i915g: Implement sRGB textures

Since we don't have them in hw we emulate them in the shader. Although not
recommended by the spec it is legit.

As a side effect we also get GL 2.1. I think this is as far as we can take
the i915.

5 files changed, 128 insertions, 12 deletions

diff --git a/src/gallium/drivers/i915/i915_context.h b/src/gallium/drivers/i915/i915_context.h
index b019c9f342a..16b0c57166d 100644
--- a/src/gallium/drivers/i915/i915_context.h
+++ b/src/gallium/drivers/i915/i915_context.h
@@ -155,6 +155,8 @@ struct i915_state
    unsigned sampler[I915_TEX_UNITS][3];
    unsigned sampler_enable_flags;
    unsigned sampler_enable_nr;
+   boolean sampler_srgb[I915_TEX_UNITS];
+   int srgb_const_offset;
 
    /* texture image buffers */
    unsigned texbuffer[I915_TEX_UNITS][2];
diff --git a/src/gallium/drivers/i915/i915_screen.c b/src/gallium/drivers/i915/i915_screen.c
index 9f293392f81..ff352035fb0 100644
--- a/src/gallium/drivers/i915/i915_screen.c
+++ b/src/gallium/drivers/i915/i915_screen.c
@@ -110,7 +110,9 @@ i915_get_shader_param(struct pipe_screen *screen, unsigned shader, enum pipe_sha
             return PIPE_MAX_VERTEX_SAMPLERS;
          else
             return 0;
-      default:
+      case PIPE_SHADER_CAP_INTEGERS:
+         return 1;
+       default:
          return draw_get_shader_param(shader, cap);
       }
    case PIPE_SHADER_FRAGMENT:
@@ -290,6 +292,7 @@ i915_is_format_supported(struct pipe_screen *screen,
 {
    static const enum pipe_format tex_supported[] = {
       PIPE_FORMAT_B8G8R8A8_UNORM,
+      PIPE_FORMAT_B8G8R8A8_SRGB,
       PIPE_FORMAT_B8G8R8X8_UNORM,
       PIPE_FORMAT_R8G8B8A8_UNORM,
       PIPE_FORMAT_R8G8B8X8_UNORM,
diff --git a/src/gallium/drivers/i915/i915_state_emit.c b/src/gallium/drivers/i915/i915_state_emit.c
index 8ab8fb8cd74..ac999792c06 100644
--- a/src/gallium/drivers/i915/i915_state_emit.c
+++ b/src/gallium/drivers/i915/i915_state_emit.c
@@ -30,6 +30,7 @@
 #include "i915_context.h"
 #include "i915_batch.h"
 #include "i915_debug.h"
+#include "i915_fpc.h"
 #include "i915_resource.h"
 
 #include "pipe/p_context.h"
@@ -313,11 +314,42 @@ emit_sampler(struct i915_context *i915)
    }
 }
 
+static boolean is_tex_instruction(uint32_t* instruction)
+{
+   uint32_t op = instruction[0] &0xFF000000;
+   return ( (op == T0_TEXLD) ||
+            (op == T0_TEXLDP) ||
+            (op == T0_TEXLDB));
+}
+
+static uint32_t tex_sampler(uint32_t* instruction)
+{
+   return ( instruction[0] & T0_SAMPLER_NR_MASK);
+}
+
+static uint additional_constants(struct i915_context *i915)
+{
+   int i;
+
+   for (i = 0 ; i < i915->fs->program_len; i+=3) {
+      if ( is_tex_instruction(i915->fs->program + i)) {
+           int sampler = tex_sampler(i915->fs->program + i);
+           assert(sampler < I915_TEX_UNITS);
+           if ( i915->current.sampler_srgb[sampler] )
+              return 1;
+      }
+   }
+   return 0;
+}
+
 static void
 validate_constants(struct i915_context *i915, unsigned *batch_space)
 {
-   *batch_space = i915->fs->num_constants ?
+   int nr = i915->fs->num_constants ?
       2 + 4*i915->fs->num_constants : 0;
+
+   nr += 4*additional_constants(i915);
+   *batch_space = nr;
 }
 
 static void
@@ -326,8 +358,11 @@ emit_constants(struct i915_context *i915)
    /* Collate the user-defined constants with the fragment shader's
     * immediates according to the constant_flags[] array.
     */
-   const uint nr = i915->fs->num_constants;
+   const uint nr = i915->fs->num_constants + additional_constants(i915);
+
+   assert(nr < I915_MAX_CONSTANT);
    if (nr) {
+      const float srgb_constants[4] = {1.0/1.055, 0.055/1.055, 2.4, 0.0822};
       uint i;
 
