#include "util/u_memory.h"
#include "util/u_math.h"
#include "tgsi/tgsi_parse.h"
-#include "lp_bld_debug.h"
-#include "lp_bld_const.h"
-#include "lp_bld_arit.h"
-#include "lp_bld_swizzle.h"
+#include "gallivm/lp_bld_debug.h"
+#include "gallivm/lp_bld_const.h"
+#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_swizzle.h"
#include "lp_bld_interp.h"
+/*
+ * The shader JIT function operates on blocks of quads.
+ * Each block has 2x2 quads and each quad has 2x2 pixels.
+ *
+ * We iterate over the quads in order 0, 1, 2, 3:
+ *
+ * #################
+ * # | # | #
+ * #---0---#---1---#
+ * # | # | #
+ * #################
+ * # | # | #
+ * #---2---#---3---#
+ * # | # | #
+ * #################
+ *
+ * Within each quad, we have four pixels which are represented in SOA
+ * order:
+ *
+ * #########
+ * # 0 | 1 #
+ * #---+---#
+ * # 2 | 3 #
+ * #########
+ *
+ * So the green channel (for example) of the four pixels is stored in
+ * a single vector register: {g0, g1, g2, g3}.
+ */
+
+
static void
attrib_name(LLVMValueRef val, unsigned attrib, unsigned chan, const char *suffix)
{
}
+/**
+ * Initialize the bld->a0, dadx, dady fields. This involves fetching
+ * those values from the arrays which are passed into the JIT function.
+ */
static void
coeffs_init(struct lp_build_interp_soa_context *bld,
LLVMValueRef a0_ptr,
unsigned chan;
for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
- unsigned mask = bld->mask[attrib];
- unsigned mode = bld->mode[attrib];
+ const unsigned mask = bld->mask[attrib];
+ const unsigned interp = bld->interp[attrib];
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
if(mask & (1 << chan)) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), attrib*NUM_CHANNELS + chan, 0);
LLVMValueRef dadx = NULL;
LLVMValueRef dady = NULL;
- switch( mode ) {
+ switch( interp ) {
case TGSI_INTERPOLATE_PERSPECTIVE:
/* fall-through */
case TGSI_INTERPOLATE_CONSTANT:
a0 = LLVMBuildLoad(builder, LLVMBuildGEP(builder, a0_ptr, &index, 1, ""), "");
a0 = lp_build_broadcast_scalar(&bld->base, a0);
- attrib_name(a0, attrib, chan, ".dady");
+ attrib_name(a0, attrib, chan, ".a0");
break;
default:
/**
- * Small vector x scale multiplication optimization.
- *
- * TODO: Should be elsewhere.
- */
-static LLVMValueRef
-coeff_multiply(struct lp_build_interp_soa_context *bld,
- LLVMValueRef coeff,
- int step)
-{
- LLVMValueRef factor;
-
- switch(step) {
- case 0:
- return bld->base.zero;
- case 1:
- return coeff;
- case 2:
- return lp_build_add(&bld->base, coeff, coeff);
- default:
- factor = lp_build_const_scalar(bld->base.type, (double)step);
- return lp_build_mul(&bld->base, coeff, factor);
- }
-}
-
-
-/**
- * Multiply the dadx and dady with the xstep and ystep respectively.
+ * Emit LLVM code to compute the fragment shader input attribute values.
+ * For example, for a color input, we'll compute red, green, blue and alpha
+ * values for the four pixels in a quad.
+ * Recall that we're operating on 4-element vectors so each arithmetic
+ * operation is operating on the four pixels in a quad.
*/
-static void
-coeffs_update(struct lp_build_interp_soa_context *bld)
-{
- unsigned attrib;
- unsigned chan;
-
- for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
- unsigned mask = bld->mask[attrib];
- unsigned mode = bld->mode[attrib];
- if (mode != TGSI_INTERPOLATE_CONSTANT) {
- for(chan = 0; chan < NUM_CHANNELS; ++chan) {
- if(mask & (1 << chan)) {
- bld->dadx[attrib][chan] = coeff_multiply(bld, bld->dadx[attrib][chan], bld->xstep);
- bld->dady[attrib][chan] = coeff_multiply(bld, bld->dady[attrib][chan], bld->ystep);
- }
- }
- }
- }
-}
-
-
static void
attribs_init(struct lp_build_interp_soa_context *bld)
{
unsigned chan;
for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
- unsigned mask = bld->mask[attrib];
- unsigned mode = bld->mode[attrib];
+ const unsigned mask = bld->mask[attrib];
+ const unsigned interp = bld->interp[attrib];
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
if(mask & (1 << chan)) {
LLVMValueRef a0 = bld->a0 [attrib][chan];
res = a0;
- if (mode != TGSI_INTERPOLATE_CONSTANT) {
+ if (interp != TGSI_INTERPOLATE_CONSTANT) {
+ /* res = res + x * dadx */
res = lp_build_add(&bld->base, res, lp_build_mul(&bld->base, x, dadx));
+ /* res = res + y * dady */
res = lp_build_add(&bld->base, res, lp_build_mul(&bld->base, y, dady));
}
- /* Keep the value of the attribue before perspective divide for faster updates */
+ /* Keep the value of the attribute before perspective divide
+ * for faster updates.
+ */
bld->attribs_pre[attrib][chan] = res;
- if (mode == TGSI_INTERPOLATE_PERSPECTIVE) {
+ if (interp == TGSI_INTERPOLATE_PERSPECTIVE) {
LLVMValueRef w = bld->pos[3];
assert(attrib != 0);
if(!oow)
}
+/**
+ * Increment the shader input attribute values.
