st/vega: Add color transformation support.

Per OpenVG 1.1.  A new shader stage is added.  It uses the first two
constants of the fragment shader for color transformation parameters.

src/gallium/state_trackers/vega/asm_fill.h

@@ -44,7 +44,7 @@ solid_fill( struct ureg_program *ureg,
             struct ureg_dst *temp,
             struct ureg_src *constant)
 {
-   ureg_MOV(ureg, *out, constant[0]);
+   ureg_MOV(ureg, *out, constant[2]);
 }
 
 static INLINE void
@@ -61,24 +61,24 @@ linear_grad( struct ureg_program *ureg,
             in[0]);
    ureg_MOV(ureg,
             ureg_writemask(temp[0], TGSI_WRITEMASK_Z),
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y));
-   ureg_DP3(ureg, temp[1], constant[2], ureg_src(temp[0]));
-   ureg_DP3(ureg, temp[2], constant[3], ureg_src(temp[0]));
-   ureg_DP3(ureg, temp[3], constant[4], ureg_src(temp[0]));
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y));
+   ureg_DP3(ureg, temp[1], constant[4], ureg_src(temp[0]));
+   ureg_DP3(ureg, temp[2], constant[5], ureg_src(temp[0]));
+   ureg_DP3(ureg, temp[3], constant[6], ureg_src(temp[0]));
    ureg_RCP(ureg, temp[3], ureg_src(temp[3]));
    ureg_MUL(ureg, temp[1], ureg_src(temp[1]), ureg_src(temp[3]));
    ureg_MUL(ureg, temp[2], ureg_src(temp[2]), ureg_src(temp[3]));
    ureg_MOV(ureg, ureg_writemask(temp[4], TGSI_WRITEMASK_X), ureg_src(temp[1]));
    ureg_MOV(ureg, ureg_writemask(temp[4], TGSI_WRITEMASK_Y), ureg_src(temp[2]));
    ureg_MUL(ureg, temp[0],
-            ureg_scalar(constant[0], TGSI_SWIZZLE_Y),
+            ureg_scalar(constant[2], TGSI_SWIZZLE_Y),
             ureg_scalar(ureg_src(temp[4]), TGSI_SWIZZLE_Y));
    ureg_MAD(ureg, temp[1],
-            ureg_scalar(constant[0], TGSI_SWIZZLE_X),
+            ureg_scalar(constant[2], TGSI_SWIZZLE_X),
             ureg_scalar(ureg_src(temp[4]), TGSI_SWIZZLE_X),
             ureg_src(temp[0]));
    ureg_MUL(ureg, temp[2], ureg_src(temp[1]),
-            ureg_scalar(constant[0], TGSI_SWIZZLE_Z));
+            ureg_scalar(constant[2], TGSI_SWIZZLE_Z));
    ureg_TEX(ureg, *out, TGSI_TEXTURE_1D, ureg_src(temp[2]), sampler[0]);
 }
 
