/**************************************************************************
*
* Copyright 2009 VMware, Inc.
- * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * Copyright 2007 VMware, Inc.
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
- * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
* ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
#include "util/u_format.h"
#include "util/u_dump.h"
#include "util/u_string.h"
-#include "util/u_simple_list.h"
+#include "util/simple_list.h"
#include "util/u_dual_blend.h"
#include "os/os_time.h"
#include "pipe/p_shader_tokens.h"
* XXX: We'll need a different path for 16 x u8
*/
assert(fs_type.width == 32);
- assert(fs_type.length <= Elements(bits));
+ assert(fs_type.length <= ARRAY_SIZE(bits));
mask_type = lp_int_type(fs_type);
/*
for (i = 0; i < fs_type.length / 4; i++) {
unsigned j = 2 * (i % 2) + (i / 2) * 8;
- bits[4*i + 0] = LLVMConstInt(i32t, 1 << (j + 0), 0);
- bits[4*i + 1] = LLVMConstInt(i32t, 1 << (j + 1), 0);
- bits[4*i + 2] = LLVMConstInt(i32t, 1 << (j + 4), 0);
- bits[4*i + 3] = LLVMConstInt(i32t, 1 << (j + 5), 0);
+ bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
+ bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
+ bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
+ bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
}
mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
}
+/**
+ * Fetch the specified lp_jit_viewport structure for a given viewport_index.
+ */
+static LLVMValueRef
+lp_llvm_viewport(LLVMValueRef context_ptr,
+ struct gallivm_state *gallivm,
+ LLVMValueRef viewport_index)
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef ptr;
+ LLVMValueRef res;
+ struct lp_type viewport_type =
+ lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
+
+ ptr = lp_jit_context_viewports(gallivm, context_ptr);
+ ptr = LLVMBuildPointerCast(builder, ptr,
+ LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
+
+ res = lp_build_pointer_get(builder, ptr, viewport_index);
+
+ return res;
+}
+
+
+static LLVMValueRef
+lp_build_depth_clamp(struct gallivm_state *gallivm,
+ LLVMBuilderRef builder,
+ struct lp_type type,
+ LLVMValueRef context_ptr,
+ LLVMValueRef thread_data_ptr,
+ LLVMValueRef z)
+{
+ LLVMValueRef viewport, min_depth, max_depth;
+ LLVMValueRef viewport_index;
+ struct lp_build_context f32_bld;
+
+ assert(type.floating);
+ lp_build_context_init(&f32_bld, gallivm, type);
+
+ /*
+ * Assumes clamping of the viewport index will occur in setup/gs. Value
+ * is passed through the rasterization stage via lp_rast_shader_inputs.
+ *
+ * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
+ * semantics.
+ */
+ viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
+ thread_data_ptr);
+
+ /*
+ * Load the min and max depth from the lp_jit_context.viewports
+ * array of lp_jit_viewport structures.
+ */
+ viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
+
+ /* viewports[viewport_index].min_depth */
+ min_depth = LLVMBuildExtractElement(builder, viewport,
+ lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH), "");
+ min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
+
+ /* viewports[viewport_index].max_depth */
+ max_depth = LLVMBuildExtractElement(builder, viewport,
+ lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH), "");
+ max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
+
+ /*
+ * Clamp to the min and max depth values for the given viewport.
