Author: post
Date: 2009-10-11 15:52:17 +0200 (Sun, 11 Oct 2009)
New Revision: 2697
Modified:
trunk/plugins/dcp/dcp.c
Log:
Added SSE2 optimized DCP huesatmap processing. Results of valdiv > 2 verified
against reference implementation. About 2x as fast.
Modified: trunk/plugins/dcp/dcp.c
===================================================================
--- trunk/plugins/dcp/dcp.c 2009-10-10 21:42:48 UTC (rev 2696)
+++ trunk/plugins/dcp/dcp.c 2009-10-11 13:52:17 UTC (rev 2697)
@@ -676,6 +676,241 @@
*v = MIN(*v * valScale, 1.0);
}
+#if defined (__SSE2__)
+
+/* SSE2 implementation, matches the reference implementation pretty closely */
+
+static void
+huesat_map_SSE2(RSHuesatMap *map, __m128 *_h, __m128 *_s, __m128 *_v)
+{
+ g_assert(RS_IS_HUESAT_MAP(map));
+
+ __m128 h = *_h;
+ __m128 s = *_s;
+ __m128 v = *_v;
+ gint i;
+ gint xfer_0[4] __attribute__ ((aligned (16)));
+ gint xfer_1[4] __attribute__ ((aligned (16)));
+
+ //TODO: Precalc BEGIN
+ gfloat hScale = (map->hue_divisions < 2) ? 0.0f : (map->hue_divisions *
(1.0f / 6.0f));
+ gfloat sScale = (gfloat) (map->sat_divisions - 1);
+ gfloat vScale = (gfloat) (map->val_divisions - 1);
+ // END precalc this.
+
+ gint _maxHueIndex0 = map->hue_divisions - 1;
+ gint _maxSatIndex0 = map->sat_divisions - 2;
+ gint _maxValIndex0 = map->val_divisions - 2;
+
+ const RS_VECTOR3 *tableBase = map->deltas;
+
+ gint _valStep = map->hue_divisions * map->sat_divisions;
+
+ __m128 hueShift;
+ __m128 satScale;
+ __m128 valScale;
+
+ if (map->val_divisions < 2)
+ {
+ __m128 hScaled = _mm_mul_ps(h, _mm_set_ps( hScale, hScale,
hScale, hScale));
+ __m128 sScaled = _mm_mul_ps(s, _mm_set_ps( sScale, sScale,
sScale, sScale));
+
+ __m128 half_ps = _mm_set_ps(0.5, 0.5, 0.5, 0.5);
+ __m128i maxHueIndex0 = _mm_set_epi32(_maxHueIndex0,
_maxHueIndex0, _maxHueIndex0, _maxHueIndex0);
+ __m128i maxSatIndex0 = _mm_set_epi32(_maxSatIndex0,
_maxSatIndex0, _maxSatIndex0, _maxSatIndex0);
+ __m128i hIndex0 = _mm_cvtps_epi32(_mm_sub_ps( hScaled, half_ps
));
+ __m128i sIndex0 = _mm_cvtps_epi32(_mm_sub_ps( sScaled, half_ps
));
+
+ sIndex0 = _mm_min_epi16(sIndex0, maxSatIndex0);
+ __m128i ones_epi32 = _mm_set_epi32(1,1,1,1);
+ __m128i hIndex1 = _mm_add_epi32(hIndex0, ones_epi32);
+
+ /* if (hIndex0 >= maxHueIndex0) */
+ __m128i hIndexMask = _mm_cmpgt_epi32( hIndex0,
_mm_sub_epi32(maxHueIndex0, ones_epi32));
+ hIndex0 = _mm_andnot_si128(hIndexMask, hIndex0);
+ /* hIndex1 = 0; */
+ hIndex1 = _mm_andnot_si128(hIndexMask, hIndex1);
+ /* hIndex0 = maxHueIndex0 */
+ hIndex0 = _mm_or_si128(hIndex0, _mm_and_si128(hIndexMask,
maxHueIndex0));
+
+ __m128 hFract1 = _mm_sub_ps( hScaled, _mm_cvtepi32_ps(hIndex0));
