diff --git a/CMakeLists.txt b/CMakeLists.txt index 828fe34..197ae5f 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -19,6 +19,9 @@ macro_log_feature(GLIB2_FOUND "GLib" "Required to build gst-plugins-vision" "htt find_package(GObject REQUIRED) macro_log_feature(GOBJECT_FOUND "GObject" "Required to build gst-plugins-vision" "http://www.gtk.org/" TRUE) +find_package(Orc REQUIRED) +macro_log_feature(ORC_FOUND "Orc" "Required library to improve performance" "http://code.entropywave.com/orc/" TRUE) + find_package(NIIMAQ) macro_log_feature(NIIMAQ_FOUND "NI-IMAQ" "Required to build National Instruments IMAQ source element" "http://www.ni.com/" FALSE) diff --git a/gst/extractcolor/CMakeLists.txt b/gst/extractcolor/CMakeLists.txt index eb84daf..cabdf08 100644 --- a/gst/extractcolor/CMakeLists.txt +++ b/gst/extractcolor/CMakeLists.txt @@ -1,14 +1,16 @@ add_definitions (-DHAVE_CONFIG_H) set (SOURCES - gstextractcolor.c) + gstextractcolor.c + gstextractcolororc-dist.c) set (HEADERS gstextractcolor.h) include_directories (AFTER ${GSTREAMER_BASE_INCLUDE_DIR} - ${GSTREAMER_VIDEO_INCLUDE_DIR}) + ${GSTREAMER_VIDEO_INCLUDE_DIR} + ${ORC_INCLUDE_DIR}) add_library (libgstextractcolor MODULE ${SOURCES} @@ -19,7 +21,8 @@ target_link_libraries (libgstextractcolor ${GOBJECT_LIBRARIES} ${GSTREAMER_LIBRARY} ${GSTREAMER_BASE_LIBRARY} - ${GSTREAMER_VIDEO_LIBRARY}) + ${GSTREAMER_VIDEO_LIBRARY} + ${ORC_LIBRARIES}) install (TARGETS libgstextractcolor LIBRARY DESTINATION lib/gstreamer-1.0) diff --git a/gst/extractcolor/gstextractcolor.c b/gst/extractcolor/gstextractcolor.c index 87cf5fa..b5d9777 100644 --- a/gst/extractcolor/gstextractcolor.c +++ b/gst/extractcolor/gstextractcolor.c @@ -42,6 +42,8 @@ #include +#include "gstextractcolororc-dist.h" + /* GstExtractColor signals and args */ enum { @@ -338,18 +340,54 @@ gst_extract_color_transform_frame (GstVideoFilter * filter, GstVideoFrame * in_frame, GstVideoFrame * out_frame) { GstExtractColor *filt = GST_EXTRACT_COLOR (filter); - GstClockTime start = - gst_clock_get_time (gst_element_get_clock (GST_ELEMENT (filt))); + GTimer *timer = NULL; guint comp = filt->component; GST_LOG_OBJECT (filt, "Performing non-inplace transform"); - /* TODO: create orc functions for these to speed them up */ +#if 0 + timer = g_timer_new (); +#endif + if (GST_VIDEO_FRAME_COMP_DEPTH (in_frame, comp) == 8) { + guint8 *src, *dst; + + src = GST_VIDEO_FRAME_PLANE_DATA (in_frame, 0); + dst = GST_VIDEO_FRAME_PLANE_DATA (out_frame, 0); + + switch (GST_VIDEO_FRAME_COMP_OFFSET (in_frame, comp)) { + case 0: + extractcolor_orc_copy32_0 (dst, GST_VIDEO_FRAME_PLANE_STRIDE (out_frame, + 0), src, GST_VIDEO_FRAME_PLANE_STRIDE (in_frame, 0), + GST_VIDEO_FRAME_WIDTH (in_frame), + GST_VIDEO_FRAME_HEIGHT (out_frame)); + break; + case 1: + extractcolor_orc_copy32_1 (dst, GST_VIDEO_FRAME_PLANE_STRIDE (out_frame, + 0), src, GST_VIDEO_FRAME_PLANE_STRIDE (in_frame, 0), + GST_VIDEO_FRAME_WIDTH (in_frame), + GST_VIDEO_FRAME_HEIGHT (out_frame)); + break; + case 2: + extractcolor_orc_copy32_2 (dst, GST_VIDEO_FRAME_PLANE_STRIDE (out_frame, + 0), src, GST_VIDEO_FRAME_PLANE_STRIDE (in_frame, 0), + GST_VIDEO_FRAME_WIDTH (in_frame), + GST_VIDEO_FRAME_HEIGHT (out_frame)); + break; + case 3: + extractcolor_orc_copy32_3 (dst, GST_VIDEO_FRAME_PLANE_STRIDE (out_frame, + 0), src, GST_VIDEO_FRAME_PLANE_STRIDE (in_frame, 0), + GST_VIDEO_FRAME_WIDTH (in_frame), + GST_VIDEO_FRAME_HEIGHT (out_frame)); + break; + default: + g_assert_not_reached (); + } +#if 0 gint x, y; - guint8 *src = GST_VIDEO_FRAME_COMP_DATA (in_frame, comp); - guint8 *dst = GST_VIDEO_FRAME_COMP_DATA (out_frame, 0); const guint pstride = GST_VIDEO_FRAME_COMP_PSTRIDE (in_frame, comp); + src = GST_VIDEO_FRAME_COMP_DATA (in_frame, comp); + dst = GST_VIDEO_FRAME_COMP_DATA (out_frame, 0); for (y = 0; y < GST_VIDEO_FRAME_COMP_HEIGHT (in_frame, comp); y++) { for (x = 0; x < GST_VIDEO_FRAME_COMP_WIDTH (in_frame, comp); x++) { dst[x] = src[x * pstride]; @@ -357,6 +395,7 @@ gst_extract_color_transform_frame (GstVideoFilter * filter, src += GST_VIDEO_FRAME_COMP_STRIDE (in_frame, comp); dst += GST_VIDEO_FRAME_COMP_STRIDE (out_frame, 0); } +#endif } else { gint x, y; guint16 *src = (guint16 *) GST_VIDEO_FRAME_COMP_DATA (in_frame, comp); @@ -371,10 +410,11 @@ gst_extract_color_transform_frame (GstVideoFilter * filter, } } - GST_LOG_OBJECT (filt, "Processing took %" G_GINT64_FORMAT "ms", - GST_TIME_AS_MSECONDS (GST_CLOCK_DIFF (start, - gst_clock_get_time (gst_element_get_clock (GST_ELEMENT - (filt)))))); +#if 0 + GST_LOG_OBJECT (filt, "Processing took %.3f ms", g_timer_elapsed (timer, + NULL) * 1000); + g_timer_destroy (timer); +#endif return GST_FLOW_OK; } diff --git a/gst/extractcolor/gstextractcolororc-dist.c b/gst/extractcolor/gstextractcolororc-dist.c new file mode 100644 index 0000000..4d4fc00 --- /dev/null +++ b/gst/extractcolor/gstextractcolororc-dist.c @@ -0,0 +1,677 @@ +#include "gstextractcolororc-dist.h" +/* autogenerated from gstextractcolororc.orc */ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#ifndef _ORC_INTEGER_TYPEDEFS_ +#define _ORC_INTEGER_TYPEDEFS_ +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#include +typedef int8_t orc_int8; +typedef int16_t orc_int16; +typedef int32_t orc_int32; +typedef int64_t orc_int64; +typedef uint8_t orc_uint8; +typedef uint16_t orc_uint16; +typedef uint32_t orc_uint32; +typedef uint64_t orc_uint64; +#define ORC_UINT64_C(x) UINT64_C(x) +#elif defined(_MSC_VER) +typedef signed __int8 orc_int8; +typedef signed __int16 orc_int16; +typedef signed __int32 orc_int32; +typedef signed __int64 orc_int64; +typedef unsigned __int8 orc_uint8; +typedef unsigned __int16 orc_uint16; +typedef unsigned __int32 orc_uint32; +typedef unsigned __int64 orc_uint64; +#define ORC_UINT64_C(x) (x##Ui64) +#define inline __inline +#else +#include +typedef signed char orc_int8; +typedef short orc_int16; +typedef int orc_int32; +typedef unsigned char orc_uint8; +typedef unsigned short orc_uint16; +typedef unsigned int orc_uint32; +#if INT_MAX == LONG_MAX +typedef long long orc_int64; +typedef unsigned long long orc_uint64; +#define ORC_UINT64_C(x) (x##ULL) +#else +typedef long orc_int64; +typedef unsigned long orc_uint64; +#define ORC_UINT64_C(x) (x##UL) +#endif +#endif +typedef union { orc_int16 i; orc_int8 x2[2]; } orc_union16; +typedef union { orc_int32 i; float f; orc_int16 x2[2]; orc_int8 x4[4]; } orc_union32; +typedef union { orc_int64 i; double f; orc_int32 x2[2]; float x2f[2]; orc_int16 x4[4]; } orc_union64; +#endif +#ifndef ORC_RESTRICT +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#define ORC_RESTRICT restrict +#elif defined(__GNUC__) && __GNUC__ >= 4 +#define ORC_RESTRICT __restrict__ +#else +#define ORC_RESTRICT +#endif +#endif + +#ifndef ORC_INTERNAL +#if defined(__SUNPRO_C) && (__SUNPRO_C >= 0x590) +#define ORC_INTERNAL __attribute__((visibility("hidden"))) +#elif defined(__SUNPRO_C) && (__SUNPRO_C >= 0x550) +#define ORC_INTERNAL __hidden +#elif defined (__GNUC__) +#define ORC_INTERNAL __attribute__((visibility("hidden"))) +#else +#define ORC_INTERNAL +#endif +#endif + + +#ifndef DISABLE_ORC +#include +#endif +void extractcolor_orc_copy32_0 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); +void extractcolor_orc_copy32_1 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); +void extractcolor_orc_copy32_2 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); +void extractcolor_orc_copy32_3 