+ */
+ return lp_build_clamp(&f32_bld, z, min_depth, max_depth);
+}
+
+
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
*/
{
const struct util_format_description *zs_format_desc = NULL;
const struct tgsi_token *tokens = shader->base.tokens;
- LLVMTypeRef vec_type;
+ struct lp_type int_type = lp_int_type(type);
+ LLVMTypeRef vec_type, int_vec_type;
LLVMValueRef mask_ptr, mask_val;
- LLVMValueRef consts_ptr;
+ LLVMValueRef consts_ptr, num_consts_ptr;
LLVMValueRef z;
LLVMValueRef z_value, s_value;
LLVMValueRef z_fb, s_fb;
zs_format_desc = util_format_description(key->zsbuf_format);
assert(zs_format_desc);
- if (!shader->info.base.writes_z) {
- if (key->alpha.enabled || shader->info.base.uses_kill) {
+ if (!shader->info.base.writes_z && !shader->info.base.writes_stencil) {
+ if (key->alpha.enabled ||
+ key->blend.alpha_to_coverage ||
+ shader->info.base.uses_kill) {
/* With alpha test and kill, can do the depth test early
* and hopefully eliminate some quads. But need to do a
* special deferred depth write once the final mask value
depth_mode = 0;
}
+ vec_type = lp_build_vec_type(gallivm, type);
+ int_vec_type = lp_build_vec_type(gallivm, int_type);
stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
-
- vec_type = lp_build_vec_type(gallivm, type);
+ /* convert scalar stencil refs into vectors */
+ stencil_refs[0] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[0]);
+ stencil_refs[1] = lp_build_broadcast(gallivm, int_vec_type, stencil_refs[1]);
consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
+ num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
lp_build_for_loop_begin(&loop_state, gallivm,
lp_build_const_int32(gallivm, 0),
z = interp->pos[2];
if (depth_mode & EARLY_DEPTH_TEST) {
+ /*
+ * Clamp according to ARB_depth_clamp semantics.
+ */
+ if (key->depth_clamp) {
+ z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
+ thread_data_ptr, z);
+ }
lp_build_depth_stencil_load_swizzled(gallivm, type,
zs_format_desc, key->resource_1d,
depth_ptr, depth_stride,
/* Build the actual shader */
lp_build_tgsi_soa(gallivm, tokens, type, &mask,
- consts_ptr, &system_values,
+ consts_ptr, num_consts_ptr, &system_values,
interp->inputs,
- outputs, sampler, &shader->info.base, NULL);
+ outputs, context_ptr, thread_data_ptr,
+ sampler, &shader->info.base, NULL);
/* Alpha test */
if (key->alpha.enabled) {
}
}
+ /* Emulate Alpha to Coverage with Alpha test */
+ if (key->blend.alpha_to_coverage) {
+ int color0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_COLOR,
+ 0);
+
+ if (color0 != -1 && outputs[color0][3]) {
+ LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
+
+ lp_build_alpha_to_coverage(gallivm, type,
+ &mask, alpha,
+ (depth_mode & LATE_DEPTH_TEST) != 0);
+ }
+ }
+
/* Late Z test */
if (depth_mode & LATE_DEPTH_TEST) {
int pos0 = find_output_by_semantic(&shader->info.base,
TGSI_SEMANTIC_POSITION,
0);
-
+ int s_out = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_STENCIL,
+ 0);
if (pos0 != -1 && outputs[pos0][2]) {
z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
}
+ /*
+ * Clamp according to ARB_depth_clamp semantics.
+ */
+ if (key->depth_clamp) {
+ z = lp_build_depth_clamp(gallivm, builder, type, context_ptr,
+ thread_data_ptr, z);
+ }
+
+ if (s_out != -1 && outputs[s_out][1]) {
+ /* there's only one value, and spec says to discard additional bits */
+ LLVMValueRef s_max_mask = lp_build_const_int_vec(gallivm, int_type, 255);
+ stencil_refs[0] = LLVMBuildLoad(builder, outputs[s_out][1], "output.s");
+ stencil_refs[0] = LLVMBuildBitCast(builder, stencil_refs[0], int_vec_type, "");
+ stencil_refs[0] = LLVMBuildAnd(builder, stencil_refs[0], s_max_mask, "");
+ stencil_refs[1] = stencil_refs[0];
+ }
lp_build_depth_stencil_load_swizzled(gallivm, type,
zs_format_desc, key->resource_1d,
src_count = num_fs * src_channels;
assert(pixels == 2 || pixels == 1);
- assert(num_fs * src_channels <= Elements(src));
+ assert(num_fs * src_channels <= ARRAY_SIZE(src));
/*
* Transpose from SoA -> AoS
dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
- lp_set_load_alignment(dst[i], dst_alignment);
+ LLVMSetAlignment(dst[i], dst_alignment);
}
}
src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
- lp_set_store_alignment(src_ptr, src_alignment);
+ LLVMSetAlignment(src_ptr, src_alignment);
}
}
*
* A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
*/
-static INLINE boolean
+static inline boolean
is_arithmetic_format(const struct util_format_description *format_desc)
{
boolean arith = false;
* to floats for blending, and furthermore has "natural" packed AoS -> unpacked
* SoA conversion.
