Add SkDefaultXform as a catch-all to handle color conversions

src/core/SkColorSpaceXform.cpp

 /*
  * Copyright 2016 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #include "SkColorPriv.h"
 #include "SkColorSpace_Base.h"
 #include "SkColorSpaceXform.h"
 
 bool compute_gamut_xform(SkMatrix44* srcToDst, const SkMatrix44& srcToXYZ,
                          const SkMatrix44& dstToXYZ) {
     if (!dstToXYZ.invert(srcToDst)) {
         return false;
     }
 
     srcToDst->postConcat(srcToXYZ);
     return true;
 }
 
 std::unique_ptr<SkColorSpaceXform> SkColorSpaceXform::New(const sk_sp<SkColorSpace>& srcSpace,
                                                           const sk_sp<SkColorSpace>& dstSpace) {
-    if (!srcSpace || !dstSpace) {
+    if (!srcSpace || !dstSpace || as_CSB(srcSpace)->colorLUT() || as_CSB(dstSpace)->colorLUT()) {

[scroggo, 2016/06/06 14:06:44]

Just making sure I understand - colorLUT is not yet supported?

[msarett, 2016/06/06 17:33:44]

Yes, this is a mix of unimplemented and invalid.  I think it's more clear if I
separate into two statements.  FWIW, I don't know that we'll ever support
colorLUTs on dstSpaces.  But it may make sense to just ignore them...

+        return nullptr;
+    }
+
+    SkMatrix44 srcToDst(SkMatrix44::kUninitialized_Constructor);
+    if (!compute_gamut_xform(&srcToDst, srcSpace->xyz(), dstSpace->xyz())) {
         return nullptr;
     }
 
     if (as_CSB(srcSpace)->gammas()->isValues() && as_CSB(dstSpace)->gammas()->isValues()) {
-        SkMatrix44 srcToDst(SkMatrix44::kUninitialized_Constructor);
-        if (!compute_gamut_xform(&srcToDst, srcSpace->xyz(), dstSpace->xyz())) {
-            return nullptr;
-        }
-
         float srcGammas[3];
         float dstGammas[3];
         srcGammas[0] = as_CSB(srcSpace)->gammas()->fRed.fValue;
         srcGammas[1] = as_CSB(srcSpace)->gammas()->fGreen.fValue;
         srcGammas[2] = as_CSB(srcSpace)->gammas()->fBlue.fValue;
         dstGammas[0] = 1.0f / as_CSB(dstSpace)->gammas()->fRed.fValue;
         dstGammas[1] = 1.0f / as_CSB(dstSpace)->gammas()->fGreen.fValue;
         dstGammas[2] = 1.0f / as_CSB(dstSpace)->gammas()->fBlue.fValue;
 
         return std::unique_ptr<SkColorSpaceXform>(
                 new SkGammaByValueXform(srcGammas, srcToDst, dstGammas));
     }
 
-    // Unimplemeted
-    return nullptr;
+    return std::unique_ptr<SkColorSpaceXform>(
+            new SkDefaultXform(as_CSB(srcSpace)->gammas(), srcToDst, as_CSB(dstSpace)->gammas()));
 }
 
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 
-SkGammaByValueXform::SkGammaByValueXform(float srcGammas[3], const SkMatrix44& srcToDst,
-                                         float dstGammas[3])
-    : fSrcToDst(srcToDst)
-{
-    memcpy(fSrcGammas, srcGammas, 3 * sizeof(float));
-    memcpy(fDstGammas, dstGammas, 3 * sizeof(float));
+static float byte_to_float(uint8_t v) {

[scroggo, 2016/06/06 14:06:44]

Should this be inline?

[msarett, 2016/06/06 17:33:44]

Yeah I think so.  "inline" is just a suggestion to the compiler (and I'm 99%
sure this will get inlined regardless of whether I suggest it or not), but at
the very least, this adds a bit of documentation.

+    return ((float) v) * (1.0f / 255.0f);
 }
 
 static uint8_t clamp_float_to_byte(float v) {
     v = v * 255.0f;
     if (v > 255.0f) {
         return 255;
     } else if (v <= 0.0f) {
         return 0;
     } else {
         return (uint8_t) (v + 0.5f);
     }
 }
 
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+SkGammaByValueXform::SkGammaByValueXform(float srcGammas[3], const SkMatrix44& srcToDst,
+                                         float dstGammas[3])
+    : fSrcToDst(srcToDst)
+{
+    memcpy(fSrcGammas, srcGammas, 3 * sizeof(float));
+    memcpy(fDstGammas, dstGammas, 3 * sizeof(float));
+}
+
 void SkGammaByValueXform::xform_RGBA_8888(uint32_t* dst, const uint32_t* src, uint32_t len) const {
     while (len-- > 0) {
         float srcFloats[3];
-        srcFloats[0] = ((*src >>  0) & 0xFF) * (1.0f / 255.0f);
-        srcFloats[1] = ((*src >>  8) & 0xFF) * (1.0f / 255.0f);
-        srcFloats[2] = ((*src >> 16) & 0xFF) * (1.0f / 255.0f);
+        srcFloats[0] = byte_to_float((*src >>  0) & 0xFF);
+        srcFloats[1] = byte_to_float((*src >>  8) & 0xFF);
+        srcFloats[2] = byte_to_float((*src >> 16) & 0xFF);
 