+ __m128 sFract1 = _mm_sub_ps( sScaled, _mm_cvtepi32_ps(sIndex0));
+ __m128 ones_ps = _mm_add_ps( half_ps, half_ps);
+
+ __m128 hFract0 = _mm_sub_ps(ones_ps, hFract1);
+ __m128 sFract0 = _mm_sub_ps(ones_ps, sFract1);
+ __m128i hueStep = _mm_set_epi32(map->sat_divisions,
map->sat_divisions, map->sat_divisions, map->sat_divisions);
+ __m128i table_offsets = _mm_add_epi32(sIndex0,
_mm_mullo_epi16(hIndex0, hueStep));
+ __m128i next_offsets = _mm_add_epi32(sIndex0,
_mm_mullo_epi16(hIndex1, hueStep));
+
+ _mm_store_si128((__m128i*)xfer_0, table_offsets);
+ _mm_store_si128((__m128i*)xfer_1, next_offsets);
+
+ const RS_VECTOR3 *entry00[4] = { tableBase + xfer_0[0],
tableBase + xfer_0[1], tableBase + xfer_0[2], tableBase + xfer_0[3]};
+ const RS_VECTOR3 *entry01[4] = { tableBase + xfer_1[0],
tableBase + xfer_1[1], tableBase + xfer_1[2], tableBase + xfer_1[3]};
+
+ __m128 hs0 = _mm_set_ps(entry00[3]->fHueShift,
entry00[2]->fHueShift, entry00[1]->fHueShift, entry00[0]->fHueShift);
+ __m128 hs1 = _mm_set_ps(entry01[3]->fHueShift,
entry01[2]->fHueShift, entry01[1]->fHueShift, entry01[0]->fHueShift);
+ __m128 hueShift0 = _mm_add_ps(_mm_mul_ps(hs0, hFract0),
_mm_mul_ps(hs1, hFract1));
+ hueShift0 = _mm_mul_ps(hueShift0, sFract0);
+
+ __m128 ss0 = _mm_set_ps(entry00[3]->fSatScale,
entry00[2]->fSatScale, entry00[1]->fSatScale, entry00[0]->fSatScale);
+ __m128 ss1 = _mm_set_ps(entry01[3]->fSatScale,
entry01[2]->fSatScale, entry01[1]->fSatScale, entry01[0]->fSatScale);
+ __m128 satScale0 = _mm_add_ps(_mm_mul_ps(ss0, hFract0),
_mm_mul_ps(ss1, hFract1));
+ satScale0 = _mm_mul_ps(satScale0, sFract0);
+
+ __m128 vs0 = _mm_set_ps(entry00[3]->fValScale,
entry00[2]->fValScale, entry00[1]->fValScale, entry00[0]->fValScale);
+ __m128 vs1 = _mm_set_ps(entry01[3]->fValScale,
entry01[2]->fValScale, entry01[1]->fValScale, entry01[0]->fValScale);
+ __m128 valScale0 = _mm_add_ps(_mm_mul_ps(vs0, hFract0),
_mm_mul_ps(vs1, hFract1));
+ valScale0 = _mm_mul_ps(valScale0, sFract0);
+
+
+ for (i = 0; i < 4; i++) {
+ entry00[i]++;
+ entry01[i]++;
+ }
+
+ hs0 = _mm_set_ps(entry00[3]->fHueShift, entry00[2]->fHueShift,
entry00[1]->fHueShift, entry00[0]->fHueShift);
+ hs1 = _mm_set_ps(entry01[3]->fHueShift, entry01[2]->fHueShift,
entry01[1]->fHueShift, entry01[0]->fHueShift);
+ __m128 hueShift1 = _mm_add_ps(_mm_mul_ps(hs0, hFract0),
_mm_mul_ps(hs1, hFract1));
+ hueShift = _mm_add_ps(hueShift0, _mm_mul_ps(hueShift1,
sFract1));
+
+ ss0 = _mm_set_ps(entry00[3]->fSatScale, entry00[2]->fSatScale,
entry00[1]->fSatScale, entry00[0]->fSatScale);
+ ss1 = _mm_set_ps(entry01[3]->fSatScale, entry01[2]->fSatScale,
entry01[1]->fSatScale, entry01[0]->fSatScale);
+ __m128 satScale1 = _mm_add_ps(_mm_mul_ps(ss0, hFract0),
