Use the caching transform prelude for non-SSE code in transform.c

Follow on from https://codereview.chromium.org/1299203002, use
the cache for the qcms_transform_data_tetra_clut_rgba s/w path
also. Note: this s/w code path is not currently used in chrome
(the SSE code path is always selected).

Before this CL:

Windows 64 bit:
 ./out/Release_x64/qcms_test.exe -i 40 -w 2880 -h 1800
Test qcms clut transforms for 40 iterations
Test image size 2880 x 1800 pixels
15.533663 (avg 0.388342) seconds qcms_transform_data_tetra_clut_rgba
4.307897 (avg 0.107697) seconds qcms_transform_data_tetra_clut_rgba_sse2
3.605858 speedup after 40 iterations

Linux 64 bit:
 ./out/Release/qcms_test -i 40 -w 2880 -h 1800
Test qcms clut transforms for 40 iterations
Test image size 2880 x 1800 pixels
7.678742 (avg 0.191969) seconds qcms_transform_data_tetra_clut_rgba
3.138837 (avg 0.078471) seconds qcms_transform_data_tetra_clut_rgba_sse2
2.446366 speedup after 40 iterations

With this CL:

Windows 64 bit:
 ./out/Release_x64/qcms_test.exe -i 40 -w 2880 -h 1800
Test qcms clut transforms for 40 iterations
Test image size 2880 x 1800 pixels
9.351532 (avg 0.233788) seconds qcms_transform_data_tetra_clut_rgba
4.337411 (avg 0.108435) seconds qcms_transform_data_tetra_clut_rgba_sse2
2.156017 speedup after 40 iterations

Linux 64 bit:
 ./out/Release/qcms_test -i 40 -w 2880 -h 1800
Test qcms clut transforms for 40 iterations
Test image size 2880 x 1800 pixels
5.444788 (avg 0.136120) seconds qcms_transform_data_tetra_clut_rgba
3.143324 (avg 0.078583) seconds qcms_transform_data_tetra_clut_rgba_sse2
1.732175 speedup after 40 iterations

Win software path 388ms -> 233ms, no change in the SSE path as
expected (it was speedup by the previous code review).

Also compared to where we were before doing tetra SSE, the s/w
path has been speedup by ~2x, and the SSE path 4.25x. Absolute
frame speed of the SSE is now faster than SkColorCubeFilter on
win-x64 (122ms).

CQ_INCLUDE_TRYBOTS=tryserver.blink:linux_blink_rel,mac_blink_rel,win_blink_rel

BUG=506607

Review URL: https://codereview.chromium.org/1391603002

Cr-Commit-Position: refs/heads/master@{#352603}

2 files changed, 26 insertions, 18 deletions

diff --git a/third_party/qcms/README.chromium b/third_party/qcms/README.chromium
index b1845fd..d05e0c1 100644
--- a/third_party/qcms/README.chromium
+++ b/third_party/qcms/README.chromium
@@ -105,6 +105,8 @@ The following changes have been made since qcms was imported:
    - https://code.google.com/p/chromium/issues/detail?id=487284
  - Caching transform prelude to save computation in V4 profiles
    - https://code.google.com/p/chromium/issues/detail?id=506607
+ - Use the caching transform prelude for non-SSE code in transform.c
+   - https://code.google.com/p/chromium/issues/detail?id=506607 
 
 For the Chromium changes, since the import, in a patch format run:
   git diff b8456f38 src
diff --git a/third_party/qcms/src/transform.c b/third_party/qcms/src/transform.c
index cbdfa59..1e707b7 100644
--- a/third_party/qcms/src/transform.c
+++ b/third_party/qcms/src/transform.c
@@ -531,6 +531,9 @@ void qcms_transform_data_tetra_clut_rgba(qcms_transform *transform, unsigned cha
 	float c0_b, c1_b, c2_b, c3_b;
 	float clut_r, clut_g, clut_b;
 
+	if (!(transform->transform_flags & TRANSFORM_FLAG_CLUT_CACHE))
+		qcms_transform_build_clut_cache(transform);
+
 	for (i = 0; i < length; i++) {
 		unsigned char in_r = *src++;
 		unsigned char in_g = *src++;
@@ -539,17 +542,17 @@ void qcms_transform_data_tetra_clut_rgba(qcms_transform *transform, unsigned cha
 
 		float linear_r = in_r/255.0f, linear_g=in_g/255.0f, linear_b = in_b/255.0f;
 
-		int x = floor(linear_r * (transform->grid_size-1));
-		int y = floor(linear_g * (transform->grid_size-1));
-		int z = floor(linear_b * (transform->grid_size-1));
+		int x = transform->floor_cache[in_r];
+		int y = transform->floor_cache[in_g];
+		int z = transform->floor_cache[in_b];
 
-		int x_n = ceil(linear_r * (transform->grid_size-1));
-		int y_n = ceil(linear_g * (transform->grid_size-1));
-		int z_n = ceil(linear_b * (transform->grid_size-1));
+		int x_n = transform->ceil_cache[in_r];
+		int y_n = transform->ceil_cache[in_g];
+		int z_n = transform->ceil_cache[in_b];
 
-		float rx = linear_r * (transform->grid_size-1) - x;
-		float ry = linear_g * (transform->grid_size-1) - y;
-		float rz = linear_b * (transform->grid_size-1) - z;
+		float rx = transform->r_cache[in_r];
+		float ry = transform->r_cache[in_g];
+		float rz = transform->r_cache[in_b];
 
 		c0_r = CLU(r_table, x, y, z);
 		c0_g = CLU(g_table, x, y, z);
@@ -655,6 +658,9 @@ static void qcms_transform_data_tetra_clut(qcms_transform *transform, unsigned c
 	float c0_b, c1_b, c2_b, c3_b;
 	float clut_r, clut_g, clut_b;
 
+	if (!(transform->transform_flags & TRANSFORM_FLAG_CLUT_CACHE))
+		qcms_transform_build_clut_cache(transform);
+
 	for (i = 0; i < length; i++) {
 		unsigned char in_r = *src++;
 		unsigned char in_g = *src++;
@@ -662,17 +668,17 @@ static void qcms_transform_data_tetra_clut(qcms_transform *transform, unsigned c
 
 		float linear_r = in_r/255.0f, linear_g=in_g/255.0f, linear_b = in_b/255.0f;
 
-		int x = floor(linear_r * (transform->grid_size-1));
-		int y = floor(linear_g * (transform->grid_size-1));
-		int z = floor(linear_b * (transform->grid_size-1));
+		int x = transform->floor_cache[in_r];
+		int y = transform->floor_cache[in_g];
+		int z = transform->floor_cache[in_b];
 
-		int x_n = ceil(linear_r * (transform->grid_size-1));
-		int y_n = ceil(linear_g * (transform->grid_size-1));
-		int z_n = ceil(linear_b * (transform->grid_size-1));
+		int x_n = transform->ceil_cache[in_r];
+		int y_n = transform->ceil_cache[in_g];
+		int z_n = transform->ceil_cache[in_b];
 
-		float rx = linear_r * (transform->grid_size-1) - x;
-		float ry = linear_g * (transform->grid_size-1) - y;
-		float rz = linear_b * (transform->grid_size-1) - z;
+		float rx = transform->r_cache[in_r];
+		float ry = transform->r_cache[in_g];
+		float rz = transform->r_cache[in_b];
 
 		c0_r = CLU(r_table, x, y, z);
 		c0_g = CLU(g_table, x, y, z);