/**************************************************************************
*
- * Copyright 2009 VMware, Inc.
+ * Copyright 2009-2010 VMware, Inc.
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* ... ... ... ... ... ... ... ... ...
*
*
- * Stencil test:
- * Two-sided stencil test is supported but probably not as efficient as
- * it could be. Currently, we use if/then/else constructs to do the
- * operations for front vs. back-facing polygons. We could probably do
- * both the front and back arithmetic then use a Select() instruction to
- * choose the result depending on polyon orientation. We'd have to
- * measure performance both ways and see which is better.
- *
* @author Jose Fonseca <jfonseca@vmware.com>
+ * @author Brian Paul <jfonseca@vmware.com>
*/
#include "pipe/p_state.h"
#include "gallivm/lp_bld_type.h"
#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_bitarit.h"
#include "gallivm/lp_bld_const.h"
+#include "gallivm/lp_bld_conv.h"
#include "gallivm/lp_bld_logic.h"
#include "gallivm/lp_bld_flow.h"
+#include "gallivm/lp_bld_intr.h"
#include "gallivm/lp_bld_debug.h"
#include "gallivm/lp_bld_swizzle.h"
/**
* Do the one or two-sided stencil test comparison.
* \sa lp_build_stencil_test_single
- * \param face an integer indicating front (+) or back (-) facing polygon.
- * If NULL, assume front-facing.
+ * \param front_facing an integer vector mask, indicating front (~0) or back
+ * (0) facing polygon. If NULL, assume front-facing.
*/
static LLVMValueRef
lp_build_stencil_test(struct lp_build_context *bld,
const struct pipe_stencil_state stencil[2],
LLVMValueRef stencilRefs[2],
LLVMValueRef stencilVals,
- LLVMValueRef face)
+ LLVMValueRef front_facing)
{
LLVMValueRef res;
assert(stencil[0].enabled);
- if (stencil[1].enabled && face) {
- /* do two-sided test */
- struct lp_build_flow_context *flow_ctx;
- struct lp_build_if_state if_ctx;
- LLVMValueRef front_facing;
- LLVMValueRef zero = LLVMConstReal(LLVMFloatType(), 0.0);
- LLVMValueRef result = bld->undef;
-
- flow_ctx = lp_build_flow_create(bld->builder);
- lp_build_flow_scope_begin(flow_ctx);
-
- lp_build_flow_scope_declare(flow_ctx, &result);
+ /* do front face test */
+ res = lp_build_stencil_test_single(bld, &stencil[0],
+ stencilRefs[0], stencilVals);
- /* front_facing = face > 0.0 */
- front_facing = LLVMBuildFCmp(bld->builder, LLVMRealUGT, face, zero, "");
-
- lp_build_if(&if_ctx, flow_ctx, bld->builder, front_facing);
- {
- result = lp_build_stencil_test_single(bld, &stencil[0],
- stencilRefs[0], stencilVals);
- }
- lp_build_else(&if_ctx);
- {
- result = lp_build_stencil_test_single(bld, &stencil[1],
- stencilRefs[1], stencilVals);
- }
- lp_build_endif(&if_ctx);
+ if (stencil[1].enabled && front_facing) {
+ /* do back face test */
+ LLVMValueRef back_res;
- lp_build_flow_scope_end(flow_ctx);
- lp_build_flow_destroy(flow_ctx);
+ back_res = lp_build_stencil_test_single(bld, &stencil[1],
+ stencilRefs[1], stencilVals);
- res = result;
- }
- else {
- /* do single-side test */
- res = lp_build_stencil_test_single(bld, &stencil[0],
- stencilRefs[0], stencilVals);
+ res = lp_build_select(bld, front_facing, res, back_res);
}
return res;
const struct pipe_stencil_state *stencil,
enum stencil_op op,
LLVMValueRef stencilRef,
- LLVMValueRef stencilVals,