@@ -94,19 +94,19 @@ radial_grad( struct ureg_program *ureg,
    ureg_MOV(ureg, ureg_writemask(temp[0], TGSI_WRITEMASK_XY), in[0]);
    ureg_MOV(ureg,
             ureg_writemask(temp[0], TGSI_WRITEMASK_Z),
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y));
-   ureg_DP3(ureg, temp[1], constant[2], ureg_src(temp[0]));
-   ureg_DP3(ureg, temp[2], constant[3], ureg_src(temp[0]));
-   ureg_DP3(ureg, temp[3], constant[4], ureg_src(temp[0]));
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y));
+   ureg_DP3(ureg, temp[1], constant[4], ureg_src(temp[0]));
+   ureg_DP3(ureg, temp[2], constant[5], ureg_src(temp[0]));
+   ureg_DP3(ureg, temp[3], constant[6], ureg_src(temp[0]));
    ureg_RCP(ureg, temp[3], ureg_src(temp[3]));
    ureg_MUL(ureg, temp[1], ureg_src(temp[1]), ureg_src(temp[3]));
    ureg_MUL(ureg, temp[2], ureg_src(temp[2]), ureg_src(temp[3]));
    ureg_MOV(ureg, ureg_writemask(temp[5], TGSI_WRITEMASK_X), ureg_src(temp[1]));
    ureg_MOV(ureg, ureg_writemask(temp[5], TGSI_WRITEMASK_Y), ureg_src(temp[2]));
-   ureg_MUL(ureg, temp[0], ureg_scalar(constant[0], TGSI_SWIZZLE_Y),
+   ureg_MUL(ureg, temp[0], ureg_scalar(constant[2], TGSI_SWIZZLE_Y),
             ureg_scalar(ureg_src(temp[5]), TGSI_SWIZZLE_Y));
    ureg_MAD(ureg, temp[1],
-            ureg_scalar(constant[0], TGSI_SWIZZLE_X),
+            ureg_scalar(constant[2], TGSI_SWIZZLE_X),
             ureg_scalar(ureg_src(temp[5]), TGSI_SWIZZLE_X), ureg_src(temp[0]));
    ureg_ADD(ureg, temp[1], ureg_src(temp[1]), ureg_src(temp[1]));
    ureg_MUL(ureg, temp[3],
@@ -118,10 +118,10 @@ radial_grad( struct ureg_program *ureg,
             ureg_src(temp[3]));
    ureg_MOV(ureg, temp[4], ureg_negate(ureg_src(temp[4])));
    ureg_MUL(ureg, temp[2],
-            ureg_scalar(constant[0], TGSI_SWIZZLE_Z),
+            ureg_scalar(constant[2], TGSI_SWIZZLE_Z),
             ureg_src(temp[4]));
    ureg_MUL(ureg, temp[0],
-            ureg_scalar(constant[1], TGSI_SWIZZLE_W),
+            ureg_scalar(constant[3], TGSI_SWIZZLE_W),
             ureg_src(temp[2]));
    ureg_MUL(ureg, temp[3], ureg_src(temp[1]), ureg_src(temp[1]));
 
@@ -130,8 +130,8 @@ radial_grad( struct ureg_program *ureg,
    ureg_RCP(ureg, temp[2], ureg_src(temp[2]));
    ureg_SUB(ureg, temp[1], ureg_src(temp[2]), ureg_src(temp[1]));
    ureg_ADD(ureg, temp[0],
-            ureg_scalar(constant[0], TGSI_SWIZZLE_Z),
-            ureg_scalar(constant[0], TGSI_SWIZZLE_Z));
+            ureg_scalar(constant[2], TGSI_SWIZZLE_Z),
+            ureg_scalar(constant[2], TGSI_SWIZZLE_Z));
    ureg_RCP(ureg, temp[0], ureg_src(temp[0]));
    ureg_MUL(ureg, temp[2], ureg_src(temp[1]), ureg_src(temp[0]));
    ureg_TEX(ureg, *out, TGSI_TEXTURE_1D, ureg_src(temp[2]), sampler[0]);
@@ -152,17 +152,17 @@ pattern( struct ureg_program *ureg,
             in[0]);
    ureg_MOV(ureg,
             ureg_writemask(temp[0], TGSI_WRITEMASK_Z),
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y));
-   ureg_DP3(ureg, temp[1], constant[2], ureg_src(temp[0]));
-   ureg_DP3(ureg, temp[2], constant[3], ureg_src(temp[0]));
-   ureg_DP3(ureg, temp[3], constant[4], ureg_src(temp[0]));
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y));
+   ureg_DP3(ureg, temp[1], constant[4], ureg_src(temp[0]));
+   ureg_DP3(ureg, temp[2], constant[5], ureg_src(temp[0]));
+   ureg_DP3(ureg, temp[3], constant[6], ureg_src(temp[0]));
    ureg_RCP(ureg, temp[3], ureg_src(temp[3]));
    ureg_MUL(ureg, temp[1], ureg_src(temp[1]), ureg_src(temp[3]));
    ureg_MUL(ureg, temp[2], ureg_src(temp[2]), ureg_src(temp[3]));
    ureg_MOV(ureg, ureg_writemask(temp[4], TGSI_WRITEMASK_X), ureg_src(temp[1]));
    ureg_MOV(ureg, ureg_writemask(temp[4], TGSI_WRITEMASK_Y), ureg_src(temp[2]));
    ureg_RCP(ureg, temp[0],
-            ureg_swizzle(constant[1],
+            ureg_swizzle(constant[3],
                          TGSI_SWIZZLE_Z,
                          TGSI_SWIZZLE_W,
                          TGSI_SWIZZLE_Z,
@@ -179,6 +179,23 @@ pattern( struct ureg_program *ureg,
    ureg_TEX(ureg, *out, TGSI_TEXTURE_2D, ureg_src(temp[1]), sampler[0]);
 }
 