*/
-static INLINE boolean
+static inline boolean
format_expands_to_float_soa(const struct util_format_description *format_desc)
{
if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
*
* e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
*/
-static INLINE void
+static inline void
lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
struct lp_type* type)
{
unsigned chan;
if (format_expands_to_float_soa(format_desc)) {
- /* just make this a 32bit uint */
+ /* just make this a uint with width of block */
type->floating = false;
type->fixed = false;
type->sign = false;
type->norm = false;
- type->width = 32;
+ type->width = format_desc->block.bits;
type->length = 1;
return;
}
*
* e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
*/
-static INLINE void
+static inline void
lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
struct lp_type* type)
{
/**
- * Scale a normalized value from src_bits to dst_bits
+ * Scale a normalized value from src_bits to dst_bits.
+ *
+ * The exact calculation is
+ *
+ * dst = iround(src * dst_mask / src_mask)
+ *
+ * or with integer rounding
+ *
+ * dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
+ *
+ * where
+ *
+ * src_mask = (1 << src_bits) - 1
+ * dst_mask = (1 << dst_bits) - 1
+ *
+ * but we try to avoid division and multiplication through shifts.
*/
-static INLINE LLVMValueRef
+static inline LLVMValueRef
scale_bits(struct gallivm_state *gallivm,
int src_bits,
int dst_bits,
LLVMValueRef result = src;
if (dst_bits < src_bits) {
- /* Scale down by LShr */
- result = LLVMBuildLShr(builder,
- src,
- lp_build_const_int_vec(gallivm, src_type, src_bits - dst_bits),
- "");
+ int delta_bits = src_bits - dst_bits;
+
+ if (delta_bits <= dst_bits) {
+ /*
+ * Approximate the rescaling with a single shift.
+ *
+ * This gives the wrong rounding.
+ */
+
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+
+ } else {
+ /*
+ * Try more accurate rescaling.
+ */
+
+ /*
+ * Drop the least significant bits to make space for the multiplication.
+ *
+ * XXX: A better approach would be to use a wider integer type as intermediate. But
+ * this is enough to convert alpha from 16bits -> 2 when rendering to
+ * PIPE_FORMAT_R10G10B10A2_UNORM.
+ */
+ result = LLVMBuildLShr(builder,
+ src,
+ lp_build_const_int_vec(gallivm, src_type, dst_bits),
+ "");
+
+
+ result = LLVMBuildMul(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
+ "");
+
+ /*
+ * Add a rounding term before the division.
+ *
+ * TODO: Handle signed integers too.
+ */
+ if (!src_type.sign) {
+ result = LLVMBuildAdd(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
+ "");
+ }
+
+ /*
+ * Approximate the division by src_mask with a src_bits shift.
+ *
+ * Given the src has already been shifted by dst_bits, all we need
+ * to do is to shift by the difference.
+ */
+
+ result = LLVMBuildLShr(builder,
+ result,
+ lp_build_const_int_vec(gallivm, src_type, delta_bits),
+ "");
+ }
+
} else if (dst_bits > src_bits) {
/* Scale up bits */
int db = dst_bits - src_bits;
return result;
}
+/**
+ * If RT is a smallfloat (needing denorms) format
+ */
+static inline int
+have_smallfloat_format(struct lp_type dst_type,
+ enum pipe_format format)
+{
+ return ((dst_type.floating && dst_type.width != 32) ||
+ /* due to format handling hacks this format doesn't have floating set
+ * here (and actually has width set to 32 too) so special case this. */
+ (format == PIPE_FORMAT_R11G11B10_FLOAT));
+}
+
/**
* Convert from memory format to blending format
* This is pretty suboptimal for this case blending in SoA would be much
* better, since conversion gets us SoA values so need to convert back.
