llvmpipe: improve polygon-offset precision
This performs the polygon offset addition after interpolation, which prevents floating-point cancellation issues completely. This does mean that we have to perform a single floating-point addition more per fragment than before, unless we also want to spend a bit in the fragment-shader variant key to avoid this. Reviewed-by: Roland Scheidegger <sroland@vmware.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12442>
This commit is contained in:
Erik Faye-Lund
committed by
Marge Bot
Marge Bot
parent
1fa61483de
commit
4d6e18b6cb
@@ -10,7 +10,6 @@ dEQP-EGL.functional.robustness.reset_context.shaders.infinite_loop.sync_status.v
|
||||
dEQP-EGL.functional.robustness.reset_context.shaders.infinite_loop.sync_status.vertex_and_fragment,Fail
|
||||
dEQP-GLES2.functional.clipping.line.wide_line_clip_viewport_center,Fail
|
||||
dEQP-GLES2.functional.clipping.line.wide_line_clip_viewport_corner,Fail
|
||||
dEQP-GLES2.functional.polygon_offset.default_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.clipping.line.wide_line_clip_viewport_center,Fail
|
||||
dEQP-GLES3.functional.clipping.line.wide_line_clip_viewport_corner,Fail
|
||||
dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_mag,Fail
|
||||
@@ -23,9 +22,6 @@ dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_dst_x,Fail
|
||||
dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_src_dst_x,Fail
|
||||
dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_src_dst_y,Fail
|
||||
dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_src_x,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.default_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.fixed24_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.float32_displacement_with_units,Fail
|
||||
dEQP-GLES31.functional.primitive_bounding_box.wide_points.global_state.vertex_tessellation_fragment.default_framebuffer_bbox_equal,Fail
|
||||
dEQP-GLES31.functional.primitive_bounding_box.wide_points.global_state.vertex_tessellation_fragment.default_framebuffer_bbox_larger,Fail
|
||||
dEQP-GLES31.functional.primitive_bounding_box.wide_points.global_state.vertex_tessellation_fragment.fbo_bbox_equal,Fail
|
||||
|
||||
@@ -413,15 +413,25 @@ attribs_update_simple(struct lp_build_interp_soa_context *bld,
|
||||
break;
|
||||
}
|
||||
|
||||
if ((attrib == 0) && (chan == 2) && !bld->depth_clamp){
|
||||
/* OpenGL requires clamping z to 0..1 range after polgon offset
|
||||
* is applied if depth-clamping isn't enabled.
|
||||
*
|
||||
* This also fixes the problem that depth values can exceed 1.0,
|
||||
* due to imprecision in the calculations.
|
||||
*/
|
||||
a = lp_build_clamp(coeff_bld, a, coeff_bld->zero, coeff_bld->one);
|
||||
if ((attrib == 0) && (chan == 2)) {
|
||||
/* add polygon-offset value, stored in the X component of a0 */
|
||||
LLVMValueRef offset =
|
||||
lp_build_extract_broadcast(gallivm, setup_bld->type,
|
||||
coeff_bld->type, bld->a0aos[0],
|
||||
lp_build_const_int32(gallivm, 0));
|
||||
a = LLVMBuildFAdd(builder, a, offset, "");
|
||||
|
||||
if (!bld->depth_clamp){
|
||||
/* OpenGL requires clamping z to 0..1 range after polgon offset
|
||||
* is applied if depth-clamping isn't enabled.
|
||||
*
|
||||
* This also fixes the problem that depth values can exceed 1.0,
|
||||
* due to imprecision in the calculations.