+static INLINE void
+color_transform( struct ureg_program *ureg,
+                 struct ureg_dst *out,
+                 struct ureg_src *in,
+                 struct ureg_src *sampler,
+                 struct ureg_dst *temp,
+                 struct ureg_src *constant)
+{
+   ureg_MUL(ureg, temp[1], ureg_src(temp[0]), constant[0]);
+   ureg_ADD(ureg, temp[1], ureg_src(temp[1]), constant[1]);
+   ureg_CLAMP(ureg, temp[1],
+              ureg_src(temp[1]),
+              ureg_scalar(constant[3], TGSI_SWIZZLE_X),
+              ureg_scalar(constant[3], TGSI_SWIZZLE_Y));
+   ureg_MOV(ureg, *out, ureg_src(temp[1]));
+}
+
 static INLINE void
 mask( struct ureg_program *ureg,
       struct ureg_dst *out,
@@ -233,10 +250,10 @@ image_stencil( struct ureg_program *ureg,
 
 #define EXTENDED_BLENDER_OVER_FUNC                                      \
    ureg_SUB(ureg, temp[3],                                              \
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y),                   \
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y),                   \
             ureg_scalar(ureg_src(temp[1]), TGSI_SWIZZLE_W));            \
    ureg_SUB(ureg, temp[4],                                              \
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y),                   \
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y),                   \
             ureg_scalar(ureg_src(temp[0]), TGSI_SWIZZLE_W));            \
    ureg_MUL(ureg, temp[3], ureg_src(temp[0]), ureg_src(temp[3]));       \
    ureg_MUL(ureg, temp[4], ureg_src(temp[1]), ureg_src(temp[4]));       \
@@ -369,11 +386,11 @@ color_bw( struct ureg_program *ureg,
                 struct ureg_src *constant)
 {
    ureg_ADD(ureg, temp[1],
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y),
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y));
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y),
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y));
    ureg_RCP(ureg, temp[2], ureg_src(temp[1]));
    ureg_ADD(ureg, temp[1],
-            ureg_scalar(constant[1], TGSI_SWIZZLE_Y),
+            ureg_scalar(constant[3], TGSI_SWIZZLE_Y),
             ureg_src(temp[2]));
    ureg_ADD(ureg, ureg_writemask(temp[2], TGSI_WRITEMASK_X),
             ureg_scalar(ureg_src(temp[0]), TGSI_SWIZZLE_X),
@@ -413,13 +430,13 @@ struct shader_asm_info {
 /* paint types */
 static const struct shader_asm_info shaders_paint_asm[] = {
    {VEGA_SOLID_FILL_SHADER, solid_fill,
-    VG_FALSE, 0, 1, 0, 0, 0, 0},
+    VG_FALSE, 2, 1, 0, 0, 0, 0},
    {VEGA_LINEAR_GRADIENT_SHADER, linear_grad,
-    VG_TRUE,  0, 5, 0, 1, 0, 5},
+    VG_TRUE,  2, 5, 0, 1, 0, 5},
    {VEGA_RADIAL_GRADIENT_SHADER, radial_grad,
-    VG_TRUE,  0, 5, 0, 1, 0, 6},
+    VG_TRUE,  2, 5, 0, 1, 0, 6},
    {VEGA_PATTERN_SHADER, pattern,
-    VG_TRUE,  1, 4, 0, 1, 0, 5}
+    VG_TRUE,  3, 4, 0, 1, 0, 5}
 };
 