*/
- assert(src_type.width == 32);
+ assert(src_type.width == 32 || src_type.width == 16);
assert(dst_type.floating);
assert(dst_type.width == 32);
assert(dst_type.length % 4 == 0);
assert(num_srcs % 4 == 0);
+ if (src_type.width == 16) {
+ /* expand 4x16bit values to 4x32bit */
+ struct lp_type type32x4 = src_type;
+ LLVMTypeRef ltype32x4;
+ unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
+ type32x4.width = 32;
+ ltype32x4 = lp_build_vec_type(gallivm, type32x4);
+ for (i = 0; i < num_fetch; i++) {
+ src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
+ }
+ src_type.width = 32;
+ }
for (i = 0; i < 4; i++) {
tmpsrc[i] = src[i];
}
assert(src_type.floating);
assert(src_type.width == 32);
assert(src_type.length % 4 == 0);
- assert(dst_type.width == 32);
+ assert(dst_type.width == 32 || dst_type.width == 16);
for (i = 0; i < num_srcs / 4; i++) {
LLVMValueRef tmpsoa[4], tmpdst;
src[i] = tmpdst;
}
}
+ if (dst_type.width == 16) {
+ struct lp_type type16x8 = dst_type;
+ struct lp_type type32x4 = dst_type;
+ LLVMTypeRef ltype16x4, ltypei64, ltypei128;
+ unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
+ type16x8.length = 8;
+ type32x4.width = 32;
+ ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
+ ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
+ ltype16x4 = lp_build_vec_type(gallivm, dst_type);
+ /* We could do vector truncation but it doesn't generate very good code */
+ for (i = 0; i < num_fetch; i++) {
+ src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
+ src[i], lp_build_zero(gallivm, type32x4));
+ src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
+ src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
+ src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
+ }
+ }
return;
}
LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
LLVMValueRef src_alpha[4 * 4];
- LLVMValueRef src1_alpha[4 * 4];
+ LLVMValueRef src1_alpha[4 * 4] = { NULL };
LLVMValueRef src_mask[4 * 4];
LLVMValueRef src[4 * 4];
LLVMValueRef src1[4 * 4];
const boolean is_1d = variant->key.resource_1d;
unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
+ LLVMValueRef fpstate = 0;
+
+ /* Get type from output format */
+ lp_blend_type_from_format_desc(out_format_desc, &row_type);
+ lp_mem_type_from_format_desc(out_format_desc, &dst_type);
+
+ /*
+ * Technically this code should go into lp_build_smallfloat_to_float
+ * and lp_build_float_to_smallfloat but due to the
+ * http://llvm.org/bugs/show_bug.cgi?id=6393
+ * llvm reorders the mxcsr intrinsics in a way that breaks the code.
+ * So the ordering is important here and there shouldn't be any
+ * llvm ir instrunctions in this function before
+ * this, otherwise half-float format conversions won't work
+ * (again due to llvm bug #6393).
+ */
+ if (have_smallfloat_format(dst_type, out_format)) {
+ /* We need to make sure that denorms are ok for half float
+ conversions */
+ fpstate = lp_build_fpstate_get(gallivm);
+ lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
+ }
mask_type = lp_int32_vec4_type();
mask_type.length = fs_type.length;
partial_mask |= !variant->opaque;
i32_zero = lp_build_const_int32(gallivm, 0);
-#if HAVE_LLVM < 0x0302
- /*
- * undef triggers a crash in LLVMBuildTrunc in convert_from_blend_type in some
- * cases (seen with r10g10b10a2, 128bit wide vectors) (only used for 1d case).
- */
- undef_src_val = lp_build_zero(gallivm, fs_type);
-#else
undef_src_val = lp_build_undef(gallivm, fs_type);
-#endif
-
-
- /* Get type from output format */
- lp_blend_type_from_format_desc(out_format_desc, &row_type);
- lp_mem_type_from_format_desc(out_format_desc, &dst_type);
row_type.length = fs_type.length;
vector_width = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
/* Convert */
lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
+ if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
+ /*
+ * since blending is done with floats, there was no conversion.
+ * However, the rules according to fixed point renderbuffers still
+ * apply, that is we must clamp inputs to 0.0/1.0.
+ * (This would apply to separate alpha conversion too but we currently
+ * force has_alpha to be true.)
+ * TODO: should skip this with "fake" blend, since post-blend conversion
+ * will clamp anyway.