_mm_mul_ps(ss1, hFract1));
+ satScale = _mm_add_ps(satScale0, _mm_mul_ps(satScale1,
sFract1));
+
+ vs0 = _mm_set_ps(entry00[3]->fValScale, entry00[2]->fValScale,
entry00[1]->fValScale, entry00[0]->fValScale);
+ vs1 = _mm_set_ps(entry01[3]->fValScale, entry01[2]->fValScale,
entry01[1]->fValScale, entry01[0]->fValScale);
+ __m128 valScale1 = _mm_add_ps(_mm_mul_ps(vs0, hFract0),
_mm_mul_ps(vs1, hFract1));
+ valScale = _mm_add_ps(valScale0, _mm_mul_ps(valScale1,
sFract1));
+
+ }
+ else
+ {
+ __m128 hScaled = _mm_mul_ps(h, _mm_set_ps( hScale, hScale,
hScale, hScale));
+ __m128 sScaled = _mm_mul_ps(s, _mm_set_ps( sScale, sScale,
sScale, sScale));
+ __m128 vScaled = _mm_mul_ps(v, _mm_set_ps( vScale, vScale,
vScale, vScale));
+
+ __m128 half_ps = _mm_set_ps(0.5, 0.5, 0.5, 0.5);
+ __m128i maxHueIndex0 = _mm_set_epi32(_maxHueIndex0,
_maxHueIndex0, _maxHueIndex0, _maxHueIndex0);
+ __m128i maxSatIndex0 = _mm_set_epi32(_maxSatIndex0,
_maxSatIndex0, _maxSatIndex0, _maxSatIndex0);
+ __m128i maxValIndex0 = _mm_set_epi32(_maxValIndex0,
_maxValIndex0, _maxValIndex0, _maxValIndex0);
+ __m128i hIndex0 = _mm_cvtps_epi32(_mm_sub_ps( hScaled, half_ps
));
+ __m128i sIndex0 = _mm_cvtps_epi32(_mm_sub_ps( sScaled, half_ps
));
+ __m128i vIndex0 = _mm_cvtps_epi32(_mm_sub_ps( vScaled, half_ps
));
+
+ // Requires that maxSatIndex0 and sIndex0 can be contained
within a 16 bit signed word.
+ sIndex0 = _mm_min_epi16(sIndex0, maxSatIndex0);
+ vIndex0 = _mm_min_epi16(vIndex0, maxValIndex0);
+ __m128i ones_epi32 = _mm_set_epi32(1,1,1,1);
+ __m128i hIndex1 = _mm_add_epi32(hIndex0, ones_epi32);
+
+ /* if (hIndex0 > (maxHueIndex0 - 1)) */
+ __m128i hIndexMask = _mm_cmpgt_epi32( hIndex0,
_mm_sub_epi32(maxHueIndex0, ones_epi32));
+ /* Make room in hIndex0 */
+ hIndex0 = _mm_andnot_si128(hIndexMask, hIndex0);
+ /* hIndex1 = 0; */
+ hIndex1 = _mm_andnot_si128(hIndexMask, hIndex1);
+ /* hIndex0 = maxHueIndex0, where hIndex0 >= (maxHueIndex0) */
+ hIndex0 = _mm_or_si128(hIndex0, _mm_and_si128(hIndexMask,
maxHueIndex0));
+
+ __m128 hFract1 = _mm_sub_ps( hScaled, _mm_cvtepi32_ps(hIndex0));
+ __m128 sFract1 = _mm_sub_ps( sScaled, _mm_cvtepi32_ps(sIndex0));
+ __m128 vFract1 = _mm_sub_ps( vScaled, _mm_cvtepi32_ps(vIndex0));
+ __m128 ones_ps = _mm_add_ps( half_ps, half_ps);
+
+ __m128 hFract0 = _mm_sub_ps(ones_ps, hFract1);
+ __m128 sFract0 = _mm_sub_ps(ones_ps, sFract1);
+ __m128 vFract0 = _mm_sub_ps(ones_ps, vFract1);
+
+ __m128i hueStep = _mm_set_epi32(map->sat_divisions,
map->sat_divisions, map->sat_divisions, map->sat_divisions);
+ __m128i valStep = _mm_set_epi32(_valStep, _valStep, _valStep,
_valStep);
+
+ // This requires that hueStep and valStep can be contained in a
16 bit signed integer.
+ __m128i table_offsets = _mm_add_epi32(sIndex0,
_mm_mullo_epi16(vIndex0, valStep));
+ __m128i next_offsets = _mm_mullo_epi16(hIndex1, hueStep);
+ next_offsets = _mm_add_epi32(next_offsets, table_offsets);
+ table_offsets = _mm_add_epi32(table_offsets,
_mm_mullo_epi16(hIndex0, hueStep));
+
+ // TODO: This will result in a store->load forward size
mismatch penalty, if possible, avoid.
+ _mm_store_si128((__m128i*)xfer_0, table_offsets);
+ _mm_store_si128((__m128i*)xfer_1, next_offsets);
+
+ const RS_VECTOR3 *entry00[4] = { tableBase + xfer_0[0],
tableBase + xfer_0[1], tableBase + xfer_0[2], tableBase + xfer_0[3]};
+ const RS_VECTOR3 *entry01[4] = { tableBase + xfer_1[0],
tableBase + xfer_1[1], tableBase + xfer_1[2], tableBase + xfer_1[3]};
+ const RS_VECTOR3 *entry10[4] = { entry00[0] + _valStep,
entry00[1] + _valStep, entry00[2] + _valStep, entry00[3] + _valStep};
+ const RS_VECTOR3 *entry11[4] = { entry01[0] + _valStep,
entry01[1] + _valStep, entry01[2] + _valStep, entry01[3] + _valStep};
+
+ __m128 hs00 = _mm_set_ps(entry00[3]->fHueShift,
entry00[2]->fHueShift, entry00[1]->fHueShift, entry00[0]->fHueShift);
+ __m128 hs01 = _mm_set_ps(entry01[3]->fHueShift,
entry01[2]->fHueShift, entry01[1]->fHueShift, entry01[0]->fHueShift);
+ __m128 hs10 = _mm_set_ps(entry10[3]->fHueShift,
entry10[2]->fHueShift, entry10[1]->fHueShift, entry10[0]->fHueShift);
+ __m128 hs11 = _mm_set_ps(entry11[3]->fHueShift,
entry11[2]->fHueShift, entry11[1]->fHueShift, entry11[0]->fHueShift);
+ __m128 hueShift0 = _mm_mul_ps(vFract0,
_mm_add_ps(_mm_mul_ps(hs00, hFract0), _mm_mul_ps(hs01, hFract1)));
+ __m128 hueShift1 = _mm_mul_ps(vFract1,
_mm_add_ps(_mm_mul_ps(hs10, hFract0), _mm_mul_ps(hs11, hFract1)));
+ hueShift = _mm_mul_ps(sFract0, _mm_add_ps(hueShift0,
hueShift1));
+
+ __m128 ss00 = _mm_set_ps(entry00[3]->fSatScale,
entry00[2]->fSatScale, entry00[1]->fSatScale, entry00[0]->fSatScale);
+ __m128 ss01 = _mm_set_ps(entry01[3]->fSatScale,
entry01[2]->fSatScale, entry01[1]->fSatScale, entry01[0]->fSatScale);
+ __m128 ss10 = _mm_set_ps(entry10[3]->fSatScale,
entry10[2]->fSatScale, entry10[1]->fSatScale, entry10[0]->fSatScale);
+ __m128 ss11 = _mm_set_ps(entry11[3]->fSatScale,
entry11[2]->fSatScale, entry11[1]->fSatScale, entry11[0]->fSatScale);
+ __m128 satScale0 = _mm_mul_ps(vFract0,
_mm_add_ps(_mm_mul_ps(ss00, hFract0), _mm_mul_ps(ss01, hFract1)));
+ __m128 satScale1 = _mm_mul_ps(vFract1,
