#include "lp_state_setup.h"
-
/* currently organized to interpolate full float[4] attributes even
* when some elements are unused. Later, can pack vertex data more
* closely.
LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
LLVMValueRef a0_0 = args->facing;
LLVMValueRef a0_0f = LLVMBuildSIToFP(builder, a0_0, float_type, "");
+ LLVMValueRef a0, face_val;
const unsigned char swizzles[4] = { PIPE_SWIZZLE_RED, PIPE_SWIZZLE_ZERO,
PIPE_SWIZZLE_ZERO, PIPE_SWIZZLE_ZERO };
/* Our face val is either 1 or 0 so we do
* face = (val * 2) - 1
* to make it 1 or -1
*/
- LLVMValueRef face_val =
+ face_val =
LLVMBuildFAdd(builder,
LLVMBuildFMul(builder, a0_0f,
lp_build_const_float(gallivm, 2.0),
""),
lp_build_const_float(gallivm, -1.0),
"facing");
- LLVMValueRef a0 = lp_build_swizzle_aos(&args->bld, face_val, swizzles);
+ face_val = lp_build_broadcast_scalar(&args->bld, face_val);
+ a0 = lp_build_swizzle_aos(&args->bld, face_val, swizzles);
store_coef(gallivm, args, slot, a0, args->bld.zero, args->bld.zero);
}
}
}
+/*
+ * FIXME: interpolation is always done wrt fb origin (0/0).
+ * However, if some (small) tri is far away from the origin and gradients
+ * are large, this can lead to HUGE errors, since the a0 value calculated
+ * here can get very large (with the actual values inside the triangle way
+ * smaller), leading to complete loss of accuracy. This could be prevented
+ * by using some point inside (or at corner) of the tri as interpolation
+ * origin, or just use barycentric interpolation (which GL suggests and is
+ * what real hw does - you can get the barycentric coordinates from the
+ * edge functions in rasterization in principle (though we skip these
+ * sometimes completely in case of tris covering a block fully,
+ * which obviously wouldn't work)).
+ */
static void
emit_coef4( struct gallivm_state *gallivm,
struct lp_setup_args *args,
LLVMValueRef a2)
{
LLVMBuilderRef b = gallivm->builder;
+ LLVMValueRef attr_0;
LLVMValueRef dy20_ooa = args->dy20_ooa;
LLVMValueRef dy01_ooa = args->dy01_ooa;
LLVMValueRef dx20_ooa = args->dx20_ooa;
/* Calculate a0 - the attribute value at the origin
*/
- LLVMValueRef dadx_x0 = LLVMBuildFMul(b, dadx, x0_center, "dadx_x0");
- LLVMValueRef dady_y0 = LLVMBuildFMul(b, dady, y0_center, "dady_y0");
- LLVMValueRef attr_v0 = LLVMBuildFAdd(b, dadx_x0, dady_y0, "attr_v0");
- LLVMValueRef attr_0 = LLVMBuildFSub(b, a0, attr_v0, "attr_0");
+ LLVMValueRef dadx_x0 = LLVMBuildFMul(b, dadx, x0_center, "dadx_x0");
+ LLVMValueRef dady_y0 = LLVMBuildFMul(b, dady, y0_center, "dady_y0");
+ LLVMValueRef attr_v0 = LLVMBuildFAdd(b, dadx_x0, dady_y0, "attr_v0");
+ attr_0 = LLVMBuildFSub(b, a0, attr_v0, "attr_0");
store_coef(gallivm, args, slot, attr_0, dadx, dady);
}
LLVMValueRef zeroi = lp_build_const_int32(gallivm, 0);
LLVMValueRef pixel_center, xy0_center, dxy01, dxy20, dyx20;
LLVMValueRef e, f, ef, ooa;
- LLVMValueRef shuffles[4];
+ LLVMValueRef shuffles[4], shuf10;
LLVMValueRef attr_pos[3];
struct lp_type typef4 = lp_type_float_vec(32, 128);
struct lp_build_context bld;
shuffles[1] = zeroi;
shuffles[2] = LLVMGetUndef(shuf_type);
shuffles[3] = LLVMGetUndef(shuf_type);
+ shuf10 = LLVMConstVector(shuffles, 4);
- dyx20 = LLVMBuildShuffleVector(b, dxy20, dxy20, LLVMConstVector(shuffles, 4), "");
+ dyx20 = LLVMBuildShuffleVector(b, dxy20, dxy20, shuf10, "");
ef = LLVMBuildFMul(b, dxy01, dyx20, "ef");
e = LLVMBuildExtractElement(b, ef, zeroi, "");