+ * TODO: could also skip this if fragment color clamping is enabled. We
+ * don't support it natively so it gets baked into the shader however, so
+ * can't really tell here.
+ */
+ struct lp_build_context f32_bld;
+ assert(row_type.floating);
+ lp_build_context_init(&f32_bld, gallivm, row_type);
+ for (i = 0; i < src_count; i++) {
+ src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
+ }
+ if (dual_source_blend) {
+ for (i = 0; i < src_count; i++) {
+ src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
+ }
+ }
+ /* probably can't be different than row_type but better safe than sorry... */
+ lp_build_context_init(&f32_bld, gallivm, blend_type);
+ blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
+ }
+
/* Extract alpha */
blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
*/
convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
+ /*
+ * FIXME: Really should get logic ops / masks out of generic blend / row
+ * format. Logic ops will definitely not work on the blend float format
+ * used for SRGB here and I think OpenGL expects this to work as expected
+ * (that is incoming values converted to srgb then logic op applied).
+ */
for (i = 0; i < src_count; ++i) {
dst[i] = lp_build_blend_aos(gallivm,
&variant->key.blend,
dst, dst_type, dst_count, dst_alignment);
}
+ if (have_smallfloat_format(dst_type, out_format)) {
+ lp_build_fpstate_set(gallivm, fpstate);
+ }
+
if (do_branch) {
lp_build_mask_end(&mask_ctx);
}
struct gallivm_state *gallivm = variant->gallivm;
const struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
- char func_name[256];
+ char func_name[64];
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
}
/* check if writes to cbuf[0] are to be copied to all cbufs */
- cbuf0_write_all = FALSE;
- for (i = 0;i < shader->info.base.num_properties; i++) {
- if (shader->info.base.properties[i].name ==
- TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
- cbuf0_write_all = TRUE;
- break;
- }
- }
+ cbuf0_write_all =
+ shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
blend_vec_type = lp_build_vec_type(gallivm, blend_type);
- util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
- shader->no, variant->no, partial_mask ? "partial" : "whole");
+ util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
+ shader->no, variant->no, partial_mask ? "partial" : "whole");
arg_types[0] = variant->jit_context_ptr_type; /* context */
arg_types[1] = int32_type; /* x */
arg_types[12] = int32_type; /* depth_stride */
func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
- arg_types, Elements(arg_types), 0);
+ arg_types, ARRAY_SIZE(arg_types), 0);
function = LLVMAddFunction(gallivm->module, func_name, func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
- for(i = 0; i < Elements(arg_types); ++i)
+ for(i = 0; i < ARRAY_SIZE(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr_ptr");
lp_build_name(depth_ptr, "depth");
- lp_build_name(thread_data_ptr, "thread_data");
lp_build_name(mask_input, "mask_input");
+ lp_build_name(thread_data_ptr, "thread_data");
lp_build_name(stride_ptr, "stride_ptr");
lp_build_name(depth_stride, "depth_stride");
LLVMPositionBuilderAtEnd(builder, block);
/* code generated texture sampling */
- sampler = lp_llvm_sampler_soa_create(key->state, context_ptr);
+ sampler = lp_llvm_sampler_soa_create(key->state);
num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
/* for 1d resources only run "upper half" of stamp */
LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
num_loop, "mask_store");
LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
+ boolean pixel_center_integer =
+ shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
/*
* The shader input interpolation info is not explicitely baked in the
gallivm,
shader->info.base.num_inputs,
inputs,
- shader->info.base.pixel_center_integer,
+ pixel_center_integer,
+ key->depth_clamp,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
- LLVMValueRef color_ptr;
- LLVMValueRef stride;
- LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