- LLVMValueRef mask)
+ LLVMValueRef stencilVals)
{
- const unsigned stencilMax = 255; /* XXX fix */
struct lp_type type = bld->type;
LLVMValueRef res;
- LLVMValueRef max = lp_build_const_int_vec(type, stencilMax);
+ LLVMValueRef max = lp_build_const_int_vec(type, 0xff);
unsigned stencil_op;
assert(type.sign);
break;
default:
assert(0 && "bad stencil op mode");
- res = NULL;
- }
-
- if (stencil->writemask != stencilMax) {
- /* compute res = (res & mask) | (stencilVals & ~mask) */
- LLVMValueRef mask = lp_build_const_int_vec(type, stencil->writemask);
- LLVMValueRef cmask = LLVMBuildNot(bld->builder, mask, "notWritemask");
- LLVMValueRef t1 = LLVMBuildAnd(bld->builder, res, mask, "t1");
- LLVMValueRef t2 = LLVMBuildAnd(bld->builder, stencilVals, cmask, "t2");
- res = LLVMBuildOr(bld->builder, t1, t2, "t1_or_t2");
+ res = bld->undef;
}
- /* only the update the vector elements enabled by 'mask' */
- res = lp_build_select(bld, mask, res, stencilVals);
-
return res;
}
LLVMValueRef stencilRefs[2],
LLVMValueRef stencilVals,
LLVMValueRef mask,
- LLVMValueRef face)
+ LLVMValueRef front_facing)
{
- assert(stencil[0].enabled);
-
- if (stencil[1].enabled && face) {
- /* do two-sided op */
- struct lp_build_flow_context *flow_ctx;
- struct lp_build_if_state if_ctx;
- LLVMValueRef front_facing;
- LLVMValueRef zero = LLVMConstReal(LLVMFloatType(), 0.0);
- LLVMValueRef result = bld->undef;
+ LLVMValueRef res;
- flow_ctx = lp_build_flow_create(bld->builder);
- lp_build_flow_scope_begin(flow_ctx);
+ assert(stencil[0].enabled);
- lp_build_flow_scope_declare(flow_ctx, &result);
+ /* do front face op */
+ res = lp_build_stencil_op_single(bld, &stencil[0], op,
+ stencilRefs[0], stencilVals);
- /* front_facing = face > 0.0 */
- front_facing = LLVMBuildFCmp(bld->builder, LLVMRealUGT, face, zero, "");
+ if (stencil[1].enabled && front_facing) {
+ /* do back face op */
+ LLVMValueRef back_res;
- lp_build_if(&if_ctx, flow_ctx, bld->builder, front_facing);
- {
- result = lp_build_stencil_op_single(bld, &stencil[0], op,
- stencilRefs[0], stencilVals, mask);
- }
- lp_build_else(&if_ctx);
- {
- result = lp_build_stencil_op_single(bld, &stencil[1], op,
- stencilRefs[1], stencilVals, mask);
- }
- lp_build_endif(&if_ctx);
+ back_res = lp_build_stencil_op_single(bld, &stencil[1], op,
+ stencilRefs[1], stencilVals);
- lp_build_flow_scope_end(flow_ctx);
- lp_build_flow_destroy(flow_ctx);
+ res = lp_build_select(bld, front_facing, res, back_res);
+ }
- return result;
+ if (stencil->writemask != 0xff) {
+ /* mask &= stencil->writemask */
+ LLVMValueRef writemask = lp_build_const_int_vec(bld->type, stencil->writemask);
+ mask = LLVMBuildAnd(bld->builder, mask, writemask, "");
+ /* res = (res & mask) | (stencilVals & ~mask) */
+ res = lp_build_select_bitwise(bld, writemask, res, stencilVals);
}
else {
- /* do single-sided op */
- return lp_build_stencil_op_single(bld, &stencil[0], op,
- stencilRefs[0], stencilVals, mask);
+ /* res = mask ? res : stencilVals */
+ res = lp_build_select(bld, mask, res, stencilVals);
}
+
+ return res;
}
}
+/**
+ * Perform the occlusion test and increase the counter.
+ * Test the depth mask. Add the number of channel which has none zero mask
+ * into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
+ * The counter will add 4.
+ *
+ * \param type holds element type of the mask vector.
+ * \param maskvalue is the depth test mask.
+ * \param counter is a pointer of the uint32 counter.
+ */
+void
+lp_build_occlusion_count(LLVMBuilderRef builder,
+ struct lp_type type,
+ LLVMValueRef maskvalue,
+ LLVMValueRef counter)
+{
+ LLVMValueRef countmask = lp_build_const_int_vec(type, 1);
+ LLVMValueRef countv = LLVMBuildAnd(builder, maskvalue, countmask, "countv");
+ LLVMTypeRef i8v16 = LLVMVectorType(LLVMInt8Type(), 16);
+ LLVMValueRef counti = LLVMBuildBitCast(builder, countv, i8v16, "counti");
+ LLVMValueRef maskarray[4] = {
+ LLVMConstInt(LLVMInt32Type(), 0, 0),
+ LLVMConstInt(LLVMInt32Type(), 4, 0),
+ LLVMConstInt(LLVMInt32Type(), 8, 0),
+ LLVMConstInt(LLVMInt32Type(), 12, 0),
+ };
+ LLVMValueRef shufflemask = LLVMConstVector(maskarray, 4);
+ LLVMValueRef shufflev = LLVMBuildShuffleVector(builder, counti, LLVMGetUndef(i8v16), shufflemask, "shufflev");
+ LLVMValueRef shuffle = LLVMBuildBitCast(builder, shufflev, LLVMInt32Type(), "shuffle");
+ LLVMValueRef count = lp_build_intrinsic_unary(builder, "llvm.ctpop.i32", LLVMInt32Type(), shuffle);
+ LLVMValueRef orig = LLVMBuildLoad(builder, counter, "orig");
+ LLVMValueRef incr = LLVMBuildAdd(builder, orig, count, "incr");
+ LLVMBuildStore(builder, incr, counter);
+}
+
+
/**
* Generate code for performing depth and/or stencil tests.
* \param format_desc description of the depth/stencil surface
* \param mask the alive/dead pixel mask for the quad (vector)
* \param stencil_refs the front/back stencil ref values (scalar)
- * \param z_src the incoming depth/stencil values (a 2x2 quad)
+ * \param z_src the incoming depth/stencil values (a 2x2 quad, float32)
* \param zs_dst_ptr pointer to depth/stencil values in framebuffer
* \param facing contains float value indicating front/back facing polygon
*/
lp_build_depth_stencil_test(LLVMBuilderRef builder,
const struct pipe_depth_state *depth,
const struct pipe_stencil_state stencil[2],
- struct lp_type type,
+ struct lp_type z_src_type,
const struct util_format_description *format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef stencil_refs[2],
LLVMValueRef z_src,
LLVMValueRef zs_dst_ptr,
- LLVMValueRef face)
+ LLVMValueRef face,
+ LLVMValueRef *zs_value,
+ boolean do_branch)
{
+ struct lp_type type;
struct lp_build_context bld;
struct lp_build_context sbld;
struct lp_type s_type;
LLVMValueRef stencil_vals = NULL;
LLVMValueRef z_bitmask = NULL, stencil_shift = NULL;
LLVMValueRef z_pass = NULL, s_pass_mask = NULL;
- LLVMValueRef orig_mask = mask->value;
+ LLVMValueRef orig_mask = lp_build_mask_value(mask);
+ LLVMValueRef front_facing = NULL;
+
+ /* Prototype a simpler path:
+ */
+ if (z_src_type.floating &&
+ format_desc->format == PIPE_FORMAT_X8Z24_UNORM &&
+ depth->enabled)
+ {
+ LLVMValueRef zscaled;
+ LLVMValueRef const_ffffff_float;
+ LLVMValueRef const_8_int;
+ LLVMTypeRef int32_vec_type;
+
+ /* We know the values in z_dst are all >= 0, so allow
+ * lp_build_compare to use signed compare intrinsics:
+ */
+ type.floating = 0;
+ type.fixed = 0;
+ type.sign = 1;
+ type.norm = 1;
+ type.width = 32;
+ type.length = z_src_type.length;
+