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); + + +/* begin Orc C target preamble */ +#define ORC_CLAMP(x,a,b) ((x)<(a) ? (a) : ((x)>(b) ? (b) : (x))) +#define ORC_ABS(a) ((a)<0 ? -(a) : (a)) +#define ORC_MIN(a,b) ((a)<(b) ? (a) : (b)) +#define ORC_MAX(a,b) ((a)>(b) ? (a) : (b)) +#define ORC_SB_MAX 127 +#define ORC_SB_MIN (-1-ORC_SB_MAX) +#define ORC_UB_MAX 255 +#define ORC_UB_MIN 0 +#define ORC_SW_MAX 32767 +#define ORC_SW_MIN (-1-ORC_SW_MAX) +#define ORC_UW_MAX 65535 +#define ORC_UW_MIN 0 +#define ORC_SL_MAX 2147483647 +#define ORC_SL_MIN (-1-ORC_SL_MAX) +#define ORC_UL_MAX 4294967295U +#define ORC_UL_MIN 0 +#define ORC_CLAMP_SB(x) ORC_CLAMP(x,ORC_SB_MIN,ORC_SB_MAX) +#define ORC_CLAMP_UB(x) ORC_CLAMP(x,ORC_UB_MIN,ORC_UB_MAX) +#define ORC_CLAMP_SW(x) ORC_CLAMP(x,ORC_SW_MIN,ORC_SW_MAX) +#define ORC_CLAMP_UW(x) ORC_CLAMP(x,ORC_UW_MIN,ORC_UW_MAX) +#define ORC_CLAMP_SL(x) ORC_CLAMP(x,ORC_SL_MIN,ORC_SL_MAX) +#define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX) +#define ORC_SWAP_W(x) ((((x)&0xffU)<<8) | (((x)&0xff00U)>>8)) +#define ORC_SWAP_L(x) ((((x)&0xffU)<<24) | (((x)&0xff00U)<<8) | (((x)&0xff0000U)>>8) | (((x)&0xff000000U)>>24)) +#define ORC_SWAP_Q(x) ((((x)&ORC_UINT64_C(0xff))<<56) | (((x)&ORC_UINT64_C(0xff00))<<40) | (((x)&ORC_UINT64_C(0xff0000))<<24) | (((x)&ORC_UINT64_C(0xff000000))<<8) | (((x)&ORC_UINT64_C(0xff00000000))>>8) | (((x)&ORC_UINT64_C(0xff0000000000))>>24) | (((x)&ORC_UINT64_C(0xff000000000000))>>40) | (((x)&ORC_UINT64_C(0xff00000000000000))>>56)) +#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset))) +#define ORC_DENORMAL(x) ((x) & ((((x)&0x7f800000) == 0) ? 0xff800000 : 0xffffffff)) +#define ORC_ISNAN(x) ((((x)&0x7f800000) == 0x7f800000) && (((x)&0x007fffff) != 0)) +#define ORC_DENORMAL_DOUBLE(x) ((x) & ((((x)&ORC_UINT64_C(0x7ff0000000000000)) == 0) ? ORC_UINT64_C(0xfff0000000000000) : ORC_UINT64_C(0xffffffffffffffff))) +#define ORC_ISNAN_DOUBLE(x) ((((x)&ORC_UINT64_C(0x7ff0000000000000)) == ORC_UINT64_C(0x7ff0000000000000)) && (((x)&ORC_UINT64_C(0x000fffffffffffff)) != 0)) +#ifndef ORC_RESTRICT +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#define ORC_RESTRICT restrict +#elif defined(__GNUC__) && __GNUC__ >= 4 +#define ORC_RESTRICT __restrict__ +#else +#define ORC_RESTRICT +#endif +#endif +/* end Orc C target preamble */ + + + +/* extractcolor_orc_copy32_0 */ +#ifdef DISABLE_ORC +void +extractcolor_orc_copy32_0 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m){ + int i; + int j; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET(s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[0]; + } + /* 2: select0wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[0]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +#else +static void +_backup_extractcolor_orc_copy32_0 (OrcExecutor * ORC_RESTRICT ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET(ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[0]; + } + /* 2: select0wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[0]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +void +extractcolor_orc_copy32_0 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static volatile int p_inited = 0; + static OrcCode *c = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcProgram *p; + +#if 1 + static const orc_uint8 bc[] = { + 1, 7, 9, 25, 101, 120, 116, 114, 97, 99, 116, 99, 111, 108, 111, 114, + 95, 111, 114, 99, 95, 99, 111, 112, 121, 51, 50, 95, 48, 11, 1, 1, + 12, 4, 4, 20, 2, 190, 32, 4, 188, 0, 32, 2, 0, + }; + p = orc_program_new_from_static_bytecode (bc); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_0); +#else + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "extractcolor_orc_copy32_0"); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_0); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_D1); +#endif + + orc_program_compile (p); + c = orc_program_take_code (p); + orc_program_free (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->arrays[ORC_VAR_A2] = c; + ex->program = 0; + + ex->n = n; + ORC_EXECUTOR_M(ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *)s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = c->exec; + func (ex); +} +#endif + + +/* extractcolor_orc_copy32_1 */ +#ifdef DISABLE_ORC +void +extractcolor_orc_copy32_1 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m){ + int i; + int j; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET(s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[0]; + } + /* 2: select1wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[1]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +#else +static void +_backup_extractcolor_orc_copy32_1 (OrcExecutor * ORC_RESTRICT ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET(ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select0lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[0]; + } + /* 2: select1wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[1]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +void +extractcolor_orc_copy32_1 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static volatile int p_inited = 0; + static OrcCode *c = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcProgram *p; + +#if 1 + static const orc_uint8 bc[] = { + 1, 7, 9, 25, 101, 120, 116, 114, 97, 99, 116, 99, 111, 108, 111, 114, + 95, 111, 114, 99, 95, 99, 111, 112, 121, 51, 50, 95, 49, 11, 1, 1, + 12, 4, 4, 20, 2, 190, 32, 4, 189, 0, 32, 2, 0, + }; + p = orc_program_new_from_static_bytecode (bc); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_1); +#else + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "extractcolor_orc_copy32_1"); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_1); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_D1); +#endif + + orc_program_compile (p); + c = orc_program_take_code (p); + orc_program_free (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->arrays[ORC_VAR_A2] = c; + ex->program = 0; + + ex->n = n; + ORC_EXECUTOR_M(ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *)s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = c->exec; + func (ex); +} +#endif + + +/* extractcolor_orc_copy32_2 */ +#ifdef DISABLE_ORC +void +extractcolor_orc_copy32_2 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m){ + int i; + int j; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET(s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[1]; + } + /* 2: select0wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[0]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +#else +static void +_backup_extractcolor_orc_copy32_2 (OrcExecutor * ORC_RESTRICT ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET(ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[1]; + } + /* 2: select0wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[0]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +void +extractcolor_orc_copy32_2 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static volatile int p_inited = 0; + static OrcCode *c = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcProgram *p; + +#if 1 + static const orc_uint8 bc[] = { + 1, 7, 9, 25, 101, 120, 116, 114, 97, 99, 116, 99, 111, 108, 111, 114, + 95, 111, 114, 99, 95, 99, 111, 112, 121, 51, 50, 95, 50, 11, 1, 1, + 12, 4, 4, 20, 2, 191, 32, 4, 188, 0, 32, 2, 0, + }; + p = orc_program_new_from_static_bytecode (bc); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_2); +#else + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "extractcolor_orc_copy32_2"); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_2); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_D1); +#endif + + orc_program_compile (p); + c = orc_program_take_code (p); + orc_program_free (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->arrays[ORC_VAR_A2] = c; + ex->program = 0; + + ex->n = n; + ORC_EXECUTOR_M(ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *)s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = c->exec; + func (ex); +} +#endif + + +/* extractcolor_orc_copy32_3 */ +#ifdef DISABLE_ORC +void +extractcolor_orc_copy32_3 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m){ + int i; + int j; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(d1, d1_stride * j); + ptr4 = ORC_PTR_OFFSET(s1, s1_stride * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[1]; + } + /* 2: select1wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[1]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +#else +static void +_backup_extractcolor_orc_copy32_3 (OrcExecutor * ORC_RESTRICT ex) +{ + int i; + int j; + int n = ex->n; + int m = ex->params[ORC_VAR_A1]; + orc_int8 * ORC_RESTRICT ptr0; + const orc_union32 * ORC_RESTRICT ptr4; + orc_union32 var33; + orc_int8 var34; + orc_union16 var35; + + for (j = 0; j < m; j++) { + ptr0 = ORC_PTR_OFFSET(ex->arrays[0], ex->params[0] * j); + ptr4 = ORC_PTR_OFFSET(ex->arrays[4], ex->params[4] * j); + + + for (i = 0; i < n; i++) { + /* 0: loadl */ + var33 = ptr4[i]; + /* 1: select1lw */ + { + orc_union32 _src; + _src.i = var33.i; + var35.i = _src.x2[1]; + } + /* 2: select1wb */ + { + orc_union16 _src; + _src.i = var35.i; + var34 = _src.x2[1]; + } + /* 3: storeb */ + ptr0[i] = var34; + } + } + +} + +void +extractcolor_orc_copy32_3 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m) +{ + OrcExecutor _ex, *ex = &_ex; + static volatile int p_inited = 0; + static OrcCode *c = 0; + void (*func) (OrcExecutor *); + + if (!p_inited) { + orc_once_mutex_lock (); + if (!p_inited) { + OrcProgram *p; + +#if 1 + static const orc_uint8 bc[] = { + 1, 7, 9, 25, 101, 120, 116, 114, 97, 99, 116, 99, 111, 108, 111, 114, + 95, 111, 114, 99, 95, 99, 111, 112, 121, 51, 50, 95, 51, 11, 1, 1, + 12, 4, 4, 20, 2, 191, 32, 4, 189, 0, 32, 2, 0, + }; + p = orc_program_new_from_static_bytecode (bc); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_3); +#else + p = orc_program_new (); + orc_program_set_2d (p); + orc_program_set_name (p, "extractcolor_orc_copy32_3"); + orc_program_set_backup_function (p, _backup_extractcolor_orc_copy32_3); + orc_program_add_destination (p, 1, "d1"); + orc_program_add_source (p, 4, "s1"); + orc_program_add_temporary (p, 2, "t1"); + + orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1, ORC_VAR_D1); + orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_D1); +#endif + + orc_program_compile (p); + c = orc_program_take_code (p); + orc_program_free (p); + } + p_inited = TRUE; + orc_once_mutex_unlock (); + } + ex->arrays[ORC_VAR_A2] = c; + ex->program = 0; + + ex->n = n; + ORC_EXECUTOR_M(ex) = m; + ex->arrays[ORC_VAR_D1] = d1; + ex->params[ORC_VAR_D1] = d1_stride; + ex->arrays[ORC_VAR_S1] = (void *)s1; + ex->params[ORC_VAR_S1] = s1_stride; + + func = c->exec; + func (ex); +} +#endif + + diff --git a/gst/extractcolor/gstextractcolororc-dist.h b/gst/extractcolor/gstextractcolororc-dist.h new file mode 100644 index 0000000..69b8b7e --- /dev/null +++ b/gst/extractcolor/gstextractcolororc-dist.h @@ -0,0 +1,92 @@ +#include +/* autogenerated from gstextractcolororc.orc */ + +#ifndef _OUT_H_ +#define _OUT_H_ + + +#ifdef __cplusplus +extern "C" { +#endif + + + +#ifndef _ORC_INTEGER_TYPEDEFS_ +#define _ORC_INTEGER_TYPEDEFS_ +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#include +typedef int8_t orc_int8; +typedef int16_t orc_int16; +typedef int32_t orc_int32; +typedef int64_t orc_int64; +typedef uint8_t orc_uint8; +typedef uint16_t orc_uint16; +typedef uint32_t orc_uint32; +typedef uint64_t orc_uint64; +#define ORC_UINT64_C(x) UINT64_C(x) +#elif defined(_MSC_VER) +typedef signed __int8 orc_int8; +typedef signed __int16 orc_int16; +typedef signed __int32 orc_int32; +typedef signed __int64 orc_int64; +typedef unsigned __int8 orc_uint8; +typedef unsigned __int16 orc_uint16; +typedef unsigned __int32 orc_uint32; +typedef unsigned __int64 orc_uint64; +#define ORC_UINT64_C(x) (x##Ui64) +#define inline __inline +#else +#include +typedef signed char orc_int8; +typedef short orc_int16; +typedef int orc_int32; +typedef unsigned char orc_uint8; +typedef unsigned short orc_uint16; +typedef unsigned int orc_uint32; +#if INT_MAX == LONG_MAX +typedef long long orc_int64; +typedef unsigned long long orc_uint64; +#define ORC_UINT64_C(x) (x##ULL) +#else +typedef long orc_int64; +typedef unsigned long orc_uint64; +#define ORC_UINT64_C(x) (x##UL) +#endif +#endif +typedef union { orc_int16 i; orc_int8 x2[2]; } orc_union16; +typedef union { orc_int32 i; float f; orc_int16 x2[2]; orc_int8 x4[4]; } orc_union32; +typedef union { orc_int64 i; double f; orc_int32 x2[2]; float x2f[2]; orc_int16 x4[4]; } orc_union64; +#endif +#ifndef ORC_RESTRICT +#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L +#define ORC_RESTRICT restrict +#elif defined(__GNUC__) && __GNUC__ >= 4 +#define ORC_RESTRICT __restrict__ +#else +#define ORC_RESTRICT +#endif +#endif + +#ifndef ORC_INTERNAL +#if defined(__SUNPRO_C) && (__SUNPRO_C >= 0x590) +#define ORC_INTERNAL __attribute__((visibility("hidden"))) +#elif defined(__SUNPRO_C) && (__SUNPRO_C >= 0x550) +#define ORC_INTERNAL __hidden +#elif defined (__GNUC__) +#define ORC_INTERNAL __attribute__((visibility("hidden"))) +#else +#define ORC_INTERNAL +#endif +#endif + +void extractcolor_orc_copy32_0 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); +void extractcolor_orc_copy32_1 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); +void extractcolor_orc_copy32_2 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); +void extractcolor_orc_copy32_3 (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int n, int m); + +#ifdef __cplusplus +} +#endif + +#endif + diff --git a/gst/extractcolor/gstextractcolororc.orc b/gst/extractcolor/gstextractcolororc.orc new file mode 100644 index 0000000..17cf07a --- /dev/null +++ b/gst/extractcolor/gstextractcolororc.orc @@ -0,0 +1,71 @@ + +.function extractcolor_orc_copy32_0 +.flags 2d +.dest 1 d guint8 +.source 4 s guint8 +.temp 2 t +select0lw t, s +select0wb d, t + + +.function extractcolor_orc_copy32_1 +.flags 2d +.dest 1 d guint8 +.source 4 s guint8 +.temp 2 t +select0lw t, s +select1wb d, t + + +.function extractcolor_orc_copy32_2 +.flags 2d +.dest 1 d guint8 +.source 4 s guint8 +.temp 2 t +select1lw t, s +select0wb d, t + + +.function extractcolor_orc_copy32_3 +.flags 2d +.dest 1 d guint8 +.source 4 s guint8 +.temp 2 t +select1lw t, s +select1wb d, t + + +.function extractcolor_orc_copy64_0 +.flags 2d +.dest 2 d guint16 +.source 8 s guint16 +.temp 4 t +select0ql t, s +select0lw d, t + + +.function extractcolor_orc_copy64_1 +.flags 2d +.dest 2 d guint16 +.source 8 s guint16 +.temp 4 t +select0ql t, s +select1lw d, t + + +.function extractcolor_orc_copy64_2 +.flags 2d +.dest 2 d guint16 +.source 8 s guint16 +.temp 4 t +select1ql t, s +select0lw d, t + + +.function extractcolor_orc_copy64_3 +.flags 2d +.dest 2 d guint16 +.source 8 s guint16 +.temp 4 t +select1ql t, s +select1lw d, t \ No newline at end of file