+ if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
+ LLVMValueRef color_ptr;
+ LLVMValueRef stride;
+ LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
- boolean do_branch = ((key->depth.enabled
- || key->stencil[0].enabled
- || key->alpha.enabled)
- && !shader->info.base.uses_kill);
+ boolean do_branch = ((key->depth.enabled
+ || key->stencil[0].enabled
+ || key->alpha.enabled)
+ && !shader->info.base.uses_kill);
- color_ptr = LLVMBuildLoad(builder,
- LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
- "");
+ color_ptr = LLVMBuildLoad(builder,
+ LLVMBuildGEP(builder, color_ptr_ptr,
+ &index, 1, ""),
+ "");
- lp_build_name(color_ptr, "color_ptr%d", cbuf);
+ lp_build_name(color_ptr, "color_ptr%d", cbuf);
- stride = LLVMBuildLoad(builder,
- LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
- "");
+ stride = LLVMBuildLoad(builder,
+ LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
+ "");
- generate_unswizzled_blend(gallivm, cbuf, variant, key->cbuf_format[cbuf],
- num_fs, fs_type, fs_mask, fs_out_color,
- context_ptr, color_ptr, stride, partial_mask, do_branch);
+ generate_unswizzled_blend(gallivm, cbuf, variant,
+ key->cbuf_format[cbuf],
+ num_fs, fs_type, fs_mask, fs_out_color,
+ context_ptr, color_ptr, stride,
+ partial_mask, do_branch);
+ }
}
LLVMBuildRetVoid(builder);
gallivm_verify_function(gallivm, function);
-
- variant->nr_instrs += lp_build_count_instructions(function);
}
debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
}
debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
+ if (key->blend.alpha_to_coverage) {
+ debug_printf("blend.alpha_to_coverage is enabled\n");
+ }
for (i = 0; i < key->nr_samplers; ++i) {
const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
debug_printf("sampler[%u] = \n", i);
struct lp_fragment_shader_variant *variant;
const struct util_format_description *cbuf0_format_desc;
boolean fullcolormask;
+ char module_name[64];
variant = CALLOC_STRUCT(lp_fragment_shader_variant);
- if(!variant)
+ if (!variant)
return NULL;
- variant->gallivm = gallivm_create();
+ util_snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
+ shader->no, shader->variants_created);
+
+ variant->gallivm = gallivm_create(module_name, lp->context);
if (!variant->gallivm) {
FREE(variant);
return NULL;
fullcolormask &&
!key->stencil[0].enabled &&
!key->alpha.enabled &&
+ !key->blend.alpha_to_coverage &&
!key->depth.enabled &&
!shader->info.base.uses_kill
- ? TRUE : FALSE;
+ ? TRUE : FALSE;
+
+ if ((shader->info.base.num_tokens <= 1) &&
+ !key->depth.enabled && !key->stencil[0].enabled) {
+ variant->ps_inv_multiplier = 0;
+ } else {
+ variant->ps_inv_multiplier = 1;
+ }
if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
lp_debug_fs_variant(variant);
gallivm_compile_module(variant->gallivm);
+ variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
+
if (variant->function[RAST_EDGE_TEST]) {
variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
gallivm_jit_function(variant->gallivm,
variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
}
+ gallivm_free_ir(variant->gallivm);
+
return variant;
}
switch (shader->info.base.input_interpolate[i]) {
case TGSI_INTERPOLATE_CONSTANT:
- shader->inputs[i].interp = LP_INTERP_CONSTANT;
- break;
+ shader->inputs[i].interp = LP_INTERP_CONSTANT;
+ break;
case TGSI_INTERPOLATE_LINEAR:
- shader->inputs[i].interp = LP_INTERP_LINEAR;
- break;
+ shader->inputs[i].interp = LP_INTERP_LINEAR;
+ break;
case TGSI_INTERPOLATE_PERSPECTIVE:
- shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
- break;
+ shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
+ break;
case TGSI_INTERPOLATE_COLOR:
- shader->inputs[i].interp = LP_INTERP_COLOR;
- break;
+ shader->inputs[i].interp = LP_INTERP_COLOR;
+ break;
default:
- assert(0);
- break;
+ assert(0);
+ break;
}
switch (shader->info.base.input_semantic_name[i]) {
case TGSI_SEMANTIC_FACE:
- shader->inputs[i].interp = LP_INTERP_FACING;
- break;
+ shader->inputs[i].interp = LP_INTERP_FACING;
+ break;
case TGSI_SEMANTIC_POSITION:
- /* Position was already emitted above
