static unsigned fs_no = 0;
-/**
- * Generate the depth /stencil test code.
- */
-static void
-generate_depth_stencil(LLVMBuilderRef builder,
- const struct lp_fragment_shader_variant_key *key,
- struct lp_type src_type,
- struct lp_build_mask_context *mask,
- LLVMValueRef stencil_refs[2],
- LLVMValueRef src,
- LLVMValueRef dst_ptr,
- LLVMValueRef facing,
- LLVMValueRef counter)
-{
- const struct util_format_description *format_desc;
- struct lp_type dst_type;
-
- if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
- return;
-
- format_desc = util_format_description(key->zsbuf_format);
- assert(format_desc);
-
- /*
- * 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(src_type.floating) {
- src_type.sign = FALSE;
- src_type.norm = TRUE;
- }
- else {
- assert(!src_type.sign);
- assert(src_type.norm);
- }
-
- /* Pick the depth type. */
- dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
-
- /* FIXME: Cope with a depth test type with a different bit width. */
- assert(dst_type.width == src_type.width);
- assert(dst_type.length == src_type.length);
-
- /* Convert fragment Z from float to integer */
- lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
-
- dst_ptr = LLVMBuildBitCast(builder,
- dst_ptr,
- LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
- lp_build_depth_stencil_test(builder,
- &key->depth,
- key->stencil,
- dst_type,
- format_desc,
- mask,
- stencil_refs,
- src,
- dst_ptr,
- facing,
- counter);
-}
-
/**
* Expand the relevent bits of mask_input to a 4-dword mask for the
}
+#define EARLY_DEPTH_TEST 0x1
+#define LATE_DEPTH_TEST 0x2
+#define EARLY_DEPTH_WRITE 0x4
+#define LATE_DEPTH_WRITE 0x8
+
+static int
+find_output_by_semantic( const struct tgsi_shader_info *info,
+ unsigned semantic,
+ unsigned index )
+{
+ int i;
+
+ for (i = 0; i < info->num_outputs; i++)
+ if (info->output_semantic_name[i] == semantic &&
+ info->output_semantic_index[i] == index)
+ return i;
+
+ return -1;
+}
+
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
LLVMValueRef mask_input,
LLVMValueRef counter)
{
+ const struct util_format_description *zs_format_desc = NULL;
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef vec_type;
LLVMValueRef consts_ptr;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
LLVMValueRef z;
+ LLVMValueRef zs_value = NULL;
LLVMValueRef stencil_refs[2];
- struct lp_build_flow_context *flow;
struct lp_build_mask_context mask;
- boolean early_depth_stencil_test;
+ boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
+ shader->info.base.num_inputs < 3 &&
+ shader->info.base.num_instructions < 8);
unsigned attrib;
unsigned chan;
unsigned cbuf;
+ unsigned depth_mode;
+
+ if (key->depth.enabled ||
+ key->stencil[0].enabled ||
+ key->stencil[1].enabled) {
+
+ 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)
+ /* 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
+ * is known.
+ */
+ depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
+ else
+ depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
+ }
+ else {
+ depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
+ }
+
+ if (!(key->depth.enabled && key->depth.writemask) &&
+ !(key->stencil[0].enabled && key->stencil[0].writemask))
+ depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
+ }
+ else {
+ depth_mode = 0;
+ }
assert(i < 4);
consts_ptr = lp_jit_context_constants(builder, context_ptr);
- flow = lp_build_flow_create(builder);
-
memset(outputs, 0, sizeof outputs);
- lp_build_flow_scope_begin(flow);
-
/* Declare the color and z variables */
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
}
/* 'mask' will control execution based on quad's pixel alive/killed state */
- lp_build_mask_begin(&mask, flow, type, *pmask);
+ lp_build_mask_begin(&mask, builder, type, *pmask);
- lp_build_interp_soa_update_pos(interp, i);
+ if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
+ lp_build_mask_check(&mask);
- /* Try to avoid the 1/w for quads where mask is zero. TODO: avoid
- * this for depth-fail quads also.