 /* image draw modes */
@@ -432,6 +449,11 @@ static const struct shader_asm_info shaders_image_asm[] = {
     VG_TRUE,  0, 0, 3, 1, 0, 2}
 };
 
+static const struct shader_asm_info shaders_color_transform_asm[] = {
+   {VEGA_COLOR_TRANSFORM_SHADER, color_transform,
+    VG_FALSE, 0, 4, 0, 0, 0, 2}
+};
+
 static const struct shader_asm_info shaders_mask_asm[] = {
    {VEGA_MASK_SHADER, mask,
     VG_TRUE,  0, 0, 1, 1, 0, 2}
@@ -440,13 +462,13 @@ static const struct shader_asm_info shaders_mask_asm[] = {
 /* extra blend modes */
 static const struct shader_asm_info shaders_blend_asm[] = {
    {VEGA_BLEND_MULTIPLY_SHADER, blend_multiply,
-    VG_TRUE,  1, 1, 2, 1, 0, 5},
+    VG_TRUE,  3, 1, 2, 1, 0, 5},
    {VEGA_BLEND_SCREEN_SHADER, blend_screen,
     VG_TRUE,  0, 0, 2, 1, 0, 4},
    {VEGA_BLEND_DARKEN_SHADER, blend_darken,
-    VG_TRUE,  1, 1, 2, 1, 0, 6},
+    VG_TRUE,  3, 1, 2, 1, 0, 6},
    {VEGA_BLEND_LIGHTEN_SHADER, blend_lighten,
-    VG_TRUE,  1, 1, 2, 1, 0, 6},
+    VG_TRUE,  3, 1, 2, 1, 0, 6},
 };
 
 /* premultiply */
@@ -460,7 +482,7 @@ static const struct shader_asm_info shaders_premultiply_asm[] = {
 /* color transform to black and white */
 static const struct shader_asm_info shaders_bw_asm[] = {
    {VEGA_BW_SHADER, color_bw,
-    VG_FALSE,  1, 1, 0, 0, 0, 3},
+    VG_FALSE,  3, 1, 0, 0, 0, 3},
 };
 
 #endif

src/gallium/state_trackers/vega/mask.c

@@ -253,7 +253,11 @@ static void mask_resource_fill(struct pipe_resource *dst,
                                VGfloat coverage)
 {
    struct vg_context *ctx = vg_current_context();
-   VGfloat color[4] = { 0.0f, 0.0f, 0.0f, coverage };
+   VGfloat fs_consts[12] = {
+      0.0f, 0.0f, 0.0f, 0.0f, /* not used */
+      0.0f, 0.0f, 0.0f, 0.0f, /* not used */
+      0.0f, 0.0f, 0.0f, coverage /* color */
+   };
    void *fs;
 
    if (x < 0) {
@@ -267,8 +271,8 @@ static void mask_resource_fill(struct pipe_resource *dst,
 
    fs = shaders_cache_fill(ctx->sc, VEGA_SOLID_FILL_SHADER);
 
-   if (renderer_filter_begin(ctx->renderer, dst, VG_FALSE,
-            ~0, NULL, NULL, 0, fs, (const void *) color, sizeof(color))) {
+   if (renderer_filter_begin(ctx->renderer, dst, VG_FALSE, ~0,
+            NULL, NULL, 0, fs, (const void *) fs_consts, sizeof(fs_consts))) {
       renderer_filter(ctx->renderer, x, y, width, height, 0, 0, 0, 0);
       renderer_filter_end(ctx->renderer);
    }

src/gallium/state_trackers/vega/shader.c

@@ -38,12 +38,14 @@
 #include "pipe/p_state.h"
 #include "util/u_inlines.h"
 #include "util/u_memory.h"
+#include "util/u_math.h"
 
-#define MAX_CONSTANTS 20
+#define MAX_CONSTANTS 28
 
 struct shader {
    struct vg_context *context;
 
+   VGboolean color_transform;
    VGboolean masking;
    struct vg_paint *paint;
    struct vg_image *image;
@@ -72,6 +74,11 @@ void shader_destroy(struct shader *shader)
    FREE(shader);
 }
 
+void shader_set_color_transform(struct shader *shader, VGboolean set)
+{
+   shader->color_transform = set;
+}
+
 void shader_set_masking(struct shader *shader, VGboolean set)
 {
    shader->masking = set;
@@ -94,10 +101,25 @@ struct vg_paint * shader_paint(struct shader *shader)
 
 static VGint setup_constant_buffer(struct shader *shader)
 {
+   const struct vg_state *state = &shader->context->state.vg;
    VGint param_bytes = paint_constant_buffer_size(shader->paint);
+   VGint i;
 
+   param_bytes += sizeof(VGfloat) * 8;
    assert(param_bytes <= sizeof(shader->constants));
-   paint_fill_constant_buffer(shader->paint, shader->constants);
+
+   if (state->color_transform) {
+      for (i = 0; i < 8; i++) {
+         VGfloat val = (i < 4) ? 127.0f : 1.0f;
+         shader->constants[i] =
+            CLAMP(state->color_transform_values[i], -val, val);
+      }
+   }
+   else {
+      memset(shader->constants, 0, sizeof(VGfloat) * 8);
+   }
+
+   paint_fill_constant_buffer(shader->paint, shader->constants + 8);
 
    return param_bytes;
 }
@@ -227,6 +249,9 @@ static void setup_shader_program(struct shader *shader)
       }
    }
 
+   if (shader->color_transform)
+      shader_id |= VEGA_COLOR_TRANSFORM_SHADER;
+
    if (shader->masking)
       shader_id |= VEGA_MASK_SHADER;
 

src/gallium/state_trackers/vega/shader.h

@@ -37,6 +37,8 @@ struct vg_image;
 struct shader *shader_create(struct vg_context *context);
 void shader_destroy(struct shader *shader);
 
+void shader_set_color_transform(struct shader *shader, VGboolean set);
+
 void shader_set_masking(struct shader *shader, VGboolean set);
 VGboolean shader_is_masking(struct shader *shader);
 

src/gallium/state_trackers/vega/shaders_cache.c

@@ -52,12 +52,13 @@
  * of the pipeline. The stages are:
  * 1) Paint generation (color/gradient/pattern)
  * 2) Image composition (normal/multiply/stencil)
- * 3) Mask
- * 4) Extended blend (multiply/screen/darken/lighten)
- * 5) Premultiply/Unpremultiply
- * 6) Color transform (to black and white)
+ * 3) Color transform
+ * 4) Mask
+ * 5) Extended blend (multiply/screen/darken/lighten)
+ * 6) Premultiply/Unpremultiply
+ * 7) Color transform (to black and white)
  */
-#define SHADER_STAGES 6
+#define SHADER_STAGES 7
 
 struct cached_shader {
    void *driver_shader;
@@ -286,6 +287,19 @@ create_shader(struct pipe_context *pipe,
    assert(idx == ((!sh || sh == VEGA_IMAGE_NORMAL_SHADER) ? 1 : 2));
 
    /* third stage */
+   sh = SHADERS_GET_COLOR_TRANSFORM_SHADER(id);
+   switch (sh) {
+   case VEGA_COLOR_TRANSFORM_SHADER:
+      shaders[idx] = &shaders_color_transform_asm[
+         (sh >> SHADERS_COLOR_TRANSFORM_SHIFT) - 1];
+      assert(shaders[idx]->id == sh);
+      idx++;
+      break;
+   default:
+      break;
+   }
+
+   /* fourth stage */
    sh = SHADERS_GET_MASK_SHADER(id);
    switch (sh) {
    case VEGA_MASK_SHADER:
@@ -297,7 +311,7 @@ create_shader(struct pipe_context *pipe,
       break;
    }
 
-   /* fourth stage */
+   /* fifth stage */
    sh = SHADERS_GET_BLEND_SHADER(id);
    switch (sh) {
    case VEGA_BLEND_MULTIPLY_SHADER:
@@ -312,7 +326,7 @@ create_shader(struct pipe_context *pipe,
       break;
    }
 
-   /* fifth stage */
+   /* sixth stage */
    sh = SHADERS_GET_PREMULTIPLY_SHADER(id);
    switch (sh) {
    case VEGA_PREMULTIPLY_SHADER:
@@ -326,7 +340,7 @@ create_shader(struct pipe_context *pipe,
       break;
    }
 
-   /* sixth stage */
+   /* seventh stage */
    sh = SHADERS_GET_BW_SHADER(id);
    switch (sh) {
    case VEGA_BW_SHADER:

src/gallium/state_trackers/vega/shaders_cache.h

@@ -35,6 +35,7 @@ struct shaders_cache;
 
 #define _SHADERS_PAINT_BITS            3
 #define _SHADERS_IMAGE_BITS            2
+#define _SHADERS_COLOR_TRANSFORM_BITS  1
 #define _SHADERS_MASK_BITS             1
 #define _SHADERS_BLEND_BITS            3
 #define _SHADERS_PREMULTIPLY_BITS      2
@@ -42,7 +43,8 @@ struct shaders_cache;
 
 #define SHADERS_PAINT_SHIFT           (0)
 #define SHADERS_IMAGE_SHIFT           (SHADERS_PAINT_SHIFT + _SHADERS_PAINT_BITS)
-#define SHADERS_MASK_SHIFT            (SHADERS_IMAGE_SHIFT + _SHADERS_IMAGE_BITS)
+#define SHADERS_COLOR_TRANSFORM_SHIFT (SHADERS_IMAGE_SHIFT + _SHADERS_IMAGE_BITS)
+#define SHADERS_MASK_SHIFT            (SHADERS_COLOR_TRANSFORM_SHIFT + _SHADERS_COLOR_TRANSFORM_BITS)
 #define SHADERS_BLEND_SHIFT           (SHADERS_MASK_SHIFT + _SHADERS_MASK_BITS)
 #define SHADERS_PREMULTIPLY_SHIFT     (SHADERS_BLEND_SHIFT + _SHADERS_BLEND_BITS)
 #define SHADERS_BW_SHIFT              (SHADERS_PREMULTIPLY_SHIFT + _SHADERS_PREMULTIPLY_BITS)
@@ -52,6 +54,7 @@ struct shaders_cache;
 
 #define SHADERS_GET_PAINT_SHADER(id)           _SHADERS_GET_STAGE(PAINT, id)
 #define SHADERS_GET_IMAGE_SHADER(id)           _SHADERS_GET_STAGE(IMAGE, id)
+#define SHADERS_GET_COLOR_TRANSFORM_SHADER(id) _SHADERS_GET_STAGE(COLOR_TRANSFORM, id)
 #define SHADERS_GET_MASK_SHADER(id)            _SHADERS_GET_STAGE(MASK, id)
 #define SHADERS_GET_BLEND_SHADER(id)           _SHADERS_GET_STAGE(BLEND, id)
 #define SHADERS_GET_PREMULTIPLY_SHADER(id)     _SHADERS_GET_STAGE(PREMULTIPLY, id)
@@ -67,6 +70,8 @@ enum VegaShaderType {
    VEGA_IMAGE_MULTIPLY_SHADER     = 2 << SHADERS_IMAGE_SHIFT,
    VEGA_IMAGE_STENCIL_SHADER      = 3 << SHADERS_IMAGE_SHIFT,
 
+   VEGA_COLOR_TRANSFORM_SHADER    = 1 <<  SHADERS_COLOR_TRANSFORM_SHIFT,
+
    VEGA_MASK_SHADER               = 1 << SHADERS_MASK_SHIFT,
 
    VEGA_BLEND_MULTIPLY_SHADER     = 1 << SHADERS_BLEND_SHIFT,

src/gallium/state_trackers/vega/vg_context.c

@@ -416,6 +416,7 @@ void vg_validate_state(struct vg_context *ctx)
 
    shader_set_masking(ctx->shader, ctx->state.vg.masking);
    shader_set_image_mode(ctx->shader, ctx->state.vg.image_mode);
+   shader_set_color_transform(ctx->shader, ctx->state.vg.color_transform);
 }
 
 VGboolean vg_object_is_valid(void *ptr, enum vg_object_type type)