- */
- shader->inputs[i].interp = LP_INTERP_POSITION;
- shader->inputs[i].src_index = 0;
- continue;
+ /* Position was already emitted above
+ */
+ shader->inputs[i].interp = LP_INTERP_POSITION;
+ shader->inputs[i].src_index = 0;
+ continue;
}
+ /* XXX this is a completely pointless index map... */
shader->inputs[i].src_index = i+1;
}
llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
struct lp_fragment_shader_variant *variant)
{
- unsigned i;
-
if (gallivm_debug & GALLIVM_DEBUG_IR) {
debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
" #%u v total cached #%u\n",
lp->nr_fs_variants);
}
- /* free all the variant's JIT'd functions */
- for (i = 0; i < Elements(variant->function); i++) {
- if (variant->function[i]) {
- gallivm_free_function(variant->gallivm,
- variant->function[i],
- variant->jit_function[i]);
- }
- }
-
gallivm_destroy(variant->gallivm);
/* remove from shader's list */
static void
llvmpipe_set_constant_buffer(struct pipe_context *pipe,
uint shader, uint index,
- struct pipe_constant_buffer *cb)
+ const struct pipe_constant_buffer *cb)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
struct pipe_resource *constants = cb ? cb->buffer : NULL;
assert(shader < PIPE_SHADER_TYPES);
- assert(index < Elements(llvmpipe->constants[shader]));
+ assert(index < ARRAY_SIZE(llvmpipe->constants[shader]));
/* note: reference counting */
util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
+ if (constants) {
+ if (!(constants->bind & PIPE_BIND_CONSTANT_BUFFER)) {
+ debug_printf("Illegal set constant without bind flag\n");
+ constants->bind |= PIPE_BIND_CONSTANT_BUFFER;
+ }
+ }
+
if (shader == PIPE_SHADER_VERTEX ||
shader == PIPE_SHADER_GEOMETRY) {
/* Pass the constants to the 'draw' module */
draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
index, data, size);
}
-
- llvmpipe->dirty |= LP_NEW_CONSTANTS;
+ else {
+ llvmpipe->dirty |= LP_NEW_FS_CONSTANTS;
+ }
if (cb && cb->user_buffer) {
pipe_resource_reference(&constants, NULL);
/**
* Return the blend factor equivalent to a destination alpha of one.
*/
-static INLINE unsigned
-force_dst_alpha_one(unsigned factor)
+static inline unsigned
+force_dst_alpha_one(unsigned factor, boolean clamped_zero)
{
switch(factor) {
case PIPE_BLENDFACTOR_DST_ALPHA:
case PIPE_BLENDFACTOR_INV_DST_ALPHA:
return PIPE_BLENDFACTOR_ZERO;
case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
- return PIPE_BLENDFACTOR_ZERO;
+ if (clamped_zero)
+ return PIPE_BLENDFACTOR_ZERO;
+ else
+ return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
}
return factor;
}
}
+ /*
+ * Propagate the depth clamp setting from the rasterizer state.
+ * depth_clip == 0 implies depth clamping is enabled.
+ *
+ * When clip_halfz is enabled, then always clamp the depth values.
+ *
+ * XXX: This is incorrect for GL, but correct for d3d10 (depth
+ * clamp is always active in d3d10, regardless if depth clip is
+ * enabled or not).
+ * (GL has an always-on [0,1] clamp on fs depth output instead
+ * to ensure the depth values stay in range. Doesn't look like
+ * we do that, though...)
+ */
+ if (lp->rasterizer->clip_halfz) {
+ key->depth_clamp = 1;
+ } else {
+ key->depth_clamp = (lp->rasterizer->depth_clip == 0) ? 1 : 0;
+ }
+
/* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
if (!lp->framebuffer.nr_cbufs ||
+ !lp->framebuffer.cbufs[0] ||
!util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
key->alpha.enabled = lp->depth_stencil->alpha.enabled;
}
}
for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
- enum pipe_format format = lp->framebuffer.cbufs[i]->format;
struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
- const struct util_format_description *format_desc;
- key->cbuf_format[i] = format;
+ if (lp->framebuffer.cbufs[i]) {
+ enum pipe_format format = lp->framebuffer.cbufs[i]->format;
+ const struct util_format_description *format_desc;
- /*
- * Figure out if this is a 1d resource. Note that OpenGL allows crazy
- * mixing of 2d textures with height 1 and 1d textures, so make sure
- * we pick 1d if any cbuf or zsbuf is 1d.