- */
+ lp_build_interp_soa_update_pos(interp, i);
z = interp->pos[2];
- early_depth_stencil_test =
- (key->depth.enabled || key->stencil[0].enabled) &&
- !key->alpha.enabled &&
- !shader->info.uses_kill &&
- !shader->info.writes_z;
-
- if (early_depth_stencil_test)
- generate_depth_stencil(builder, key,
- type, &mask,
- stencil_refs, z, depth_ptr, facing, counter);
+ if (depth_mode & EARLY_DEPTH_TEST) {
+ lp_build_depth_stencil_test(builder,
+ &key->depth,
+ key->stencil,
+ type,
+ zs_format_desc,
+ &mask,
+ stencil_refs,
+ z,
+ depth_ptr, facing,
+ &zs_value,
+ !simple_shader);
+
+ if (depth_mode & EARLY_DEPTH_WRITE) {
+ lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
+ }
+ }
lp_build_interp_soa_update_inputs(interp, i);
-
+
+ /* Build the actual shader */
lp_build_tgsi_soa(builder, tokens, type, &mask,
consts_ptr, interp->pos, interp->inputs,
- outputs, sampler, &shader->info);
+ outputs, sampler, &shader->info.base);
- /* loop over fragment shader outputs/results */
- for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
- for(chan = 0; chan < NUM_CHANNELS; ++chan) {
- if(outputs[attrib][chan]) {
- LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
- lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
-
- switch (shader->info.output_semantic_name[attrib]) {
- case TGSI_SEMANTIC_COLOR:
- {
- unsigned cbuf = shader->info.output_semantic_index[attrib];
-
- lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
-
- /* Alpha test */
- /* XXX: should only test the final assignment to alpha */
- if (cbuf == 0 && chan == 3 && key->alpha.enabled) {
- LLVMValueRef alpha = out;
- LLVMValueRef alpha_ref_value;
- alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
- alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
- lp_build_alpha_test(builder, key->alpha.func, type,
- &mask, alpha, alpha_ref_value);
- }
-
- LLVMBuildStore(builder, out, color[cbuf][chan]);
- break;
- }
-
- case TGSI_SEMANTIC_POSITION:
- if(chan == 2)
- z = out;
- break;
- }
- }
+
+ /* Alpha test */
+ if (key->alpha.enabled) {
+ int color0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_COLOR,
+ 0);
+
+ if (color0 != -1) {
+ LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
+ LLVMValueRef alpha_ref_value;
+
+ alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
+ alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
+
+ lp_build_alpha_test(builder, key->alpha.func, type,
+ &mask, alpha, alpha_ref_value,
+ (depth_mode & LATE_DEPTH_TEST) != 0);
}
}
- if (!early_depth_stencil_test)
- generate_depth_stencil(builder, key,
- type, &mask,
- stencil_refs, z, depth_ptr, facing, counter);
+ /* Late Z test */
+ if (depth_mode & LATE_DEPTH_TEST) {
+ int pos0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_POSITION,
+ 0);
+
+ if (pos0 != -1) {
+ z = LLVMBuildLoad(builder, outputs[pos0][2], "z");
+ lp_build_name(z, "output%u.%u.%c", i, pos0, "xyzw"[chan]);
+ }
- lp_build_mask_end(&mask);
+ lp_build_depth_stencil_test(builder,
+ &key->depth,
+ key->stencil,
+ type,
+ zs_format_desc,
+ &mask,
+ stencil_refs,
+ z,
+ depth_ptr, facing,
+ &zs_value,
+ !simple_shader);
+ /* Late Z write */
+ if (depth_mode & LATE_DEPTH_WRITE) {
+ lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
+ }
+ }
+ else if ((depth_mode & EARLY_DEPTH_TEST) &&
+ (depth_mode & LATE_DEPTH_WRITE))
+ {
+ /* Need to apply a reduced mask to the depth write. Reload the
+ * depth value, update from zs_value with the new mask value and
+ * write that out.
+ */
+ lp_build_deferred_depth_write(builder,
+ type,
+ zs_format_desc,
+ &mask,
+ depth_ptr,
+ zs_value);
+ }
- lp_build_flow_scope_end(flow);
- lp_build_flow_destroy(flow);
+ /* Color write */
+ for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
+ {
+ if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR)
+ {
+ unsigned cbuf = shader->info.base.output_semantic_index[attrib];
+ for(chan = 0; chan < NUM_CHANNELS; ++chan)
+ {
+ /* XXX: just initialize outputs to point at colors[] and
+ * skip this.