|
||||
*/
|
||||
a = lp_build_clamp(coeff_bld, a, coeff_bld->zero, coeff_bld->one);
|
||||
}
|
||||
}
|
||||
|
||||
bld->attribs[attrib][chan] = a;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -201,7 +201,7 @@ lp_twoside(struct gallivm_state *gallivm,
|
||||
|
||||
}
|
||||
|
||||
static void
|
||||
static LLVMValueRef
|
||||
lp_do_offset_tri(struct gallivm_state *gallivm,
|
||||
struct lp_setup_args *args,
|
||||
const struct lp_setup_variant_key *key,
|
||||
@@ -215,9 +215,7 @@ lp_do_offset_tri(struct gallivm_state *gallivm,
|
||||
struct lp_build_context int_scalar_bld;
|
||||
struct lp_build_context *bld = &args->bld;
|
||||
LLVMValueRef zoffset, mult;
|
||||
LLVMValueRef z0_new, z1_new, z2_new;
|
||||
LLVMValueRef dzdxdzdy, dzdx, dzdy, dzxyz20, dyzzx01, dyzzx01_dzxyz20, dzx01_dyz20;
|
||||
LLVMValueRef z0z1, z0z1z2;
|
||||
LLVMValueRef max, max_value, res12;
|
||||
LLVMValueRef shuffles[4];
|
||||
LLVMTypeRef shuf_type = LLVMInt32TypeInContext(gallivm->context);
|
||||
@@ -323,29 +321,7 @@ lp_do_offset_tri(struct gallivm_state *gallivm,
|
||||
zoffset);
|
||||
}
|
||||
|
||||
/* yuck */
|
||||
shuffles[0] = twoi;
|
||||
shuffles[1] = lp_build_const_int32(gallivm, 6);
|
||||
shuffles[2] = LLVMGetUndef(shuf_type);
|
||||
shuffles[3] = LLVMGetUndef(shuf_type);
|
||||
z0z1 = LLVMBuildShuffleVector(b, attribv[0], attribv[1], LLVMConstVector(shuffles, 4), "");
|
||||
shuffles[0] = zeroi;
|
||||
shuffles[1] = onei;
|
||||
shuffles[2] = lp_build_const_int32(gallivm, 6);
|
||||
shuffles[3] = LLVMGetUndef(shuf_type);
|
||||
z0z1z2 = LLVMBuildShuffleVector(b, z0z1, attribv[2], LLVMConstVector(shuffles, 4), "");
|
||||
zoffset = lp_build_broadcast_scalar(bld, zoffset);
|
||||
|
||||
z0z1z2 = LLVMBuildFAdd(b, z0z1z2, zoffset, "");
|
||||
|
||||
/* insert into args->a0.z, a1.z, a2.z:
|
||||
*/
|
||||
z0_new = LLVMBuildExtractElement(b, z0z1z2, zeroi, "");
|
||||
z1_new = LLVMBuildExtractElement(b, z0z1z2, onei, "");
|
||||
z2_new = LLVMBuildExtractElement(b, z0z1z2, twoi, "");
|
||||
attribv[0] = LLVMBuildInsertElement(b, attribv[0], z0_new, twoi, "");
|
||||
attribv[1] = LLVMBuildInsertElement(b, attribv[1], z1_new, twoi, "");
|
||||
attribv[2] = LLVMBuildInsertElement(b, attribv[2], z2_new, twoi, "");
|
||||
return zoffset;
|
||||
}
|
||||
|
||||
static void
|
||||
@@ -653,6 +629,7 @@ init_args(struct gallivm_state *gallivm,
|
||||
LLVMValueRef e, f, ef, ooa;
|
||||
LLVMValueRef shuffles[4], shuf10;
|
||||
LLVMValueRef attr_pos[3];
|
||||
LLVMValueRef polygon_offset;
|
||||
struct lp_type typef4 = lp_type_float_vec(32, 128);
|
||||
struct lp_build_context bld;
|
||||
|
||||
@@ -693,7 +670,9 @@ init_args(struct gallivm_state *gallivm,
|
||||
|
||||
/* tri offset calc shares a lot of arithmetic, do it here */
|
||||
if (key->pgon_offset_scale != 0.0f || key->pgon_offset_units != 0.0f) {
|
||||
lp_do_offset_tri(gallivm, args, key, ooa, dxy01, dxy20, attr_pos);
|
||||
polygon_offset = lp_do_offset_tri(gallivm, args, key, ooa, dxy01, dxy20, attr_pos);
|
||||
} else {
|
||||
polygon_offset = lp_build_const_float(gallivm, 0.0f);
|
||||
}
|
||||
|
||||
dxy20 = LLVMBuildFMul(b, dxy20, ooa, "");
|
||||