- */
- if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[0]->texture)) {
- key->resource_1d = TRUE;
- }
+ key->cbuf_format[i] = format;
- format_desc = util_format_description(format);
- assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
- format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
+ /*
+ * Figure out if this is a 1d resource. Note that OpenGL allows crazy
+ * mixing of 2d textures with height 1 and 1d textures, so make sure
+ * we pick 1d if any cbuf or zsbuf is 1d.
+ */
+ if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
+ key->resource_1d = TRUE;
+ }
- /*
- * Mask out color channels not present in the color buffer.
- */
- blend_rt->colormask &= util_format_colormask(format_desc);
+ format_desc = util_format_description(format);
+ assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
+ format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
- /*
- * Disable blend for integer formats.
- */
- if (util_format_is_pure_integer(format)) {
- blend_rt->blend_enable = 0;
- }
+ /*
+ * Mask out color channels not present in the color buffer.
+ */
+ blend_rt->colormask &= util_format_colormask(format_desc);
- /*
- * Our swizzled render tiles always have an alpha channel, but the linear
- * render target format often does not, so force here the dst alpha to be
- * one.
- *
- * This is not a mere optimization. Wrong results will be produced if the
- * dst alpha is used, the dst format does not have alpha, and the previous
- * rendering was not flushed from the swizzled to linear buffer. For
- * example, NonPowTwo DCT.
- *
- * TODO: This should be generalized to all channels for better
- * performance, but only alpha causes correctness issues.
- *
- * Also, force rgb/alpha func/factors match, to make AoS blending easier.
- */
- if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
- format_desc->swizzle[3] == format_desc->swizzle[0]) {
- blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
- blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
- blend_rt->alpha_func = blend_rt->rgb_func;
- blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
- blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
+ /*
+ * Disable blend for integer formats.
+ */
+ if (util_format_is_pure_integer(format)) {
+ blend_rt->blend_enable = 0;
+ }
+
+ /*
+ * Our swizzled render tiles always have an alpha channel, but the
+ * linear render target format often does not, so force here the dst
+ * alpha to be one.
+ *
+ * This is not a mere optimization. Wrong results will be produced if
+ * the dst alpha is used, the dst format does not have alpha, and the
+ * previous rendering was not flushed from the swizzled to linear
+ * buffer. For example, NonPowTwo DCT.
+ *
+ * TODO: This should be generalized to all channels for better
+ * performance, but only alpha causes correctness issues.
+ *
+ * Also, force rgb/alpha func/factors match, to make AoS blending
+ * easier.
+ */
+ if (format_desc->swizzle[3] > PIPE_SWIZZLE_W ||
+ format_desc->swizzle[3] == format_desc->swizzle[0]) {
+ /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
+ boolean clamped_zero = !util_format_is_float(format) &&
+ !util_format_is_snorm(format);
+ blend_rt->rgb_src_factor =
+ force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
+ blend_rt->rgb_dst_factor =
+ force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
+ blend_rt->alpha_func = blend_rt->rgb_func;
+ blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
+ blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
+ }
+ }
+ else {
+ /* no color buffer for this fragment output */
+ key->cbuf_format[i] = PIPE_FORMAT_NONE;
+ blend_rt->colormask = 0x0;
+ blend_rt->blend_enable = 0;
}
}
LP_COUNT_ADD(llvm_compile_time, dt);
LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
- llvmpipe_variant_count++;
-
/* Put the new variant into the list */
if (variant) {
insert_at_head(&shader->variants, &variant->list_item_local);
llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
}
+
+/*
+ * Rasterization is disabled if there is no pixel shader and
+ * both depth and stencil testing are disabled:
+ * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
+ */
+boolean
+llvmpipe_rasterization_disabled(struct llvmpipe_context *lp)
+{
+ boolean null_fs = !lp->fs || lp->fs->info.base.num_tokens <= 1;
+
+ return (null_fs &&
+ !lp->depth_stencil->depth.enabled &&
+ !lp->depth_stencil->stencil[0].enabled);
+}