aboutsummaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers/i915/i915_state_emit.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/drivers/i915/i915_state_emit.c')
-rw-r--r--src/gallium/drivers/i915/i915_state_emit.c128
1 files changed, 118 insertions, 10 deletions
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) |
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-15 08:04:35 +0000