyYschZtnqlXYt7gWnrw3yfm9IU+XRokg9Bq2Tl3QSpP4 WL69PuZeeFzfp9KsN/qa6psoNfeApUVmoLcXLxgdNuiowGkr+pCYzHBN1wA/ZdcfUUYWsxLdr bUlsMw5tVijqKF7H6aVFHxCgv6i3I3RdKOuQ3yDov9pJ7slSvKuuCaO1q5V2AIpMnLz5NGkOV x5rH6kdmppNvRY6GZ73s8zs60Al85NKUhpFoaFKIrBorbOwIn8Y5yWwL9sXQGCHB/AcPeImkX h+lkTy7merP1RXsnAIonC9SjpVurURF7lso0XDsXXuKc2OVC9SHSjTJjdZHPHR6zbhUf+Yrqp d4+JsCJkNa2LQEgOURZjDUlxZRSZ2H/3JSTUUm4f7InV1/sRd6s2ueJKB5qu8Y3qIiTmiuut1 Yo47U4+0uDMXIqc5l+bhCzazEACFw/0pvx5cGGqXrGSqcgLu1doN0LtvUHA+ZX Subject: [FFmpeg-devel] [PATCH v3] libswscale/ppc: VSX-optimize 9-16 bit yuv2planeX 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 Errors-To: ffmpeg-devel-bounces@ffmpeg.org Sender: "ffmpeg-devel" ./ffmpeg_g -f rawvideo -pix_fmt rgb24 -s hd1080 -i /dev/zero -pix_fmt yuv420p16be \ -s 1920x1728 -f null -vframes 100 -v error -nostats - 9-14 bit funcs get about 6x speedup, 16-bit gets about 15x. Fate passes, each format tested with an image to video conversion. Only POWER8 includes 32-bit vector multiplies, so POWER7 is locked out of the 16-bit function. This includes the vec_mulo/mule functions too, not just vmuluwm. yuv420p9le 12341 UNITS in planarX, 130976 runs, 96 skips 73752 UNITS in planarX, 131066 runs, 6 skips yuv420p9be 12364 UNITS in planarX, 131025 runs, 47 skips 73001 UNITS in planarX, 131055 runs, 17 skips yuv420p10le 12386 UNITS in planarX, 131042 runs, 30 skips 72735 UNITS in planarX, 131062 runs, 10 skips yuv420p10be 12337 UNITS in planarX, 131045 runs, 27 skips 72734 UNITS in planarX, 131057 runs, 15 skips yuv420p12le 12236 UNITS in planarX, 131058 runs, 14 skips 73029 UNITS in planarX, 131062 runs, 10 skips yuv420p12be 12218 UNITS in planarX, 130973 runs, 99 skips 72402 UNITS in planarX, 131069 runs, 3 skips yuv420p14le 12168 UNITS in planarX, 131067 runs, 5 skips 72480 UNITS in planarX, 131069 runs, 3 skips yuv420p14be 12358 UNITS in planarX, 130948 runs, 124 skips 73772 UNITS in planarX, 131063 runs, 9 skips yuv420p16le 10439 UNITS in planarX, 130911 runs, 161 skips 157923 UNITS in planarX, 131068 runs, 4 skips yuv420p16be 10463 UNITS in planarX, 130874 runs, 198 skips 154405 UNITS in planarX, 131061 runs, 11 skips Signed-off-by: Lauri Kasanen --- v2: Separate macros so that yuv2plane1_16_vsx remains available for power7 v3: Remove accidental tabs, switch to HAVE_POWER8 from configure + runtime check As far as I can tell, for HAVE_POWER8 to be defined, -march has to be at least power8, meaning with the current setup such a binary wouldn't run on POWER7. However using the configure define lets it be disabled in configure like Michael pointed out, and having the runtime check doesn't hurt any (it allows for future splits like on x86, where one binary can run on low cpu but use higher ISA if available). libswscale/ppc/swscale_ppc_template.c | 4 +- libswscale/ppc/swscale_vsx.c | 195 +++++++++++++++++++++++++++++++++- 2 files changed, 193 insertions(+), 6 deletions(-) diff --git a/libswscale/ppc/swscale_ppc_template.c b/libswscale/ppc/swscale_ppc_template.c index 00e4b99..11decab 100644 --- a/libswscale/ppc/swscale_ppc_template.c +++ b/libswscale/ppc/swscale_ppc_template.c @@ -21,7 +21,7 @@ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ -static void FUNC(yuv2planeX_16)(const int16_t *filter, int filterSize, +static void FUNC(yuv2planeX_8_16)(const int16_t *filter, int filterSize, const int16_t **src, uint8_t *dest, const uint8_t *dither, int offset, int x) { @@ -88,7 +88,7 @@ static void FUNC(yuv2planeX)(const int16_t *filter, int filterSize, yuv2planeX_u(filter, filterSize, src, dest, dst_u, dither, offset, 0); for (i = dst_u; i < dstW - 15; i += 16) - FUNC(yuv2planeX_16)(filter, filterSize, src, dest + i, dither, + FUNC(yuv2planeX_8_16)(filter, filterSize, src, dest + i, dither, offset, i); yuv2planeX_u(filter, filterSize, src, dest, dstW, dither, offset, i); diff --git a/libswscale/ppc/swscale_vsx.c b/libswscale/ppc/swscale_vsx.c index 70da6ae..77680f8 100644 --- a/libswscale/ppc/swscale_vsx.c +++ b/libswscale/ppc/swscale_vsx.c @@ -83,6 +83,8 @@ #include "swscale_ppc_template.c" #undef FUNC +#undef vzero + #endif /* !HAVE_BIGENDIAN */ static void yuv2plane1_8_u(const int16_t *src, uint8_t *dest, int dstW, @@ -180,6 +182,76 @@ static void yuv2plane1_nbps_vsx(const int16_t *src, uint16_t *dest, int dstW, yuv2plane1_nbps_u(src, dest, dstW, big_endian, output_bits, i); } +static void yuv2planeX_nbps_u(const int16_t *filter, int filterSize, + const int16_t **src, uint16_t *dest, int dstW, + int big_endian, int output_bits, int start) +{ + int i; + int shift = 11 + 16 - output_bits; + + for (i = start; i < dstW; i++) { + int val = 1 << (shift - 1); + int j; + + for (j = 0; j < filterSize; j++) + val += src[j][i] * filter[j]; + + output_pixel(&dest[i], val); + } +} + +static void yuv2planeX_nbps_vsx(const int16_t *filter, int filterSize, + const int16_t **src, uint16_t *dest, int dstW, + int big_endian, int output_bits) +{ + const int dst_u = -(uintptr_t)dest & 7; + const int shift = 11 + 16 - output_bits; + const int add = (1 << (shift - 1)); + const int clip = (1 << output_bits) - 1; + const uint16_t swap = big_endian ? 8 : 0; + const vector uint32_t vadd = (vector uint32_t) {add, add, add, add}; + const vector uint32_t vshift = (vector uint32_t) {shift, shift, shift, shift}; + const vector uint16_t vswap = (vector uint16_t) {swap, swap, swap, swap, swap, swap, swap, swap}; + const vector uint16_t vlargest = (vector uint16_t) {clip, clip, clip, clip, clip, clip, clip, clip}; + const vector int16_t vzero = vec_splat_s16(0); + const vector uint8_t vperm = (vector uint8_t) {0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15}; + vector int16_t vfilter[MAX_FILTER_SIZE], vin; + vector uint16_t v; + vector uint32_t vleft, vright, vtmp; + int i, j; + + for (i = 0; i < filterSize; i++) { + vfilter[i] = (vector int16_t) {filter[i], filter[i], filter[i], filter[i], + filter[i], filter[i], filter[i], filter[i]}; + } + + yuv2planeX_nbps_u(filter, filterSize, src, dest, dst_u, big_endian, output_bits, 0); + + for (i = dst_u; i < dstW - 7; i += 8) { + vleft = vright = vadd; + + for (j = 0; j < filterSize; j++) { + vin = vec_vsx_ld(0, &src[j][i]); + vtmp = (vector uint32_t) vec_mule(vin, vfilter[j]); + vleft = vec_add(vleft, vtmp); + vtmp = (vector uint32_t) vec_mulo(vin, vfilter[j]); + vright = vec_add(vright, vtmp); + } + + vleft = vec_sra(vleft, vshift); + vright = vec_sra(vright, vshift); + v = vec_packsu(vleft, vright); + v = (vector uint16_t) vec_max((vector int16_t) v, vzero); + v = vec_min(v, vlargest); + v = vec_rl(v, vswap); + v = vec_perm(v, v, vperm); + vec_st(v, 0, &dest[i]); + } + + yuv2planeX_nbps_u(filter, filterSize, src, dest, dstW, big_endian, output_bits, i); +} + + #undef output_pixel #define output_pixel(pos, val, bias, signedness) \ @@ -234,7 +306,97 @@ static void yuv2plane1_16_vsx(const int32_t *src, uint16_t *dest, int dstW, yuv2plane1_16_u(src, dest, dstW, big_endian, output_bits, i); } +#ifdef HAVE_POWER8 + +static void yuv2planeX_16_u(const int16_t *filter, int filterSize, + const int32_t **src, uint16_t *dest, int dstW, + int big_endian, int output_bits, int start) +{ + int i; + int shift = 15; + + for (i = start; i < dstW; i++) { + int val = 1 << (shift - 1); + int j; + + /* range of val is [0,0x7FFFFFFF], so 31 bits, but with lanczos/spline + * filters (or anything with negative coeffs, the range can be slightly + * wider in both directions. To account for this overflow, we subtract + * a constant so it always fits in the signed range (assuming a + * reasonable filterSize), and re-add that at the end. */ + val -= 0x40000000; + for (j = 0; j < filterSize; j++) + val += src[j][i] * (unsigned)filter[j]; + + output_pixel(&dest[i], val, 0x8000, int); + } +} + +static void yuv2planeX_16_vsx(const int16_t *filter, int filterSize, + const int32_t **src, uint16_t *dest, int dstW, + int big_endian, int output_bits) +{ + const int dst_u = -(uintptr_t)dest & 7; + const int shift = 15; + const int bias = 0x8000; + const int add = (1 << (shift - 1)) - 0x40000000; + const uint16_t swap = big_endian ? 8 : 0; + const vector uint32_t vadd = (vector uint32_t) {add, add, add, add}; + const vector uint32_t vshift = (vector uint32_t) {shift, shift, shift, shift}; + const vector uint16_t vswap = (vector uint16_t) {swap, swap, swap, swap, swap, swap, swap, swap}; + const vector uint16_t vbias = (vector uint16_t) {bias, bias, bias, bias, bias, bias, bias, bias}; + vector int32_t vfilter[MAX_FILTER_SIZE]; + vector uint16_t v; + vector uint32_t vleft, vright, vtmp; + vector int32_t vin32l, vin32r; + int i, j; + + for (i = 0; i < filterSize; i++) { + vfilter[i] = (vector int32_t) {filter[i], filter[i], filter[i], filter[i]}; + } + + yuv2planeX_16_u(filter, filterSize, src, dest, dst_u, big_endian, output_bits, 0); + + for (i = dst_u; i < dstW - 7; i += 8) { + vleft = vright = vadd; + + for (j = 0; j < filterSize; j++) { + vin32l = vec_vsx_ld(0, &src[j][i]); + vin32r = vec_vsx_ld(0, &src[j][i + 4]); + +#ifdef __GNUC__ + // GCC does not support vmuluwm yet. Bug open. + __asm__("vmuluwm %0, %1, %2" : "=v"(vtmp) : "v"(vin32l), "v"(vfilter[j])); + vleft = vec_add(vleft, vtmp); + __asm__("vmuluwm %0, %1, %2" : "=v"(vtmp) : "v"(vin32r), "v"(vfilter[j])); + vright = vec_add(vright, vtmp); +#else + // No idea which compilers this works in, untested. Copied from libsimdpp + vtmp = vec_vmuluwm(vin32l, vfilter[j]); + vleft = vec_add(vleft, vtmp); + vtmp = vec_vmuluwm(vin32r, vfilter[j]); + vright = vec_add(vright, vtmp); +#endif + } + + vleft = vec_sra(vleft, vshift); + vright = vec_sra(vright, vshift); + v = (vector uint16_t) vec_packs((vector int32_t) vleft, (vector int32_t) vright); + v = vec_add(v, vbias); + v = vec_rl(v, vswap); + vec_st(v, 0, &dest[i]); + } + + yuv2planeX_16_u(filter, filterSize, src, dest, dstW, big_endian, output_bits, i); +} + +#endif /* HAVE_POWER8 */ + #define yuv2NBPS(bits, BE_LE, is_be, template_size, typeX_t) \ + yuv2NBPS1(bits, BE_LE, is_be, template_size, typeX_t) \ + yuv2NBPSX(bits, BE_LE, is_be, template_size, typeX_t) + +#define yuv2NBPS1(bits, BE_LE, is_be, template_size, typeX_t) \ static void yuv2plane1_ ## bits ## BE_LE ## _vsx(const int16_t *src, \ uint8_t *dest, int dstW, \ const uint8_t *dither, int offset) \ @@ -243,6 +405,16 @@ static void yuv2plane1_ ## bits ## BE_LE ## _vsx(const int16_t *src, \ (uint16_t *) dest, dstW, is_be, bits); \ } +#define yuv2NBPSX(bits, BE_LE, is_be, template_size, typeX_t) \ +static void yuv2planeX_ ## bits ## BE_LE ## _vsx(const int16_t *filter, int filterSize, \ + const int16_t **src, uint8_t *dest, int dstW, \ + const uint8_t *dither, int offset)\ +{ \ + yuv2planeX_## template_size ## _vsx(filter, \ + filterSize, (const typeX_t **) src, \ + (uint16_t *) dest, dstW, is_be, bits); \ +} + yuv2NBPS( 9, BE, 1, nbps, int16_t) yuv2NBPS( 9, LE, 0, nbps, int16_t) yuv2NBPS(10, BE, 1, nbps, int16_t) @@ -251,8 +423,13 @@ yuv2NBPS(12, BE, 1, nbps, int16_t) yuv2NBPS(12, LE, 0, nbps, int16_t) yuv2NBPS(14, BE, 1, nbps, int16_t) yuv2NBPS(14, LE, 0, nbps, int16_t) -yuv2NBPS(16, BE, 1, 16, int32_t) -yuv2NBPS(16, LE, 0, 16, int32_t) + +yuv2NBPS1(16, BE, 1, 16, int32_t) +yuv2NBPS1(16, LE, 0, 16, int32_t) +#ifdef HAVE_POWER8 +yuv2NBPSX(16, BE, 1, 16, int32_t) +yuv2NBPSX(16, LE, 0, 16, int32_t) +#endif #endif /* !HAVE_BIGENDIAN */ @@ -262,8 +439,9 @@ av_cold void ff_sws_init_swscale_vsx(SwsContext *c) { #if HAVE_VSX enum AVPixelFormat dstFormat = c->dstFormat; + const int cpu_flags = av_get_cpu_flags(); - if (!(av_get_cpu_flags() & AV_CPU_FLAG_VSX)) + if (!(cpu_flags & AV_CPU_FLAG_VSX)) return; #if !HAVE_BIGENDIAN @@ -286,20 +464,29 @@ av_cold void ff_sws_init_swscale_vsx(SwsContext *c) #if !HAVE_BIGENDIAN case 9: c->yuv2plane1 = isBE(dstFormat) ? yuv2plane1_9BE_vsx : yuv2plane1_9LE_vsx; + c->yuv2planeX = isBE(dstFormat) ? yuv2planeX_9BE_vsx : yuv2planeX_9LE_vsx; break; case 10: c->yuv2plane1 = isBE(dstFormat) ? yuv2plane1_10BE_vsx : yuv2plane1_10LE_vsx; + c->yuv2planeX = isBE(dstFormat) ? yuv2planeX_10BE_vsx : yuv2planeX_10LE_vsx; break; case 12: c->yuv2plane1 = isBE(dstFormat) ? yuv2plane1_12BE_vsx : yuv2plane1_12LE_vsx; + c->yuv2planeX = isBE(dstFormat) ? yuv2planeX_12BE_vsx : yuv2planeX_12LE_vsx; break; case 14: c->yuv2plane1 = isBE(dstFormat) ? yuv2plane1_14BE_vsx : yuv2plane1_14LE_vsx; + c->yuv2planeX = isBE(dstFormat) ? yuv2planeX_14BE_vsx : yuv2planeX_14LE_vsx; break; case 16: c->yuv2plane1 = isBE(dstFormat) ? yuv2plane1_16BE_vsx : yuv2plane1_16LE_vsx; +#ifdef HAVE_POWER8 + if (cpu_flags & AV_CPU_FLAG_POWER8) { + c->yuv2planeX = isBE(dstFormat) ? yuv2planeX_16BE_vsx : yuv2planeX_16LE_vsx; + } +#endif /* HAVE_POWER8 */ break; -#endif +#endif /* !HAVE_BIGENDIAN */ } } #endif /* HAVE_VSX */