         // Convert to linear.
         srcFloats[0] = pow(srcFloats[0], fSrcGammas[0]);
         srcFloats[1] = pow(srcFloats[1], fSrcGammas[1]);
         srcFloats[2] = pow(srcFloats[2], fSrcGammas[2]);
 
         // Convert to dst gamut.
         float dstFloats[3];
         dstFloats[0] = srcFloats[0] * fSrcToDst.getFloat(0, 0) +
                        srcFloats[1] * fSrcToDst.getFloat(1, 0) +
@@ skipping 12 matching lines @@
 
         *dst = SkPackARGB32NoCheck(((*src >> 24) & 0xFF),
                                    clamp_float_to_byte(dstFloats[0]),
                                    clamp_float_to_byte(dstFloats[1]),
                                    clamp_float_to_byte(dstFloats[2]));
 
         dst++;
         src++;
     }
 }
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Interpolating lookup in a variably sized table.
+float interp_lut(uint8_t byte, float* table, size_t tableSize) {
+    float index = byte * (1.0f / 255.0f) * (tableSize - 1);

[scroggo, 2016/06/06 14:06:44]

nit:

    byte * (1.0f / 255.0f)

Won't "byte" get promoted to float, and then this is the the same as saying

    byte_to_float(byte) * (tableSize - 1)

?

[msarett, 2016/06/06 17:33:43]

Yes it is, thanks!

+    float diff = index - floor(index);
+    return table[(int) floor(index)] * (1.0f - diff) + table[(int) ceil(index)] * diff;

[scroggo, 2016/06/06 14:06:44]

Note: We have macros for sk_float_floor2int and sk_ceil_floor2int. Maybe you
should use them here?

[msarett, 2016/06/06 17:33:43]

Agreed, done.

+}
+
+// Inverse table lookup.  Ex: what index corresponds to the input value?  This will
+// have strange results when the table is non-increasing.  But any sane gamma

[scroggo, 2016/06/06 14:06:44]

Have you seen any insane gamma functions? In that case, do we just return a
strange result, or do we avoid calling this method?

[msarett, 2016/06/06 17:33:44]

I have not seen any insane gammas.  If we did encounter some, I think we would
return a strange result.

In general, I have a lot of work left to do on destination profiles with strange
gammas.  I'm not even if sure if or how often these profiles exist.  This code
fills a gap in the implementation (I think we'll need this because qcms supports
it), but is quite a bit less efficient/robust than qcms I think.

Maybe it would be better for me to not include this at all (until I can do a
better job)?  And just do strange src gammas in this CL?  But I think it's
alright to leave it here, and add more verbose FIXMEs.  Having it does help with
testing.

[scroggo, 2016/06/06 17:40:50]

FIXME sounds good to me.

+// function will be increasing.
+float interp_lut_inv(float input, float* table, size_t tableSize) {
+    if (input <= table[0]) {
+        return table[0];
+    } else if (input >= table[tableSize - 1]) {
+        return 1.0f;
+    }
+
+    for (uint32_t i = 1; i < tableSize; i++) {

[scroggo, 2016/06/06 14:06:44]

Would it be worth it to use something faster than a linear search?

[msarett, 2016/06/06 17:33:44]

If we verify that this function is increasing beforehand, we could/should use a
better search.  I think this entire strategy is flawed though - I think it's
more likely that I'll change this implementation than try to optimize it.

[scroggo, 2016/06/06 17:40:50]