@@ -708,7 +687,22 @@ init_args(struct gallivm_state *gallivm,
|
||||
args->x0_center = lp_build_extract_broadcast(gallivm, typef4, typef4, xy0_center, zeroi);
|
||||
args->y0_center = lp_build_extract_broadcast(gallivm, typef4, typef4, xy0_center, onei);
|
||||
|
||||
emit_linear_coef(gallivm, args, 0, attr_pos);
|
||||
LLVMValueRef coeffs[3];
|
||||
calc_coef4(gallivm, args,
|
||||
attr_pos[0], attr_pos[1], attr_pos[2],
|
||||
coeffs);
|
||||
|
||||
/* This is a bit sneaky:
|
||||
* Because we observe that the X component of A0 is otherwise unused,
|
||||
* we can overwrite it with the computed polygon-offset value, to make
|
||||
* sure it's available in the fragment shader without having to change
|
||||
* the interface (which is error-prone).
|
||||
*/
|
||||
coeffs[0] = LLVMBuildInsertElement(b, coeffs[0], polygon_offset,
|
||||
lp_build_const_int32(gallivm, 0), "");
|
||||
|
||||
store_coef(gallivm, args, 0,
|
||||
coeffs[0], coeffs[1], coeffs[2]);
|
||||
}
|
||||
|
||||
/**
|
||||
|
||||
@@ -3,7 +3,6 @@ dEQP-GLES2.functional.clipping.line.wide_line_clip_viewport_corner,Fail
|
||||
dEQP-GLES2.functional.clipping.point.wide_point_clip,Fail
|
||||
dEQP-GLES2.functional.clipping.point.wide_point_clip_viewport_center,Fail
|
||||
dEQP-GLES2.functional.clipping.point.wide_point_clip_viewport_corner,Fail
|
||||
dEQP-GLES2.functional.polygon_offset.default_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.clipping.line.wide_line_clip_viewport_center,Fail
|
||||
dEQP-GLES3.functional.clipping.line.wide_line_clip_viewport_corner,Fail
|
||||
dEQP-GLES3.functional.clipping.point.wide_point_clip,Fail
|
||||
@@ -19,9 +18,6 @@ dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_dst_x,Fail
|
||||
dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_src_dst_x,Fail
|
||||
dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_src_dst_y,Fail
|
||||
dEQP-GLES3.functional.fbo.blit.rect.nearest_consistency_min_reverse_src_x,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.default_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.fixed24_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.float32_displacement_with_units,Fail
|
||||
dEQP-GLES31.functional.draw_buffers_indexed.random.max_required_draw_buffers.4,Fail
|
||||
dEQP-GLES31.functional.draw_buffers_indexed.random.max_implementation_draw_buffers.8,Fail
|
||||
dEQP-GLES31.functional.image_load_store.buffer.image_size.writeonly_12,Fail
|
||||
|
||||
@@ -1,9 +1,6 @@
|
||||
dEQP-GLES2.functional.clipping.point.wide_point_clip,Fail
|
||||
dEQP-GLES2.functional.clipping.point.wide_point_clip_viewport_center,Fail
|
||||
dEQP-GLES2.functional.clipping.point.wide_point_clip_viewport_corner,Fail
|
||||
dEQP-GLES2.functional.polygon_offset.default_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.default_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.polygon_offset.float32_displacement_with_units,Fail
|
||||
dEQP-GLES3.functional.clipping.line.wide_line_clip_viewport_center,Fail
|
||||
dEQP-GLES3.functional.clipping.line.wide_line_clip_viewport_corner,Fail
|
||||
dEQP-GLES3.functional.clipping.point.wide_point_clip,Fail
|
||||
|
||||
Reference in New Issue
Block a user