       OUT_BATCH( _3DSTATE_PIXEL_SHADER_CONSTANTS | (nr * 4) );
@@ -340,9 +375,16 @@ emit_constants(struct i915_context *i915)
             c = (uint *) i915_buffer(i915->constants[PIPE_SHADER_FRAGMENT])->data;
             c += 4 * i;
          }
-         else {
+         else if (i < i915->fs->num_constants) {
             /* emit program constant */
             c = (uint *) i915->fs->constants[i];
+         } else {
+            /* emit constants for sRGB */
+
+            /* save const position in context for use in shader emit */
+            i915->current.srgb_const_offset = i;
+
+            c = (uint *) srgb_constants;
          }
 #if 0 /* debug */
          {
@@ -363,18 +405,74 @@ emit_constants(struct i915_context *i915)
 static void
 validate_program(struct i915_context *i915, unsigned *batch_space)
 {
-   uint additional_size = i915->current.target_fixup_format ? 1 : 0;
+   uint additional_size = 0, i;
+
+   additional_size += i915->current.target_fixup_format ? 3 : 0;
 
-   /* we need more batch space if we want to emulate rgba framebuffers */
-   *batch_space = i915->fs->decl_len + i915->fs->program_len + 3 * additional_size;
+   for (i = 0 ; i < i915->fs->program_len; i+=3)
+      if ( is_tex_instruction(i915->fs->program + i) &&
+           i915->current.sampler_srgb[tex_sampler(i915->fs->program+i)] )
+         additional_size += 3 * 8 /* 8 instructions for srgb emulation */;
+
+   /* we need more batch space if we want to emulate rgba framebuffers
+    * or sRGB textures */
+   *batch_space = i915->fs->decl_len + i915->fs->program_len + additional_size;
+}
+
+static void emit_instruction(struct i915_context *i915,
+                             int op,
+                             int dst_mask,
+			     int dst_reg,
+                             int src0_reg,
+                             int src1_reg,
+                             int src2_reg)
+{
+   OUT_BATCH(op |
+             dst_mask |
+             0 | /* saturate */
+             A0_DEST(dst_reg) |
+             A0_SRC0(src0_reg)
+             );
+   OUT_BATCH(A1_SRC0(src0_reg) | A1_SRC1(src1_reg));
+   OUT_BATCH(A2_SRC1(src1_reg) | A2_SRC2(src2_reg));
+}
+
+static void
+emit_srgb_fixup(struct i915_context *i915,
+                uint *program)
+{
+   int dst_reg =
+       (program[0] & UREG_TYPE_NR_MASK) >> UREG_A0_DEST_SHIFT_LEFT;
+   int dst_mask = program[0] & A0_DEST_CHANNEL_ALL;
+   int cst_idx = i915->current.srgb_const_offset;
+   int cst0_reg = swizzle(UREG(REG_TYPE_CONST, cst_idx), X, X, X, X);
+   int cst1_reg = swizzle(UREG(REG_TYPE_CONST, cst_idx), Y, Y, Y, Y);
+   int cst2_reg = swizzle(UREG(REG_TYPE_CONST, cst_idx), Z, Z, Z, Z);
+   int t1_reg = UREG(REG_TYPE_R, 1);
+   int t1x_reg = swizzle(UREG(REG_TYPE_R, 1), X, X, X, X);
+   int t1y_reg = swizzle(UREG(REG_TYPE_R, 1), Y, Y, Y, Y);
+   int t1z_reg = swizzle(UREG(REG_TYPE_R, 1), Z, Z, Z, Z);
+   
+   emit_instruction(i915, A0_MAD, A0_DEST_CHANNEL_ALL, t1_reg, dst_reg, cst0_reg, cst1_reg);
+   emit_instruction(i915, A0_LOG, A0_DEST_CHANNEL_X, t1_reg, t1x_reg, 0, 0);
+   emit_instruction(i915, A0_LOG, A0_DEST_CHANNEL_Y, t1_reg, t1y_reg, 0, 0);
+   emit_instruction(i915, A0_LOG, A0_DEST_CHANNEL_Z, t1_reg, t1z_reg, 0, 0);
+   emit_instruction(i915, A0_MUL, A0_DEST_CHANNEL_ALL, t1_reg, t1_reg, cst2_reg, 0);
+   emit_instruction(i915, A0_EXP, dst_mask & A0_DEST_CHANNEL_X, dst_reg, t1x_reg, 0, 0);
+   emit_instruction(i915, A0_EXP, dst_mask & A0_DEST_CHANNEL_Y, dst_reg, t1y_reg, 0, 0);
+   emit_instruction(i915, A0_EXP, dst_mask & A0_DEST_CHANNEL_Z, dst_reg, t1z_reg, 0, 0);
 }
 
 static void
 emit_program(struct i915_context *i915)
 {
-   uint need_target_fixup = i915->current.target_fixup_format ? 1 : 0;
+   uint additional_size = 0;
    uint i;
 