_mm_add_ps(_mm_mul_ps(ss10, hFract0), _mm_mul_ps(ss11, hFract1)));
+ satScale = _mm_mul_ps(sFract0, _mm_add_ps(satScale0,
satScale1));
+
+ __m128 vs00 = _mm_set_ps(entry00[3]->fValScale,
entry00[2]->fValScale, entry00[1]->fValScale, entry00[0]->fValScale);
+ __m128 vs01 = _mm_set_ps(entry01[3]->fValScale,
entry01[2]->fValScale, entry01[1]->fValScale, entry01[0]->fValScale);
+ __m128 vs10 = _mm_set_ps(entry10[3]->fValScale,
entry10[2]->fValScale, entry10[1]->fValScale, entry10[0]->fValScale);
+ __m128 vs11 = _mm_set_ps(entry11[3]->fValScale,
entry11[2]->fValScale, entry11[1]->fValScale, entry11[0]->fValScale);
+ __m128 valScale0 = _mm_mul_ps(vFract0,
_mm_add_ps(_mm_mul_ps(vs00, hFract0), _mm_mul_ps(vs01, hFract1)));
+ __m128 valScale1 = _mm_mul_ps(vFract1,
_mm_add_ps(_mm_mul_ps(vs10, hFract0), _mm_mul_ps(vs11, hFract1)));
+ valScale = _mm_mul_ps(sFract0, _mm_add_ps(valScale0,
valScale1));
+
+ for (i = 0; i < 4; i++) {
+ entry00[i]++;
+ entry01[i]++;
+ entry10[i]++;
+ entry11[i]++;
+ }
+
+ hs00 = _mm_set_ps(entry00[3]->fHueShift, entry00[2]->fHueShift,
entry00[1]->fHueShift, entry00[0]->fHueShift);
+ hs01 = _mm_set_ps(entry01[3]->fHueShift, entry01[2]->fHueShift,
entry01[1]->fHueShift, entry01[0]->fHueShift);
+ hs10 = _mm_set_ps(entry10[3]->fHueShift, entry10[2]->fHueShift,
entry10[1]->fHueShift, entry10[0]->fHueShift);
+ hs11 = _mm_set_ps(entry11[3]->fHueShift, entry11[2]->fHueShift,
entry11[1]->fHueShift, entry11[0]->fHueShift);
+ hueShift0 = _mm_mul_ps(vFract0, _mm_add_ps(_mm_mul_ps(hs00,
hFract0), _mm_mul_ps(hs01, hFract1)));
+ hueShift1 = _mm_mul_ps(vFract1, _mm_add_ps(_mm_mul_ps(hs10,
hFract0), _mm_mul_ps(hs11, hFract1)));
+ hueShift = _mm_add_ps(hueShift, _mm_mul_ps(sFract1,
_mm_add_ps(hueShift0, hueShift1)));
+
+ ss00 = _mm_set_ps(entry00[3]->fSatScale, entry00[2]->fSatScale,
entry00[1]->fSatScale, entry00[0]->fSatScale);
+ ss01 = _mm_set_ps(entry01[3]->fSatScale, entry01[2]->fSatScale,
entry01[1]->fSatScale, entry01[0]->fSatScale);
+ ss10 = _mm_set_ps(entry10[3]->fSatScale, entry10[2]->fSatScale,
entry10[1]->fSatScale, entry10[0]->fSatScale);
+ ss11 = _mm_set_ps(entry11[3]->fSatScale, entry11[2]->fSatScale,
entry11[1]->fSatScale, entry11[0]->fSatScale);
+ satScale0 = _mm_mul_ps(vFract0, _mm_add_ps(_mm_mul_ps(ss00,
hFract0), _mm_mul_ps(ss01, hFract1)));
+ satScale1 = _mm_mul_ps(vFract1, _mm_add_ps(_mm_mul_ps(ss10,
hFract0), _mm_mul_ps(ss11, hFract1)));
+ satScale = _mm_add_ps(satScale, _mm_mul_ps(sFract1,