+ * This is called when we move from one quad to the next.
+ */
static void
-attribs_update(struct lp_build_interp_soa_context *bld)
+attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
{
LLVMValueRef oow = NULL;
unsigned attrib;
unsigned chan;
+ assert(quad_index < 4);
+
for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
- unsigned mask = bld->mask[attrib];
- unsigned mode = bld->mode[attrib];
+ const unsigned mask = bld->mask[attrib];
+ const unsigned interp = bld->interp[attrib];
- if (mode != TGSI_INTERPOLATE_CONSTANT) {
+ if (interp != TGSI_INTERPOLATE_CONSTANT) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
if(mask & (1 << chan)) {
LLVMValueRef dadx = bld->dadx[attrib][chan];
res = bld->attribs_pre[attrib][chan];
- if(bld->xstep)
+ if (quad_index == 1 || quad_index == 3) {
+ /* top-right or bottom-right quad */
+ /* build res = res + dadx + dadx */
res = lp_build_add(&bld->base, res, dadx);
+ res = lp_build_add(&bld->base, res, dadx);
+ }
- if(bld->ystep)
+ if (quad_index == 2 || quad_index == 3) {
+ /* bottom-left or bottom-right quad */
+ /* build res = res + dady + dady */
res = lp_build_add(&bld->base, res, dady);
+ res = lp_build_add(&bld->base, res, dady);
+ }
- bld->attribs_pre[attrib][chan] = res;
-
- if (mode == TGSI_INTERPOLATE_PERSPECTIVE) {
+ if (interp == TGSI_INTERPOLATE_PERSPECTIVE) {
LLVMValueRef w = bld->pos[3];
assert(attrib != 0);
if(!oow)
}
+/**
+ * Update quad position values when moving to the next quad.
+ */
static void
-pos_update(struct lp_build_interp_soa_context *bld)
+pos_update(struct lp_build_interp_soa_context *bld, int quad_index)
{
LLVMValueRef x = bld->attribs[0][0];
LLVMValueRef y = bld->attribs[0][1];
+ const int xstep = 2, ystep = 2;
- if(bld->xstep)
- x = lp_build_add(&bld->base, x, lp_build_const_scalar(bld->base.type, bld->xstep));
+ if (quad_index == 1 || quad_index == 3) {
+ /* top-right or bottom-right quad in block */
+ /* build x += xstep */
+ x = lp_build_add(&bld->base, x,
+ lp_build_const_vec(bld->base.type, xstep));
+ }
- if(bld->ystep)
- y = lp_build_add(&bld->base, y, lp_build_const_scalar(bld->base.type, bld->ystep));
+ if (quad_index == 2) {
+ /* bottom-left quad in block */
+ /* build y += ystep */
+ y = lp_build_add(&bld->base, y,
+ lp_build_const_vec(bld->base.type, ystep));
+ /* build x -= xstep */
+ x = lp_build_sub(&bld->base, x,
+ lp_build_const_vec(bld->base.type, xstep));
+ }
lp_build_name(x, "pos.x");
lp_build_name(y, "pos.y");
}
+/**
+ * Initialize fragment shader input attribute info.
+ */
void
lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
const struct tgsi_token *tokens,
+ boolean flatshade,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef a0_ptr,
LLVMValueRef dadx_ptr,
LLVMValueRef dady_ptr,
LLVMValueRef x0,
- LLVMValueRef y0,
- int xstep,
- int ystep)
+ LLVMValueRef y0)
{
struct tgsi_parse_context parse;
struct tgsi_full_declaration *decl;
/* Position */
bld->num_attribs = 1;
bld->mask[0] = TGSI_WRITEMASK_ZW;
- bld->mode[0] = TGSI_INTERPOLATE_LINEAR;
+ bld->interp[0] = TGSI_INTERPOLATE_LINEAR;
/* Inputs */
tgsi_parse_init( &parse, tokens );
unsigned first, last, mask;
unsigned attrib;
- first = decl->DeclarationRange.First;
- last = decl->DeclarationRange.Last;
+ first = decl->Range.First;
+ last = decl->Range.Last;
mask = decl->Declaration.UsageMask;
for( attrib = first; attrib <= last; ++attrib ) {
bld->mask[1 + attrib] = mask;
- bld->mode[1 + attrib] = decl->Declaration.Interpolate;
+
+ /* XXX: have mesa set INTERP_CONSTANT in the fragment
+ * shader.
+ */
+ if (decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
+ flatshade)
+ bld->interp[1 + attrib] = TGSI_INTERPOLATE_CONSTANT;
+ else
+ bld->interp[1 + attrib] = decl->Declaration.Interpolate;
}
bld->num_attribs = MAX2(bld->num_attribs, 1 + last + 1);
case TGSI_TOKEN_TYPE_INSTRUCTION:
case TGSI_TOKEN_TYPE_IMMEDIATE:
+ case TGSI_TOKEN_TYPE_PROPERTY:
break;
default:
pos_init(bld, x0, y0);
attribs_init(bld);
-
- bld->xstep = xstep;
- bld->ystep = ystep;
-
- coeffs_update(bld);
}
/**
- * Advance the position and inputs with the xstep and ystep.
+ * Advance the position and inputs to the given quad within the block.
*/
void
-lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld)
+lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld,
+ int quad_index)
{
- pos_update(bld);
+ assert(quad_index < 4);
+
+ pos_update(bld, quad_index);
- attribs_update(bld);
+ attribs_update(bld, quad_index);
}