+ */
+ LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
+ lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
+ LLVMBuildStore(builder, out, color[cbuf][chan]);
+ }
+ }
+ }
- *pmask = mask.value;
+ if (counter)
+ lp_build_occlusion_count(builder, type,
+ lp_build_mask_value(&mask), counter);
+ *pmask = lp_build_mask_end(&mask);
}
LLVMValueRef context_ptr,
LLVMValueRef mask,
LLVMValueRef *src,
- LLVMValueRef dst_ptr)
+ LLVMValueRef dst_ptr,
+ boolean do_branch)
{
struct lp_build_context bld;
- struct lp_build_flow_context *flow;
struct lp_build_mask_context mask_ctx;
LLVMTypeRef vec_type;
LLVMValueRef const_ptr;
lp_build_context_init(&bld, builder, type);
- flow = lp_build_flow_create(builder);
-
- /* we'll use this mask context to skip blending if all pixels are dead */
- lp_build_mask_begin(&mask_ctx, flow, type, mask);
+ lp_build_mask_begin(&mask_ctx, builder, type, mask);
+ if (do_branch)
+ lp_build_mask_check(&mask_ctx);
vec_type = lp_build_vec_type(type);
}
lp_build_mask_end(&mask_ctx);
- lp_build_flow_destroy(flow);
}
LLVMValueRef blend_mask;
LLVMValueRef function;
LLVMValueRef facing;
+ const struct util_format_description *zs_format_desc;
unsigned num_fs;
unsigned i;
unsigned chan;
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
- arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
+ arg_types[8] = LLVMPointerType(LLVMInt8Type(), 0); /* depth */
arg_types[9] = LLVMInt32Type(); /* mask_input */
arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
+ zs_format_desc = util_format_description(key->zsbuf_format);
+
for(i = 0; i < num_fs; ++i) {
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+ LLVMValueRef depth_offset = LLVMConstInt(LLVMInt32Type(),
+ i*fs_type.length*zs_format_desc->block.bits/8,
+ 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
- depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
+ depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
generate_fs(lp, shader, key,
builder,
/*
* Blending.
*/
- generate_blend(&key->blend,
- rt,
- builder,
- blend_type,
- context_ptr,
- blend_mask,
- blend_in_color,
- color_ptr);
+ {
+ /* Could the 4x4 have been killed?
+ */
+ boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
+ !key->alpha.enabled &&
+ !shader->info.base.uses_kill);
+
+ generate_blend(&key->blend,
+ rt,
+ builder,
+ blend_type,
+ context_ptr,
+ blend_mask,
+ blend_in_color,
+ color_ptr,
+ do_branch);
+ }
}
#ifdef PIPE_ARCH_X86
/* Apply optimizations to LLVM IR */
LLVMRunFunctionPassManager(screen->pass, function);
- if (gallivm_debug & GALLIVM_DEBUG_IR) {
+ if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
/* Print the LLVM IR to stderr */
lp_debug_dump_value(function);
debug_printf("\n");
variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
- if (gallivm_debug & GALLIVM_DEBUG_ASM) {
+ if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
lp_disassemble(f);
}
lp_func_delete_body(function);
!key->stencil[0].enabled &&
!key->alpha.enabled &&
!key->depth.enabled &&
- !shader->info.uses_kill
+ !shader->info.base.uses_kill
? TRUE : FALSE;
make_empty_list(&shader->variants);
/* get/save the summary info for this shader */
- tgsi_scan_shader(templ->tokens, &shader->info);
+ lp_build_tgsi_info(templ->tokens, &shader->info);
/* we need to keep a local copy of the tokens */
shader->base.tokens = tgsi_dup_tokens(templ->tokens);
return NULL;
}
- nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
+ nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
sampler[nr_samplers]);
debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
tgsi_dump(templ->tokens, 0);
debug_printf("usage masks:\n");
- for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) {
- unsigned usage_mask = shader->info.input_usage_mask[attrib];
+ for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
+ unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
debug_printf(" IN[%u].%s%s%s%s\n",
attrib,
usage_mask & TGSI_WRITEMASK_X ? "x" : "",
/* This value will be the same for all the variants of a given shader:
*/
- key->nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
+ key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
for(i = 0; i < key->nr_samplers; ++i) {
- if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
+ if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
lp_sampler_static_state(&key->sampler[i],
lp->fragment_sampler_views[i],
lp->sampler[i]);
projects / mesa.git / blobdiff