_mm_add_ps(satScale0, satScale1)));
+
+ vs00 = _mm_set_ps(entry00[3]->fValScale, entry00[2]->fValScale,
entry00[1]->fValScale, entry00[0]->fValScale);
+ vs01 = _mm_set_ps(entry01[3]->fValScale, entry01[2]->fValScale,
entry01[1]->fValScale, entry01[0]->fValScale);
+ vs10 = _mm_set_ps(entry10[3]->fValScale, entry10[2]->fValScale,
entry10[1]->fValScale, entry10[0]->fValScale);
+ vs11 = _mm_set_ps(entry11[3]->fValScale, entry11[2]->fValScale,
entry11[1]->fValScale, entry11[0]->fValScale);
+ valScale0 = _mm_mul_ps(vFract0, _mm_add_ps(_mm_mul_ps(vs00,
hFract0), _mm_mul_ps(vs01, hFract1)));
+ valScale1 = _mm_mul_ps(vFract1, _mm_add_ps(_mm_mul_ps(vs10,
hFract0), _mm_mul_ps(vs11, hFract1)));
+ valScale = _mm_add_ps(valScale, _mm_mul_ps(sFract1,
_mm_add_ps(valScale0, valScale1)));
+ }
+
+ __m128 mul_hue = _mm_set_ps(6.0f / 360.0f, 6.0f / 360.0f, 6.0f /
360.0f, 6.0f / 360.0f);
+ __m128 ones_ps = _mm_set_ps(1.0, 1.0, 1.0, 1.0);
+ hueShift = _mm_mul_ps(hueShift, mul_hue);
+ s = _mm_min_ps(ones_ps, _mm_mul_ps(s, satScale));
+ v = _mm_min_ps(ones_ps, _mm_mul_ps(v, valScale));
+ h = _mm_add_ps(h, hueShift);
+ *_h = h;
+ *_s = s;
+ *_v = v;
+}
+
+#endif // defined __SSE2__
+
/* RefBaselineRGBTone() */
void
rgb_tone(gfloat *_r, gfloat *_g, gfloat *_b, const gfloat * const tone_lut)
@@ -814,7 +1049,6 @@
__m128 r, g, b, r2, g2, b2;
__m128i zero;
- printf("Using SSE2\n");
int xfer[4] __attribute__ ((aligned (16)));
const gfloat exposure_comp = pow(2.0, dcp->exposure);
@@ -882,14 +1116,7 @@
if (dcp->huesatmap)
{
- gfloat* h_p = (gfloat*)&h;
- gfloat* s_p = (gfloat*)&s;
- gfloat* v_p = (gfloat*)&v;
-
- huesat_map(dcp->huesatmap, &h_p[0], &s_p[0],
&v_p[0]);
- huesat_map(dcp->huesatmap, &h_p[1], &s_p[1],
&v_p[1]);
- huesat_map(dcp->huesatmap, &h_p[2], &s_p[2],
&v_p[2]);
- huesat_map(dcp->huesatmap, &h_p[3], &s_p[3],
&v_p[3]);
+ huesat_map_SSE2(dcp->huesatmap, &h, &s, &v);
}
/* Exposure */
@@ -933,12 +1160,7 @@
if (dcp->looktable) {
- gfloat* h_p = (gfloat*)&h;
- gfloat* s_p = (gfloat*)&s;
- huesat_map(dcp->looktable, &h_p[0], &s_p[0],
&v_p[0]);
- huesat_map(dcp->looktable, &h_p[1], &s_p[1],
&v_p[1]);
- huesat_map(dcp->looktable, &h_p[2], &s_p[2],
&v_p[2]);
- huesat_map(dcp->looktable, &h_p[3], &s_p[3],
&v_p[3]);
+ huesat_map_SSE2(dcp->looktable, &h, &s, &v);
}
_______________________________________________
Rawstudio-commit mailing list
[email protected]
http://rawstudio.org/cgi-bin/mailman/listinfo/rawstudio-commit
