#include "lp_state_setup.h"
+/** Setup shader number (for debugging) */
+static unsigned setup_no = 0;
+
/* currently organized to interpolate full float[4] attributes even
* when some elements are unused. Later, can pack vertex data more
static void
store_coef(struct gallivm_state *gallivm,
- struct lp_setup_args *args,
- unsigned slot,
- LLVMValueRef a0,
- LLVMValueRef dadx,
- LLVMValueRef dady)
+ struct lp_setup_args *args,
+ unsigned slot,
+ LLVMValueRef a0,
+ LLVMValueRef dadx,
+ LLVMValueRef dady)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef idx = lp_build_const_int32(gallivm, slot);
LLVMBuildStore(builder,
- a0,
- LLVMBuildGEP(builder, args->a0, &idx, 1, ""));
+ a0,
+ LLVMBuildGEP(builder, args->a0, &idx, 1, ""));
LLVMBuildStore(builder,
- dadx,
- LLVMBuildGEP(builder, args->dadx, &idx, 1, ""));
+ dadx,
+ LLVMBuildGEP(builder, args->dadx, &idx, 1, ""));
LLVMBuildStore(builder,
- dady,
- LLVMBuildGEP(builder, args->dady, &idx, 1, ""));
+ dady,
+ LLVMBuildGEP(builder, args->dady, &idx, 1, ""));
}
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;
lp_build_context_init(&bld, gallivm, typef4);
+ args->bld = bld;
/* The internal position input is in slot zero:
*/
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, "");
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);
- args->bld = bld;
-
emit_linear_coef(gallivm, args, 0, attr_pos);
}
struct lp_setup_variant *variant = NULL;
struct gallivm_state *gallivm;
struct lp_setup_args args;
- char func_name[256];
+ char func_name[64];
LLVMTypeRef vec4f_type;
LLVMTypeRef func_type;
LLVMTypeRef arg_types[7];
if (variant == NULL)
goto fail;
- variant->gallivm = gallivm = gallivm_create();
+ variant->no = setup_no++;
+
+ util_snprintf(func_name, sizeof(func_name), "setup_variant_%u",
+ variant->no);
+
+ variant->gallivm = gallivm = gallivm_create(func_name, lp->context);
if (!variant->gallivm) {
goto fail;
}
memcpy(&variant->key, key, key->size);
variant->list_item_global.base = variant;
- util_snprintf(func_name, sizeof(func_name), "fs%u_setup%u",
- 0, variant->no);
-
/* Currently always deal with full 4-wide vertex attributes from
* the vertices.
*/
if (!variant->jit_function)
goto fail;
+ gallivm_free_ir(variant->gallivm);
+
/*
* Update timing information:
*/
LP_COUNT_ADD(llvm_compile_time, t1 - t0);
LP_COUNT_ADD(nr_llvm_compiles, 1);
}
-
+
return variant;
fail:
if (variant) {
- if (variant->function) {
- gallivm_free_function(gallivm,
- variant->function,
- variant->jit_function);
- }
if (variant->gallivm) {
gallivm_destroy(variant->gallivm);
}
FREE(variant);
}
-
+
return NULL;
}
static void
lp_make_setup_variant_key(struct llvmpipe_context *lp,
- struct lp_setup_variant_key *key)
+ struct lp_setup_variant_key *key)
{
struct lp_fragment_shader *fs = lp->fs;
unsigned i;
assert(sizeof key->inputs[0] == sizeof(uint));
-
+
key->num_inputs = fs->info.base.num_inputs;
key->flatshade_first = lp->rasterizer->flatshade_first;
key->pixel_center_half = lp->rasterizer->half_pixel_center;
for (i = 0; i < key->num_inputs; i++) {
if (key->inputs[i].interp == LP_INTERP_COLOR) {
if (lp->rasterizer->flatshade)
- key->inputs[i].interp = LP_INTERP_CONSTANT;
- else
- key->inputs[i].interp = LP_INTERP_PERSPECTIVE;
+ key->inputs[i].interp = LP_INTERP_CONSTANT;
+ else
+ key->inputs[i].interp = LP_INTERP_PERSPECTIVE;
}
}
static void
remove_setup_variant(struct llvmpipe_context *lp,
- struct lp_setup_variant *variant)
+ struct lp_setup_variant *variant)
{
if (gallivm_debug & GALLIVM_DEBUG_IR) {
debug_printf("llvmpipe: del setup_variant #%u total %u\n",
- variant->no, lp->nr_setup_variants);
- }
-
- if (variant->function) {
- gallivm_free_function(variant->gallivm,
- variant->function,
- variant->jit_function);
+ variant->no, lp->nr_setup_variants);
}
if (variant->gallivm) {
foreach(li, &lp->setup_variants_list) {
if(li->base->key.size == key->size &&
- memcmp(&li->base->key, key, key->size) == 0) {
+ memcmp(&li->base->key, key, key->size) == 0) {
variant = li->base;
break;
}
}
else {
if (lp->nr_setup_variants >= LP_MAX_SETUP_VARIANTS) {
- cull_setup_variants(lp);
+ cull_setup_variants(lp);
}
variant = generate_setup_variant(key, lp);
if (variant) {
insert_at_head(&lp->setup_variants_list, &variant->list_item_global);
lp->nr_setup_variants++;
- llvmpipe_variant_count++;
}
}
- lp_setup_set_setup_variant(lp->setup,
- variant);
+ lp_setup_set_setup_variant(lp->setup, variant);
}
void
}
void
-lp_dump_setup_coef( const struct lp_setup_variant_key *key,
- const float (*sa0)[4],
- const float (*sdadx)[4],
- const float (*sdady)[4])
+lp_dump_setup_coef(const struct lp_setup_variant_key *key,
+ const float (*sa0)[4],
+ const float (*sdadx)[4],
+ const float (*sdady)[4])
{
int i, slot;
float dady = sdady[0][i];
debug_printf("POS.%c: a0 = %f, dadx = %f, dady = %f\n",
- "xyzw"[i],
- a0, dadx, dady);
+ "xyzw"[i], a0, dadx, dady);
}
for (slot = 0; slot < key->num_inputs; slot++) {
unsigned usage_mask = key->inputs[slot].usage_mask;
for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
- if (usage_mask & (1 << i)) {
- float a0 = sa0 [1 + slot][i];
- float dadx = sdadx[1 + slot][i];
- float dady = sdady[1 + slot][i];
-
- debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",
- slot,
- "xyzw"[i],
- a0, dadx, dady);
- }
+ if (usage_mask & (1 << i)) {
+ float a0 = sa0 [1 + slot][i];
+ float dadx = sdadx[1 + slot][i];
+ float dady = sdady[1 + slot][i];
+
+ debug_printf("IN[%u].%c: a0 = %f, dadx = %f, dady = %f\n",
+ slot, "xyzw"[i], a0, dadx, dady);
+ }
}
}
}
projects / mesa.git / blobdiff