Browse Source

i965: Delete previous workaround for textureGrad with shadow samplers.

It had many problems:
- The shadow comparison was done post-filtering.
- It required state-dependent recompiles whenever the comparison
  function changed.
- It didn't even work: many cases hit assertion failures.
- I never implemented it for the VS.

The new lowering pass which converts textureGrad to textureLod by
computing the LOD value works much better.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
tags/rgb10_a2ui-v3
Kenneth Graunke 13 years ago
parent
commit
b546aebae9

+ 0
- 3
src/mesa/drivers/dri/i965/brw_fs.cpp View File

@@ -2163,9 +2163,6 @@ brw_fs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
key.clamp_fragment_color = true;

for (int i = 0; i < BRW_MAX_TEX_UNIT; i++) {
if (fp->Base.ShadowSamplers & (1 << i))
key.tex.compare_funcs[i] = GL_LESS;

/* FINISHME: depth compares might use (0,0,0,W) for example */
key.tex.swizzles[i] = SWIZZLE_XYZW;
}

+ 5
- 57
src/mesa/drivers/dri/i965/brw_fs_visitor.cpp View File

@@ -712,7 +712,7 @@ fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
/* g0 header. */
mlen = 1;

if (ir->shadow_comparitor && ir->op != ir_txd) {
if (ir->shadow_comparitor) {
for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i), coordinate);
coordinate.reg_offset++;
@@ -931,7 +931,7 @@ fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
}
mlen += vector_elements * reg_width;

if (ir->shadow_comparitor && ir->op != ir_txd) {
if (ir->shadow_comparitor) {
mlen = MAX2(mlen, header_present + 4 * reg_width);

ir->shadow_comparitor->accept(this);
@@ -1038,7 +1038,7 @@ fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
base_mrf--;
}

if (ir->shadow_comparitor && ir->op != ir_txd) {
if (ir->shadow_comparitor) {
ir->shadow_comparitor->accept(this);
emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
mlen += reg_width;
@@ -1163,19 +1163,6 @@ fs_visitor::visit(ir_texture *ir)
int sampler = _mesa_get_sampler_uniform_value(ir->sampler, prog, &fp->Base);
sampler = fp->Base.SamplerUnits[sampler];

/* Our hardware doesn't have a sample_d_c message, so shadow compares
* for textureGrad/TXD need to be emulated with instructions.
*/
bool hw_compare_supported = ir->op != ir_txd;
if (ir->shadow_comparitor && !hw_compare_supported) {
assert(c->key.tex.compare_funcs[sampler] != GL_NONE);
/* No need to even sample for GL_ALWAYS or GL_NEVER...bail early */
if (c->key.tex.compare_funcs[sampler] == GL_ALWAYS)
return swizzle_result(ir, fs_reg(1.0f), sampler);
else if (c->key.tex.compare_funcs[sampler] == GL_NEVER)
return swizzle_result(ir, fs_reg(0.0f), sampler);
}

if (ir->coordinate)
ir->coordinate->accept(this);
fs_reg coordinate = this->result;
@@ -1325,47 +1312,8 @@ fs_visitor::visit(ir_texture *ir)

inst->sampler = sampler;

if (ir->shadow_comparitor) {
if (hw_compare_supported) {
inst->shadow_compare = true;
} else {
ir->shadow_comparitor->accept(this);
fs_reg ref = this->result;

fs_reg value = dst;
dst = fs_reg(this, glsl_type::vec4_type);

/* FINISHME: This needs to be done pre-filtering. */

uint32_t conditional = 0;
switch (c->key.tex.compare_funcs[sampler]) {
/* GL_ALWAYS and GL_NEVER were handled at the top of the function */
case GL_LESS: conditional = BRW_CONDITIONAL_L; break;
case GL_GREATER: conditional = BRW_CONDITIONAL_G; break;
case GL_LEQUAL: conditional = BRW_CONDITIONAL_LE; break;
case GL_GEQUAL: conditional = BRW_CONDITIONAL_GE; break;
case GL_EQUAL: conditional = BRW_CONDITIONAL_EQ; break;
case GL_NOTEQUAL: conditional = BRW_CONDITIONAL_NEQ; break;
default: assert(!"Should not get here: bad shadow compare function");
}

/* Use conditional moves to load 0 or 1 as the result */
this->current_annotation = "manual shadow comparison";
for (int i = 0; i < 4; i++) {
inst = emit(BRW_OPCODE_MOV, dst, fs_reg(0.0f));

inst = emit(BRW_OPCODE_CMP, reg_null_f, ref, value);
inst->conditional_mod = conditional;

inst = emit(BRW_OPCODE_MOV, dst, fs_reg(1.0f));
inst->predicated = true;

dst.reg_offset++;
value.reg_offset++;
}
dst.reg_offset = 0;
}
}
if (ir->shadow_comparitor)
inst->shadow_compare = true;

swizzle_result(ir, dst, sampler);
}

+ 0
- 12
src/mesa/drivers/dri/i965/brw_program.h View File

@@ -28,18 +28,6 @@
* Sampler information needed by VS, WM, and GS program cache keys.
*/
struct brw_sampler_prog_key_data {
/**
* Per-sampler comparison functions:
*
* If comparison mode is GL_COMPARE_R_TO_TEXTURE, then this is set to one
* of GL_NEVER, GL_LESS, GL_EQUAL, GL_LEQUAL, GL_GREATER, GL_NOTEQUAL,
* GL_GEQUAL, or GL_ALWAYS. Otherwise (comparison mode is GL_NONE), this
* field is irrelevant so it's left as GL_NONE (0).
*
* While this is a GLenum, all possible values fit in 16-bits.
*/
uint16_t compare_funcs[BRW_MAX_TEX_UNIT];

/**
* EXT_texture_swizzle and DEPTH_TEXTURE_MODE swizzles.
*/

+ 0
- 3
src/mesa/drivers/dri/i965/brw_wm.c View File

@@ -366,9 +366,6 @@ brw_populate_sampler_prog_key_data(struct gl_context *ctx,

if (img->_BaseFormat == GL_DEPTH_COMPONENT ||
img->_BaseFormat == GL_DEPTH_STENCIL) {
if (sampler->CompareMode == GL_COMPARE_R_TO_TEXTURE_ARB)
key->compare_funcs[i] = sampler->CompareFunc;

/* We handle GL_DEPTH_TEXTURE_MODE here instead of as surface format
* overrides because shadow comparison always returns the result of
* the comparison in all channels anyway.

Loading…
Cancel
Save