+   /* count how much additional space we'll need */
+   validate_program(i915, &additional_size);
+   additional_size -= i915->fs->decl_len + i915->fs->program_len;
+
    /* we should always have, at least, a pass-through program */
    assert(i915->fs->program_len > 0);
 
@@ -382,7 +480,7 @@ emit_program(struct i915_context *i915)
    {
       /* first word has the size, we have to adjust that */
       uint size = (i915->fs->decl[0]);
-      size += need_target_fixup * 3;
+      size += additional_size;
       OUT_BATCH(size);
    }
 
@@ -390,11 +488,21 @@ emit_program(struct i915_context *i915)
       OUT_BATCH(i915->fs->decl[i]);
 
    /* output the program */
-   for (i = 0 ; i < i915->fs->program_len; i++)
+   assert(i915->fs->program_len % 3 == 0);
+   for (i = 0 ; i < i915->fs->program_len; i+=3) {
       OUT_BATCH(i915->fs->program[i]);
+      OUT_BATCH(i915->fs->program[i+1]);
+      OUT_BATCH(i915->fs->program[i+2]);
+      
+      /* TEX fixup for sRGB */
+      if ( is_tex_instruction(i915->fs->program+i) &&
+           i915->current.sampler_srgb[tex_sampler(i915->fs->program+i)] )
+         emit_srgb_fixup(i915, i915->fs->program);
+
+   }
 
    /* we emit an additional mov with swizzle to fake RGBA framebuffers */
-   if (need_target_fixup) {
+   if (i915->current.target_fixup_format) {
       /* mov out_color, out_color.zyxw */
       OUT_BATCH(A0_MOV |
                 (REG_TYPE_OC << A0_DEST_TYPE_SHIFT) |
diff --git a/src/gallium/drivers/i915/i915_state_sampler.c b/src/gallium/drivers/i915/i915_state_sampler.c
index 19d3c83ffcd..9aba68afbca 100644
--- a/src/gallium/drivers/i915/i915_state_sampler.c
+++ b/src/gallium/drivers/i915/i915_state_sampler.c
@@ -215,6 +215,7 @@ static uint translate_texture_format(enum pipe_format pipeFormat,
    case PIPE_FORMAT_B10G10R10A2_UNORM:
       return MAPSURF_32BIT | MT_32BIT_ARGB2101010;
    case PIPE_FORMAT_B8G8R8A8_UNORM:
+   case PIPE_FORMAT_B8G8R8A8_SRGB:
       return MAPSURF_32BIT | MT_32BIT_ARGB8888;
    case PIPE_FORMAT_B8G8R8X8_UNORM:
       return MAPSURF_32BIT | MT_32BIT_XRGB8888;
@@ -309,6 +310,8 @@ static void update_map(struct i915_context *i915,
    assert(depth);
 
    format = translate_texture_format(pt->format, view);
+   i915->current.sampler_srgb[unit] = ( pt->format == PIPE_FORMAT_B8G8R8A8_SRGB ||
+                                        pt->format == PIPE_FORMAT_L8_SRGB );
    pitch = tex->stride;
 
    assert(format);
diff --git a/src/gallium/drivers/i915/i915_state_static.c b/src/gallium/drivers/i915/i915_state_static.c
index 9587dec4b90..4d020975576 100644
--- a/src/gallium/drivers/i915/i915_state_static.c
+++ b/src/gallium/drivers/i915/i915_state_static.c
@@ -31,7 +31,6 @@
 #include "i915_screen.h"
 
 
-
 /***********************************************************************
  * Update framebuffer state
  */
@@ -39,6 +38,7 @@ static unsigned translate_format(enum pipe_format format)
 {
    switch (format) {
    case PIPE_FORMAT_B8G8R8A8_UNORM:
+   case PIPE_FORMAT_B8G8R8A8_SRGB:
    case PIPE_FORMAT_B8G8R8X8_UNORM:
    case PIPE_FORMAT_R8G8B8A8_UNORM:
    case PIPE_FORMAT_R8G8B8X8_UNORM: