Module: wine Branch: master Commit: e8c9588e2ad4168380fda86e7ebc3dfd56c3e863 URL: https://source.winehq.org/git/wine.git/?a=commit;h=e8c9588e2ad4168380fda86e7...
Author: Stefan Dösinger stefan@codeweavers.com Date: Mon May 16 12:58:05 2022 -0500
wined3d: Support CPU clear of float16 formats.
Signed-off-by: Stefan Dösinger stefan@codeweavers.com Signed-off-by: Zebediah Figura zfigura@codeweavers.com Signed-off-by: Alexandre Julliard julliard@winehq.org
---
dlls/wined3d/surface.c | 63 ----------------------------------- dlls/wined3d/utils.c | 25 ++++++++++++++ dlls/wined3d/wined3d_private.h | 75 ++++++++++++++++++++++++++++++++++++++---- 3 files changed, 93 insertions(+), 70 deletions(-)
diff --git a/dlls/wined3d/surface.c b/dlls/wined3d/surface.c index 0d4809ea55b..29bdef9844f 100644 --- a/dlls/wined3d/surface.c +++ b/dlls/wined3d/surface.c @@ -39,69 +39,6 @@ static void get_color_masks(const struct wined3d_format *format, uint32_t *masks masks[2] = wined3d_mask_from_size(format->blue_size) << format->blue_offset; }
-/* See also float_16_to_32() in wined3d_private.h */ -static inline unsigned short float_32_to_16(const float *in) -{ - int exp = 0; - float tmp = fabsf(*in); - unsigned int mantissa; - unsigned short ret; - - /* Deal with special numbers */ - if (*in == 0.0f) - return 0x0000; - if (isnan(*in)) - return 0x7c01; - if (isinf(*in)) - return (*in < 0.0f ? 0xfc00 : 0x7c00); - - if (tmp < (float)(1u << 10)) - { - do - { - tmp = tmp * 2.0f; - exp--; - } while (tmp < (float)(1u << 10)); - } - else if (tmp >= (float)(1u << 11)) - { - do - { - tmp /= 2.0f; - exp++; - } while (tmp >= (float)(1u << 11)); - } - - mantissa = (unsigned int)tmp; - if (tmp - mantissa >= 0.5f) - ++mantissa; /* Round to nearest, away from zero. */ - - exp += 10; /* Normalize the mantissa. */ - exp += 15; /* Exponent is encoded with excess 15. */ - - if (exp > 30) /* too big */ - { - ret = 0x7c00; /* INF */ - } - else if (exp <= 0) - { - /* exp == 0: Non-normalized mantissa. Returns 0x0000 (=0.0) for too small numbers. */ - while (exp <= 0) - { - mantissa = mantissa >> 1; - ++exp; - } - ret = mantissa & 0x3ff; - } - else - { - ret = (exp << 10) | (mantissa & 0x3ff); - } - - ret |= ((*in < 0.0f ? 1 : 0) << 15); /* Add the sign */ - return ret; -} - static void convert_r32_float_r16_float(const BYTE *src, BYTE *dst, DWORD pitch_in, DWORD pitch_out, unsigned int w, unsigned int h) { diff --git a/dlls/wined3d/utils.c b/dlls/wined3d/utils.c index 6d785c461d4..892020e2aa9 100644 --- a/dlls/wined3d/utils.c +++ b/dlls/wined3d/utils.c @@ -6137,6 +6137,31 @@ void wined3d_format_convert_from_float(const struct wined3d_format *format, return; }
+ if ((format->flags[WINED3D_GL_RES_TYPE_TEX_2D] & WINED3DFMT_FLAG_FLOAT) && format->red_size == 16) + { + uint16_t *ret_s = ret; + + switch (format->byte_count) + { + case 8: + ret_s[3] = float_32_to_16(&color->a); + ret_s[2] = float_32_to_16(&color->b); + /* fall through */ + case 4: + ret_s[1] = float_32_to_16(&color->g); + /* fall through */ + case 2: + ret_s[0] = float_32_to_16(&color->r); + break; + + default: + ERR("Unexpected byte count %u, format %s.\n", format->byte_count, debug_d3dformat(format_id)); + break; + } + + return; + } + FIXME("Conversion for format %s not implemented.\n", debug_d3dformat(format_id)); }
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h index 37cec7dd2c3..1d8bb40cec4 100644 --- a/dlls/wined3d/wined3d_private.h +++ b/dlls/wined3d/wined3d_private.h @@ -354,13 +354,12 @@ static inline GLenum wined3d_gl_min_mip_filter(enum wined3d_texture_filter_type return minMipLookup[min_filter].mip[mip_filter]; }
-/* float_16_to_32() and float_32_to_16() (see implementation in - * surface_base.c) convert 16 bit floats in the FLOAT16 data type - * to standard C floats and vice versa. They do not depend on the encoding - * of the C float, so they are platform independent, but slow. On x86 and - * other IEEE 754 compliant platforms the conversion can be accelerated by - * bit shifting the exponent and mantissa. There are also some SSE-based - * assembly routines out there. +/* float_16_to_32() and float_32_to_16() convert 16 bit floats in the + * FLOAT16 data type to standard C floats and vice versa. They do not + * depend on the encoding of the C float, so they are platform independent, + * but slow. On x86 and other IEEE 754 compliant platforms the conversion + * can be accelerated by bit shifting the exponent and mantissa. There are + * also some SSE-based assembly routines out there. * * See GL_NV_half_float for a reference of the FLOAT16 / GL_HALF format */ @@ -404,6 +403,68 @@ static inline float float_24_to_32(DWORD in) } }
+static inline unsigned short float_32_to_16(const float *in) +{ + int exp = 0; + float tmp = fabsf(*in); + unsigned int mantissa; + unsigned short ret; + + /* Deal with special numbers */ + if (*in == 0.0f) + return 0x0000; + if (isnan(*in)) + return 0x7c01; + if (isinf(*in)) + return (*in < 0.0f ? 0xfc00 : 0x7c00); + + if (tmp < (float)(1u << 10)) + { + do + { + tmp = tmp * 2.0f; + exp--; + } while (tmp < (float)(1u << 10)); + } + else if (tmp >= (float)(1u << 11)) + { + do + { + tmp /= 2.0f; + exp++; + } while (tmp >= (float)(1u << 11)); + } + + mantissa = (unsigned int)tmp; + if (tmp - mantissa >= 0.5f) + ++mantissa; /* Round to nearest, away from zero. */ + + exp += 10; /* Normalize the mantissa. */ + exp += 15; /* Exponent is encoded with excess 15. */ + + if (exp > 30) /* too big */ + { + ret = 0x7c00; /* INF */ + } + else if (exp <= 0) + { + /* exp == 0: Non-normalized mantissa. Returns 0x0000 (=0.0) for too small numbers. */ + while (exp <= 0) + { + mantissa = mantissa >> 1; + ++exp; + } + ret = mantissa & 0x3ff; + } + else + { + ret = (exp << 10) | (mantissa & 0x3ff); + } + + ret |= ((*in < 0.0f ? 1 : 0) << 15); /* Add the sign */ + return ret; +} + static inline unsigned int wined3d_popcount(unsigned int x) { #if defined(__MINGW32__)