+ int32_vec_type = LLVMVectorType(LLVMInt32Type(), z_src_type.length);
+
+ const_8_int = lp_build_const_int_vec(type, 8);
+ const_ffffff_float = lp_build_const_vec(z_src_type, (float)0xffffff);
+
+ zscaled = LLVMBuildFMul(builder, z_src, const_ffffff_float, "zscaled");
+ z_src = LLVMBuildFPToSI(builder, zscaled, int32_vec_type, "z_src");
+
+ /* Load current z/stencil value from z/stencil buffer */
+ z_dst = LLVMBuildLoad(builder, zs_dst_ptr, "zsbufval");
+ z_dst = LLVMBuildLShr(builder, z_dst, const_8_int, "z_dst");
+
+ /* compare src Z to dst Z, returning 'pass' mask */
+ z_pass = lp_build_compare(builder,
+ type,
+ depth->func, z_src, z_dst);
+
+ lp_build_mask_update(mask, z_pass);
+
+ if (do_branch)
+ lp_build_mask_check(mask);
+
+ /* No need to worry about old stencil contents, just blend the
+ * old and new values and shift into the correct position for
+ * storage.
+ */
+ if (depth->writemask) {
+ type.sign = 1;
+ lp_build_context_init(&bld, builder, type);
+
+ z_dst = lp_build_select(&bld, lp_build_mask_value(mask), z_src, z_dst);
+ z_dst = LLVMBuildShl(builder, z_dst, const_8_int, "z_dst");
+ *zs_value = z_dst;
+ }
+
+ return;
+ }
+
+ /*
+ * Depths are expected to be between 0 and 1, even if they are stored in
+ * floats. Setting these bits here will ensure that the lp_build_conv() call
+ * below won't try to unnecessarily clamp the incoming values.
+ */
+ if(z_src_type.floating) {
+ z_src_type.sign = FALSE;
+ z_src_type.norm = TRUE;
+ }
+ else {
+ assert(!z_src_type.sign);
+ assert(z_src_type.norm);
+ }
+
+ /* Pick the depth type. */
+ type = lp_depth_type(format_desc, z_src_type.width*z_src_type.length);
+
+ /* FIXME: Cope with a depth test type with a different bit width. */
+ assert(type.width == z_src_type.width);
+ assert(type.length == z_src_type.length);
+
+ /* Convert fragment Z from float to integer */
+ lp_build_conv(builder, z_src_type, type, &z_src, 1, &z_src, 1);
+
+ zs_dst_ptr = LLVMBuildBitCast(builder,
+ zs_dst_ptr,
+ LLVMPointerType(lp_build_vec_type(type), 0), "");
+
+
/* Sanity checking */
{
}
}
-
if (stencil[0].enabled) {
+
+ if (face) {
+ LLVMValueRef zero = LLVMConstReal(LLVMFloatType(), 0.0);
+
+ /* front_facing = face > 0.0 ? ~0 : 0 */
+ front_facing = LLVMBuildFCmp(builder, LLVMRealUGT, face, zero, "");
+ front_facing = LLVMBuildSExt(builder, front_facing,
+ LLVMIntType(bld.type.length*bld.type.width),
+ "");
+ front_facing = LLVMBuildBitCast(builder, front_facing,
+ bld.int_vec_type, "");
+ }
+
/* convert scalar stencil refs into vectors */
stencil_refs[0] = lp_build_broadcast_scalar(&bld, stencil_refs[0]);
stencil_refs[1] = lp_build_broadcast_scalar(&bld, stencil_refs[1]);
s_pass_mask = lp_build_stencil_test(&sbld, stencil,
- stencil_refs, stencil_vals, face);
+ stencil_refs, stencil_vals,
+ front_facing);
/* apply stencil-fail operator */
{
- LLVMValueRef s_fail_mask = lp_build_andc(&bld, orig_mask, s_pass_mask);
+ LLVMValueRef s_fail_mask = lp_build_andnot(&bld, orig_mask, s_pass_mask);
stencil_vals = lp_build_stencil_op(&sbld, stencil, S_FAIL_OP,
stencil_refs, stencil_vals,
- s_fail_mask, face);
+ s_fail_mask, front_facing);
}
}
* buffer values. Don't need to update Z buffer values.
*/
lp_build_mask_update(mask, z_pass);
+
+ if (do_branch) {
+ lp_build_mask_check(mask);
+ do_branch = FALSE;
+ }
}
if (depth->writemask) {
- if(z_bitmask)
- z_bitmask = LLVMBuildAnd(builder, mask->value, z_bitmask, "");
- else
- z_bitmask = mask->value;
+ LLVMValueRef zselectmask = lp_build_mask_value(mask);
+
+ /* mask off bits that failed Z test */
+ zselectmask = LLVMBuildAnd(builder, zselectmask, z_pass, "");
+
+ /* mask off bits that failed stencil test */
+ if (s_pass_mask) {
+ zselectmask = LLVMBuildAnd(builder, zselectmask, s_pass_mask, "");
+ }
+
+ /* if combined Z/stencil format, mask off the stencil bits */
+ if (z_bitmask) {
+ zselectmask = LLVMBuildAnd(builder, zselectmask, z_bitmask, "");
+ }
- z_dst = lp_build_select(&bld, z_bitmask, z_src, z_dst);
+ /* Mix the old and new Z buffer values.
+ * z_dst[i] = (zselectmask[i] & z_src[i]) | (~zselectmask[i] & z_dst[i])
+ */
+ z_dst = lp_build_select_bitwise(&bld, zselectmask, z_src, z_dst);
}
if (stencil[0].enabled) {
LLVMValueRef z_fail_mask, z_pass_mask;
/* apply Z-fail operator */
- z_fail_mask = lp_build_andc(&bld, orig_mask, z_pass);
+ z_fail_mask = lp_build_andnot(&bld, orig_mask, z_pass);
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_FAIL_OP,
stencil_refs, stencil_vals,
- z_fail_mask, face);
+ z_fail_mask, front_facing);
/* apply Z-pass operator */
z_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, z_pass, "");
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
- z_pass_mask, face);
+ z_pass_mask, front_facing);
}
}
else {
s_pass_mask = LLVMBuildAnd(bld.builder, orig_mask, s_pass_mask, "");
stencil_vals = lp_build_stencil_op(&sbld, stencil, Z_PASS_OP,
stencil_refs, stencil_vals,
- s_pass_mask, face);
+ s_pass_mask, front_facing);
}
/* The Z bits are already in the right place but we may need to shift the
else
zs_dst = stencil_vals;
- LLVMBuildStore(builder, zs_dst, zs_dst_ptr);
+ *zs_value = zs_dst;
}
if (s_pass_mask)
if (depth->enabled && stencil[0].enabled)
lp_build_mask_update(mask, z_pass);
+
+ if (do_branch)
+ lp_build_mask_check(mask);
+
+}
+
+
+
+void
+lp_build_deferred_depth_write(LLVMBuilderRef builder,
+ struct lp_type z_src_type,
+ const struct util_format_description *format_desc,
+ struct lp_build_mask_context *mask,
+ LLVMValueRef zs_dst_ptr,
+ LLVMValueRef zs_value)
+{
+ struct lp_type type;
+ struct lp_build_context bld;
+ LLVMValueRef z_dst;
+
+ /* XXX: pointlessly redo type logic:
+ */
+ type = lp_depth_type(format_desc, z_src_type.width*z_src_type.length);
+ lp_build_context_init(&bld, builder, type);
+
+ z_dst = LLVMBuildLoad(builder, zs_dst_ptr, "zsbufval");
+ z_dst = lp_build_select(&bld, lp_build_mask_value(mask), zs_value, z_dst);
+
+ LLVMBuildStore(builder, z_dst, zs_dst_ptr);
}