sgtm

+        if (table[i] >= input) {
+            // We are guaranteed that input is greater than table[i - 1].
+            float diff = input - table[i - 1];
+            float distance = table[i] - table[i - 1];
+            float index = (i - 1) + diff / distance;
+            return index / (tableSize - 1);
+        }
+    }
+
+    // Should be unreachable, since we'll return before the loop if input is
+    // larger than the last entry.
+    SkASSERT(false);
+    return 0.0f;
+}
+
+SkDefaultXform::SkDefaultXform(const sk_sp<SkGammas>& srcGammas, const SkMatrix44& srcToDst,
+                               const sk_sp<SkGammas>& dstGammas)
+    : fSrcGammas(srcGammas)
+    , fSrcToDst(srcToDst)
+    , fDstGammas(dstGammas)
+{}
+
+void SkDefaultXform::xform_RGBA_8888(uint32_t* dst, const uint32_t* src, uint32_t len) const {
+    while (len-- > 0) {
+        // Convert to linear.
+        // FIXME (msarett):
+        // Rather than support three different strategies of transforming gamma, QCMS
+        // builds a 256 entry float lookup table from the gamma info.  This handles
+        // the gamma transform and the conversion from bytes to floats.  This may
+        // be simpler and faster than our current approach.
+        float srcFloats[3];
+        for (int i = 0; i < 3; i++) {
+            const SkGammaCurve& gamma = (*fSrcGammas)[i];
+            uint8_t byte = (*src >> (8 * i)) & 0xFF;
+            if (gamma.isValue()) {
+                srcFloats[i] = pow(byte_to_float(byte), gamma.fValue);
+            } else if (gamma.isTable()) {
+                srcFloats[i] = interp_lut(byte, gamma.fTable.get(), gamma.fTableSize);
+            } else {
+                SkASSERT(gamma.isParametric());
+                float component = byte_to_float(byte);
+                if (component < gamma.fD) {
+                    // Y = E * X + F
+                    srcFloats[i] = gamma.fE * component + gamma.fF;
+                } else {
+                    // Y = (A * X + B)^G + C
+                    srcFloats[i] = pow(gamma.fA * component + gamma.fB, gamma.fG) + gamma.fC;
+                }
+            }
+        }
+
+        // Convert to dst gamut.
+        float dstFloats[3];
+        dstFloats[0] = srcFloats[0] * fSrcToDst.getFloat(0, 0) +
+                       srcFloats[1] * fSrcToDst.getFloat(1, 0) +
+                       srcFloats[2] * fSrcToDst.getFloat(2, 0) + fSrcToDst.getFloat(3, 0);
+        dstFloats[1] = srcFloats[0] * fSrcToDst.getFloat(0, 1) +
+                       srcFloats[1] * fSrcToDst.getFloat(1, 1) +
+                       srcFloats[2] * fSrcToDst.getFloat(2, 1) + fSrcToDst.getFloat(3, 1);
+        dstFloats[2] = srcFloats[0] * fSrcToDst.getFloat(0, 2) +
+                       srcFloats[1] * fSrcToDst.getFloat(1, 2) +
+                       srcFloats[2] * fSrcToDst.getFloat(2, 2) + fSrcToDst.getFloat(3, 2);
+
+        // Convert to dst gamma.
+        // FIXME (msarett):
+        // Rather than support three different strategies of transforming inverse gamma,
+        // QCMS builds a large float lookup table from the gamma info.  Is this faster or
+        // better than our approach?
+        for (int i = 0; i < 3; i++) {
+            const SkGammaCurve& gamma = (*fDstGammas)[i];
+            if (gamma.isValue()) {
+                dstFloats[i] = pow(dstFloats[i], 1.0f / gamma.fValue);
+            } else if (gamma.isTable()) {
+                // FIXME (msarett):
+                // An inverse table lookup is particularly strange and non-optimal.
+                dstFloats[i] = interp_lut_inv(dstFloats[i], gamma.fTable.get(), gamma.fTableSize);
+            } else {
+                SkASSERT(gamma.isParametric());
+                // We need to take the inverse of a piecewise function.  Assume that
+                // the gamma function is continuous, or this won't make much sense
+                // anyway.
+                // Plug in |fD| to the first equation to calculate the new piecewise
+                // interval.  Then simply use the inverse of the original functions.
+                float interval = gamma.fE * gamma.fD + gamma.fF;
+
+                // FIXME (msarett):
+                // Is this math safe?  Are we at risk of dividing by zero?

[scroggo, 2016/06/06 14:06:44]

The risk is dividing by E or G. Those were read from the input. Did you correct
them if they were zero? (Is zero wrong?) If not, the stream can be arbitrary -
including garbage - so we need to catch it somehow.

[msarett, 2016/06/06 17:33:44]

You're right.  I added some checks and removed the FIXME.  Also fixed a bug
where I forgot to divide by A.  This passed the test because A = 1 in the test.

+                if (dstFloats[i] < interval) {
+                    // X = (Y - F) / E
+                    dstFloats[i] = (dstFloats[i] - gamma.fF) / gamma.fE;
+                } else {
+                    // X = (Y - C)^(1 / G) - B
+                    dstFloats[i] = pow(dstFloats[i] - gamma.fC, 1.0f / gamma.fG) - gamma.fB;
+                }
+            }
+        }
+
+        *dst = SkPackARGB32NoCheck(((*src >> 24) & 0xFF),

[scroggo, 2016/06/06 14:06:44]

Why is it safe to use the NoCheck version? Is this unpremultiplied? (I don't see
that documented anywhere.)

[msarett, 2016/06/06 17:33:44]

This function is (currently) opaque->opaque or unpremul->unpremul.  There is a
comment on the base class declaration of xform_RGBA_8888() that states that the
output is not premultiplied.

When we do add a premultiplication option to this function, it is probably most
correct to premultiply the linear floats before applying the dst gamma.

+                                   clamp_float_to_byte(dstFloats[0]),
+                                   clamp_float_to_byte(dstFloats[1]),
+                                   clamp_float_to_byte(dstFloats[2]));
+
+        dst++;
+        src++;
+    }
+}